diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS
index 2802f461b6..a9694036cf 100644
--- a/.github/CODEOWNERS
+++ b/.github/CODEOWNERS
@@ -6,10 +6,10 @@
 /unsloth/models/rl_replacements.py @Datta0 @pluesclues @danielhanchen
 /unsloth/trainer.py @danielhanchen
 /unsloth/models/sentence_transformer.py @Etherll @danielhanchen
-/unsloth/save.py @rolandtannous @danielhanchen
+/unsloth/save.py @danielhanchen
 /unsloth/tokenizer_utils.py @mmathew23 @danielhanchen
-/unsloth/chat_templates.py @rolandtannous @danielhanchen
-/unsloth/ollama_template_mappers.py @rolandtannous @danielhanchen
+/unsloth/chat_templates.py @danielhanchen
+/unsloth/ollama_template_mappers.py @danielhanchen
 /unsloth/kernels/moe/*.py @Datta0
 /unsloth/import_fixes.py @danielhanchen
 /unsloth/device_type.py @danielhanchen
@@ -45,14 +45,14 @@
 /unsloth/utils/hf_hub.py @mmathew23
 /unsloth/utils/packing.py @mmathew23
 
-/cli/ @rolandtannous @Manan17
-/studio/frontend/ @Shine1i @rolandtannous @Manan17
+/cli/ @Manan17
+/studio/frontend/ @Shine1i @Manan17
 /studio/frontend/public/ @Shine1i
-/studio/backend/ @rolandtannous
-/studio/backend/core/data_recipe/ @rolandtannous
-/studio/backend/tests/ @rolandtannous @danielhanchen
-/tests/ @rolandtannous @danielhanchen
-/scripts/ @rolandtannous @danielhanchen
+/studio/backend/
+/studio/backend/core/data_recipe/
+/studio/backend/tests/ @danielhanchen
+/tests/ @danielhanchen
+/scripts/ @danielhanchen
 
 # Snapshot data for the notebook linter / Colab oracle. Drift in these
 # files changes the pin floor for every Unsloth notebook, so refreshes
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index c6c2e1fc37..f7c338d76b 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -333,6 +333,17 @@ jobs:
         run: |
           python -m pytest -v --tb=short tests/test_callback_signature_drift.py
 
+      - name: generation correctness guards (HARD GATE)
+        # Deterministic CPU guards, each validated to fail on its pre-fix code:
+        # leftpad = batched left-padded generation (#1066/#3699, fixed by
+        # #2216 + #4100; staging proof: unsloth-staging-2 PRs 170/172);
+        # rope_scaling_drift = config.rope_scaling dropped by replaced rotary
+        # classes (#2405). AST checks run first so import breakage cannot mask them.
+        run: |
+          python -m pytest -v --tb=short \
+            tests/utils/test_prepare_inputs_leftpad.py \
+            tests/utils/test_rope_scaling_drift.py
+
       - name: unsloth Bucket-A — CPU tests not in Repo tests (CPU)
         # CPU tests across 6 files under tests/saving/, tests/utils/, tests/python/
         # that Repo tests (CPU) --ignores. AST/protobuf/regex plus tiny CPU model
diff --git a/.github/workflows/lint-ci.yml b/.github/workflows/lint-ci.yml
index 8cb3535075..bd859a6e9e 100644
--- a/.github/workflows/lint-ci.yml
+++ b/.github/workflows/lint-ci.yml
@@ -105,29 +105,37 @@ jobs:
         # exactly where a hoist refactor lives, and it skips brand-new
         # files whose re-export imports would otherwise look "unused".
         #
-        # actions/checkout uses fetch-depth: 1, so the base branch is not
-        # present locally. Fetch the single base commit with an explicit
-        # refspec so origin/<base> is reliably created (a bare
-        # `git fetch origin <ref>` only updates FETCH_HEAD in some
-        # configs). Two-dot diff avoids needing a merge-base on a shallow
-        # clone.
+        # Diff against the true merge-base, not the base tip. A two-dot
+        # diff against the tip re-lints every file the base branch
+        # changed after the PR branched, comparing newer base code
+        # (BEFORE) against the PR's older snapshot (AFTER) - a
+        # time-reversed comparison that flags the base branch's own
+        # refactors as blockers on PRs that never touched those files.
+        # The compare API returns the merge-base without needing local
+        # history, and fetching that single commit by SHA keeps the
+        # shallow (fetch-depth: 1) clone.
         if: github.event_name == 'pull_request'
+        env:
+          GH_TOKEN: ${{ github.token }}
         run: |
-          git fetch --no-tags --depth=1 origin \
-            "${{ github.base_ref }}:refs/remotes/origin/${{ github.base_ref }}"
+          MERGE_BASE=$(gh api \
+            "repos/${{ github.repository }}/compare/${{ github.event.pull_request.base.sha }}...${{ github.event.pull_request.head.sha }}" \
+            --jq .merge_base_commit.sha)
+          git fetch --no-tags --depth=1 origin "$MERGE_BASE"
           mapfile -t CHANGED < <(
             git diff --name-only --diff-filter=M \
-              "origin/${{ github.base_ref }}" HEAD -- '*.py' \
+              "$MERGE_BASE" HEAD -- '*.py' \
               | grep -vE '(^|/)(unsloth_compiled_cache|node_modules|build|dist)/' || true
           )
           if [ "${#CHANGED[@]}" -eq 0 ]; then
             echo "no in-place-modified Python files to check"
             exit 0
           fi
+          printf 'merge base: %s\n' "$MERGE_BASE"
           printf 'checking %d file(s):\n' "${#CHANGED[@]}"
           printf '  %s\n' "${CHANGED[@]}"
           python scripts/verify_import_hoist.py \
-            --before "origin/${{ github.base_ref }}" --after HEAD "${CHANGED[@]}"
+            --before "$MERGE_BASE" --after HEAD "${CHANGED[@]}"
 
       - name: No leftover debugger / pdb / breakpoint calls
         # Catches the "I'll just stick a breakpoint() here" mistake
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 75940832a0..221e86f235 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -231,33 +231,14 @@ jobs:
             tests/studio/test_is_mlx_dispatch_gate.py \
             tests/studio/test_mlx_training_worker_behaviors.py
 
-      # Studio prebuilt llama.cpp install + GGUF inference. Drives the
-      # exact path Studio's setup.sh takes on macOS: invokes
-      # studio/install_llama_prebuilt.py with --published-repo
-      # ggml-org/llama.cpp and --published-release-tag b9049 (the
-      # latest llama.cpp release at the time this step was added; bump
-      # via UNSLOTH_LLAMA_TAG / DEFAULT_LLAMA_TAG when refreshing).
-      # The installer downloads llama-b9049-bin-macos-arm64.tar.gz,
-      # which is the universal Apple Silicon (arm64) build -- the
-      # same artifact works on M1/M2/M3/M4 because llama.cpp compiles
-      # against the ARMv8.2 baseline.
-      #
-      # The b9049 release also publishes:
-      #   - llama-b9049-bin-macos-arm64-kleidiai.tar.gz
-      #     KleidiAI dispatches at runtime; on M1 it falls back where
-      #     ISA features (e.g. I8MM) are missing, so this asset also
-      #     runs on M1 -- Studio just doesn't choose it by default.
-      #   - llama-b9049-bin-macos-x64.tar.gz
-      #     Intel-only; would only run on M1 via Rosetta 2 emulation,
-      #     which we explicitly avoid.
-      #   - iOS XCFramework
-      #     iOS-app build artifact, unrelated to a macOS desktop CI.
-      #
-      # After install, downloads a small published GGUF
-      # (unsloth/gemma-3-270m-it-GGUF, Q4_K_M) from HuggingFace and
-      # runs the prebuilt llama-cli on it. Asserts the prompt echo
-      # appears in stdout. If the install fails OR the binary exits
-      # non-zero, that's an Unsloth/Studio bug.
+      # Studio prebuilt llama.cpp install + GGUF inference. Mirrors the
+      # path Studio's setup.sh takes on macOS since #5963: plan against
+      # the unslothai/llama.cpp fork's latest release, which ships the
+      # bin-macos-arm64 bundle plus the llama-prebuilt-manifest.json the
+      # default policy reads. After install, downloads a small published
+      # GGUF (unsloth/gemma-3-270m-it-GGUF, Q4_K_M) and validates
+      # llama-server /completion end to end. An install failure or a
+      # non-zero binary exit is an Unsloth/Studio bug.
       - name: Studio prebuilt llama.cpp install + GGUF inference (Mac M1)
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
@@ -272,20 +253,12 @@ jobs:
           set -euo pipefail
           INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
           rm -rf "$INSTALL_DIR"
-          # --simple-policy is required when --published-repo points
-          # at upstream ggml-org/llama.cpp; that repo doesn't ship the
-          # llama-prebuilt-manifest.json asset Studio's default policy
-          # expects, so the simple platform-specific policy maps
-          # Darwin+arm64 -> bin-macos-arm64 directly. studio/setup.sh
-          # passes both --published-repo ggml-org/llama.cpp AND
-          # --simple-policy automatically on macOS, so this CI step
-          # exercises the same code path users hit when they run
-          # `curl -fsSL https://unsloth.ai/install.sh | sh`.
+          # Mirror studio/setup.sh on macOS (the install.sh user path):
+          # it plans against the unslothai/llama.cpp fork's latest
+          # release with no policy or tag flags.
           python studio/install_llama_prebuilt.py \
             --install-dir "$INSTALL_DIR" \
-            --published-repo ggml-org/llama.cpp \
-            --published-release-tag b9049 \
-            --simple-policy
+            --published-repo unslothai/llama.cpp
 
           # Studio bundles only llama-server + llama-quantize from the
           # prebuilt (not llama-cli) -- inference goes through
diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml
index 53514e2ce1..b196805cf7 100644
--- a/.github/workflows/studio-api-smoke.yml
+++ b/.github/workflows/studio-api-smoke.yml
@@ -77,7 +77,7 @@ jobs:
           path: hf-cache
           # Same key as studio-ui-smoke.yml so the two jobs share a
           # single GGUF download across CI.
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -88,17 +88,19 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-frontend-ci.yml b/.github/workflows/studio-frontend-ci.yml
index 1270a57ef6..b42086f191 100644
--- a/.github/workflows/studio-frontend-ci.yml
+++ b/.github/workflows/studio-frontend-ci.yml
@@ -17,6 +17,8 @@ on:
       - 'studio/frontend/**'
       - 'scripts/check_frontend_dep_removal.py'
       - 'tests/studio/test_frontend_dep_removal.py'
+      - 'scripts/sync_allow_scripts_pins.py'
+      - 'tests/studio/test_sync_allow_scripts_pins.py'
       - '.github/workflows/studio-frontend-ci.yml'
   push:
     branches: [main, pip]
@@ -60,6 +62,19 @@ jobs:
         with:
           node-version: '22'
 
+      # node 22 bundles npm 10.x, which predates allowScripts. Move to the
+      # 11.x line and fail loudly if the gate is still missing, so the
+      # strict flag below can never silently degrade into a warning.
+      - name: Upgrade npm to 11.x (allowScripts enforcement)
+        working-directory: ${{ github.workspace }}
+        run: |
+          npm install -g npm@^11 --no-fund --no-audit
+          V=$(npm -v)
+          case "$V" in
+            11.1[6-9].*|11.[2-9][0-9].*|1[2-9].*) echo "npm $V has allowScripts" ;;
+            *) echo "::error::npm $V lacks allowScripts (need >=11.16)"; exit 1 ;;
+          esac
+
       # Run the structural lockfile scan BEFORE npm ci. A compromised
       # tarball runs its `prepare` / `postinstall` during `npm ci`,
       # so any catch has to fire upstream of that. The scanner is
@@ -68,14 +83,23 @@ jobs:
         working-directory: ${{ github.workspace }}
         run: python3 scripts/lockfile_supply_chain_audit.py
 
+      # Dependency bumps strand the version-pinned allowScripts entries.
+      # The paired pre-commit hook auto-fixes PRs; this is the backstop.
+      - name: allowScripts pins must match the lockfile
+        working-directory: ${{ github.workspace }}
+        run: |
+          python3 tests/studio/test_sync_allow_scripts_pins.py
+          python3 scripts/sync_allow_scripts_pins.py --check
+
       - name: Lockfile must agree with package.json (npm ci is strict)
-        # Lifecycle scripts (esbuild native-binary postinstall, etc.) are
-        # required for `vite build`. The pre-install lockfile structural
-        # audit (lockfile_supply_chain_audit.py) is the practical defence
-        # against the npm postinstall-dropper class -- it fires BEFORE any
-        # tarball runs, on the injection pattern itself rather than an
-        # advisory-DB lookup.
-        run: npm ci --no-fund --no-audit
+        # The vite 8 chain (rolldown, lightningcss, tailwind oxide) ships napi
+        # binaries with no install scripts. The only script-bearing deps are
+        # covered by `allowScripts` in package.json (npm >=11.16, default in
+        # npm 12). The pre-install lockfile audit above stays the first line
+        # of defence -- it fires before any tarball can run code.
+        # --strict-allow-scripts: any unreviewed install script hard-fails
+        # the job; the sync hook keeps the pins fresh after bumps.
+        run: npm ci --strict-allow-scripts --no-fund --no-audit
 
       - name: npm ci must not have modified the working tree
         working-directory: ${{ github.workspace }}
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index 3a4c76a2bb..cffb33f71d 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -20,7 +20,7 @@
 #        enable_tools / enabled_tools, and enable_thinking on/off.
 #
 #   3. JSON, images
-#        Qwen3-VL-2B-Instruct UD-IQ2_XXS (~570 MiB) + mmproj-F16 (~780 MiB).
+#        Qwen3-VL-2B-Instruct UD-Q4_K_XL (~1.1 GiB) + mmproj-F16 (~780 MiB).
 #        response_format JSON-schema decoding and OpenAI image_url
 #        (data URI) plus Anthropic source/base64 image inputs.
 #
@@ -91,7 +91,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -102,17 +102,19 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -375,6 +377,7 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -792,8 +795,14 @@ jobs:
     timeout-minutes: 30
     env:
       GGUF_REPO: unsloth/Qwen3-VL-2B-Instruct-GGUF
-      GGUF_VARIANT: UD-IQ2_XXS
-      GGUF_FILE: Qwen3-VL-2B-Instruct-UD-IQ2_XXS.gguf
+      # UD-Q4_K_XL, not UD-IQ2_XXS: at 2-bit the temp-0 answer to the JSON
+      # step's capital-of-France probe flips with the host's SIMD kernels
+      # (GitHub runners deterministically answered France while other CPUs
+      # answer Paris; seeds do not rescue it, 1/5 Paris at temp 0.7). The
+      # Q4 quant answered Paris 13/13 across temps and seeds on the same
+      # runners, so the hard Paris assertion below stays reliable.
+      GGUF_VARIANT: UD-Q4_K_XL
+      GGUF_FILE: Qwen3-VL-2B-Instruct-UD-Q4_K_XL.gguf
       MMPROJ_FILE: mmproj-F16.gguf
       STUDIO_PORT: '18890'
       HF_HOME: ${{ github.workspace }}/hf-cache
@@ -823,7 +832,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2
 
       - name: Prime HF_HOME with the GGUF + mmproj
         id: prime-hf
@@ -835,17 +844,19 @@ jobs:
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$MMPROJ_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }} (model + mmproj)
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml
index 8839b559fd..412726538c 100644
--- a/.github/workflows/studio-mac-api-smoke.yml
+++ b/.github/workflows/studio-mac-api-smoke.yml
@@ -62,7 +62,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -73,17 +73,19 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index e08ac6ca68..c794a34acd 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -85,7 +85,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -96,6 +96,7 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       # Save partial caches on cancel/timeout -- hf download resumes by
       # content hash. `outcome != skipped` keeps cache-hit a no-op.
@@ -104,11 +105,12 @@ jobs:
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -361,6 +363,7 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -749,6 +752,7 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-install-matrix.yml b/.github/workflows/studio-mac-install-matrix.yml
index 4e2722d1cd..da944d4b5c 100644
--- a/.github/workflows/studio-mac-install-matrix.yml
+++ b/.github/workflows/studio-mac-install-matrix.yml
@@ -63,6 +63,7 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 0176b0a168..4f9f94b534 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -62,7 +62,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -73,17 +73,19 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-update-smoke.yml b/.github/workflows/studio-mac-update-smoke.yml
index 1eca227096..f554a16415 100644
--- a/.github/workflows/studio-mac-update-smoke.yml
+++ b/.github/workflows/studio-mac-update-smoke.yml
@@ -62,6 +62,7 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -73,6 +74,7 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -91,6 +93,7 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml
index 455fe4b7e1..de106e201f 100644
--- a/.github/workflows/studio-ui-smoke.yml
+++ b/.github/workflows/studio-ui-smoke.yml
@@ -76,7 +76,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -87,17 +87,19 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-update-smoke.yml b/.github/workflows/studio-update-smoke.yml
index 057aeacbd4..307bb51972 100644
--- a/.github/workflows/studio-update-smoke.yml
+++ b/.github/workflows/studio-update-smoke.yml
@@ -71,6 +71,7 @@ jobs:
         # prebuilt path falls back to source build.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -85,6 +86,7 @@ jobs:
         # idempotency regressed.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -107,6 +109,7 @@ jobs:
         # the first one.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml
index eee61516c7..78efe918ac 100644
--- a/.github/workflows/studio-windows-api-smoke.yml
+++ b/.github/workflows/studio-windows-api-smoke.yml
@@ -69,7 +69,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -80,13 +80,14 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
         shell: pwsh
@@ -123,6 +124,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 6148856016..a772a6d102 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -101,7 +101,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -114,6 +114,7 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME cache for ${{ env.GGUF_REPO }}
         # Only write a fresh cache entry when we actually rebuilt the
@@ -123,7 +124,7 @@ jobs:
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
         shell: pwsh
@@ -160,6 +161,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -504,6 +506,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -879,7 +882,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2
 
       - name: Prime HF_HOME with the GGUF + mmproj
         id: prime-hf
@@ -891,13 +894,14 @@ jobs:
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$MMPROJ_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME cache for ${{ env.GGUF_REPO }} (model + mmproj)
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2
 
       - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
         shell: pwsh
@@ -934,6 +938,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml
index e3c8642122..40d8e530cd 100644
--- a/.github/workflows/studio-windows-ui-smoke.yml
+++ b/.github/workflows/studio-windows-ui-smoke.yml
@@ -85,7 +85,7 @@ jobs:
         continue-on-error: true
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Prime HF_HOME with the GGUF
         id: prime-hf
@@ -96,13 +96,14 @@ jobs:
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
           bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
 
       - name: Save HF_HOME for ${{ env.GGUF_REPO }}
         if: always() && steps.prime-hf.outcome == 'success'
         uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
         with:
           path: hf-cache
-          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v1
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
       - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
         shell: pwsh
@@ -143,6 +144,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 redirects ALL PowerShell streams (stdout, stderr,
diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index aa3e35f052..4a4806cfb1 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -133,6 +133,7 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -179,6 +180,7 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -197,6 +199,7 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index 56d32c9312..cffbf73cd5 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -19,3 +19,15 @@ repos:
         exclude: '(chat_templates|ollama_template_mappers|_auto_install|mapper)\.py$'
         additional_dependencies:
           - ruff==0.6.9
+      # Re-pins allowScripts entries after dependency bumps. pre-commit.ci
+      # pushes the fix to PR branches, Dependabot's included, so stale pins
+      # heal without a human in the loop.
+      - id: sync-allow-scripts-pins
+        name: Sync allowScripts pins with the frontend lockfile
+        # `python <script>` not a direct exec: autofix commits can drop the
+        # executable bit, which kills shebang-style entries.
+        entry: python scripts/sync_allow_scripts_pins.py
+        args: [--fix]
+        language: python
+        files: ^studio/frontend/(package\.json|package-lock\.json)$
+        pass_filenames: false
diff --git a/install.ps1 b/install.ps1
index 57887a1ed4..d2312c98ec 100644
--- a/install.ps1
+++ b/install.ps1
@@ -158,6 +158,10 @@ function Install-UnslothStudio {
 
     # UNSLOTH_PYTHON pins the version (mirrors install.sh --python); default 3.13.
     $PythonVersion = if ($env:UNSLOTH_PYTHON) { $env:UNSLOTH_PYTHON } else { "3.13" }
+    # python.org fallback patch, used only when winget is unavailable/broken AND
+    # the live python.org listing can't be fetched. The installer URL scheme is
+    # stable so an older patch still installs. Bump alongside $PythonVersion.
+    $PythonFallbackFullVersion = "3.13.13"
 
     # Resolve install destinations. Priority: UNSLOTH_STUDIO_HOME, then
     # STUDIO_HOME alias, then USERPROFILE-redirect, then default.
@@ -868,6 +872,7 @@ shell.Run cmd, 0, False
             try {
                 $wshell = New-Object -ComObject WScript.Shell
                 $createdShortcutCount = 0
+                $createdShortcutPaths = @()
                 foreach ($linkPath in @($desktopLink, $startMenuLink)) {
                     if (-not $linkPath -or [string]::IsNullOrWhiteSpace($linkPath)) { continue }
                     try {
@@ -881,12 +886,46 @@ shell.Run cmd, 0, False
                         }
                         $shortcut.Save()
                         $createdShortcutCount++
+                        $createdShortcutPaths += $linkPath
                     } catch {
                         substep "could not create shortcut at ${linkPath}: $($_.Exception.Message)" "Yellow"
                     }
                 }
                 if ($createdShortcutCount -gt 0) {
                     substep "Created Unsloth Studio shortcut"
+                    # Force Explorer to re-read each new shortcut's icon so it renders
+                    # immediately instead of a stale/generic entry (a same-name .lnk
+                    # recreated across reinstalls keeps Explorer's cached per-item icon).
+                    # The reliable, non-disruptive fix (no explorer restart) is a per-item
+                    # SHChangeNotify SHCNE_UPDATEITEM + SHCNF_PATHW per .lnk; the global
+                    # SHCNE_ASSOCCHANGED broadcast alone does NOT recover a stale item.
+                    # Also clear the on-disk icon cache (covers heavier staleness).
+                    try { & "$env:SystemRoot\System32\ie4uinit.exe" -ClearIconCache 2>$null } catch {}
+                    try { & "$env:SystemRoot\System32\ie4uinit.exe" -show 2>$null } catch {}
+                    try {
+                        Add-Type -Namespace UnslothShell -Name IconRefresh -MemberDefinition '[System.Runtime.InteropServices.DllImport("shell32.dll", CharSet = System.Runtime.InteropServices.CharSet.Unicode)] public static extern void SHChangeNotify(int eventId, uint flags, string item1, System.IntPtr item2);' -ErrorAction SilentlyContinue
+                        # SHCNE_UPDATEITEM (0x00002000) + SHCNF_PATHW (0x0005) per shortcut
+                        foreach ($scPath in $createdShortcutPaths) {
+                            try { [UnslothShell.IconRefresh]::SHChangeNotify(0x00002000, 0x0005, $scPath, [System.IntPtr]::Zero) } catch {}
+                        }
+                        # SHCNE_ASSOCCHANGED (0x08000000) global refresh (belt-and-suspenders)
+                        [UnslothShell.IconRefresh]::SHChangeNotify(0x08000000, 0, $null, [System.IntPtr]::Zero)
+                    } catch {}
+                    # Win11's Start Menu (StartMenuExperienceHost) keeps its OWN
+                    # pre-rendered tile-icon cache that ie4uinit/explorer restart do NOT
+                    # invalidate, so a rewritten same-name shortcut shows the old tile
+                    # until the host restarts. Drop only the render caches (NEVER
+                    # start2.bin -- the pinned layout) and let the host rebuild.
+                    # Best-effort; Win10 has no such host (Test-Path skips it).
+                    try {
+                        $smehTemp = Join-Path $env:LOCALAPPDATA "Packages\Microsoft.Windows.StartMenuExperienceHost_cw5n1h2txyewy\TempState"
+                        if (Test-Path -LiteralPath $smehTemp) {
+                            Get-ChildItem -LiteralPath $smehTemp -Filter "TileCache_*" -ErrorAction SilentlyContinue |
+                                Remove-Item -Force -ErrorAction SilentlyContinue
+                            Remove-Item -LiteralPath (Join-Path $smehTemp "StartUnifiedTileModelCache.dat") -Force -ErrorAction SilentlyContinue
+                            Stop-Process -Name StartMenuExperienceHost -Force -ErrorAction SilentlyContinue
+                        }
+                    } catch {}
                 } else {
                     substep "no Unsloth Studio shortcuts were created" "Yellow"
                 }
@@ -958,10 +997,13 @@ shell.Run cmd, 0, False
     function Find-CompatiblePython {
         # Try the Python Launcher first (most reliable on Windows)
         # py.exe resolves to the standard CPython install, not conda.
-        $pyLauncher = Get-Command py -CommandType Application -ErrorAction SilentlyContinue
-        if ($pyLauncher -and $pyLauncher.Source -notmatch $script:CondaSkipPattern) {
-            # Prefer the requested $PythonVersion, then newest-first fallback.
-            $minors = @($PythonVersion) + (@("3.13", "3.12", "3.11") | Where-Object { $_ -ne $PythonVersion })
+        # Prefer the requested $PythonVersion, then newest-first fallback.
+        $minors = @($PythonVersion) + (@("3.13", "3.12", "3.11") | Where-Object { $_ -ne $PythonVersion })
+        # Enumerate every py.exe on PATH with -All (Windows PowerShell 5.1
+        # returns only the first launcher without it) and search each for a
+        # supported, non-conda interpreter.
+        foreach ($pyLauncher in @(Get-Command py -All -CommandType Application -ErrorAction SilentlyContinue)) {
+            if ($pyLauncher.Source -match $script:CondaSkipPattern) { continue }
             foreach ($minor in $minors) {
                 try {
                     $out = & $pyLauncher.Source "-$minor" --version 2>&1 | Out-String
@@ -999,6 +1041,79 @@ shell.Run cmd, 0, False
         return $null
     }
 
+    # ── Fallback: install CPython directly from python.org ──
+    # Used when winget is unavailable or fails (notably msstore cert-pinning error
+    # 0x8a15005e, which aborts `winget install` unless --source winget is given).
+    # Downloads the official installer and runs it silently as a per-user install
+    # (no UAC), putting python.exe + the py launcher on PATH. Mirrors the uv ->
+    # astral.sh fallback below. Returns @{ Version; Path } or $null.
+    function Install-PythonFromPythonOrg {
+        # python.org ships one installer per architecture.
+        $archSuffix = switch (Get-TauriDiagArch) {
+            "x86_64" { "-amd64" }
+            "arm64"  { "-arm64" }
+            "x86"    { "" }
+            default  { $null }
+        }
+        if ($null -eq $archSuffix) {
+            substep "No python.org installer is available for this architecture." "Yellow"
+            return $null
+        }
+
+        # Resolve the latest $PythonVersion.x patch from the python.org listing,
+        # falling back to a same-minor version if the listing cannot be fetched.
+        # Use the pinned full version only when it matches the requested minor so a
+        # non-default UNSLOTH_PYTHON (e.g. 3.12) doesn't silently install 3.13.
+        $full = if ($PythonFallbackFullVersion -like "$PythonVersion.*") { $PythonFallbackFullVersion } else { "$PythonVersion.0" }
+        try {
+            $listing = [string](Invoke-RestMethod -Uri "https://www.python.org/ftp/python/" -UseBasicParsing -TimeoutSec 20)
+            $patches = [regex]::Matches($listing, ([regex]::Escape($PythonVersion) + '\.(\d+)/')) |
+                ForEach-Object { [int]$_.Groups[1].Value } | Sort-Object -Descending
+            if ($patches.Count -gt 0) { $full = "$PythonVersion.$($patches[0])" }
+        } catch {}
+
+        $file = "python-$full$archSuffix.exe"
+        $url  = "https://www.python.org/ftp/python/$full/$file"
+        $dest = Join-Path ([System.IO.Path]::GetTempPath()) $file
+        substep "downloading Python $full from python.org..." "Yellow"
+        try {
+            Invoke-WebRequest -Uri $url -OutFile $dest -UseBasicParsing
+        } catch {
+            substep "python.org download failed: $($_.Exception.Message)" "Yellow"
+            return $null
+        }
+
+        # Per-user install => no UAC. PrependPath puts python + py on PATH;
+        # Include_launcher installs py.exe (preferred by Find-CompatiblePython).
+        substep "installing Python $full (silent, per-user)..."
+        $installArgs = @(
+            "/quiet",
+            "InstallAllUsers=0",
+            "PrependPath=1",
+            "Include_launcher=1",
+            # Launcher per-user too: Include_launcher defaults InstallLauncherAllUsers=1,
+            # which needs admin and would break this non-admin per-user fallback.
+            "InstallLauncherAllUsers=0",
+            "Include_pip=1",
+            "AssociateFiles=0",
+            "Shortcuts=0"
+        )
+        $rc = 1
+        try {
+            $proc = Start-Process -FilePath $dest -ArgumentList $installArgs -Wait -PassThru
+            $rc = $proc.ExitCode
+        } catch {
+            substep "python.org installer failed to start: $($_.Exception.Message)" "Yellow"
+        } finally {
+            Remove-Item -LiteralPath $dest -Force -ErrorAction SilentlyContinue
+        }
+        if ($rc -ne 0) {
+            substep "python.org installer exited with code $rc." "Yellow"
+        }
+        Refresh-SessionPath
+        return (Find-CompatiblePython)
+    }
+
     # ── Install Python if no compatible version (3.11-3.13) found ──
     # Find-CompatiblePython returns @{ Version = "3.13"; Path = "C:\...\python.exe" } or $null.
     Write-TauriLog "STEP" "Installing Python"
@@ -1008,47 +1123,62 @@ shell.Run cmd, 0, False
         step "python" "Python $($DetectedPython.Version) already installed"
     }
     if (-not $DetectedPython) {
-        if (-not $script:WingetAvailable) {
-            Write-Host "[ERROR] No compatible Python (3.11-3.13) found and winget is unavailable on this host." -ForegroundColor Red
-            Write-Host "        Install Python $PythonVersion from https://www.python.org/downloads/" -ForegroundColor Yellow
-            Write-Host "        and re-run this installer (make sure 'Add Python to PATH' is checked)." -ForegroundColor Yellow
-            return (Exit-InstallFailure "winget required to install Python on this host")
-        }
         substep "installing Python ${PythonVersion}..."
         $pythonPackageId = "Python.Python.$PythonVersion"
-        # Temporarily lower ErrorActionPreference so that winget stderr
-        # (progress bars, warnings) does not become a terminating error
-        # on PowerShell 5.1 where native-command stderr is ErrorRecord.
-        $prevEAP = $ErrorActionPreference
-        $ErrorActionPreference = "Continue"
-        try {
-            winget install -e --id $pythonPackageId --accept-package-agreements --accept-source-agreements
-            $wingetExit = $LASTEXITCODE
-        } catch { $wingetExit = 1 }
-        $ErrorActionPreference = $prevEAP
-        Refresh-SessionPath
+        $wingetExit = $null
 
-        # Re-detect after install (PATH may have changed)
-        $DetectedPython = Find-CompatiblePython
-
-        if (-not $DetectedPython) {
-            # Python still not functional after winget -- force reinstall.
-            # This handles both real failures AND "already installed" codes where
-            # winget thinks Python is present but it's not actually on PATH
-            # (e.g. user partially uninstalled, or installed via a different method).
-            substep "Python not found on PATH after winget. Retrying with --force..." "Yellow"
+        if ($script:WingetAvailable) {
+            # --source winget avoids the msstore source, which can fail with
+            # cert-pinning error 0x8a15005e and abort the whole `winget install`
+            # (winget then demands --source). Python and uv both live in the
+            # winget source, so pinning it is correct and faster.
+            #
+            # Lower ErrorActionPreference so winget stderr (progress/warnings) is
+            # not a terminating error on PS 5.1 (native stderr is ErrorRecord).
+            $prevEAP = $ErrorActionPreference
             $ErrorActionPreference = "Continue"
             try {
-                winget install -e --id $pythonPackageId --accept-package-agreements --accept-source-agreements --force
+                winget install -e --id $pythonPackageId --source winget --accept-package-agreements --accept-source-agreements
                 $wingetExit = $LASTEXITCODE
             } catch { $wingetExit = 1 }
             $ErrorActionPreference = $prevEAP
             Refresh-SessionPath
+
+            # Re-detect after install (PATH may have changed)
             $DetectedPython = Find-CompatiblePython
+
+            if (-not $DetectedPython) {
+                # Python still not functional after winget -- force reinstall.
+                # This handles both real failures AND "already installed" codes where
+                # winget thinks Python is present but it's not actually on PATH
+                # (e.g. user partially uninstalled, or installed via a different method).
+                substep "Python not found on PATH after winget. Retrying with --force..." "Yellow"
+                $ErrorActionPreference = "Continue"
+                try {
+                    winget install -e --id $pythonPackageId --source winget --accept-package-agreements --accept-source-agreements --force
+                    $wingetExit = $LASTEXITCODE
+                } catch { $wingetExit = 1 }
+                $ErrorActionPreference = $prevEAP
+                Refresh-SessionPath
+                $DetectedPython = Find-CompatiblePython
+            }
+        }
+
+        # Fall back to python.org if winget is unavailable OR couldn't install a
+        # working Python (missing/broken winget, msstore cert errors --source
+        # winget can't fix). Keeps the install automatic instead of failing out.
+        if (-not $DetectedPython) {
+            if ($script:WingetAvailable) {
+                substep "winget could not install Python -- falling back to python.org..." "Yellow"
+            } else {
+                substep "winget is unavailable -- installing Python from python.org..." "Yellow"
+            }
+            $DetectedPython = Install-PythonFromPythonOrg
         }
 
         if (-not $DetectedPython) {
-            Write-Host "[ERROR] Python installation failed (exit code $wingetExit)" -ForegroundColor Red
+            $exitNote = if ($null -ne $wingetExit) { " (winget exit code $wingetExit)" } else { "" }
+            Write-Host "[ERROR] Python installation failed$exitNote" -ForegroundColor Red
             Write-Host "        Please install Python $PythonVersion manually from https://www.python.org/downloads/" -ForegroundColor Yellow
             Write-Host "        Make sure to check 'Add Python to PATH' during installation." -ForegroundColor Yellow
             Write-Host "        Then re-run this installer." -ForegroundColor Yellow
@@ -1068,7 +1198,7 @@ shell.Run cmd, 0, False
         if ($script:WingetAvailable) {
             $prevEAP = $ErrorActionPreference
             $ErrorActionPreference = "Continue"
-            try { winget install --id=astral-sh.uv -e --accept-package-agreements --accept-source-agreements } catch {}
+            try { winget install --id=astral-sh.uv -e --source winget --accept-package-agreements --accept-source-agreements } catch {}
             $ErrorActionPreference = $prevEAP
             Refresh-SessionPath
         }
@@ -1244,14 +1374,107 @@ shell.Run cmd, 0, False
         try { [System.IO.File]::WriteAllText((Join-Path $VenvDir ".unsloth-studio-owned"), "") } catch {}
     }
 
+    # ── Helper: run amd-smi without triggering a UAC elevation prompt ──
+    # amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing
+    # DiskPart UAC prompt mid-install (Studio backend amd.py hits the same).
+    # __COMPAT_LAYER=RunAsInvoker forces it (and helpers it spawns) to run
+    # un-elevated; on failure the WMI name -> gfx fallback still resolves the arch.
+    function Invoke-AmdSmiNoElevate {
+        param(
+            [Parameter(Mandatory = $true, Position = 0)][string]$Exe,
+            [Parameter(Position = 1)][string[]]$SmiArgs = @(),
+            [int]$TimeoutSec = 30
+        )
+        # RunAsInvoker blocks the auto-elevation/UAC prompt; the timeout bounds a
+        # flaky amd-smi that can otherwise spin for minutes (30s mirrors amd.py).
+        $prevCompat = [Environment]::GetEnvironmentVariable('__COMPAT_LAYER', 'Process')
+        $env:__COMPAT_LAYER = 'RunAsInvoker'
+        try {
+            # [Process]::Start, NOT Start-Process -PassThru: the latter leaves
+            # .ExitCode $null after WaitForExit on PS 5.1, so $LASTEXITCODE (checked
+            # by callers) reads non-zero and kills detection. Async reads drain the
+            # pipes (no deadlock); amd-smi args have no spaces so a plain join is safe.
+            $psi = New-Object System.Diagnostics.ProcessStartInfo
+            $psi.FileName = $Exe
+            $psi.Arguments = ($SmiArgs -join ' ')
+            $psi.UseShellExecute = $false
+            $psi.RedirectStandardOutput = $true
+            $psi.RedirectStandardError = $true
+            $psi.CreateNoWindow = $true
+            $proc = [System.Diagnostics.Process]::Start($psi)
+            $outTask = $proc.StandardOutput.ReadToEndAsync()
+            $errTask = $proc.StandardError.ReadToEndAsync()
+            if (-not $proc.WaitForExit($TimeoutSec * 1000)) {
+                try { $proc.Kill() } catch {}
+                $global:LASTEXITCODE = 124
+                return ""
+            }
+            $global:LASTEXITCODE = $proc.ExitCode
+            return ($outTask.Result + "`n" + $errTask.Result)
+        } catch {
+            $global:LASTEXITCODE = 1
+            return ""
+        } finally {
+            if ($null -eq $prevCompat) {
+                Remove-Item Env:__COMPAT_LAYER -ErrorAction SilentlyContinue
+            } else {
+                $env:__COMPAT_LAYER = $prevCompat
+            }
+        }
+    }
+
+    # ── Helper: run nvidia-smi under a timeout ──
+    # A wedged NVIDIA driver can make nvidia-smi block during init or after a
+    # reset; WaitForExit bounds it (mirrors Invoke-AmdSmiNoElevate) so detection
+    # cannot hang the installer. No RunAsInvoker compat layer: nvidia-smi does
+    # not auto-elevate. Returns combined stdout+stderr; "" on timeout/failure.
+    function Invoke-NvidiaSmiBounded {
+        param(
+            [Parameter(Mandatory = $true, Position = 0)][string]$Exe,
+            [Parameter(Position = 1)][string[]]$SmiArgs = @(),
+            [int]$TimeoutSec = 10
+        )
+        try {
+            $psi = New-Object System.Diagnostics.ProcessStartInfo
+            $psi.FileName = $Exe
+            $psi.Arguments = ($SmiArgs -join ' ')
+            $psi.UseShellExecute = $false
+            $psi.RedirectStandardOutput = $true
+            $psi.RedirectStandardError = $true
+            $psi.CreateNoWindow = $true
+            $proc = [System.Diagnostics.Process]::Start($psi)
+            $outTask = $proc.StandardOutput.ReadToEndAsync()
+            $errTask = $proc.StandardError.ReadToEndAsync()
+            if (-not $proc.WaitForExit($TimeoutSec * 1000)) {
+                try { $proc.Kill() } catch {}
+                $global:LASTEXITCODE = 124
+                return ""
+            }
+            $global:LASTEXITCODE = $proc.ExitCode
+            return ($outTask.Result + "`n" + $errTask.Result)
+        } catch {
+            $global:LASTEXITCODE = 1
+            return ""
+        }
+    }
+
+    # ── Helper: nvidia-smi -L lists at least one real GPU ──
+    # Exit code 0 alone is not enough: a stale/driverless nvidia-smi can exit 0
+    # while listing no GPU, which would mark an AMD host NVIDIA and suppress
+    # ROCm detection. Require a "GPU <n>:" data row.
+    function Test-NvidiaSmiHasGpu {
+        param([Parameter(Mandatory = $true)][string]$Exe)
+        $out = Invoke-NvidiaSmiBounded $Exe @('-L')
+        return ($LASTEXITCODE -eq 0 -and $out -match '(?m)^GPU\s+\d+:')
+    }
+
     # ── Detect GPU (robust: PATH + hardcoded fallback paths, mirrors setup.ps1) ──
     $HasNvidiaSmi = $false
     $NvidiaSmiExe = $null
     try {
         $nvSmiCmd = Get-Command nvidia-smi -ErrorAction SilentlyContinue
-        if ($nvSmiCmd) {
-            & $nvSmiCmd.Source *> $null
-            if ($LASTEXITCODE -eq 0) { $HasNvidiaSmi = $true; $NvidiaSmiExe = $nvSmiCmd.Source }
+        if ($nvSmiCmd -and (Test-NvidiaSmiHasGpu $nvSmiCmd.Source)) {
+            $HasNvidiaSmi = $true; $NvidiaSmiExe = $nvSmiCmd.Source
         }
     } catch {}
     if (-not $HasNvidiaSmi) {
@@ -1261,8 +1484,7 @@ shell.Run cmd, 0, False
         )) {
             if (Test-Path $p) {
                 try {
-                    & $p *> $null
-                    if ($LASTEXITCODE -eq 0) { $HasNvidiaSmi = $true; $NvidiaSmiExe = $p; break }
+                    if (Test-NvidiaSmiHasGpu $p) { $HasNvidiaSmi = $true; $NvidiaSmiExe = $p; break }
                 } catch {}
             }
         }
@@ -1317,11 +1539,21 @@ shell.Run cmd, 0, False
                 }
             } catch {}
         }
-        if (-not $HasROCm) {
+        # On hosts without a working HIP runtime amd-smi elevates a child at runtime,
+        # popping a UAC/DiskPart prompt RunAsInvoker can't suppress (manifest is
+        # asInvoker). So only probe when a HIP SDK is present (hipinfo found ->
+        # un-elevated) or the user opts in; else fall through to WMI name inference
+        # (enough to pick ROCm wheels + lemonade llama.cpp).
+        # An explicit opt-out (UNSLOTH_ENABLE_AMD_SMI=0/false/no/off) wins over the
+        # HIP-SDK heuristic: a HIP SDK binary with a broken runtime can still pop the
+        # prompt, so $HipSdkInstalled must NOT silently re-enable it.
+        $amdSmiOptOut = $env:UNSLOTH_ENABLE_AMD_SMI -match '^(?i)(0|false|no|off)$'
+        $amdSmiAllowed = (-not $amdSmiOptOut) -and ($HipSdkInstalled -or ($env:UNSLOTH_ENABLE_AMD_SMI -match '^(?i)(1|true|yes|on)$'))
+        if (-not $HasROCm -and $amdSmiAllowed) {
             $amdSmiExe = Get-Command "amd-smi" -ErrorAction SilentlyContinue
             if ($amdSmiExe) {
                 try {
-                    $smiOut = & $amdSmiExe.Source list 2>&1 | Out-String
+                    $smiOut = Invoke-AmdSmiNoElevate $amdSmiExe.Source @('list')
                     if ($LASTEXITCODE -eq 0 -and $smiOut -match "(?im)^GPU\s*[:\[]\s*\d") {
                         $HasROCm = $true
                         # Mirror the hipinfo path: collect all gfx tokens in enumeration
@@ -1336,7 +1568,7 @@ shell.Run cmd, 0, False
                             # Attempt 2: 'static --asic' exposes ASIC details on ROCm 6+,
                             # including the GFX target needed for wheel index selection.
                             $smiAsicOut = ""
-                            try { $smiAsicOut = & $amdSmiExe.Source static --asic 2>&1 | Out-String } catch {}
+                            try { $smiAsicOut = Invoke-AmdSmiNoElevate $amdSmiExe.Source @('static','--asic') } catch {}
                             $_asicGfxTokens = @([regex]::Matches($smiAsicOut, "(?i)\b(gfx\d+[a-z]?)\b") | ForEach-Object { $_.Groups[1].Value.ToLower() })
                             if ($_asicGfxTokens.Count -gt 0) {
                                 $ROCmGfxArch = if ($_smiVisIdx -lt $_asicGfxTokens.Count) { $_asicGfxTokens[$_smiVisIdx] } else { $_asicGfxTokens[0] }
@@ -1360,9 +1592,12 @@ shell.Run cmd, 0, False
             } catch {}
         }
         # ── Arch resolution: env-var override → name inference ──────────────
-        # Covers users whose amd-smi is too old to report the GFX target and
-        # who don't have hipinfo (HIP-runtime-only, common on Strix Halo / iGPU).
-        if ($HasROCm -and -not $ROCmGfxArch) {
+        # Runs even when the hipinfo/amd-smi probe could NOT confirm a runtime
+        # ($HasROCm false): the gfx arch inferred from the WMI GPU name lets the
+        # studio setup forward --rocm-gfx and pull a GPU-accelerated (lemonade)
+        # llama.cpp, which bundles its own ROCm runtime. PyTorch's ROCm wheels
+        # still require a confirmed HIP SDK -- they stay gated on $HasROCm below.
+        if (-not $ROCmGfxArch) {
             # 1. Manual override: set UNSLOTH_ROCM_GFX_ARCH=gfx1151 before running.
             if ($env:UNSLOTH_ROCM_GFX_ARCH) {
                 $ROCmGfxArch = $env:UNSLOTH_ROCM_GFX_ARCH.Trim().ToLower()
@@ -1370,15 +1605,20 @@ shell.Run cmd, 0, False
                 substep "gfx arch from UNSLOTH_ROCM_GFX_ARCH env override: $ROCmGfxArch" "Cyan"
             }
             # 2. Best-effort name → arch lookup from marketing name (amd-smi / WMI).
+            #    Targets only arches the lemonade-sdk ROCm prebuilts cover
+            #    (gfx120X/110X/1151/1150/103X); unknown names fall back to CPU.
             elseif ($ROCmGpuLabel) {
                 $nameArchTable = @(
-                    @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4
-                    @{ P = "9070|9060";                                            A = "gfx1200" }  # RDNA 4
-                    @{ P = "8060S|890M|Strix Halo|HX 37[05]|HX 38[05]|AI 9 HX";  A = "gfx1151" }  # RDNA 3.5 iGPU (Strix Halo / Radeon 8060S retail)
-                    @{ P = "880M|Strix Point|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]"; A = "gfx1150" } # RDNA 3.5 iGPU (Strix Point)
-                    @{ P = "RX 7900|RX 7800|RX 7700(?! S)";                       A = "gfx1100" }  # RDNA 3 desktop
-                    @{ P = "RX 7600";                                              A = "gfx1102" }  # RDNA 3
-                    @{ P = "780M|760M|740M|Phoenix";                               A = "gfx1103" }  # RDNA 3 iGPU (Phoenix)
+                    @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4 (RX 9070 XT / 9080)
+                    @{ P = "9070|9060";                                           A = "gfx1200" }  # RDNA 4 (RX 9070 / 9060)
+                    @{ P = "8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+                    @{ P = "890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1150" }  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+                    @{ P = "RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700"; A = "gfx1100" }  # RDNA 3 desktop/workstation (Navi 31)
+                    @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710"; A = "gfx1102" }  # RDNA 3 (Navi 33)
+                    @{ P = "780M|760M|740M|Phoenix|Hawk Point|Z1 Extreme|Z2 Extreme"; A = "gfx1103" }  # RDNA 3 iGPU (Phoenix / Hawk Point)
+                    @{ P = "RX 6900|RX 6800|RX 6750|RX 6700|PRO W6800|PRO W6900";  A = "gfx1030" }  # RDNA 2 (Navi 21) -- lemonade gfx103X
+                    @{ P = "RX 6650|RX 6600|PRO W6600|PRO W6650";                  A = "gfx1032" }  # RDNA 2 (Navi 23) -- lemonade gfx103X
+                    @{ P = "RX 6500|RX 6400|RX 6300|PRO W6400|PRO W6500";          A = "gfx1034" }  # RDNA 2 (Navi 24) -- lemonade gfx103X
                 )
                 foreach ($row in $nameArchTable) {
                     if ($ROCmGpuLabel -match $row.P) {
@@ -1419,11 +1659,11 @@ shell.Run cmd, 0, False
                     }
                 } catch {}
             }
-            if (-not $ROCmVersion) {
+            if (-not $ROCmVersion -and $amdSmiAllowed) {
                 $amdSmiVer = Get-Command "amd-smi" -ErrorAction SilentlyContinue
                 if ($amdSmiVer) {
                     try {
-                        $smiVerOut = & $amdSmiVer.Source version 2>&1 | Out-String
+                        $smiVerOut = Invoke-AmdSmiNoElevate $amdSmiVer.Source @('version')
                         if ($LASTEXITCODE -eq 0 -and $smiVerOut -match 'ROCm version:\s*(\d+\.\d+)') {
                             $ROCmVersion = $Matches[1]
                         }
@@ -1433,6 +1673,46 @@ shell.Run cmd, 0, False
         }
     }
 
+    # ── Optional WSL-ROCm driver hint ────────────────────────────────────────
+    # An AMD GPU can also be used inside WSL2, but only with Adrenalin >= 26.2.2
+    # (first production ROCDXG/WSL release); native Windows GPU works with any
+    # recent driver. We can't auto-install it (AMD referrer-gates downloads, no
+    # winget package), so just point at AMD's page. Shown only when the installed
+    # driver predates 26.2.2 (Feb 2026); suppress with UNSLOTH_SKIP_AMD_DRIVER_HINT=1.
+    function Show-AmdWslDriverHint {
+        if ($env:UNSLOTH_SKIP_AMD_DRIVER_HINT) { return }
+        try {
+            $amd = Get-CimInstance Win32_VideoController -ErrorAction SilentlyContinue |
+                Where-Object { $_.Name -match 'AMD|Radeon' } | Select-Object -First 1
+            if (-not $amd) { return }
+            $drvDate = $null
+            try {
+                if ($amd.DriverDate -is [datetime]) {
+                    # Get-CimInstance returns DriverDate already parsed.
+                    $drvDate = $amd.DriverDate
+                } elseif ($amd.DriverDate) {
+                    # Get-WmiObject style WMI datetime string.
+                    $drvDate = [Management.ManagementDateTimeConverter]::ToDateTime([string]$amd.DriverDate)
+                }
+            } catch {}
+            # Older than 26.2.2 (Feb 2026) => can't expose the GPU to WSL ROCm.
+            # Unreadable date => still show the hint (informational, suppressible).
+            if ($drvDate -and $drvDate -ge (Get-Date '2026-02-01')) { return }
+            substep "Tip: to use this GPU inside WSL too, install AMD Adrenalin 26.2.2+ (for WSL2)." "Cyan"
+            substep "  Your current driver predates it; native Windows GPU is unaffected. Get it from AMD:" "Cyan"
+            substep "    https://www.amd.com/en/resources/support-articles/release-notes/RN-RAD-WIN-26-2-2.html" "Cyan"
+            substep "  Then reboot and run this installer inside an Ubuntu-24.04 WSL distro." "Cyan"
+            # If WSL isn't installed yet, point at the command that provisions it
+            # (best-effort; wsl.exe absent => no WSL).
+            $hasWsl = $false
+            try { $hasWsl = [bool](Get-Command wsl.exe -ErrorAction SilentlyContinue) } catch {}
+            if (-not $hasWsl) {
+                substep "  No WSL yet? Install it first:  wsl --install -d Ubuntu-24.04" "Cyan"
+            }
+            substep "  (suppress: set UNSLOTH_SKIP_AMD_DRIVER_HINT=1)" "Cyan"
+        } catch {}
+    }
+
     if ($HasNvidiaSmi) {
         step "gpu" "NVIDIA GPU detected"
     } elseif ($HasROCm) {
@@ -1449,15 +1729,24 @@ shell.Run cmd, 0, False
         substep "       This is a driver issue, not an SDK issue." "Yellow"
         substep "       Ensure the ROCm compute driver is installed alongside the display driver:" "Yellow"
         substep "       https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
+    } elseif ($ROCmGfxArch) {
+        # Known arch: Studio setup installs AMD's bundled-runtime ROCm PyTorch wheels
+        # (repo.amd.com), which ship their own runtime -- HIP SDK optional.
+        step "gpu" "AMD ROCm ($ROCmGfxArch)" "Cyan"
+        substep "Detected: $ROCmGpuLabel" "Cyan"
+        substep "GPU PyTorch uses AMD's bundled-runtime ROCm wheels -- HIP SDK not required (optional)." "Cyan"
     } elseif ($ROCmGpuLabel) {
-        step "gpu" "AMD GPU detected -- HIP SDK not found" "Yellow"
+        step "gpu" "AMD GPU detected -- arch unknown" "Yellow"
         substep "Detected: $ROCmGpuLabel" "Yellow"
-        substep "Install the HIP SDK for ROCm GPU inference:" "Yellow"
+        substep "Could not determine the GPU arch -- install the HIP SDK or set" "Yellow"
+        substep "UNSLOTH_ROCM_GFX_ARCH to enable GPU ROCm PyTorch:" "Yellow"
         substep "https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
     } else {
         step "gpu" "none (chat-only / GGUF)" "Yellow"
         substep "Training and GPU inference require an NVIDIA or AMD ROCm GPU." "Yellow"
     }
+    # On an AMD GPU (no NVIDIA), surface the optional WSL-ROCm driver hint.
+    if (-not $HasNvidiaSmi -and ($ROCmGfxArch -or $ROCmGpuLabel)) { Show-AmdWslDriverHint }
 
     # ── Choose the correct PyTorch index URL based on driver CUDA version ──
     # Mirrors Get-PytorchCudaTag in setup.ps1.
@@ -1465,7 +1754,7 @@ shell.Run cmd, 0, False
         $baseUrl = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/') } else { "https://download.pytorch.org/whl" }
         if (-not $NvidiaSmiExe) { return "$baseUrl/cpu" }
         try {
-            $output = & $NvidiaSmiExe 2>&1 | Out-String
+            $output = Invoke-NvidiaSmiBounded $NvidiaSmiExe
             # Newer NVIDIA drivers (e.g. 610.x on Windows) print
             # "CUDA UMD Version: X.Y" instead of the legacy "CUDA Version: X.Y".
             # Accept both spellings so we don't fall through to the cu126 default.
@@ -1873,16 +2162,24 @@ shell.Run cmd, 0, False
     # ── Print CPU-only hint when no GPU detected ──
     if (-not $SkipTorch -and -not $ROCmIndexUrl -and $TorchIndexUrl -like "*/cpu") {
         Write-Host ""
-        if ($HipSdkInstalled -and -not $HasROCm) {
-            substep "Installing CPU-only PyTorch (HIP SDK found but GPU not ROCm-accessible)." "Yellow"
-        } elseif ($ROCmGpuLabel) {
-            substep "Installing CPU-only PyTorch (ROCm wheels require the HIP SDK)." "Yellow"
+        if ($ROCmGfxArch) {
+            # Known AMD arch: install.ps1 lays down CPU PyTorch as a base, then
+            # setup.ps1 swaps in AMD's bundled-runtime GPU ROCm wheels (no HIP SDK).
+            substep "Installing CPU PyTorch as a base -- Studio setup installs GPU ROCm" "Cyan"
+            substep "wheels for $ROCmGfxArch next (bundled runtime; HIP SDK not required)." "Cyan"
         } else {
-            substep "No NVIDIA GPU detected." "Yellow"
+            if ($HipSdkInstalled -and -not $HasROCm) {
+                substep "Installing CPU-only PyTorch (HIP SDK found but GPU not ROCm-accessible)." "Yellow"
+            } elseif ($ROCmGpuLabel) {
+                substep "Installing CPU-only PyTorch (AMD GPU arch unknown -- install the HIP SDK" "Yellow"
+                substep "or set UNSLOTH_ROCM_GFX_ARCH to enable GPU ROCm)." "Yellow"
+            } else {
+                substep "No NVIDIA GPU detected." "Yellow"
+            }
+            substep "Installing CPU-only PyTorch. If you only need GGUF chat/inference," "Yellow"
+            substep "re-run with --no-torch for a faster, lighter install:" "Yellow"
+            substep ".\install.ps1 --no-torch" "Yellow"
         }
-        substep "Installing CPU-only PyTorch. If you only need GGUF chat/inference," "Yellow"
-        substep "re-run with --no-torch for a faster, lighter install:" "Yellow"
-        substep ".\install.ps1 --no-torch" "Yellow"
         Write-Host ""
     }
 
@@ -1923,7 +2220,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.1" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -1937,7 +2234,7 @@ shell.Run cmd, 0, False
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.1" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -1984,7 +2281,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.1" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython pydantic }
@@ -1996,7 +2293,7 @@ shell.Run cmd, 0, False
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.1" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.3" unsloth-zoo }
         } else {
             $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2024,7 +2321,7 @@ shell.Run cmd, 0, False
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.6.1" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.6.3" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
@@ -2139,6 +2436,11 @@ shell.Run cmd, 0, False
     $studioArgs = @('studio', 'setup')
     if ($script:UnslothVerbose) { $studioArgs += '--verbose' }
     $env:UNSLOTH_INSTALL_ROLLBACK_MANAGED = "1"
+    # Hand the venv interpreter to setup.ps1 so it reuses the Python we already
+    # resolved and built the venv with, instead of re-probing the system (which
+    # can trip over an unsupported `python` 3.14 or a Store stub on PATH even
+    # though the venv is fine). setup.ps1 Test-Path-guards this before use.
+    $env:UNSLOTH_SETUP_PYTHON = Join-Path $VenvDir "Scripts\python.exe"
     try {
         & $UnslothExe @studioArgs
         $setupExit = $LASTEXITCODE
@@ -2149,6 +2451,7 @@ shell.Run cmd, 0, False
             Remove-Item Env:UNSLOTH_STUDIO_HOME -ErrorAction SilentlyContinue
         }
         Remove-Item Env:UNSLOTH_INSTALL_ROLLBACK_MANAGED -ErrorAction SilentlyContinue
+        Remove-Item Env:UNSLOTH_SETUP_PYTHON -ErrorAction SilentlyContinue
     }
     if ($setupExit -ne 0) {
         Write-Host "[ERROR] unsloth studio setup failed (exit code $setupExit)" -ForegroundColor Red
diff --git a/install.sh b/install.sh
index 6b43bde722..4f062b9f44 100755
--- a/install.sh
+++ b/install.sh
@@ -1207,6 +1207,15 @@ STUB_EOF
         # Escape single quotes for PowerShell single-quoted string embedding
         _css_sc_args_ps=$(printf '%s' "$_css_sc_args" | sed "s/'/''/g")
 
+        # DISTINCT shortcut name so the WSL launcher never clobbers a native
+        # install's "Unsloth Studio.lnk" in the same folder. Per-distro suffix.
+        if [ -n "$_css_distro" ]; then
+            _css_lnk_name="Unsloth Studio (WSL - ${_css_distro}).lnk"
+        else
+            _css_lnk_name="Unsloth Studio (WSL).lnk"
+        fi
+        _css_lnk_name_ps=$(printf '%s' "$_css_lnk_name" | sed "s/'/''/g")
+
         # Create shortcuts via a temp PowerShell script to avoid escaping issues
         _css_ps1_tmp=$(mktemp /tmp/unsloth-shortcut-XXXXXX.ps1 2>/dev/null) || true
         if [ -n "$_css_ps1_tmp" ]; then
@@ -1214,19 +1223,57 @@ STUB_EOF
 \$WshShell = New-Object -ComObject WScript.Shell
 \$targetExe = (Get-Command '$_css_sc_target' -ErrorAction SilentlyContinue).Source
 if (-not \$targetExe) { exit 1 }
+# Best-effort: fetch the Unsloth icon to a stable Windows path (shared with a
+# native install if one exists) so the WSL shortcut shows the proper icon.
+\$iconDir = Join-Path \$env:LOCALAPPDATA 'Unsloth Studio'
+\$iconPath = Join-Path \$iconDir 'unsloth.ico'
+if (-not (Test-Path -LiteralPath \$iconPath)) {
+    try {
+        New-Item -ItemType Directory -Force -Path \$iconDir | Out-Null
+        Invoke-WebRequest -Uri 'https://raw.githubusercontent.com/unslothai/unsloth/main/studio/frontend/public/unsloth.ico' -OutFile \$iconPath -UseBasicParsing -ErrorAction Stop
+    } catch {}
+}
+\$hasIcon = \$false
+if (Test-Path -LiteralPath \$iconPath) {
+    try { \$b = [System.IO.File]::ReadAllBytes(\$iconPath); if (\$b.Length -ge 4 -and \$b[0] -eq 0 -and \$b[1] -eq 0 -and \$b[2] -eq 1 -and \$b[3] -eq 0) { \$hasIcon = \$true } } catch {}
+}
 \$locations = @(
     [Environment]::GetFolderPath('Desktop'),
     (Join-Path \$env:APPDATA 'Microsoft\Windows\Start Menu\Programs')
 )
+\$created = @()
 foreach (\$dir in \$locations) {
     if (-not \$dir -or -not (Test-Path \$dir)) { continue }
-    \$linkPath = Join-Path \$dir 'Unsloth Studio.lnk'
+    \$linkPath = Join-Path \$dir '$_css_lnk_name_ps'
     \$shortcut = \$WshShell.CreateShortcut(\$linkPath)
     \$shortcut.TargetPath = \$targetExe
     \$shortcut.Arguments = '$_css_sc_args_ps'
-    \$shortcut.Description = 'Launch Unsloth Studio'
+    \$shortcut.Description = 'Launch Unsloth Studio (WSL)'
+    if (\$hasIcon) { \$shortcut.IconLocation = "\$iconPath,0" }
     \$shortcut.Save()
+    \$created += \$linkPath
 }
+# Force Explorer to re-read EACH new shortcut's icon so it renders immediately
+# instead of a stale/blank (generic) icon. The reliable, NON-disruptive fix
+# (no explorer restart) is a PER-ITEM SHChangeNotify(SHCNE_UPDATEITEM,
+# SHCNF_PATHW, <lnk>) -- the global SHCNE_ASSOCCHANGED alone does not recover a
+# stale item. Also clear the on-disk icon cache for heavier staleness.
+try { & "\$env:SystemRoot\System32\ie4uinit.exe" -ClearIconCache } catch {}
+try { & "\$env:SystemRoot\System32\ie4uinit.exe" -show } catch {}
+try {
+    Add-Type -Namespace UnslothShell -Name IconRefresh -MemberDefinition '[System.Runtime.InteropServices.DllImport("shell32.dll", CharSet = System.Runtime.InteropServices.CharSet.Unicode)] public static extern void SHChangeNotify(int e, uint f, string a, System.IntPtr b);' -ErrorAction SilentlyContinue
+    foreach (\$p in \$created) { try { [UnslothShell.IconRefresh]::SHChangeNotify(0x00002000, 0x0005, \$p, [System.IntPtr]::Zero) } catch {} }
+    [UnslothShell.IconRefresh]::SHChangeNotify(0x08000000, 0, \$null, [System.IntPtr]::Zero)
+} catch {}
+# Win11 Start Menu keeps its own tile-icon cache (preserve start2.bin).
+try {
+    \$smeh = Join-Path \$env:LOCALAPPDATA 'Packages\Microsoft.Windows.StartMenuExperienceHost_cw5n1h2txyewy\TempState'
+    if (Test-Path -LiteralPath \$smeh) {
+        Get-ChildItem -LiteralPath \$smeh -Filter 'TileCache_*' -ErrorAction SilentlyContinue | Remove-Item -Force -ErrorAction SilentlyContinue
+        Remove-Item -LiteralPath (Join-Path \$smeh 'StartUnifiedTileModelCache.dat') -Force -ErrorAction SilentlyContinue
+        Stop-Process -Name StartMenuExperienceHost -Force -ErrorAction SilentlyContinue
+    }
+} catch {}
 WSLPS1_EOF
 
             # Convert WSL path to Windows path for powershell.exe
@@ -1236,6 +1283,13 @@ WSLPS1_EOF
             fi
             rm -f "$_css_ps1_tmp"
         fi
+        # If WSL interop is disabled (powershell.exe "Exec format error"), the
+        # shortcut wasn't created; tell the user how to launch / re-enable it.
+        if [ "$_css_created" -ne 1 ]; then
+            substep "Couldn't create the Windows shortcut (WSL interop may be disabled)." "$C_WARN"
+            substep "  Launch Studio from Windows:  wsl -d \"$_css_distro\" -- bash -lc 'unsloth studio'" "$C_WARN"
+            substep "  (re-enable shortcuts: turn WSL interop back on, e.g. run 'wsl --shutdown' then reopen WSL.)" "$C_WARN"
+        fi
     fi
 
     if [ "$_css_created" -eq 1 ]; then
@@ -1650,9 +1704,28 @@ _find_no_torch_runtime() {
 }
 
 # ── AMD ROCm GPU detection helper ──
+# WSL2 ROCDXG: the system rocminfo enumerates the GPU over /dev/dxg only when
+# HSA_ENABLE_DXG_DETECTION=1 (a no-op on bare metal), and /opt/rocm/bin can be
+# off PATH outside login shells (the profile.d drop-in). Seed both before any
+# rocminfo probe or a ROCDXG WSL host is misdetected as CPU-only.
+_ensure_rocm_probe_env() {
+    export HSA_ENABLE_DXG_DETECTION="${HSA_ENABLE_DXG_DETECTION:-1}"
+    if ! command -v rocminfo >/dev/null 2>&1 && [ -x /opt/rocm/bin/rocminfo ]; then
+        PATH="$PATH:/opt/rocm/bin"
+    fi
+}
+
 # Returns 0 if an AMD GPU is present. Checks rocminfo, amd-smi, then sysfs
 # KFD topology (env-var-independent fallback for when HIP/ROCR_VISIBLE_DEVICES hides devices).
+# Always returns 1 (false) when an NVIDIA GPU is present: blocks every
+# detection path (rocminfo, amd-smi, KFD sysfs) from producing a false
+# positive on NVIDIA-only or NVIDIA-primary hosts, even when ROCm tools
+# are co-installed.
 _has_amd_rocm_gpu() {
+    _ensure_rocm_probe_env
+    if _has_usable_nvidia_gpu; then
+        return 1
+    fi
     if command -v rocminfo >/dev/null 2>&1 && \
        rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9][0-9]/{found=1} END{exit !found}'; then
         return 0
@@ -1660,27 +1733,71 @@ _has_amd_rocm_gpu() {
          amd-smi list 2>/dev/null | awk '/^GPU[[:space:]]*[:\[][[:space:]]*[0-9]/{ found=1 } END{ exit !found }'; then
         return 0
     elif [ -e /dev/kfd ] && \
-         awk '/gpu_id/{ if ($2+0 > 0) found=1 } END{ exit !found }' \
+         awk 'FNR==1{ gpu=0; amd=0 } /gpu_id/{ gpu=($2+0>0) } /vendor_id/{ amd=($2==4098) } \
+              gpu && amd { found=1 } END{ exit !found }' \
              /sys/class/kfd/kfd/topology/nodes/*/properties 2>/dev/null; then
+        # vendor_id 4098 = 0x1002 (AMD). NVIDIA open kernel module (driver
+        # 560+) can register KFD topology nodes with non-zero gpu_id but
+        # vendor_id 4318 (0x10DE). Require AMD vendor to avoid misrouting
+        # NVIDIA-only hosts to the ROCm install path.
         return 0
     fi
     return 1
 }
 
+# ── Bounded command runner ──
+# Runs a command under a 10s timeout when the `timeout` binary is available,
+# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
+# driver init or after a reset) from hanging the installer: a timed-out probe
+# exits nonzero and is treated exactly like a failed probe. No-op semantics on
+# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
+_run_bounded() {
+    if command -v timeout >/dev/null 2>&1; then
+        timeout 10 "$@"
+    else
+        "$@"
+    fi
+}
+
+# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
+# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
+# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
+# var, so the probes below cannot see the distinction on their own.
+_cvd_hides_nvidia() {
+    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
+    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
+    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
+}
+
 # ── NVIDIA usable-GPU helper ──
-# Returns 0 (true) only if nvidia-smi is present AND actually lists a GPU.
-# Prevents AMD-only hosts with a stale nvidia-smi on PATH from being routed
-# into the CUDA branch.
+# Returns 0 (true) if an NVIDIA GPU is present and usable.
+# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
+# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
+# -- handles PATH gaps, subprocess timeouts, and driver init races that
+# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
+# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
+# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
 _has_usable_nvidia_gpu() {
+    if _cvd_hides_nvidia; then
+        return 1
+    fi
     _nvsmi=""
     if command -v nvidia-smi >/dev/null 2>&1; then
         _nvsmi="nvidia-smi"
     elif [ -x "/usr/bin/nvidia-smi" ]; then
         _nvsmi="/usr/bin/nvidia-smi"
-    else
-        return 1
     fi
-    "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'
+    if [ -n "$_nvsmi" ]; then
+        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+            return 0
+        fi
+    fi
+    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
+    if [ -d /proc/driver/nvidia/gpus ] && \
+       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
+        return 0
+    fi
+    return 1
 }
 
 # ── Detect GPU and choose PyTorch index URL ──
@@ -1697,14 +1814,16 @@ get_torch_index_url() {
     # packages) is not sufficient: otherwise an AMD-only host would
     # silently install CUDA wheels.
     _smi=""
+    _nvidia_detected=0
     if _has_usable_nvidia_gpu; then
+        _nvidia_detected=1
         if command -v nvidia-smi >/dev/null 2>&1; then
             _smi="nvidia-smi"
         elif [ -x "/usr/bin/nvidia-smi" ]; then
             _smi="/usr/bin/nvidia-smi"
         fi
     fi
-    if [ -z "$_smi" ]; then
+    if [ "$_nvidia_detected" -eq 0 ]; then
         # No NVIDIA GPU -- check for AMD ROCm GPU.
         # PyTorch only publishes ROCm wheels for linux-x86_64; skip the
         # ROCm branch entirely on aarch64 / arm64 / other architectures
@@ -1781,7 +1900,11 @@ get_torch_index_url() {
     # of the legacy "CUDA Version: X.Y"; accept both with two BRE expressions
     # (POSIX sed does not support "?" without -E).  The two patterns are
     # mutually exclusive per line, so head -1 picks the first emitted match.
-    _cuda_ver=$(LC_ALL=C $_smi 2>/dev/null \
+    # Bound the call (a wedged nvidia-smi would otherwise hang here) and force
+    # the C locale for stable parsing. LC_ALL is exported inside this command
+    # substitution subshell so it reaches nvidia-smi through _run_bounded
+    # without depending on `env`; the export is scoped to the subshell.
+    _cuda_ver=$(export LC_ALL=C; _run_bounded "$_smi" 2>/dev/null \
         | sed -n \
             -e 's/.*CUDA UMD Version:[[:space:]]*\([0-9][0-9]*\.[0-9][0-9]*\).*/\1/p' \
             -e 's/.*CUDA Version:[[:space:]]*\([0-9][0-9]*\.[0-9][0-9]*\).*/\1/p' \
@@ -1914,8 +2037,139 @@ _pick_radeon_wheel() {
     esac
 }
 
+# ── ROCm-on-WSL bootstrap for AMD Strix Halo (gfx1151) ───────────────────────
+# No-op everywhere except: WSL + GPU wanted + no usable GPU yet + /dev/dxg +
+# Strix Halo APU. Every other config (NVIDIA, native-Linux ROCm, macOS, Windows,
+# CPU, non-Strix WSL) skips it and normal detection runs unchanged. NEVER aborts
+# the installer -- always returns 0. Runs the idempotent helper (ROCm 7.2 +
+# librocdxg), then sources the env it persisted so detection finds the GPU.
+_maybe_bootstrap_rocm_wsl() {
+    [ "${OS:-}" = "wsl" ] || return 0
+    [ "${SKIP_TORCH:-false}" = "false" ] || return 0
+    [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
+    # Leave any already-usable GPU completely alone (NVIDIA, or working ROCm).
+    if _has_usable_nvidia_gpu; then return 0; fi
+    # "Usable ROCm" here = rocminfo enumerates the gfx1151 agent. Don't use the
+    # generic _has_amd_rocm_gpu: its broad gfx match accepts "gfx11-generic" and
+    # would skip this bootstrap while the real GPU is still unusable. awk consumes
+    # all input, so rocminfo isn't SIGPIPE'd like `grep -q` would under pipefail.
+    _ensure_rocm_probe_env
+    if command -v rocminfo >/dev/null 2>&1 && \
+       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx1151/{found=1} END{exit !found}'; then
+        return 0
+    fi
+    # WSL GPU passthrough device must exist (present on any WSL2 GPU host).
+    [ -e /dev/dxg ] || return 0
+    # Only Strix Halo (gfx1151): rocminfo can't tell us the arch yet, so match
+    # the CPU model string WSL exposes (e.g. "AMD Ryzen AI Max+ ... Radeon 8060S").
+    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    command -v bash >/dev/null 2>&1 || return 0
+
+    # Fast path: already configured (librocdxg present) but launched from a
+    # non-login shell so the persisted env wasn't loaded -- just load it.
+    if [ -e /opt/rocm/lib/librocdxg.so ] || [ -e /opt/rocm/lib64/librocdxg.so ]; then
+        if [ -r /etc/profile.d/unsloth-rocm-wsl.sh ]; then
+            # shellcheck disable=SC1091
+            . /etc/profile.d/unsloth-rocm-wsl.sh || true
+        else
+            # librocdxg present but the env drop-in is gone (e.g. a Studio
+            # uninstall removed it while keeping shared ROCm). Restore the FULL
+            # env inline (so rocminfo is on PATH) and recreate the drop-in.
+            _rw_rocm=/opt/rocm
+            export HSA_ENABLE_DXG_DETECTION=1
+            export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+            export PATH="${_rw_rocm}/bin:${PATH}"
+            export LD_LIBRARY_PATH="${_rw_rocm}/lib:${LD_LIBRARY_PATH:-}"
+            # Persist the drop-in so later non-login Studio launches get the env
+            # too. /etc/profile.d is root-owned: a plain redirect fails for a
+            # non-root reinstall (ROCm would silently disappear after this shell),
+            # so tee through sudo when not root. Best-effort -- the current shell
+            # already has the env, so the install proceeds either way.
+            _rw_dropin="$(
+                printf '# >>> Unsloth ROCm-on-WSL (gfx1151) >>>\n'
+                printf 'export HSA_ENABLE_DXG_DETECTION=1\n'
+                printf 'export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1\n'
+                printf 'export PATH="%s/bin:${PATH}"\n' "${_rw_rocm}"
+                printf 'export LD_LIBRARY_PATH="%s/lib:${LD_LIBRARY_PATH:-}"\n' "${_rw_rocm}"
+                printf '# <<< Unsloth ROCm-on-WSL (gfx1151) <<<\n'
+            )"
+            if [ "$(id -u)" = "0" ]; then
+                printf '%s\n' "$_rw_dropin" > /etc/profile.d/unsloth-rocm-wsl.sh 2>/dev/null || true
+            elif command -v sudo >/dev/null 2>&1; then
+                printf '%s\n' "$_rw_dropin" | sudo tee /etc/profile.d/unsloth-rocm-wsl.sh >/dev/null 2>&1 || true
+            fi
+        fi
+        return 0
+    fi
+
+    echo ""
+    substep "Detected AMD Strix Halo (Radeon 8000S) in WSL with no ROCm runtime yet." "$C_WARN"
+    substep "Setting up ROCm-on-WSL (ROCm 7.2 + librocdxg) automatically to enable this GPU."
+    substep "One-time, uses sudo and a large download. (skip: re-run with UNSLOTH_SKIP_ROCM_WSL_SETUP=1)"
+
+    # Locate the helper: prefer the copy shipped beside install.sh, else fetch it.
+    _rw_helper="${_REPO_ROOT:-.}/scripts/install_rocm_wsl_strixhalo.sh"
+    _rw_tmp=""
+    if [ ! -r "$_rw_helper" ]; then
+        _rw_tmp="$(mktemp 2>/dev/null || echo /tmp/_unsloth_rocm_wsl.sh)"
+        if download "https://raw.githubusercontent.com/unslothai/unsloth/main/scripts/install_rocm_wsl_strixhalo.sh" "$_rw_tmp" 2>/dev/null; then
+            _rw_helper="$_rw_tmp"
+        else
+            substep "Could not fetch the ROCm-on-WSL helper; using CPU fallback." "$C_WARN"
+            [ -n "$_rw_tmp" ] && rm -f "$_rw_tmp"
+            return 0
+        fi
+    fi
+
+    # Consent: the narrow guarded case is exactly the GPU setup the user ran the
+    # installer for, so it proceeds AUTOMATICALLY by default (works with no TTY,
+    # e.g. `curl ... | sh`). Opt out via UNSLOTH_SKIP_ROCM_WSL_SETUP=1 (top of
+    # function). The Tauri app drives its own consent UI, so under TAURI_MODE it
+    # only runs when the app passes UNSLOTH_ROCM_WSL_AUTO=1; else surface and wait.
+    _rw_go=1
+    if [ "${TAURI_MODE:-false}" = "true" ] && [ "${UNSLOTH_ROCM_WSL_AUTO:-0}" != "1" ]; then
+        tauri_log "ROCM_WSL_AVAILABLE" "strixhalo"
+        substep "Enable the GPU from the desktop app (or set UNSLOTH_ROCM_WSL_AUTO=1)." "$C_WARN"
+        _rw_go=0
+    fi
+
+    if [ "$_rw_go" = "1" ]; then
+        # Helper does its own sudo + is idempotent. SMOKE_TEST=0: install.sh
+        # installs torch itself right after, into the real venv.
+        if UNSLOTH_WSL_SMOKE_TEST=0 bash "$_rw_helper"; then
+            # Pull the helper's persisted env into THIS shell so detection
+            # (rocminfo) now enumerates the GPU and routes to gfx1151.
+            if [ -r /etc/profile.d/unsloth-rocm-wsl.sh ]; then
+                # shellcheck disable=SC1091
+                . /etc/profile.d/unsloth-rocm-wsl.sh || true
+            fi
+            substep "ROCm-on-WSL ready; continuing with GPU install." "$C_OK"
+        else
+            substep "ROCm-on-WSL setup did not complete; falling back to CPU-only." "$C_WARN"
+        fi
+    fi
+    [ -n "$_rw_tmp" ] && rm -f "$_rw_tmp"
+    return 0
+}
+_maybe_bootstrap_rocm_wsl || true
+
 TORCH_INDEX_URL=$(get_torch_index_url)
 
+# Export the resolved torch backend ("cuda", "rocm", or "cpu") so that
+# downstream scripts (setup.sh -> install_python_stack.py) know what was
+# chosen here and can skip ROCm-specific repair steps on CUDA/CPU hosts.
+# Classify on the FINAL path segment only: a custom UNSLOTH_PYTORCH_MIRROR
+# whose base path happens to contain "rocm" or "gfx" must not mislabel a
+# cu*/cpu index as ROCm (radeon repo URLs end in rocm-rel-X.Y/, Strix
+# overrides in gfxNNNN/, so the trailing slash is stripped first).
+_torch_index_leaf="${TORCH_INDEX_URL%/}"
+_torch_index_leaf="${_torch_index_leaf##*/}"
+case "$_torch_index_leaf" in
+    rocm*|gfx*) export UNSLOTH_TORCH_BACKEND="rocm" ;;
+    cpu)        export UNSLOTH_TORCH_BACKEND="cpu"  ;;
+    *)          export UNSLOTH_TORCH_BACKEND="cuda" ;;
+esac
+
 # rocm7.2 ships torch 2.11.0 -- adjust the constraint to allow it.
 # All other ROCm tags and CUDA stay within <2.11.0.
 case "$TORCH_INDEX_URL" in
@@ -2018,6 +2272,7 @@ if _has_usable_nvidia_gpu; then
     step "gpu" "NVIDIA GPU detected"
 elif case "$TORCH_INDEX_URL" in */rocm*|*/gfx*) true ;; *) false ;; esac; then
     # Probe gfx arch for the display label, honouring HIP_VISIBLE_DEVICES
+    _ensure_rocm_probe_env
     _gpu_disp_gfx_all=""
     _gpu_disp_mkt=""
     if command -v rocminfo >/dev/null 2>&1; then
@@ -2048,14 +2303,20 @@ elif case "$TORCH_INDEX_URL" in */rocm*|*/gfx*) true ;; *) false ;; esac; then
         substep "gfx arch from UNSLOTH_ROCM_GFX_ARCH env override: $_gpu_disp_gfx"
     # Name-based arch inference when tools don't report gfx (mirrors install.ps1 nameArchTable)
     elif [ -z "$_gpu_disp_gfx" ] && [ -n "$_gpu_disp_mkt" ]; then
+        # Kept in sync with the nameArchTable in install.ps1 / setup.ps1.
+        # gfx1102 matched BEFORE gfx1100 so the spaceless "RX 7700S" lands on
+        # gfx1102 (bash case has no negative lookahead like the PS tables).
         case "$_gpu_disp_mkt" in
-            *"9070 XT"*|*9080*)                                                     _gpu_disp_gfx="gfx1201" ;;  # RDNA 4
-            *9070*|*9060*)                                                          _gpu_disp_gfx="gfx1200" ;;  # RDNA 4
-            *"8060S"*|*"890M"*|*"Strix Halo"*|*"HX 37"*|*"HX 38"*|*"AI 9 HX"*)  _gpu_disp_gfx="gfx1151" ;;  # RDNA 3.5 iGPU
-            *"880M"*|*"Strix Point"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*)        _gpu_disp_gfx="gfx1150" ;;  # RDNA 3.5 iGPU
-            *"RX 7900"*|*"RX 7800"*|*"RX 7700"*)                                  _gpu_disp_gfx="gfx1100" ;;  # RDNA 3 desktop
-            *"RX 7600"*)                                                           _gpu_disp_gfx="gfx1102" ;;  # RDNA 3
-            *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*)                               _gpu_disp_gfx="gfx1103" ;;  # RDNA 3 iGPU
+            *"9070 XT"*|*9080*)                                                                            _gpu_disp_gfx="gfx1201" ;;  # RDNA 4
+            *9070*|*9060*)                                                                                 _gpu_disp_gfx="gfx1200" ;;  # RDNA 4
+            *"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _gpu_disp_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+            *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _gpu_disp_gfx="gfx1150" ;;  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+            *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*)                  _gpu_disp_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
+            *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*)                  _gpu_disp_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
+            *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*|*"Hawk Point"*|*"Z1 Extreme"*|*"Z2 Extreme"*)            _gpu_disp_gfx="gfx1103" ;;  # RDNA 3 iGPU (Phoenix / Hawk Point)
+            *"RX 6900"*|*"RX 6800"*|*"RX 6750"*|*"RX 6700"*|*"PRO W6800"*|*"PRO W6900"*)                    _gpu_disp_gfx="gfx1030" ;;  # RDNA 2 (Navi 21)
+            *"RX 6650"*|*"RX 6600"*|*"PRO W6600"*|*"PRO W6650"*)                                            _gpu_disp_gfx="gfx1032" ;;  # RDNA 2 (Navi 23)
+            *"RX 6500"*|*"RX 6400"*|*"RX 6300"*|*"PRO W6400"*|*"PRO W6500"*)                                _gpu_disp_gfx="gfx1034" ;;  # RDNA 2 (Navi 24)
         esac
         if [ -n "$_gpu_disp_gfx" ]; then
             substep "gfx arch inferred from GPU name: $_gpu_disp_gfx"
@@ -2089,7 +2350,34 @@ case "$TORCH_INDEX_URL" in
     */cpu)
         if [ "$SKIP_TORCH" = false ] && [ "$OS" != "macos" ]; then
             substep "No GPU detected -- installing CPU-only PyTorch." "$C_WARN"
-            substep "AMD ROCm users: see https://docs.unsloth.ai/get-started/install-and-update/amd"
+            if [ "$OS" = "wsl" ]; then
+                # WSL + no GPU detected (detection above found nothing). Common
+                # cause: an AMD GPU whose ROCm-on-WSL runtime isn't exposed yet --
+                # /dev/dxg present (graphics) but no ROCm runtime.
+                _wsl_ubu_ver=""
+                [ -r /etc/os-release ] && _wsl_ubu_ver=$(. /etc/os-release 2>/dev/null; printf '%s' "${VERSION_ID:-}")
+                if [ -e /dev/dxg ]; then
+                    substep "A GPU is plumbed into WSL (/dev/dxg) but no ROCm runtime is exposed to it." "$C_WARN"
+                fi
+                substep "For an AMD GPU, ROCm-on-WSL currently needs ALL of:"
+                substep "  1. AMD Adrenalin Edition 26.1.1+ on Windows (26.2.2+ for Strix Halo / Ryzen AI Max+)."
+                substep "     Older drivers lack production ROCDXG/WSL support, so ROCm can't see the GPU."
+                substep "     Get it from AMD (open in a browser -- direct downloads are referrer-gated):"
+                substep "       https://www.amd.com/en/resources/support-articles/release-notes/RN-RAD-WIN-26-2-2.html"
+                substep "  2. ROCm 7.2.1 + librocdxg inside WSL (with HSA_ENABLE_DXG_DETECTION=1)."
+                substep "  3. A WSL distro AMD supports for ROCm -- Ubuntu 24.04 is the known-good one."
+                if [ -n "$_wsl_ubu_ver" ] && [ "$_wsl_ubu_ver" != "24.04" ]; then
+                    substep "  This distro is Ubuntu $_wsl_ubu_ver, which AMD may not support for ROCm-on-WSL yet." "$C_WARN"
+                fi
+                substep "Set up the GPU in WSL with a dedicated Ubuntu 24.04 distro:"
+                substep "  wsl --install Ubuntu-24.04        # run in Windows PowerShell, then reopen WSL"
+                substep "  # then re-run this installer inside Ubuntu-24.04 -- it will detect the GPU."
+                substep "AMD ROCm-on-WSL docs: https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/"
+                substep "Strix Halo (gfx1151): this installer auto-offers ROCm-on-WSL setup once the"
+                substep "  driver is current; or run unsloth/scripts/install_rocm_wsl_strixhalo.sh yourself."
+            else
+                substep "AMD ROCm users: see https://docs.unsloth.ai/get-started/install-and-update/amd"
+            fi
             substep "Re-run with --no-torch for GGUF-only (faster, no PyTorch):"
             substep "  curl -fsSL https://unsloth.ai/install.sh | sh -s -- --no-torch"
         fi
@@ -2117,7 +2405,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.1" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2130,7 +2418,7 @@ if [ "$_MIGRATED" = true ]; then
     else
         run_install_cmd "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.1" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2334,7 +2622,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.1" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2352,7 +2640,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.1" unsloth-zoo
+            --upgrade-package unsloth "unsloth>=2026.6.3" unsloth-zoo
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2408,7 +2696,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.6.1" --torch-backend=auto
+        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.6.3" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
diff --git a/pyproject.toml b/pyproject.toml
index 14f6f87704..3b279d1045 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -26,6 +26,7 @@ classifiers = [
 ]
 dependencies = [
     "typer",
+    "rich",
     "pydantic",
     "pyyaml",
     "nest-asyncio",
@@ -70,7 +71,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.1",
+    "unsloth_zoo>=2026.6.3",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -91,7 +92,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.1",
+    "unsloth_zoo>=2026.6.3",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -581,7 +582,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.1",
+    "unsloth_zoo>=2026.6.3",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/scripts/install_rocm_wsl_strixhalo.sh b/scripts/install_rocm_wsl_strixhalo.sh
new file mode 100644
index 0000000000..5ef9ee386a
--- /dev/null
+++ b/scripts/install_rocm_wsl_strixhalo.sh
@@ -0,0 +1,289 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# ──────────────────────────────────────────────────────────────────────────────
+# Enable ROCm-on-WSL for AMD Strix Halo (Radeon 8060S / gfx1151)
+# ──────────────────────────────────────────────────────────────────────────────
+# install.sh already routes gfx1151 to the right ROCm wheels once a ROCm runtime
+# is present; what it does NOT do is install AMD's ROCm userspace + the WSL DXG
+# bridge. This helper automates that Linux-side prerequisite on Ubuntu 24.04
+# WSL2 and is invoked by install.sh when it sees a Strix Halo APU in WSL (via
+# /dev/dxg) but no ROCm runtime yet. Fully idempotent (re-run just re-verifies).
+#
+# Manual, admin-gated Windows prerequisite: an AMD Adrenalin driver with
+# production ROCDXG/WSL support (26.2.2+). install.ps1 offers to update it. Once
+# installed + rebooted, /dev/dxg is exposed to WSL and this script builds the rest.
+#
+# HOW ROCDXG WORKS (and why older /usr/lib/wsl/lib notes are wrong): librocdxg.so
+# is AMD's user-mode bridge between the Linux HSA runtime and the Windows driver
+# over /dev/dxg. The STANDARD hsa-rocr runtime (NOT the gone "roc4wsl" package)
+# loads it when HSA_ENABLE_DXG_DETECTION=1. No hsa/rocm libs need injecting into
+# /usr/lib/wsl/lib (it holds only d3d12/dxcore), yet rocminfo enumerates gfx1151
+# fine -- so we gate on /dev/dxg, not on WSL lib injection.
+#
+# KNOWN CAVEAT (ROCm/ROCm#6022): librocdxg can cap usable ROCm VRAM at the WSL
+# VM's RAM (.wslconfig [wsl2] memory=) on some BIOS UMA layouts, and amd-smi
+# doesn't work in WSL. On OOM below capacity, raise memory= (then wsl --shutdown)
+# and watch GPU use from Windows. Large-UMA BIOS exposes the full pool regardless.
+#
+# Verified on Ryzen AI Max+ PRO 395 / Radeon 8060S (gfx1151) with ROCm 7.2.1 +
+# Ubuntu 24.04 + WSL2 + Adrenalin. These pins MOVE; bump + re-verify on newer ROCm.
+# ──────────────────────────────────────────────────────────────────────────────
+set -euo pipefail
+
+# ── Tunables (override via env) ──────────────────────────────────────────────
+ROCM_VER="${UNSLOTH_WSL_ROCM_VER:-7.2.1}"            # ROCm release to install
+GFX="gfx1151"
+LIBROCDXG_REF="${UNSLOTH_LIBROCDXG_REF:-develop}"    # ROCm/librocdxg git ref to build
+# AMD's gfx1151 wheel index (same one install.sh uses); only for the smoke test.
+TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${GFX}/"
+# Optional torch smoke test (throwaway venv). OFF by default: install.sh installs
+# torch itself into the real venv right after, so a duplicate download is wasteful.
+SMOKE_TEST="${UNSLOTH_WSL_SMOKE_TEST:-0}"
+# REQUIRED constraint -- without it pip prefers PyPI's newer CUDA torch over the
+# gfx1151 ROCm wheel. 2.11 carries AMD's real gfx1151 fix (matches install.sh).
+TORCH_CONSTRAINT="${UNSLOTH_WSL_TORCH_CONSTRAINT:-torch>=2.11.0,<2.12.0}"
+ROCM_DIR=""                                          # resolved after install
+
+say()  { printf '\n\033[1;36m== %s\033[0m\n' "$*"; }
+note() { printf '   %s\n' "$*"; }
+die()  { printf '\n\033[1;31m[BLOCKED] %s\033[0m\n' "$*" >&2; exit 1; }
+
+# sudo only if not already root (WSL distros often run as root)
+SUDO=""
+if [ "$(id -u)" -ne 0 ]; then
+    command -v sudo >/dev/null 2>&1 || die "Need root or sudo to install ROCm."
+    SUDO="sudo"
+fi
+
+# ── Windows 11 SDK (headers for the librocdxg build) ─────────────────────────
+# librocdxg's cmake build needs the Windows SDK 'shared' headers, which live on
+# the Windows HOST under C:\Program Files (x86)\Windows Kits\10\Include\<ver>\.
+_WIN_SDK_INC_BASE="/mnt/c/Program Files (x86)/Windows Kits/10/Include"
+
+# Print the newest installed SDK include dir with 'shared' headers, or nothing.
+# find + read loop (not `for ... in $(ls)`) since the base path has a space.
+_find_win_sdk() {
+    [ -d "$_WIN_SDK_INC_BASE" ] || return 0
+    while IFS= read -r _inc; do
+        [ -n "$_inc" ] || continue
+        if [ -d "$_inc/shared" ]; then printf '%s' "$_inc"; return 0; fi
+    done < <(find "$_WIN_SDK_INC_BASE" -mindepth 1 -maxdepth 1 -type d 2>/dev/null | sort -Vr)
+    return 0
+}
+
+# Best-effort: install the Windows 11 SDK on the Windows HOST via winget so the
+# build has its headers with no manual step. Elevates -> ONE UAC prompt; headers
+# appear under /mnt/c immediately (no reboot). Never fatal -- failure falls
+# through to a manual-install message. Opt out: UNSLOTH_SKIP_WIN_SDK_INSTALL=1.
+_install_windows_sdk_via_winget() {
+    [ "${UNSLOTH_SKIP_WIN_SDK_INSTALL:-0}" = "1" ] && { note "Skipping Windows SDK auto-install (UNSLOTH_SKIP_WIN_SDK_INSTALL=1)."; return 0; }
+    command -v powershell.exe >/dev/null 2>&1 || return 0
+    # `command -v` succeeds even with WSL interop OFF (.exe on PATH but fails
+    # with "Exec format error"); verify it actually executes.
+    powershell.exe -NoProfile -Command "exit 0" >/dev/null 2>&1 || return 0
+    if ! powershell.exe -NoProfile -Command "if (Get-Command winget -ErrorAction SilentlyContinue) { exit 0 } else { exit 1 }" >/dev/null 2>&1; then
+        note "winget not available on the Windows host -- cannot auto-install the Windows SDK."
+        return 0
+    fi
+    say "Installing the Windows 11 SDK on the Windows host via winget"
+    note "librocdxg needs its headers. Approve the UAC prompt on the Windows desktop."
+    note "One-time (~1-3 GB download); opt out with UNSLOTH_SKIP_WIN_SDK_INSTALL=1."
+    # Newest SDK first, then a fallback. Header presence is the source of truth
+    # (re-check each attempt), not winget's exit code. </dev/null so winget never
+    # consumes a piped `curl | sh` stdin.
+    for _sdk_id in Microsoft.WindowsSDK.10.0.26100 Microsoft.WindowsSDK.10.0.22621; do
+        note "winget install ${_sdk_id} ..."
+        # --source winget: pin the community source so a broken default msstore
+        # source (the cert failure this PR fixes) can't abort SDK resolution.
+        powershell.exe -NoProfile -Command "winget install --id ${_sdk_id} -e --source winget --accept-source-agreements --accept-package-agreements --disable-interactivity" </dev/null || true
+        if [ -n "$(_find_win_sdk)" ]; then
+            note "Windows SDK headers present after install."
+            return 0
+        fi
+    done
+    note "Automatic Windows SDK install did not complete."
+    return 0
+}
+
+# ── PREFLIGHT ────────────────────────────────────────────────────────────────
+say "Preflight checks"
+
+# shellcheck disable=SC1091
+. /etc/os-release 2>/dev/null || true
+if [ "${VERSION_ID:-}" != "24.04" ]; then
+    die "This targets Ubuntu 24.04 (found '${VERSION_ID:-unknown}'). AMD's ROCm-on-WSL supports 24.04; create a dedicated distro:  wsl --install Ubuntu-24.04  (do not run on 26.04 -- ROCm 7.2 does not target it yet)."
+fi
+
+if [ ! -e /dev/dxg ]; then
+    die "/dev/dxg missing -- WSL GPU paravirtualization not present. Ensure this is WSL2 (not WSL1) on a recent Windows build, and that an AMD GPU + ROCDXG-capable Adrenalin driver is installed on the Windows host (then reboot)."
+fi
+note "Ubuntu 24.04 + /dev/dxg present."
+# Don't block on hsa/rocm libs in /usr/lib/wsl/lib: a working ROCDXG setup
+# doesn't need them (only d3d12/dxcore). Real readiness is checked via rocminfo.
+
+# ── Step 1: build/runtime prerequisites ──────────────────────────────────────
+say "Installing build prerequisites"
+export DEBIAN_FRONTEND=noninteractive
+$SUDO apt-get update -y
+# `make` is explicit: cmake shells out to it but Ubuntu only *recommends* it, so
+# minimal images lack it and the librocdxg `make -j` build would fail.
+$SUDO apt-get install -y cmake make gcc g++ git wget gpg ca-certificates python3-venv python3-pip
+
+# ── Step 2: ROCm ${ROCM_VER} userspace (no DKMS -- WSL uses the Windows driver) ─
+say "Installing ROCm ${ROCM_VER} userspace"
+if ! command -v rocminfo >/dev/null 2>&1 && [ ! -x /opt/rocm/bin/rocminfo ]; then
+    # Direct apt-repo install (leaner than amdgpu-install; repo is indexed by
+    # ROCm version, e.g. .../apt/7.2.1).
+    $SUDO mkdir -p /etc/apt/keyrings
+    wget -qO- https://repo.radeon.com/rocm/rocm.gpg.key \
+        | gpg --dearmor | $SUDO tee /etc/apt/keyrings/rocm.gpg >/dev/null
+    echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${ROCM_VER} noble main" \
+        | $SUDO tee /etc/apt/sources.list.d/rocm.list >/dev/null
+    printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' \
+        | $SUDO tee /etc/apt/preferences.d/rocm-pin-600 >/dev/null
+    $SUDO apt-get update -y
+    # rocm-libs pulls everything torch links at runtime (rocblas, hipblas,
+    # miopen-hip, rccl, ...); hsa-rocr + rocminfo come as deps. Large (~5 GB
+    # download / ~23 GB installed).
+    $SUDO apt-get install -y rocm-libs rocminfo hip-runtime-amd
+else
+    note "ROCm already present -- skipping apt install."
+fi
+
+# Resolve the real ROCm dir and ensure the canonical /opt/rocm symlink. apt lays
+# ROCm under /opt/rocm-<ver> and rocm-core symlinks /opt/rocm -> that; repair if
+# an earlier partial run left /opt/rocm as a real dir blocking the symlink.
+_real="$(ls -d /opt/rocm-* 2>/dev/null | sort -V | tail -1 || true)"
+if [ -n "$_real" ] && [ ! -L /opt/rocm ] && [ -d /opt/rocm ]; then
+    # /opt/rocm is a real dir blocking the symlink. Only treat it as a removable
+    # stray stub if it's NOT a real ROCm install (a real one has bin/rocminfo /
+    # bin/hipcc / .info/version) -- this protects a user's pre-existing ROCm. Even
+    # then we MOVE IT ASIDE, never rm -rf, so a wrong guess can't lose data.
+    if [ -e /opt/rocm/bin/rocminfo ] || [ -e /opt/rocm/bin/hipcc ] || [ -e /opt/rocm/.info/version ]; then
+        note "/opt/rocm is a real ROCm install -- leaving it untouched (will install librocdxg into it)."
+    else
+        note "Moving stray /opt/rocm stub aside -> $_real (not deleting it)"
+        $SUDO cp -an /opt/rocm/. "$_real"/ 2>/dev/null || true
+        $SUDO mv /opt/rocm "/opt/rocm.unsloth-stub-bak.$(date +%s)" 2>/dev/null || true
+        [ -e /opt/rocm ] || $SUDO ln -s "$_real" /opt/rocm
+    fi
+elif [ -n "$_real" ] && [ ! -e /opt/rocm ]; then
+    $SUDO ln -s "$_real" /opt/rocm
+fi
+if [ -L /opt/rocm ] || [ -d /opt/rocm ]; then ROCM_DIR="/opt/rocm"; else ROCM_DIR="$_real"; fi
+{ [ -n "$ROCM_DIR" ] && [ -d "$ROCM_DIR" ]; } || die "ROCm not found under /opt after install."
+note "ROCm at ${ROCM_DIR}"
+
+# ── Step 3: build librocdxg (DXG <-> HSA bridge; not yet an apt package) ──────
+say "Building librocdxg (${LIBROCDXG_REF})"
+if [ -e "${ROCM_DIR}/lib/librocdxg.so" ]; then
+    note "librocdxg already installed -- skipping build."
+else
+    # Discover the newest installed Win11 SDK (version differs per machine). If
+    # absent, auto-install via winget (one UAC prompt) and re-discover; only if
+    # that ALSO fails do we stop with manual instructions.
+    _win_sdk="$(_find_win_sdk)"
+    if [ -z "$_win_sdk" ]; then
+        note "Windows 11 SDK headers not found -- attempting automatic install..."
+        _install_windows_sdk_via_winget
+        _win_sdk="$(_find_win_sdk)"
+    fi
+    [ -n "$_win_sdk" ] || die "Windows 11 SDK headers not found under 'C:\\Program Files (x86)\\Windows Kits\\10\\Include\\*\\shared', and the automatic winget install did not complete. Install it on the Windows host (e.g. 'winget install Microsoft.WindowsSDK.10.0.26100') and re-run."
+    note "Windows SDK: ${_win_sdk}"
+    _src="${HOME}/.unsloth/librocdxg"
+    rm -rf "$_src"
+    git clone --depth 1 --branch "$LIBROCDXG_REF" https://github.com/ROCm/librocdxg.git "$_src" \
+        || git clone "https://github.com/ROCm/librocdxg.git" "$_src"
+    (
+        cd "$_src"
+        git checkout "$LIBROCDXG_REF" 2>/dev/null || true
+        mkdir -p build && cd build
+        cmake .. -DWIN_SDK="${_win_sdk}/shared"
+        make -j"$(nproc)"
+        $SUDO make install
+    )
+fi
+# Ensure soname symlinks resolve to whatever version was built (e.g. 1.2.0).
+_dxg_real="$(ls -1 "${ROCM_DIR}"/lib/librocdxg.so.*.* 2>/dev/null | sort -V | tail -1 || true)"
+if [ -n "$_dxg_real" ]; then
+    _dxg_base="$(basename "$_dxg_real")"                       # librocdxg.so.1.2.0
+    _dxg_major="$(printf '%s' "$_dxg_base" | sed -E 's/librocdxg\.so\.([0-9]+).*/\1/')"
+    $SUDO ln -sf "$_dxg_base" "${ROCM_DIR}/lib/librocdxg.so.${_dxg_major}"
+    $SUDO ln -sf "librocdxg.so.${_dxg_major}" "${ROCM_DIR}/lib/librocdxg.so"
+fi
+echo "${ROCM_DIR}/lib" | $SUDO tee /etc/ld.so.conf.d/rocm.conf >/dev/null
+$SUDO ldconfig
+
+# ── Step 4: persist environment (system-wide so Studio's worker inherits it) ──
+say "Persisting ROCm-on-WSL environment"
+_envfile="/etc/profile.d/unsloth-rocm-wsl.sh"
+$SUDO tee "$_envfile" >/dev/null <<EOF
+# >>> Unsloth ROCm-on-WSL (gfx1151) >>>
+export HSA_ENABLE_DXG_DETECTION=1
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+export PATH="${ROCM_DIR}/bin:\${PATH}"
+export LD_LIBRARY_PATH="${ROCM_DIR}/lib:\${LD_LIBRARY_PATH:-}"
+# <<< Unsloth ROCm-on-WSL (gfx1151) <<<
+EOF
+# also drop into ~/.bashrc for interactive shells
+if [ -n "${HOME:-}" ] && ! grep -q "Unsloth ROCm-on-WSL" "${HOME}/.bashrc" 2>/dev/null; then
+    cat "$_envfile" >> "${HOME}/.bashrc"
+fi
+# export into the current process so verification below works immediately
+export HSA_ENABLE_DXG_DETECTION=1
+export PATH="${ROCM_DIR}/bin:${PATH}"
+export LD_LIBRARY_PATH="${ROCM_DIR}/lib:${LD_LIBRARY_PATH:-}"
+
+# ── Step 5: verify the runtime enumerates the GPU ────────────────────────────
+say "Verifying rocminfo sees ${GFX}"
+# Capture rocminfo into a var BEFORE grepping: piping into `grep -q` SIGPIPEs
+# rocminfo on first match, which under `set -o pipefail` turns a successful match
+# into a pipeline failure. Match the gfx1151 ISA "Name:" agent exactly (not a
+# broad gfx1[0-9]) so a generic fallback ISA or unrelated RDNA GPU can't pass.
+_rocminfo_out="$(rocminfo 2>/dev/null || true)"
+if ! printf '%s\n' "$_rocminfo_out" | grep -qE "Name:[[:space:]]*${GFX}([^0-9]|$)"; then
+    printf '%s\n' "$_rocminfo_out" | head -25 >&2 || true
+    die "rocminfo did not enumerate a ${GFX} GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
+fi
+# Display-only summary: best-effort (|| true) so head's early pipe-close under
+# `set -o pipefail` can't fail the bootstrap after verification already passed.
+printf '%s\n' "$_rocminfo_out" | grep -E 'Marketing Name|Device Type|Compute Unit' | grep -iE "Radeon|GPU|Compute" | head -3 || true
+note "ROCm-on-WSL runtime is live for ${GFX}."
+
+# ── Step 6 (optional): torch smoke test from the gfx1151 index ───────────────
+if [ "$SMOKE_TEST" = "1" ]; then
+    say "Smoke-testing PyTorch on ${GFX} (throwaway venv)"
+    _venv="${HOME}/.unsloth/rocm-smoketest"
+    rm -rf "$_venv"; python3 -m venv "$_venv"
+    "$_venv/bin/pip" install --quiet --upgrade pip
+    # gfx1151 index is primary (torch + triton); PyPI only an extra for pure-py
+    # deps. The constraint keeps pip on the ROCm wheel, not a newer PyPI CUDA torch.
+    "$_venv/bin/pip" install --index-url "$TORCH_INDEX" \
+        --extra-index-url https://pypi.org/simple "$TORCH_CONSTRAINT" || \
+        die "torch install from ${TORCH_INDEX} failed."
+    "$_venv/bin/python" - <<'PY'
+import torch
+ok = torch.cuda.is_available()
+print("torch:", torch.__version__, "| cuda(rocm) available:", ok)
+if ok:
+    print("device:", torch.cuda.get_device_name(0))
+    free, total = torch.cuda.mem_get_info(0)
+    print(f"mem: free={free/1e9:.1f} GB total={total/1e9:.1f} GB")
+    import time
+    a = torch.randn(4096, 4096, device="cuda", dtype=torch.float16)
+    b = torch.randn(4096, 4096, device="cuda", dtype=torch.float16)
+    torch.cuda.synchronize(); t0 = time.time()
+    for _ in range(10): c = a @ b
+    torch.cuda.synchronize()
+    print(f"matmul ok ({(time.time()-t0)/10*1e3:.1f} ms/iter)")
+raise SystemExit(0 if ok else 1)
+PY
+    rm -rf "$_venv"
+fi
+
+say "Done."
+note "ROCm-on-WSL is ready for ${GFX}. If you ran this standalone, install Unsloth"
+note "in THIS distro and it will detect the GPU automatically:"
+note "  curl -fsSL https://unsloth.ai/install.sh | sh"
diff --git a/scripts/sync_allow_scripts_pins.py b/scripts/sync_allow_scripts_pins.py
new file mode 100644
index 0000000000..1d9a075ab4
--- /dev/null
+++ b/scripts/sync_allow_scripts_pins.py
@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+"""Keep `allowScripts` pins in studio/frontend/package.json in sync with
+package-lock.json.
+
+`npm approve-scripts` writes version-pinned entries ("pkg@1.2.3": true).
+A dependency bump strands the pin, so the approval (or denial) silently
+stops matching and the package's install scripts fall back to
+"unreviewed". This tool re-pins existing entries to the versions the
+lockfile actually resolves; it never adds or removes entries, so
+approving a brand-new script-bearing package stays a human decision.
+
+Usage:
+  python scripts/sync_allow_scripts_pins.py --check   # CI: exit 1 on drift
+  python scripts/sync_allow_scripts_pins.py --fix     # rewrite package.json
+
+Pinned keys follow npm's allowScripts grammar: "name@1.2.3" or
+"name@1.2.3 || 1.2.4". Bare names (no version) match every version and
+are left alone. Entries whose range is not an exact-version disjunction
+(wildcards, tags) are left alone too.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_DIR = REPO_ROOT / "studio" / "frontend"
+
+EXACT_VERSION_RE = re.compile(r"^\d+\.\d+\.\d+(?:[-+][0-9A-Za-z.+-]+)?$")
+
+
+def split_spec(key: str) -> tuple[str, str | None]:
+    """'@scope/name@1.2.3' -> ('@scope/name', '1.2.3'); bare names -> (key, None)."""
+    if key.startswith("@"):
+        rest = key[1:]
+        if "@" not in rest:
+            return key, None
+        name, rng = rest.split("@", 1)
+        return "@" + name, rng
+    if "@" not in key:
+        return key, None
+    name, rng = key.split("@", 1)
+    return name, rng
+
+
+def is_exact_disjunction(rng: str) -> bool:
+    parts = [p.strip() for p in rng.split("||")]
+    return all(EXACT_VERSION_RE.match(p) for p in parts) and bool(parts)
+
+
+def version_sort_key(version: str) -> tuple:
+    release = version.split("-", 1)[0].split("+", 1)[0]
+    return tuple(int(x) for x in release.split(".")), version
+
+
+def script_versions_from_lock(lock: dict) -> dict[str, list[str]]:
+    """Map package name -> sorted versions that carry install scripts."""
+    out: dict[str, set[str]] = {}
+    for path, meta in (lock.get("packages") or {}).items():
+        if not path or not meta.get("hasInstallScript"):
+            continue
+        name = path.rsplit("node_modules/", 1)[-1]
+        version = meta.get("version")
+        if name and version:
+            out.setdefault(name, set()).add(version)
+    return {n: sorted(vs, key = version_sort_key) for n, vs in out.items()}
+
+
+def desired_key(name: str, versions: list[str]) -> str:
+    return f"{name}@{' || '.join(versions)}"
+
+
+def compute_renames(policy: dict, lock_versions: dict[str, list[str]]) -> dict[str, str]:
+    renames: dict[str, str] = {}
+    for key in policy:
+        name, rng = split_spec(key)
+        if rng is None or not is_exact_disjunction(rng):
+            continue  # bare name or non-exact spec: matches by name, never stale
+        versions = lock_versions.get(name)
+        if not versions:
+            continue  # package gone or script-free now: stale pin is inert
+        want = desired_key(name, versions)
+        if key != want:
+            renames[key] = want
+    return renames
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(description = __doc__)
+    mode = ap.add_mutually_exclusive_group(required = True)
+    mode.add_argument("--check", action = "store_true", help = "exit 1 if pins are stale")
+    mode.add_argument("--fix", action = "store_true", help = "rewrite package.json in place")
+    ap.add_argument(
+        "--dir",
+        type = Path,
+        default = DEFAULT_DIR,
+        help = "directory holding package.json + package-lock.json",
+    )
+    args = ap.parse_args(argv)
+
+    pkg_path = args.dir / "package.json"
+    lock_path = args.dir / "package-lock.json"
+    if not pkg_path.exists() or not lock_path.exists():
+        print(f"sync-allow-scripts: nothing to do ({args.dir} has no package.json + lockfile)")
+        return 0
+
+    pkg = json.loads(pkg_path.read_text(encoding = "utf-8"))
+    policy = pkg.get("allowScripts")
+    if not isinstance(policy, dict) or not policy:
+        print("sync-allow-scripts: no allowScripts policy in package.json, nothing to do")
+        return 0
+
+    lock = json.loads(lock_path.read_text(encoding = "utf-8"))
+    renames = compute_renames(policy, script_versions_from_lock(lock))
+
+    if not renames:
+        print(f"sync-allow-scripts: {len(policy)} allowScripts entries in sync with the lockfile")
+        return 0
+
+    for old, new in renames.items():
+        print(f'  stale pin: "{old}" -> "{new}"')
+
+    if args.check:
+        print(
+            "sync-allow-scripts: pins are stale; run "
+            "`python scripts/sync_allow_scripts_pins.py --fix` and commit the result"
+        )
+        return 1
+
+    pkg["allowScripts"] = {renames.get(k, k): v for k, v in policy.items()}
+    pkg_path.write_text(json.dumps(pkg, indent = 2, ensure_ascii = False) + "\n", encoding = "utf-8")
+    print(
+        f"sync-allow-scripts: re-pinned {len(renames)} entr{'y' if len(renames) == 1 else 'ies'} in {pkg_path}"
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1
index bcec58caf1..e70b8d413f 100644
--- a/scripts/uninstall.ps1
+++ b/scripts/uninstall.ps1
@@ -16,15 +16,19 @@ function Uninstall-UnslothStudio {
     function _Step { param([string]$Msg) Write-Host $Msg }
     function _Substep { param([string]$Msg, [string]$Color = "Gray") Write-Host "  $Msg" -ForegroundColor $Color }
 
-    # Remove a file/dir/symlink only if it exists. Idempotent.
+    # Remove a file/dir/symlink if present. Idempotent; retries since a just-killed
+    # process can briefly hold a handle (Windows refuses the delete until released).
     function _RemovePath {
         param([string]$Path)
         if ([string]::IsNullOrWhiteSpace($Path)) { return }
-        if (Test-Path -LiteralPath $Path) {
+        if (-not (Test-Path -LiteralPath $Path)) { return }
+        for ($attempt = 1; $attempt -le 3; $attempt++) {
             try {
                 Remove-Item -LiteralPath $Path -Recurse -Force -ErrorAction Stop
                 _Substep "removed: $Path" "Green"
+                return
             } catch {
+                if ($attempt -lt 3) { Start-Sleep -Milliseconds 700; continue }
                 _Substep "could not remove: $Path ($($_.Exception.Message))" "Yellow"
             }
         }
@@ -236,9 +240,58 @@ function Uninstall-UnslothStudio {
         } catch { }
     }
 
+    # Stop processes that would block deleting the paths we remove. Unlike
+    # _StopStudioProcesses (venv exe only), this also catches llama-server/llama-cli,
+    # the unsloth.exe shim, and orphaned mp workers under SYSTEM python holding a
+    # venv DLL (an open DLL handle blocks the dir delete) -- found by scanning each
+    # candidate's loaded modules, not just its image path.
+    function _StopProcessesLockingRoots {
+        param([string[]]$Roots)
+        $clean = @($Roots | Where-Object { $_ } | ForEach-Object { $_.TrimEnd('\','/') })
+        if ($clean.Count -eq 0) { return }
+        $underRoot = {
+            param($p)
+            if (-not $p) { return $false }
+            foreach ($r in $clean) { if ($p -ieq $r -or $p -ilike "$r\*") { return $true } }
+            return $false
+        }
+        # 1. Image path under a target root (venv python, shim, llama-server).
+        try {
+            foreach ($proc in (Get-CimInstance Win32_Process -ErrorAction SilentlyContinue)) {
+                if ((& $underRoot $proc.ExecutablePath)) {
+                    try { Stop-Process -Id $proc.ProcessId -Force -ErrorAction SilentlyContinue } catch { }
+                }
+            }
+        } catch { }
+        # 2. A loaded module under a target root (orphaned mp-fork python holding a
+        #    venv DLL). Scoped to names that load our DLLs to keep the scan fast.
+        try {
+            $cands = Get-Process -Name python, pythonw, unsloth, llama-server, llama-cli -ErrorAction SilentlyContinue
+            foreach ($proc in $cands) {
+                $hit = $false
+                try {
+                    foreach ($m in $proc.Modules) { if ((& $underRoot $m.FileName)) { $hit = $true; break } }
+                } catch { }  # access denied enumerating modules -> skip
+                if ($hit) { try { Stop-Process -Id $proc.Id -Force -ErrorAction SilentlyContinue } catch { } }
+            }
+        } catch { }
+    }
+
     # Default install root + default data dir.
     $defaultStudioHome = if ($env:USERPROFILE) { Join-Path $env:USERPROFILE ".unsloth\studio" } else { $null }
     $defaultDataDir = if ($env:LOCALAPPDATA) { Join-Path $env:LOCALAPPDATA "Unsloth Studio" } else { $null }
+    # Default-mode ~/.unsloth holds a SHARED llama.cpp build + .cache that are
+    # siblings of studio (not under it), so deleting <studio> misses them -- handle
+    # explicitly. No-op in env/custom mode (nested under the custom root, removed
+    # with it). A user-set UNSLOTH_LLAMA_CPP_PATH is left alone.
+    $defaultUnslothHome = if ($env:USERPROFILE) { Join-Path $env:USERPROFILE ".unsloth" } else { $null }
+    $defaultLlamaCpp = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome "llama.cpp" } else { $null }
+    $defaultCache = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome ".cache" } else { $null }
+    # llama.cpp atomic-install staging root (install_llama_prebuilt.py .staging,
+    # sibling of the install dir). Usually pruned after activate, but an interrupted
+    # build can leave a "<name>.staging-XXXX" tree; removing it lets the empty-dir
+    # cleanup of ~/.unsloth below succeed. No-op in env/custom mode and when absent.
+    $defaultStaging = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome ".staging" } else { $null }
 
     # Build known-root list FIRST so the port-file kill can verify ownership.
     $customRoots = @(_CustomStudioRoots)
@@ -255,6 +308,9 @@ function Uninstall-UnslothStudio {
         _StopByPortFile -PortFile (Join-Path $r "share\studio.port") -KnownRoots $knownRoots
     }
     _StopStudioProcesses -KnownRoots $knownRoots
+    # Also stop anything holding a handle on the exact paths we delete (llama-server,
+    # the CLI shim, an mp-fork python with a venv DLL) so the dir delete isn't refused.
+    _StopProcessesLockingRoots -Roots (@($knownRoots) + @($defaultDataDir, $defaultLlamaCpp, $defaultCache))
 
     # ── Remove custom-root install trees ──
     _Step "Removing data and install directories..."
@@ -273,6 +329,16 @@ function Uninstall-UnslothStudio {
     if ($defaultStudioHome) { _RemovePath $defaultStudioHome }
     # Default data dir.
     if ($defaultDataDir) { _RemovePath $defaultDataDir }
+    # Default-mode shared llama.cpp build + cache (siblings of studio under
+    # ~/.unsloth). No-op in env/custom mode and when absent.
+    if ($defaultLlamaCpp) { _RemovePath $defaultLlamaCpp }
+    if ($defaultCache) { _RemovePath $defaultCache }
+    if ($defaultStaging) { _RemovePath $defaultStaging }
+    # Drop ~/.unsloth itself, but ONLY if now empty -- never nuke unrelated content.
+    if ($defaultUnslothHome -and (Test-Path -LiteralPath $defaultUnslothHome) -and
+        -not (Get-ChildItem -LiteralPath $defaultUnslothHome -Force -ErrorAction SilentlyContinue)) {
+        _RemovePath $defaultUnslothHome
+    }
 
     # ── Remove desktop and Start Menu shortcuts ──
     _Step "Removing desktop and Start Menu shortcuts..."
@@ -283,6 +349,18 @@ function Uninstall-UnslothStudio {
     if ($env:APPDATA) {
         _RemovePath (Join-Path $env:APPDATA "Microsoft\Windows\Start Menu\Programs\Unsloth Studio.lnk")
     }
+    # Invalidate the Win11 Start Menu tile cache so the removed shortcut's tile
+    # disappears promptly instead of lingering stale (mirrors install.ps1's
+    # New-StudioShortcuts). Preserves start2.bin (the pin layout).
+    try {
+        $smehTemp = Join-Path $env:LOCALAPPDATA "Packages\Microsoft.Windows.StartMenuExperienceHost_cw5n1h2txyewy\TempState"
+        if (Test-Path -LiteralPath $smehTemp) {
+            Get-ChildItem -LiteralPath $smehTemp -Filter "TileCache_*" -ErrorAction SilentlyContinue |
+                Remove-Item -Force -ErrorAction SilentlyContinue
+            Remove-Item -LiteralPath (Join-Path $smehTemp "StartUnifiedTileModelCache.dat") -Force -ErrorAction SilentlyContinue
+            Stop-Process -Name StartMenuExperienceHost -Force -ErrorAction SilentlyContinue
+        }
+    } catch { }
 
     # ── Clean user PATH and registry backup ──
     _Step "Cleaning user PATH and registry..."
diff --git a/scripts/uninstall.sh b/scripts/uninstall.sh
index d40fa57159..8abc8bf3b9 100755
--- a/scripts/uninstall.sh
+++ b/scripts/uninstall.sh
@@ -212,10 +212,24 @@ _custom_studio_roots | while IFS= read -r _custom_root; do
     _remove_path "$_custom_root"
 done
 _remove_path "$HOME/.unsloth/studio"
-# CUDA llama.cpp from provision_llama_cuda.sh (+ the fetched script). Clears the
-# native-Linux build dir, or on WSL the symlink to the build install.ps1 removes.
+# Default-mode shared llama.cpp build + cache are siblings of studio (not removed
+# by deleting it). No-op in env/custom mode (they nest under the custom root) and
+# when absent. A user-set UNSLOTH_LLAMA_CPP_PATH is intentionally kept.
 _remove_path "$HOME/.unsloth/llama.cpp"
+# provision_llama_cuda.sh fetched by the WoA/Spark CUDA-build path (install.ps1
+# background build + direct-WSL setup.sh). No-op when absent.
 _remove_path "$HOME/.unsloth/provision_llama_cuda.sh"
+_remove_path "$HOME/.unsloth/.cache"
+# llama.cpp atomic-install staging root (install_llama_prebuilt.py .staging).
+# Normally pruned after activate, but an interrupted build can leave it behind;
+# removing it lets the rmdir below succeed. No-op in env/custom mode and absent.
+_remove_path "$HOME/.unsloth/.staging"
+# ROCm-on-WSL helper artifacts (librocdxg build clone + smoke-test venv). No-op
+# where they don't exist; removing them lets the rmdir below succeed.
+_remove_path "$HOME/.unsloth/librocdxg"
+_remove_path "$HOME/.unsloth/rocm-smoketest"
+# Drop ~/.unsloth only if now empty (rmdir refuses non-empty, so user content is kept).
+rmdir "$HOME/.unsloth" 2>/dev/null || true
 _remove_path "$HOME/.local/share/unsloth"
 # CLI shim: only the symlink Studio created, never a pip-installed file.
 _remove_cli_shim
@@ -248,22 +262,50 @@ case "$_os" in
     Linux)
         if [ "$_is_wsl" = "1" ]; then
             echo "Removing WSL Windows-side shortcuts..."
-            # install.sh creates 'Unsloth Studio.lnk' on the Windows Desktop and
-            # Start Menu Programs folder via powershell.exe; mirror that path.
-            if command -v powershell.exe >/dev/null 2>&1; then
+            # install.sh creates per-distro 'Unsloth Studio (WSL - <distro>).lnk'
+            # on the Windows Desktop + Start Menu via powershell.exe. Scope removal
+            # to THIS distro (passed as $args[0]) so a multi-distro install keeps the
+            # other distros' launchers; the TARGET=wsl.exe check still spares a
+            # native install's "Unsloth Studio.lnk". Prefer powershell.exe; test it
+            # can EXECUTE (`command -v` succeeds even with interop OFF -- .exe then
+            # fails "Exec format error", common on systemd-enabled distros).
+            _wsl_distro="${WSL_DISTRO_NAME:-}"
+            _ps_ran=0
+            if command -v powershell.exe >/dev/null 2>&1 && \
+               powershell.exe -NoProfile -Command "exit 0" >/dev/null 2>&1; then
+                _ps_ran=1
+                # Inject the distro into the command: a -Command string does not
+                # receive trailing tokens as $args. WSL distro names are safe to
+                # embed (no quotes/$/backtick).
                 # shellcheck disable=SC2016
-                # $env:APPDATA is a PowerShell expansion; intentionally literal at shell level.
-                powershell.exe -NoProfile -Command '
+                # $env:APPDATA/$distro are PowerShell-side; $_wsl_distro is injected from shell.
+                powershell.exe -NoProfile -Command '$distro = "'"$_wsl_distro"'";
                     $dirs = @(
                         [Environment]::GetFolderPath("Desktop"),
                         (Join-Path $env:APPDATA "Microsoft\Windows\Start Menu\Programs")
                     );
+                    $ws = New-Object -ComObject WScript.Shell;
                     foreach ($d in $dirs) {
-                        if (-not $d) { continue }
-                        $p = Join-Path $d "Unsloth Studio.lnk";
-                        if (Test-Path -LiteralPath $p) { Remove-Item -LiteralPath $p -Force }
+                        if (-not $d -or -not (Test-Path -LiteralPath $d)) { continue }
+                        Get-ChildItem -LiteralPath $d -Filter "Unsloth Studio*.lnk" -ErrorAction SilentlyContinue | ForEach-Object {
+                            try {
+                                $sc = $ws.CreateShortcut($_.FullName);
+                                if ("$($sc.TargetPath) $($sc.Arguments)" -notmatch "wsl\.exe") { return }
+                                # When the distro is known, require the per-distro
+                                # name for this distro or its -d "<distro>" argument
+                                # so launchers for other distros are not removed.
+                                if ($distro) {
+                                    $nameMatch = ($_.Name -eq "Unsloth Studio (WSL - $distro).lnk");
+                                    $argMatch  = ($sc.Arguments -match ("-d\s+`"?" + [regex]::Escape($distro) + "`"?"));
+                                    if (-not ($nameMatch -or $argMatch)) { return }
+                                }
+                                Remove-Item -LiteralPath $_.FullName -Force -ErrorAction SilentlyContinue
+                            } catch { }
+                        }
                     }
-                    # WSL-fallback native shim/launcher dir (%LOCALAPPDATA%\Unsloth) + its PATH entry.
+                    # WoA WSL-fallback (install.ps1) native shim/launcher dir
+                    # (%LOCALAPPDATA%\Unsloth) + its PATH entry. install.ps1 created the
+                    # shim; clean it here too so a WSL-side bash uninstall is complete.
                     $ud = if ($env:LOCALAPPDATA) { Join-Path $env:LOCALAPPDATA "Unsloth" } else { $null };
                     if ($ud) {
                         $shim = (Join-Path $ud "bin").TrimEnd("\","/");
@@ -272,6 +314,61 @@ case "$_os" in
                         if (Test-Path -LiteralPath $ud) { Remove-Item -LiteralPath $ud -Recurse -Force -ErrorAction SilentlyContinue }
                     }' >/dev/null 2>&1 || true
             fi
+            # Fallback when powershell.exe can't run (interop disabled): remove the
+            # WSL .lnk files via drvfs. The "Unsloth Studio (WSL..." name is
+            # WSL-specific, so a native install's "Unsloth Studio.lnk" never matches.
+            if [ "$_ps_ran" = "0" ]; then
+                for _drive in /mnt/c /mnt/d /mnt/e; do
+                    [ -d "$_drive/Users" ] || continue
+                    for _udir in "$_drive"/Users/*; do
+                        [ -d "$_udir" ] || continue
+                        for _scdir in \
+                            "$_udir/Desktop" \
+                            "$_udir/OneDrive/Desktop" \
+                            "$_udir"/OneDrive*/Desktop \
+                            "$_udir/AppData/Roaming/Microsoft/Windows/Start Menu/Programs"; do
+                            [ -d "$_scdir" ] || continue
+                            if [ -n "$_wsl_distro" ]; then
+                                # Exact per-distro name (no glob) so other distros survive.
+                                _lnk="$_scdir/Unsloth Studio (WSL - ${_wsl_distro}).lnk"
+                                [ -e "$_lnk" ] && rm -f "$_lnk" 2>/dev/null && echo "  removed: $_lnk" || true
+                            else
+                                # Distro unknown: fall back to the broad WSL prefix.
+                                for _lnk in "$_scdir"/"Unsloth Studio (WSL"*.lnk; do
+                                    [ -e "$_lnk" ] && rm -f "$_lnk" 2>/dev/null && echo "  removed: $_lnk" || true
+                                done
+                            fi
+                        done
+                    done
+                done
+            fi
+            # ── ROCm-on-WSL config (install_rocm_wsl_strixhalo.sh) ──
+            # Remove Unsloth's own ROCDXG config (the env it persisted). The system
+            # ROCm userspace is a shared prereq (like CUDA) and is LEFT IN PLACE by
+            # default; set UNSLOTH_UNINSTALL_ROCM=1 to remove it too.
+            echo "Removing ROCm-on-WSL config..."
+            _sudo=""
+            if [ "$_uid" != "0" ] && command -v sudo >/dev/null 2>&1; then _sudo="sudo"; fi
+            $_sudo rm -f /etc/profile.d/unsloth-rocm-wsl.sh 2>/dev/null || true
+            if [ -f "$HOME/.bashrc" ] && grep -q "Unsloth ROCm-on-WSL" "$HOME/.bashrc" 2>/dev/null; then
+                _bk=$(mktemp 2>/dev/null || echo "$HOME/.bashrc.unsloth.tmp")
+                if sed '/# >>> Unsloth ROCm-on-WSL/,/# <<< Unsloth ROCm-on-WSL/d' "$HOME/.bashrc" > "$_bk" 2>/dev/null; then
+                    cat "$_bk" > "$HOME/.bashrc" 2>/dev/null || true
+                    echo "  cleaned ROCm-on-WSL block from ~/.bashrc"
+                fi
+                rm -f "$_bk" 2>/dev/null || true
+            fi
+            if [ "${UNSLOTH_UNINSTALL_ROCM:-0}" = "1" ]; then
+                echo "  removing system ROCm (UNSLOTH_UNINSTALL_ROCM=1)..."
+                $_sudo rm -f /etc/apt/sources.list.d/rocm.list /etc/apt/preferences.d/rocm-pin-600 \
+                    /etc/apt/keyrings/rocm.gpg /etc/ld.so.conf.d/rocm.conf 2>/dev/null || true
+                $_sudo sh -c 'rm -rf /opt/rocm /opt/rocm-*' 2>/dev/null || true
+                if command -v ldconfig >/dev/null 2>&1; then $_sudo ldconfig 2>/dev/null || true; fi
+            elif [ -d /opt/rocm ]; then
+                echo "  Note: ROCm userspace (/opt/rocm*) left in place (shared prereq)."
+                echo "        Remove it by re-running with UNSLOTH_UNINSTALL_ROCM=1, or manually:"
+                echo "          sudo rm -rf /opt/rocm /opt/rocm-* && sudo ldconfig"
+            fi
         fi
         echo "Removing Linux .desktop entry..."
         _remove_path "$HOME/.local/share/applications/unsloth-studio.desktop"
diff --git a/studio/backend/cloudflare_tunnel.py b/studio/backend/cloudflare_tunnel.py
new file mode 100644
index 0000000000..7c088b6e1e
--- /dev/null
+++ b/studio/backend/cloudflare_tunnel.py
@@ -0,0 +1,302 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Free Cloudflare quick tunnel for Studio's 0.0.0.0 launches.
+
+The raw http://<ip>:<port> is often unreachable (https-vs-http, blocked ports,
+closed security groups); a cloudflared quick tunnel gives a free
+https://*.trycloudflare.com URL that works anywhere, with no account or domain.
+
+Best-effort throughout: any failure collapses to "no URL" and Studio keeps
+running. Stdlib only (back-end imports are lazy) so it is safe to import early.
+"""
+
+from __future__ import annotations
+
+import os
+import platform
+import re
+import shutil
+import subprocess
+import sys
+import threading
+from pathlib import Path
+from typing import Optional, Tuple
+
+# cloudflared logs the quick-tunnel URL; match only the URL so we do not depend
+# on the surrounding wording, which Cloudflare may change.
+_URL_RE = re.compile(r"https://[A-Za-z0-9-]+\.trycloudflare\.com")
+
+_RELEASE_BASE = "https://github.com/cloudflare/cloudflared/releases/latest/download"
+
+_URL_TIMEOUT = 15.0  # seconds to wait for the public URL before giving up
+_DOWNLOAD_TIMEOUT = 60  # urlopen timeout for the one-time binary download
+
+
+def _windows_hidden_kwargs() -> dict:
+    """Suppress a child console window on Windows; no-op elsewhere."""
+    if sys.platform != "win32":
+        return {}
+    flags = getattr(subprocess, "CREATE_NO_WINDOW", 0)
+    return {"creationflags": flags} if flags else {}
+
+
+def _asset_name() -> Optional[Tuple[str, bool]]:
+    """(release asset filename, is_tgz) for this OS/arch, or None if unsupported."""
+    system = platform.system().lower()
+    machine = platform.machine().lower()
+    is_x64 = machine in ("x86_64", "amd64", "x64")
+    is_arm64 = machine in ("aarch64", "arm64")
+    is_x86 = machine in ("i386", "i686", "x86")
+    if system == "linux":
+        if is_x64:
+            return ("cloudflared-linux-amd64", False)
+        if is_arm64:
+            return ("cloudflared-linux-arm64", False)
+    elif system == "darwin":
+        if is_arm64:
+            return ("cloudflared-darwin-arm64.tgz", True)
+        if is_x64:
+            return ("cloudflared-darwin-amd64.tgz", True)
+    elif system == "windows":
+        if is_x64:
+            return ("cloudflared-windows-amd64.exe", False)
+        if is_x86:
+            return ("cloudflared-windows-386.exe", False)
+    return None
+
+
+def _cache_path() -> Optional[Path]:
+    """studio_bin_root()/cloudflared(.exe), or None if the studio home is unresolvable."""
+    try:
+        from utils.paths.storage_roots import studio_bin_root  # lazy: backend-only import
+    except Exception:
+        return None
+    name = "cloudflared.exe" if sys.platform == "win32" else "cloudflared"
+    return studio_bin_root() / name
+
+
+def find_cloudflared() -> Optional[str]:
+    """Locate an existing cloudflared: PATH first, then the Studio bin cache."""
+    on_path = shutil.which("cloudflared")
+    if on_path:
+        return on_path
+    cached = _cache_path()
+    if cached is not None and cached.is_file() and os.access(cached, os.X_OK):
+        return str(cached)
+    return None
+
+
+def _download(url: str, dest: Path) -> bool:
+    """Download url to dest via urllib (temp file + atomic rename). Best-effort -> bool."""
+    import tempfile
+    import urllib.request
+
+    tmp_path: Optional[Path] = None
+    try:
+        dest.parent.mkdir(parents = True, exist_ok = True)
+        with tempfile.NamedTemporaryFile(
+            prefix = dest.name + ".tmp-", dir = dest.parent, delete = False
+        ) as handle:
+            tmp_path = Path(handle.name)
+            # GitHub's CDN 403s the default Python-urllib User-Agent.
+            req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
+            with urllib.request.urlopen(req, timeout = _DOWNLOAD_TIMEOUT) as response:
+                shutil.copyfileobj(response, handle)
+        if tmp_path.stat().st_size == 0:
+            raise RuntimeError("empty download")
+        os.replace(tmp_path, dest)
+        return True
+    except Exception:
+        if tmp_path is not None:
+            try:
+                tmp_path.unlink(missing_ok = True)
+            except Exception:
+                pass
+        return False
+
+
+def _extract_tgz_member(tgz_path: Path, dest: Path) -> bool:
+    """Extract just the `cloudflared` member from a darwin .tgz to dest.
+
+    Rejects absolute paths and `..` traversal so a hostile archive cannot write
+    outside dest. Best-effort -> bool.
+    """
+    import tarfile
+    try:
+        with tarfile.open(tgz_path, "r:gz") as tar:
+            member = None
+            for m in tar.getmembers():
+                if not m.isfile() or os.path.basename(m.name) != "cloudflared":
+                    continue
+                if m.name.startswith("/") or ".." in Path(m.name).parts:
+                    continue
+                member = m
+                break
+            if member is None:
+                return False
+            src = tar.extractfile(member)
+            if src is None:
+                return False
+            with src, open(dest, "wb") as out:
+                shutil.copyfileobj(src, out)
+        return True
+    except Exception:
+        return False
+
+
+def ensure_cloudflared() -> Optional[str]:
+    """Return a cloudflared path, downloading + caching the binary once if missing."""
+    existing = find_cloudflared()
+    if existing:
+        return existing
+    asset = _asset_name()
+    cached = _cache_path()
+    if asset is None or cached is None:
+        return None
+    name, is_tgz = asset
+    url = f"{_RELEASE_BASE}/{name}"
+    try:
+        cached.parent.mkdir(parents = True, exist_ok = True)
+        if is_tgz:
+            tgz = cached.with_suffix(".tgz")
+            if not _download(url, tgz) or not _extract_tgz_member(tgz, cached):
+                tgz.unlink(missing_ok = True)
+                return None
+            tgz.unlink(missing_ok = True)
+        elif not _download(url, cached):
+            return None
+        if sys.platform != "win32":
+            os.chmod(cached, 0o755)
+        return str(cached)
+    except Exception:
+        return None
+
+
+class CloudflareTunnel:
+    """A cloudflared quick tunnel to http://localhost:<port>. Best-effort throughout.
+
+    Use localhost (not the wildcard bind) as the tunnel origin so cloudflared's
+    upstream stays local-only.
+    """
+
+    def __init__(self, port: int, binary: str):
+        self.port = port
+        self.binary = binary
+        self._proc: Optional[subprocess.Popen] = None
+        self._lock = threading.Lock()
+        self._url_event = threading.Event()
+        self.url: Optional[str] = None
+        self.error: Optional[str] = None
+
+    def start(self) -> None:
+        cmd = [
+            self.binary,
+            "tunnel",
+            "--url",
+            f"http://localhost:{self.port}",
+            "--no-autoupdate",
+        ]
+        proc = subprocess.Popen(
+            cmd,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
+            stdin = subprocess.DEVNULL,
+            text = True,
+            errors = "replace",
+            bufsize = 1,
+            **_windows_hidden_kwargs(),
+        )
+        with self._lock:
+            self._proc = proc
+        threading.Thread(
+            target = self._reader, args = (proc,), name = "cloudflared-reader", daemon = True
+        ).start()
+
+    def _reader(self, proc: subprocess.Popen) -> None:
+        # Drain cloudflared's output, capture the first trycloudflare URL, and
+        # keep draining so it never blocks on a full pipe.
+        try:
+            if proc.stdout is not None:
+                for line in proc.stdout:
+                    if self.url is None:
+                        match = _URL_RE.search(line)
+                        if match:
+                            self.url = match.group(0)
+                            self._url_event.set()
+        except Exception:
+            pass
+        finally:
+            if self.url is None:
+                self.error = "cloudflared exited before emitting a tunnel URL"
+                self._url_event.set()
+
+    def wait_for_url(self, timeout: float = _URL_TIMEOUT) -> Optional[str]:
+        self._url_event.wait(timeout)
+        return self.url
+
+    def stop(self) -> None:
+        """Terminate the tunnel. Idempotent and safe to call from a signal handler."""
+        with self._lock:
+            proc, self._proc = self._proc, None
+        if proc is None:
+            return
+        try:
+            if proc.poll() is None:
+                proc.terminate()
+                try:
+                    proc.wait(timeout = 5)
+                except subprocess.TimeoutExpired:
+                    proc.kill()
+                    try:
+                        proc.wait(timeout = 5)
+                    except Exception:
+                        pass
+        except Exception:
+            pass
+
+
+# Single serving process per Studio launch, so one module-level tunnel handle is
+# enough; the lock guards the start/stop/shutdown races.
+_active_tunnel: Optional[CloudflareTunnel] = None
+_active_lock = threading.Lock()
+
+
+def start_studio_tunnel(port: int, timeout: float = _URL_TIMEOUT) -> Optional[str]:
+    """Start a quick tunnel and return its public URL, or None (best-effort).
+
+    On any failure (no binary, no URL within timeout, early crash) the tunnel is
+    stopped and None is returned, so the caller prints a hint and continues.
+    """
+    global _active_tunnel
+    binary = ensure_cloudflared()
+    if not binary:
+        return None
+    tunnel = CloudflareTunnel(port, binary)
+    # Register before start/wait so a shutdown during the URL wait can stop it.
+    with _active_lock:
+        prior, _active_tunnel = _active_tunnel, tunnel
+    if prior is not None:
+        prior.stop()
+    try:
+        tunnel.start()
+        url = tunnel.wait_for_url(timeout)
+    except Exception:
+        url = None
+    if url:
+        return url
+    # No URL (or crash): drop it unless a concurrent shutdown already replaced it.
+    with _active_lock:
+        if _active_tunnel is tunnel:
+            _active_tunnel = None
+    tunnel.stop()
+    return None
+
+
+def stop_studio_tunnel() -> None:
+    """Terminate the active tunnel, if any. Idempotent."""
+    global _active_tunnel
+    with _active_lock:
+        tunnel, _active_tunnel = _active_tunnel, None
+    if tunnel is not None:
+        tunnel.stop()
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index a7d44b992d..cae001c34d 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -469,9 +469,24 @@ def _apply_mistral_reasoning_controls(
 
 
 # Shared client reused across all requests for HTTP connection pooling.
-# Auth headers and timeouts are per-request, so one client handles every
-# provider without storing credentials.
-_http_client = httpx.AsyncClient()
+# Auth headers and timeouts are passed per-request, so a single client
+# handles every provider without storing credentials.
+def _create_shared_http_client() -> httpx.AsyncClient:
+    # Unsupported env proxy schemes (socks:// etc) raise at construction and
+    # would crash Studio startup (#6090); retry ignoring env proxies instead.
+    try:
+        return httpx.AsyncClient()
+    except (ImportError, ValueError) as exc:
+        exc_str = str(exc)
+        if "Unknown scheme for proxy URL" not in exc_str and "socksio" not in exc_str:
+            raise
+        logger.warning(
+            "Ignoring unsupported environment proxy for the shared HTTP client: %s", exc_str
+        )
+        return httpx.AsyncClient(trust_env = False)
+
+
+_http_client = _create_shared_http_client()
 
 
 # Cap per-image fetch well below Gemini's ~20 MB total request budget.
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 55c2c551a3..2b9517692f 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -25,6 +25,7 @@ from utils.hardware import (
     get_visible_gpu_count,
 )
 from core.inference.audio_codecs import AudioCodecManager
+from core.inference.runtime_context import runtime_context_length
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -405,6 +406,10 @@ class InferenceBackend:
 
                 # Reject CPU/disk offload for audio models too
                 raise_if_offloaded(self.models[model_name]["model"], device_map, "Inference")
+                self.models[model_name]["context_length"] = runtime_context_length(
+                    self.models[model_name].get("model"),
+                    max_seq_length,
+                )
 
                 self.active_model_name = model_name
                 self.loading_models.discard(model_name)
@@ -485,6 +490,10 @@ class InferenceBackend:
                 self.models[model_name]["tokenizer"] = tokenizer
 
             raise_if_offloaded(self.models[model_name]["model"], device_map, "Inference")
+            self.models[model_name]["context_length"] = runtime_context_length(
+                self.models[model_name].get("model"),
+                max_seq_length,
+            )
 
             self._load_chat_template_info(model_name)
 
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 953c1d3141..b83e4e6961 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -23,7 +23,7 @@ import sys
 import threading
 import time
 from pathlib import Path
-from typing import Generator, Iterable, List, Optional
+from typing import Callable, Generator, Iterable, List, Optional
 
 import httpx
 
@@ -481,6 +481,54 @@ def _is_mtp_model_name(model_identifier: Optional[str], gguf_path: Optional[str]
     return False
 
 
+def _is_companion_gguf_path(path: str) -> bool:
+    """True for a non-main GGUF: vision mmproj or a separate MTP drafter
+    (repo-root ``mtp-*.gguf`` or the ``MTP/`` subdir copies, Gemma 4).
+
+    Mirrors hub.utils.gguf so variant resolution never picks a companion as
+    the main model -- e.g. a Gemma ``Q8_0`` request must not resolve to the
+    ``MTP/...-Q8_0-MTP.gguf`` drafter, which sorts ahead of the real weight.
+    """
+    p = path.lower()
+    if not p.endswith(".gguf"):
+        return False
+    if "mmproj" in p:
+        return True
+    name = p.rsplit("/", 1)[-1]
+    return name.startswith("mtp-") or "/mtp/" in f"/{p}"
+
+
+# Below this many B params, draft-mtp regresses vs spec-off (bench in
+# _build_speculative_flags); auto mode drops MTP under it.
+_MTP_MIN_SIZE_B = 3.0
+
+# Context-fit VRAM budget: tighter than _GPU_PIN_VRAM_FRACTION (0.95) on
+# purpose -- over-promising context OOMs at runtime (#5106).
+_CTX_FIT_VRAM_FRACTION = 0.90
+
+# Extra VRAM fraction reserved when MTP will engage: the draft model's
+# weights, KV cache, and compute buffers live outside the main model's
+# estimate. Applied to BOTH the ctx-fit budget and the GPU pin thresholds --
+# tightening only the fit lets a load whose weights land between the two
+# fractions pin without any room for the drafter.
+_MTP_VRAM_RESERVE_FRAC = 0.05
+
+
+def _auto_mode_drops_mtp(
+    req_mode: Optional[str],
+    size_b: Optional[float],
+    *,
+    has_separate_drafter: bool = False,
+) -> bool:
+    """Auto mode drops MTP below _MTP_MIN_SIZE_B for an embedded draft head
+    (its per-token cost regresses there); a separate drafter (Gemma) is a tiny
+    standalone model that still speeds up below 3B, so it never drops. Forced
+    mtp / mtp+ngram engage regardless of size."""
+    if has_separate_drafter:
+        return False
+    return req_mode == "auto" and size_b is not None and size_b < _MTP_MIN_SIZE_B
+
+
 def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool:
     """User passed --spec-type / --spec-default? llama-server takes one
     --spec-type (comma-separated to chain), so suppress auto-emit."""
@@ -625,6 +673,14 @@ class LlamaCppBackend:
         self._model_identifier: Optional[str] = None
         self._gguf_path: Optional[str] = None
         self._hf_repo: Optional[str] = None
+        # Separate MTP drafter launched with the current model; reload-dedup
+        # key so a drafter that appears next to the weights forces a reload.
+        self._mtp_draft_path: Optional[str] = None
+        # Why MTP was disabled on the last load that asked for it (auto on an
+        # MTP model, or forced mtp / mtp+ngram), else None. Drives the "update
+        # llama.cpp" hint in the UI. "binary_no_mtp" / "binary_outdated" ->
+        # a newer prebuilt would help; "runtime_error" -> it may not.
+        self._spec_fallback_reason: Optional[str] = None
         self._hf_variant: Optional[str] = None
         self._is_vision: bool = False
         self._healthy = False
@@ -677,6 +733,9 @@ class LlamaCppBackend:
         # Wraps load_model() end-to-end so concurrent loads serialise and never
         # coexist as two llama-server processes (#5401).
         self._serial_load_lock = threading.Lock()
+        # Set by the in-app updater while it swaps prebuilt binaries; load_model()
+        # rejects fast so no server starts from a half-swapped binary.
+        self._llama_update_in_progress = False
         # Last extra_args / requested n_ctx, preserved across unload so the chat
         # UI's /unload+/load Apply path can inherit them (#5401).
         # ``_extra_args_source`` records the (model_identifier, hf_variant) the
@@ -732,6 +791,19 @@ class LlamaCppBackend:
     def hf_variant(self) -> Optional[str]:
         return self._hf_variant
 
+    @property
+    def gguf_path(self) -> Optional[str]:
+        return self._gguf_path
+
+    @property
+    def mtp_draft_path(self) -> Optional[str]:
+        return self._mtp_draft_path
+
+    @property
+    def spec_fallback_reason(self) -> Optional[str]:
+        """Why MTP was disabled on the last MTP-requesting load, else None."""
+        return self._spec_fallback_reason
+
     @property
     def extra_args(self) -> Optional[List[str]]:
         """Extra llama-server flags from the last load (a copy). None =
@@ -1081,6 +1153,8 @@ class LlamaCppBackend:
                 "ngram_mod_flavor": None,
                 "supports_ngram_mod": False,
                 "spec_draft_n_max_flag": None,
+                "supports_kv_unified": False,
+                "supports_fit_ctx": False,
             }
         try:
             mtime = int(Path(bin_path).stat().st_mtime)
@@ -1094,6 +1168,8 @@ class LlamaCppBackend:
         mtp_token: Optional[str] = None
         ngram_mod_flavor: Optional[str] = None
         spec_draft_n_max_flag: Optional[str] = None
+        supports_kv_unified = False
+        supports_fit_ctx = False
         try:
             result = subprocess.run(
                 [bin_path, "--help"],
@@ -1181,6 +1257,9 @@ class LlamaCppBackend:
                 spec_draft_n_max_flag = "--spec-draft-n-max"
             elif _is_real("--draft-max"):
                 spec_draft_n_max_flag = "--draft-max"
+
+            supports_kv_unified = _is_real("--kv-unified")
+            supports_fit_ctx = _is_real("--fit-ctx")
         except (OSError, subprocess.SubprocessError) as exc:
             logger.debug(f"llama-server --help probe failed: {exc}")
 
@@ -1191,6 +1270,8 @@ class LlamaCppBackend:
             "ngram_mod_flavor": ngram_mod_flavor,
             "supports_ngram_mod": ngram_mod_flavor is not None,
             "spec_draft_n_max_flag": spec_draft_n_max_flag,
+            "supports_kv_unified": supports_kv_unified,
+            "supports_fit_ctx": supports_fit_ctx,
         }
         cls._capability_cache[cache_key] = info
         return info
@@ -1484,13 +1565,16 @@ class LlamaCppBackend:
 
     @staticmethod
     def _select_gpus(
-        model_size_bytes: int, gpus: list[tuple[int, int]]
+        model_size_bytes: int,
+        gpus: list[tuple[int, int]],
+        usable_fraction: Optional[float] = None,
     ) -> tuple[Optional[list[int]], bool]:
         """Pick GPU(s) for a model from estimated VRAM and free memory.
 
         ``model_size_bytes`` should include weights and estimated KV cache.
-        ``_GPU_PIN_VRAM_FRACTION`` provides headroom for compute buffers,
-        CUDA context, and other runtime overhead.
+        ``usable_fraction`` (default ``_GPU_PIN_VRAM_FRACTION``) provides
+        headroom for compute buffers, CUDA context, and other runtime
+        overhead; callers lower it when MTP reserves VRAM for a draft model.
 
         Returns (gpu_indices, use_fit):
           - ([1], False)       fits on 1 GPU at the headroom threshold
@@ -1501,7 +1585,8 @@ class LlamaCppBackend:
             return None, True
 
         model_size_mib = model_size_bytes / (1024 * 1024)
-        usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION
+        if usable_fraction is None:
+            usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION
 
         # Sort GPUs by free memory descending
         ranked = sorted(gpus, key = lambda g: g[1], reverse = True)
@@ -1749,8 +1834,8 @@ class LlamaCppBackend:
             ctx_checkpoints = ctx_checkpoints,
         )
 
-        # MTP needs a tighter budget; drop from 0.90 to 0.85.
-        budget_frac = 0.85 if mtp_engaged else 0.90
+        # MTP engaged: carve the drafter's reserve out of the fit budget.
+        budget_frac = _CTX_FIT_VRAM_FRACTION - (_MTP_VRAM_RESERVE_FRAC if mtp_engaged else 0.0)
         budget_bytes = available_mib * 1024 * 1024 * budget_frac
         model_footprint = model_size_bytes
 
@@ -1808,7 +1893,9 @@ class LlamaCppBackend:
             from huggingface_hub import get_paths_info, list_repo_files
 
             files = list_repo_files(hf_repo, token = hf_token)
-            gguf_files = [f for f in files if f.endswith(".gguf") and "mmproj" not in f.lower()]
+            gguf_files = [
+                f for f in files if f.endswith(".gguf") and not _is_companion_gguf_path(f)
+            ]
             if not gguf_files:
                 return None
 
@@ -2169,7 +2256,11 @@ class LlamaCppBackend:
                     r"(?<![a-zA-Z0-9])" + re.escape(variant_lower) + r"(?![a-zA-Z0-9])"
                 )
                 gguf_files = sorted(
-                    f for f in files if f.endswith(".gguf") and boundary.search(f.lower())
+                    f
+                    for f in files
+                    if f.endswith(".gguf")
+                    and boundary.search(f.lower())
+                    and not _is_companion_gguf_path(f)
                 )
                 if gguf_files:
                     gguf_filename = gguf_files[0]
@@ -2199,7 +2290,7 @@ class LlamaCppBackend:
                         matches = sorted(
                             p.relative_to(snap).as_posix()
                             for p in snap.rglob("*.gguf")
-                            if "mmproj" not in p.name.lower()
+                            if not _is_companion_gguf_path(p.relative_to(snap).as_posix())
                             and boundary.search(p.relative_to(snap).as_posix().lower())
                         )
                         if not matches:
@@ -2303,7 +2394,7 @@ class LlamaCppBackend:
                                 for f in all_gguf_files
                                 if f.startswith(_prefix)
                                 and f != gguf_filename
-                                and "mmproj" not in f.lower()
+                                and not _is_companion_gguf_path(f)
                             )
                         else:
                             gguf_extra_shards = []
@@ -2357,6 +2448,58 @@ class LlamaCppBackend:
             logger.info(f"GGUF downloaded in {dl_elapsed:.1f}s: {local_path}")
         return local_path
 
+    def _download_companion_gguf(
+        self,
+        *,
+        hf_repo: str,
+        hf_token: Optional[str],
+        pick: Callable[[list[str]], Optional[str]],
+        label: str,
+    ) -> Optional[str]:
+        """Resolve and fetch a companion GGUF (mmproj / MTP drafter) by name.
+
+        Tries the live repo file list, then the local HF cache snapshots
+        (offline, same fallback as _download_gguf), then hf_hub_download.
+        Runs WITHOUT self._lock (like _download_gguf); honors _cancel_event so
+        an /unload between the main download and here skips the fetch.
+        """
+        if self._cancel_event.is_set():
+            return None
+
+        target: Optional[str] = None
+        try:
+            from huggingface_hub import list_repo_files
+            target = pick(list_repo_files(hf_repo, token = hf_token))
+        except Exception as e:
+            logger.debug(f"Could not list repo files for {label}: {e}")
+
+        if target is None:
+            try:
+                from utils.models.model_config import _iter_hf_cache_snapshots
+                for snap in _iter_hf_cache_snapshots(hf_repo):
+                    rel_files = [p.relative_to(snap).as_posix() for p in snap.rglob("*.gguf")]
+                    target = pick(rel_files)
+                    if target is not None:
+                        logger.info("Resolved %s %s from local HF cache", label, target)
+                        break
+            except Exception as e:
+                logger.debug(f"Offline cache lookup for {label} failed: {e}")
+
+        if target is None or self._cancel_event.is_set():
+            return None
+
+        try:
+            from huggingface_hub import hf_hub_download
+            logger.info(f"Downloading {label}: {hf_repo}/{target}")
+            return hf_hub_download(
+                repo_id = hf_repo,
+                filename = target,
+                token = hf_token,
+            )
+        except Exception as e:
+            logger.warning(f"Could not download {label}: {e}")
+            return None
+
     def _download_mmproj(
         self,
         *,
@@ -2382,43 +2525,43 @@ class LlamaCppBackend:
                     return f
             return mmproj_files[0]
 
-        target: Optional[str] = None
-        try:
-            from huggingface_hub import list_repo_files
-            target = _pick_mmproj(list_repo_files(hf_repo, token = hf_token))
-        except Exception as e:
-            logger.debug(f"Could not list repo files for mmproj: {e}")
+        return self._download_companion_gguf(
+            hf_repo = hf_repo,
+            hf_token = hf_token,
+            pick = _pick_mmproj,
+            label = "mmproj",
+        )
 
-        # Offline: resolve mmproj from the local HF cache snapshot, same as
-        # _download_gguf's offline fallback above.
-        if target is None:
-            try:
-                from utils.models.model_config import _iter_hf_cache_snapshots
-                for snap in _iter_hf_cache_snapshots(hf_repo):
-                    rel_files = [p.relative_to(snap).as_posix() for p in snap.rglob("*.gguf")]
-                    target = _pick_mmproj(rel_files)
-                    if target is not None:
-                        logger.info("Resolved mmproj %s from local HF cache", target)
-                        break
-            except Exception as e:
-                logger.debug(f"Offline cache lookup for mmproj failed: {e}")
+    def _download_mtp(
+        self,
+        *,
+        hf_repo: str,
+        hf_token: Optional[str] = None,
+    ) -> Optional[str]:
+        """Download the separate MTP drafter (speculative head) from a GGUF repo.
 
-        if target is None:
-            return None
+        Targets the repo-root ``mtp-*.gguf`` companion -- the Q8_0 drafter
+        unsloth mirrors there for llama.cpp ``-hf`` auto-discovery (smallest,
+        recommended for speculation). Repos that bake the MTP head into the
+        main GGUF (e.g. Qwen) ship no such sibling and this returns None. The
+        higher-precision copies under ``MTP/`` are for explicit selection and
+        are intentionally skipped. Returns the local path, or None.
+        """
 
-        try:
-            from huggingface_hub import hf_hub_download
-
-            logger.info(f"Downloading mmproj: {hf_repo}/{target}")
-            local_path = hf_hub_download(
-                repo_id = hf_repo,
-                filename = target,
-                token = hf_token,
+        def _pick_mtp(candidates: list[str]) -> Optional[str]:
+            mtp_files = sorted(
+                f
+                for f in candidates
+                if f.lower().endswith(".gguf") and Path(f).name.lower().startswith("mtp-")
             )
-            return local_path
-        except Exception as e:
-            logger.warning(f"Could not download mmproj: {e}")
-            return None
+            return mtp_files[0] if mtp_files else None
+
+        return self._download_companion_gguf(
+            hf_repo = hf_repo,
+            hf_token = hf_token,
+            pick = _pick_mtp,
+            label = "MTP drafter",
+        )
 
     def _resolve_launch_mmproj_path(
         self, *, model_path: str, mmproj_path: Optional[str]
@@ -2446,6 +2589,19 @@ class LlamaCppBackend:
 
         return str(mmproj)
 
+    def _resolve_launch_mtp_path(self, *, mtp_draft_path: Optional[str]) -> Optional[str]:
+        """Return mtp_draft_path iff it exists on disk, else None.
+
+        No family check needed: the drafter is only ever auto-resolved from
+        the same repo as the main GGUF (see _download_mtp).
+        """
+        if not mtp_draft_path:
+            return None
+        if not Path(mtp_draft_path).is_file():
+            logger.warning(f"MTP drafter file not found: {mtp_draft_path}")
+            return None
+        return str(mtp_draft_path)
+
     # ── Lifecycle ─────────────────────────────────────────────────
 
     # GGUF ``general.architecture`` values for diffusion / image models.
@@ -2543,6 +2699,107 @@ class LlamaCppBackend:
             "Check that the GGUF file is valid and you have enough memory."
         )
 
+    @staticmethod
+    def _is_projector_incompatibility(output: str) -> bool:
+        """True when llama-server aborted because it cannot load the model's
+        vision/audio projector (mmproj), typically an installed llama.cpp
+        that predates the projector format. Conservative: only matches
+        projector-format errors so unrelated failures (OOM, bad GGUF, port
+        bind, ...) keep their own handling, and a bare 'clip'/'mmproj'
+        mention in a normal startup log does not match.
+        """
+        text = (output or "").lower()
+        if any(
+            m in text
+            for m in (
+                "unknown projector type",
+                "unsupported projector",
+                "unsupported mmproj",
+            )
+        ):
+            return True
+        # Builds that phrase it via clip.cpp without the exact words above.
+        return (
+            "clip" in text
+            and "projector" in text
+            and ("unknown" in text or "unsupported" in text or "not supported" in text)
+        )
+
+    @staticmethod
+    def _strip_mmproj_args(cmd: list[str]) -> list[str]:
+        """Return cmd without the '--mmproj <path>' pair (text-only retry).
+        Every other flag is preserved; a no-op when --mmproj is absent.
+        """
+        out: list[str] = []
+        skip_value = False
+        for tok in cmd:
+            if skip_value:
+                skip_value = False
+                continue
+            if tok == "--mmproj":
+                skip_value = True
+                continue
+            out.append(tok)
+        return out
+
+    def _start_llama_process(self, cmd: list[str], env: dict) -> None:
+        """Spawn llama-server from cmd and start draining its output.
+
+        Caller holds self._lock. Resets the stdout buffer, opens a fresh
+        per-attempt tee log, launches the process, and starts the drain
+        thread. Used for the initial start and the text-only mmproj retry.
+        """
+        # Defensive kill: if a concurrent load slipped past Phase 1
+        # (because its `self._process` was None at the time) and already
+        # stored a Popen handle here, drop that orphan before we overwrite
+        # the reference. See issue #5161.
+        self._kill_process()
+
+        self._stdout_lines = []
+        # Tee llama-server output to a dedicated log file so a post-mortem
+        # in CI (or after a remote-debug session) has the full subprocess
+        # trail even when the parent only stored the last 50 lines.
+        self._llama_log_fh = None
+        try:
+            log_dir = _swa_cache_path().parent / "logs" / "llama-server"
+            log_dir.mkdir(parents = True, exist_ok = True)
+            self._llama_log_path = log_dir / f"llama-{int(time.time())}-port-{self._port}.log"
+            self._llama_log_fh = open(
+                self._llama_log_path,
+                "w",
+                encoding = "utf-8",
+                buffering = 1,
+            )
+            logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
+        except OSError as e:
+            # Best-effort; never block the load on logging.
+            logger.debug(f"Could not open llama-server log file: {e}")
+            self._llama_log_path = None
+
+        # Log the argv per attempt (the text-only mmproj retry re-enters here
+        # with --mmproj stripped), redacting the API key.
+        _log_cmd = list(cmd)
+        if "--api-key" in _log_cmd:
+            _ki = _log_cmd.index("--api-key") + 1
+            if _ki < len(_log_cmd):
+                _log_cmd[_ki] = "<redacted>"
+        logger.info(f"Starting llama-server: {' '.join(_log_cmd)}")
+
+        self._process = subprocess.Popen(
+            cmd,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
+            text = True,
+            env = env,
+            **_windows_hidden_subprocess_kwargs(),
+        )
+
+        # Start background thread to drain stdout and prevent pipe deadlock
+        self._stdout_thread = threading.Thread(
+            target = self._drain_stdout, daemon = True, name = "llama-stdout"
+        )
+        self._stdout_thread.start()
+
     def load_model(
         self,
         *,
@@ -2550,6 +2807,8 @@ class LlamaCppBackend:
         gguf_path: Optional[str] = None,
         # Vision projection (mmproj) for local vision models
         mmproj_path: Optional[str] = None,
+        # Separate MTP drafter for local Gemma loads (HF loads auto-resolve it)
+        mtp_draft_path: Optional[str] = None,
         # HF mode: let llama-server download via -hf "repo:quant"
         hf_repo: Optional[str] = None,
         hf_variant: Optional[str] = None,
@@ -2578,10 +2837,15 @@ class LlamaCppBackend:
         # Serialise the whole load so concurrent /load calls never leave two
         # llama-server processes alive (#5401 / #5161). Doesn't block /unload.
         with self._serial_load_lock:
+            # In-app update swapping binaries: refuse fast (set under this lock,
+            # so any in-flight load has drained) instead of using a half-swapped one.
+            if getattr(self, "_llama_update_in_progress", False):
+                raise RuntimeError("llama.cpp is updating; try again in a moment.")
             # Duplicate /load that raced past the route check: do nothing if the
             # live server already satisfies this request.
             if self._already_in_target_state(
                 gguf_path = gguf_path,
+                mtp_draft_path = mtp_draft_path,
                 model_identifier = model_identifier,
                 hf_variant = hf_variant,
                 n_ctx = n_ctx,
@@ -2653,6 +2917,8 @@ class LlamaCppBackend:
                 )
 
             # ── Phase 2: download (NO lock held, so cancel can proceed) ──
+            # mtp_draft_path arrives set for local Gemma loads (detected
+            # sibling); for -hf loads it's None here and resolved just below.
             # Scope HF_HUB_OFFLINE to the download block only when DNS is
             # dead; cleanup runs even on exception so a transient hiccup
             # can't quarantine future loads.
@@ -2669,6 +2935,23 @@ class LlamaCppBackend:
                             hf_repo = hf_repo,
                             hf_token = hf_token,
                         )
+                    # Auto-download the separate MTP drafter (e.g. Gemma) when
+                    # the requested spec mode can use it. Repos with the head
+                    # baked into the main GGUF (Qwen) have no mtp- sibling and
+                    # this no-ops, so the size gate stays out of it: a separate
+                    # drafter speeds up even sub-3B (Gemma E2B), and the resolver
+                    # below decides the final emission. Skipped only when the
+                    # user disabled MTP or drives --spec-type manually.
+                    _spec_canon = _canonicalize_spec_mode(speculative_type) or "auto"
+                    if (
+                        not mtp_draft_path
+                        and _spec_canon in ("auto", "mtp", "mtp+ngram")
+                        and not _extra_args_set_spec_type(extra_args)
+                    ):
+                        mtp_draft_path = self._download_mtp(
+                            hf_repo = hf_repo,
+                            hf_token = hf_token,
+                        )
             elif gguf_path:
                 if not Path(gguf_path).is_file():
                     raise FileNotFoundError(f"GGUF file not found: {gguf_path}")
@@ -2736,12 +3019,19 @@ class LlamaCppBackend:
                     # Will MTP engage on this load? If so, auto-fit reserves
                     # extra VRAM for the draft model. Mirrors
                     # _build_speculative_flags' resolver: forced mtp / mtp+ngram
-                    # always engage; auto only on an MTP GGUF >= 3B; ngram /
-                    # ngram-simple / off never engage MTP.
+                    # always engage; auto only on an MTP model >= 3B; ngram /
+                    # ngram-simple / off never engage MTP. A separate drafter
+                    # (Gemma) counts as an MTP model just like a baked-in head.
                     _mtp_canonical = _canonicalize_spec_mode(speculative_type)
                     _mtp_effective = _mtp_canonical or "auto"
                     _mtp_size_for_fit = _extract_model_size_b(model_identifier)
-                    _mtp_sub_3b_for_fit = _mtp_size_for_fit is not None and _mtp_size_for_fit < 3.0
+                    # Sub-3B drops MTP only for an embedded head; a separate
+                    # drafter (Gemma) engages and needs its VRAM reserved.
+                    _mtp_sub_3b_for_fit = (
+                        _mtp_size_for_fit is not None
+                        and _mtp_size_for_fit < _MTP_MIN_SIZE_B
+                        and not bool(mtp_draft_path)
+                    )
                     _mtp_will_engage = bool(
                         not _extra_args_set_spec_type(extra_args)
                         and (
@@ -2751,6 +3041,7 @@ class LlamaCppBackend:
                                 and (
                                     bool(self._nextn_predict_layers)
                                     or _is_mtp_model_name(model_identifier, model_path)
+                                    or bool(mtp_draft_path)
                                 )
                                 and not _mtp_sub_3b_for_fit
                             )
@@ -2766,6 +3057,11 @@ class LlamaCppBackend:
                     #     context, since multi-GPU is slower.
                     gpu_indices, use_fit = None, True
                     explicit_ctx = requested_ctx > 0
+                    # MTP draft model lives outside the main estimates; carve
+                    # its reserve out of every fit budget and pin threshold so
+                    # a load can't pin into the drafter's headroom.
+                    _mtp_reserve = _MTP_VRAM_RESERVE_FRAC if _mtp_will_engage else 0.0
+                    _pin_fraction = self._GPU_PIN_VRAM_FRACTION - _mtp_reserve
 
                     if gpus and self._can_estimate_kv() and effective_ctx > 0:
                         # Largest hardware-aware cap from the native context
@@ -2790,7 +3086,7 @@ class LlamaCppBackend:
                                     capped, cache_type_kv, n_parallel = n_parallel
                                 )
                                 total_mib = (model_size + kv) / (1024 * 1024)
-                                if total_mib <= pool_mib * 0.90:
+                                if total_mib <= pool_mib * (_CTX_FIT_VRAM_FRACTION - _mtp_reserve):
                                     best_cap = max(best_cap, capped)
                             if best_cap > 0:
                                 max_available_ctx = best_cap
@@ -2807,13 +3103,15 @@ class LlamaCppBackend:
                             requested_total = model_size + self._estimate_kv_cache_bytes(
                                 effective_ctx, cache_type_kv, n_parallel = n_parallel
                             )
-                            gpu_indices, use_fit = self._select_gpus(requested_total, gpus)
+                            gpu_indices, use_fit = self._select_gpus(
+                                requested_total, gpus, usable_fraction = _pin_fraction
+                            )
                             # No silent shrink: effective_ctx stays == requested_ctx.
                         else:
                             # Auto context: prefer fewer GPUs, cap to fit. Same
                             # headroom threshold as _select_gpus (#5106).
                             ranked = sorted(gpus, key = lambda g: g[1], reverse = True)
-                            pin_fraction = self._GPU_PIN_VRAM_FRACTION
+                            pin_fraction = _pin_fraction
                             for n_gpus in range(1, len(ranked) + 1):
                                 subset = ranked[:n_gpus]
                                 pool_mib = sum(free for _, free in subset)
@@ -2861,7 +3159,9 @@ class LlamaCppBackend:
                             "Falling back to file-size-only GPU selection",
                             model_size_gb = round(model_size / (1024**3), 2),
                         )
-                        gpu_indices, use_fit = self._select_gpus(model_size, gpus)
+                        gpu_indices, use_fit = self._select_gpus(
+                            model_size, gpus, usable_fraction = _pin_fraction
+                        )
                         if use_fit and not explicit_ctx:
                             # Weights don't fit on any subset; default UI to 4096
                             # so the slider isn't on an unusable native ctx.
@@ -2941,6 +3241,16 @@ class LlamaCppBackend:
                     # Fits on selected GPU(s) -- offload all layers
                     cmd.extend(["-ngl", "-1"])
 
+                cmd.extend(
+                    self._ctx_integrity_flags(
+                        n_parallel,
+                        use_fit,
+                        requested_ctx,
+                        effective_ctx,
+                        self.probe_server_capabilities(binary),
+                    )
+                )
+
                 # -1 = llama.cpp auto-detect (physical cores). Pass explicitly
                 # so we don't inherit llama-server's internal default, which
                 # has varied (hardware concurrency incl. hyperthreads on some
@@ -2978,6 +3288,9 @@ class LlamaCppBackend:
 
                 # Speculative decoding. See _build_speculative_flags for the
                 # mode resolution, benchmarks, and llama.cpp references.
+                launch_mtp_draft_path = self._resolve_launch_mtp_path(
+                    mtp_draft_path = mtp_draft_path,
+                )
                 spec_flags = self._build_speculative_flags(
                     speculative_type = speculative_type,
                     spec_draft_n_max = spec_draft_n_max,
@@ -2986,7 +3299,11 @@ class LlamaCppBackend:
                     model_path = model_path,
                     gpus = bool(gpus),
                     binary = binary,
+                    mtp_draft_path = launch_mtp_draft_path,
                 )
+                # Remember where the spec block sits so a drafter-load failure
+                # can be retried with these flags swapped out (see below).
+                _spec_start = len(cmd)
                 cmd.extend(spec_flags)
 
                 # Apply custom chat template override if provided.
@@ -3162,7 +3479,7 @@ class LlamaCppBackend:
 
                 # Pin to selected GPU(s). On ROCm, narrowing only
                 # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full
-                # HIP/ROCR set, so set those too.
+                # set, so set HIP_VISIBLE_DEVICES too.
                 if gpu_indices is not None:
                     pinned = ",".join(str(i) for i in gpu_indices)
                     env["CUDA_VISIBLE_DEVICES"] = pinned
@@ -3170,56 +3487,116 @@ class LlamaCppBackend:
                         import torch as _torch
                         if getattr(_torch.version, "hip", None) is not None:
                             env["HIP_VISIBLE_DEVICES"] = pinned
-                            env["ROCR_VISIBLE_DEVICES"] = pinned
+                            # Do NOT also set ROCR_VISIBLE_DEVICES to the same
+                            # value. ROCR_VISIBLE_DEVICES filters at the HSA/ROCr
+                            # layer and HIP_VISIBLE_DEVICES at the HIP layer, so
+                            # setting both with the same physical indices applies
+                            # the mask twice: ROCR reduces the visible set and
+                            # re-indexes it from 0, then HIP indexes into the
+                            # already-reduced set. A single non-zero pin (e.g.
+                            # "1") then points out of range at the HIP layer, HIP
+                            # enumerates 0 devices, and llama.cpp falls back to
+                            # CPU ("ggml_cuda_init: no ROCm-capable device is
+                            # detected"). The HIP mask alone narrows correctly;
+                            # clear any inherited ROCR mask so it can't double up.
+                            env.pop("ROCR_VISIBLE_DEVICES", None)
                     except Exception as e:
                         logger.debug("Failed to set ROCm visibility env vars for child: %s", e)
 
-                # Defensive kill: drop an orphan Popen a concurrent load may
-                # have stored before we overwrite the reference (#5161).
-                self._kill_process()
+                # Captured before any text-only fallback strips it from cmd.
+                launched_with_mmproj = "--mmproj" in cmd
 
-                self._stdout_lines = []
-                # Tee llama-server output to a dedicated log file so a
-                # post-mortem has the full trail even when the parent only kept
-                # the last 50 lines. Path is under the studio home.
-                self._llama_log_fh = None
-                try:
-                    log_dir = _swa_cache_path().parent / "logs" / "llama-server"
-                    log_dir.mkdir(parents = True, exist_ok = True)
-                    self._llama_log_path = (
-                        log_dir / f"llama-{int(time.time())}-port-{self._port}.log"
-                    )
-                    self._llama_log_fh = open(
-                        self._llama_log_path,
-                        "w",
-                        encoding = "utf-8",
-                        buffering = 1,
-                    )
-                    logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
-                except OSError as e:
-                    # Best-effort; never block the load on logging.
-                    logger.debug(f"Could not open llama-server log file: {e}")
-                    self._llama_log_path = None
-                self._process = subprocess.Popen(
-                    cmd,
-                    stdout = subprocess.PIPE,
-                    stderr = subprocess.STDOUT,
-                    text = True,
-                    env = env,
-                    **_windows_hidden_subprocess_kwargs(),
-                )
+                # One-shot --fit off retry: recent llama.cpp runs a "fitting
+                # params to device memory" step by default (--fit defaults to
+                # 'on') even when -ngl is explicit. That step has aborted on
+                # some ROCm hosts (ggml-cuda.cu ROCm error during worst-case
+                # estimation, e.g. MTP + mmproj models on gfx1151). When
+                # Studio's own VRAM math already placed the model
+                # (use_fit=False), the step is redundant second-guessing --
+                # retry once with --fit off before declaring the load failed.
+                # Never retry when fit was requested (use_fit) or the caller
+                # passed an explicit fit flag via extra args.
+                def _spawn_and_wait(run_cmd, *, label = ""):
+                    """Start llama-server with run_cmd and wait for health.
 
-                # Background thread to drain stdout (prevents pipe deadlock)
-                self._stdout_thread = threading.Thread(
-                    target = self._drain_stdout, daemon = True, name = "llama-stdout"
-                )
-                self._stdout_thread.start()
+                    Retries once with --fit off when the first attempt
+                    crashes during startup and run_cmd is eligible (see
+                    _fit_off_retry_eligible).
+                    """
+                    _fit_retry_allowed = self._fit_off_retry_eligible(run_cmd, use_fit)
+                    for _spawn_attempt in (0, 1):
+                        # Defensive kill: drop an orphan Popen a concurrent load may
+                        # have stored before we overwrite the reference (#5161).
+                        # Also reaps the crashed first attempt on the retry pass.
+                        self._kill_process()
+
+                        self._stdout_lines = []
+                        # Tee llama-server output to a dedicated log file so a
+                        # post-mortem has the full trail even when the parent only
+                        # kept the last 50 lines. Path is under the studio home.
+                        # ``label`` (MTP fallback) and the attempt index (--fit
+                        # off retry) keep a respawn within the same epoch second
+                        # from truncating the crash log a retry warning just
+                        # pointed the user at.
+                        self._llama_log_fh = None
+                        try:
+                            log_dir = _swa_cache_path().parent / "logs" / "llama-server"
+                            log_dir.mkdir(parents = True, exist_ok = True)
+                            self._llama_log_path = log_dir / (
+                                f"llama-{int(time.time())}{label}-port-{self._port}"
+                                f"-try{_spawn_attempt}.log"
+                            )
+                            self._llama_log_fh = open(
+                                self._llama_log_path,
+                                "w",
+                                encoding = "utf-8",
+                                buffering = 1,
+                            )
+                            logger.info(f"llama-server stdout/stderr -> {self._llama_log_path}")
+                        except OSError as e:
+                            # Best-effort; never block the load on logging.
+                            logger.debug(f"Could not open llama-server log file: {e}")
+                            self._llama_log_path = None
+                        self._process = subprocess.Popen(
+                            run_cmd,
+                            stdout = subprocess.PIPE,
+                            stderr = subprocess.STDOUT,
+                            text = True,
+                            env = env,
+                            **_windows_hidden_subprocess_kwargs(),
+                        )
+
+                        # Background thread to drain stdout (prevents pipe deadlock)
+                        self._stdout_thread = threading.Thread(
+                            target = self._drain_stdout, daemon = True, name = "llama-stdout"
+                        )
+                        self._stdout_thread.start()
+                        if self._wait_for_health(timeout = 600.0):
+                            return True
+                        _startup_crashed = (
+                            self._process.poll() is not None and self._process.returncode != 0
+                        )
+                        if _spawn_attempt == 0 and _fit_retry_allowed and _startup_crashed:
+                            logger.warning(
+                                "llama-server crashed during startup (exit code %s) "
+                                "with the default memory-fit step enabled; Studio "
+                                "already verified the model fits, retrying once "
+                                "with --fit off. Crash log: %s",
+                                self._process.returncode,
+                                self._llama_log_path,
+                            )
+                            run_cmd = [*run_cmd, "--fit", "off"]
+                            continue
+                        return False
 
                 # Store the resolved on-disk path, not the caller's kwarg: in
                 # HF mode gguf_path is None and ``model_path`` is what
-                # llama-server mmap's, which downstream consumers need.
+                # llama-server mmap's, which downstream consumers need. Must be
+                # set BEFORE the spawn: load_progress() reads _gguf_path for
+                # the mmap progress total while the health wait runs.
                 self._gguf_path = model_path
                 self._hf_repo = hf_repo
+                self._mtp_draft_path = launch_mtp_draft_path
                 # For local GGUF files, extract variant from filename if absent
                 if hf_variant:
                     self._hf_variant = hf_variant
@@ -3239,20 +3616,107 @@ class LlamaCppBackend:
                 self._effective_context_length = (
                     effective_ctx if effective_ctx > 0 else self._context_length
                 )
+                self._reconcile_effective_ctx_with_server()
                 self._max_context_length = (
                     max_available_ctx if max_available_ctx > 0 else self._effective_context_length
                 )
 
-                # Wait for llama-server to become healthy.
-                if not self._wait_for_health(timeout = 600.0):
-                    self._kill_process()
-                    raise RuntimeError(
-                        self._classify_llama_start_failure(
-                            "\n".join(self._stdout_lines[-50:]),
-                            gguf_path,
-                            self._model_identifier,
+                healthy = _spawn_and_wait(cmd)
+                # Any MTP request can abort the server: a separate drafter
+                # (Gemma) on a binary that predates its arch, or an embedded
+                # head (Qwen) the binary cannot build. Retry once with the
+                # spec slice replaced by --spec-default so the main model still
+                # loads. Gate on the spec block (not the drafter path, which
+                # off/ngram local loads also carry) and keep
+                # _requested_spec_mode so a duplicate /load doesn't thrash. The
+                # cancel check stops an /unload-killed attempt respawning.
+                _spec_requested_mtp = any("mtp" in str(t).lower() for t in spec_flags)
+                if not healthy and _spec_requested_mtp and not self._cancel_event.is_set():
+                    # Blame the binary only when the output shows MTP itself
+                    # failing (unknown arch / draft or context build); an
+                    # unrelated crash (e.g. OOM) gets a neutral message.
+                    _lo = "\n".join(self._stdout_lines).lower()
+                    # Only an unknown architecture proves the prebuilt predates
+                    # this MTP model (an update fixes it). The memory/context
+                    # build failures are generic (VRAM / ctx pressure), where an
+                    # update may not help, so classify those as runtime_error.
+                    _arch_unsupported = "unknown model architecture" in _lo
+                    if (
+                        _arch_unsupported
+                        or "failed to measure draft model memory" in _lo
+                        or "failed to measure mtp context memory" in _lo
+                        or "failed to create llama_context" in _lo
+                    ):
+                        _retry_reason = (
+                            "the prebuilt may predate it; retrying without "
+                            "speculative decoding -- run `unsloth studio "
+                            "update` for MTP"
                         )
+                        self._spec_fallback_reason = (
+                            "binary_outdated" if _arch_unsupported else "runtime_error"
+                        )
+                    else:
+                        _retry_reason = (
+                            "retrying without speculative decoding in case MTP is the cause"
+                        )
+                        self._spec_fallback_reason = "runtime_error"
+                    _drafter = (
+                        Path(launch_mtp_draft_path).name
+                        if launch_mtp_draft_path
+                        else "embedded head"
                     )
+                    logger.warning(
+                        "llama-server failed to start with MTP (%s); %s.",
+                        _drafter,
+                        _retry_reason,
+                    )
+                    self._kill_process()
+                    fallback_cmd = (
+                        cmd[:_spec_start]
+                        + ["--spec-default"]
+                        + cmd[_spec_start + len(spec_flags) :]
+                    )
+                    healthy = _spawn_and_wait(fallback_cmd, label = "-retry")
+                    if healthy:
+                        self._speculative_type = "default"
+
+                # A vision GGUF launched with --mmproj can abort when the
+                # installed llama.cpp is too old for the model's projector
+                # ("Unknown projector type"); in that one case retry once
+                # text-only rather than failing the whole load.
+                if not healthy:
+                    out = "\n".join(self._stdout_lines[-50:])
+                    self._kill_process()
+                    if launched_with_mmproj and self._is_projector_incompatibility(out):
+                        logger.warning(
+                            "llama-server could not load this model's vision "
+                            "projector (--mmproj). The installed llama.cpp build is "
+                            "likely too old for it. Loading text-only for this "
+                            "session; run 'unsloth studio update' to enable vision."
+                        )
+                        cmd = self._strip_mmproj_args(cmd)
+                        self._is_vision = False
+                        self._mmproj_has_audio = False
+                        self._start_llama_process(cmd, env)
+                        if not self._wait_for_health(timeout = 600.0):
+                            self._kill_process()
+                            raise RuntimeError(
+                                "Vision projector incompatible with this llama.cpp "
+                                "build, and the text-only retry also failed: "
+                                + self._classify_llama_start_failure(
+                                    "\n".join(self._stdout_lines[-50:]),
+                                    gguf_path,
+                                    self._model_identifier,
+                                )
+                            )
+                    else:
+                        raise RuntimeError(
+                            self._classify_llama_start_failure(
+                                out,
+                                gguf_path,
+                                self._model_identifier,
+                            )
+                        )
 
                 self._healthy = True
 
@@ -3341,6 +3805,7 @@ class LlamaCppBackend:
         model_path: Optional[str],
         gpus: bool,
         binary: Optional[str],
+        mtp_draft_path: Optional[str] = None,
     ) -> List[str]:
         """Return the llama-server flag list for the requested spec mode.
 
@@ -3365,32 +3830,50 @@ class LlamaCppBackend:
               https://github.com/ggml-org/llama.cpp/pull/18471
               MTP guide: unsloth.ai/docs/models/qwen3.6#mtp-guide
 
-        Sub-3B dense MTP regresses vs spec-off: the draft head's per-token
-        cost exceeds the acceptance savings at this scale. Q4_K_XL clean
-        bench (each prompt once after an unrelated warmup) on B200 + x86 CPU:
+        Sub-3B dense MTP regresses vs spec-off when the head is baked into the
+        main GGUF (Qwen): the draft head's per-token cost exceeds the
+        acceptance savings at this scale. Q4_K_XL clean bench (each prompt once
+        after an unrelated warmup) on B200 + x86 CPU:
           0.8B GPU: draft-mtp n=2 = 0.58x vs OFF; ngram-only = 1.10x
           2B   GPU: draft-mtp n=2 = 0.82x vs OFF; OFF or ngram = 1.00x
           0.8B CPU: chained n=2   = 0.86x vs OFF; ngram-only = 1.19x
           2B   CPU: chained n=2   = 0.83x vs OFF; ngram-only = 1.01x
           4B+ GPU/CPU: spec on is a net win (1.08x-1.46x).
+        A separate drafter (Gemma's root mtp-*.gguf) is a different, cheaper
+        mechanism that wins even below 3B, so it is exempt from the sub-3B drop
+        (``mtp_draft_path`` set -> not too small). B200 Q4_K_XL bench, draft-mtp
+        n=2 vs OFF: gemma-4-E2B (2B) = 1.21x, accept ~0.65 (vs ngram = 1.00x);
+        gemma-4-E4B (4B) and 12B engage as usual.
         Auto falls back to ngram-mod (zero-VRAM, near-zero idle cost on
-        diverse content); forced MTP variants engage anyway and just log a
-        warning per the user's choice.
+        diverse content) for an embedded sub-3B head; forced MTP on a model
+        with no head/drafter defaults back (mtp -> spec-default, mtp+ngram ->
+        ngram-mod) since llama-server aborts otherwise; a drafter the binary
+        cannot build (older prebuilt, or a CUDA kernel limit) aborts the spawn
+        and the load retries once without speculative decoding.
         """
         flags: List[str] = []
         # Reset; emit branches re-set on the resolved emission.
         self._spec_draft_n_max = None
         self._speculative_type = None
+        self._spec_fallback_reason = None
 
         # Canonical UI-facing requested mode (legacy values mapped via
         # _canonicalize_spec_mode).
         canonical_mode = _canonicalize_spec_mode(speculative_type)
-        is_mtp_model = bool(self._nextn_predict_layers) or (
-            _is_mtp_model_name(model_identifier, model_path)
+        # MTP signals: head baked into the main GGUF (Qwen, via metadata or
+        # name), or a separate drafter resolved from the repo (Gemma).
+        is_mtp_model = (
+            bool(self._nextn_predict_layers)
+            or _is_mtp_model_name(model_identifier, model_path)
+            or bool(mtp_draft_path)
         )
         user_owns_spec_type = _extra_args_set_spec_type(extra_args)
         _mtp_size_b = _extract_model_size_b(model_identifier)
-        _mtp_too_small = _mtp_size_b is not None and _mtp_size_b < 3.0
+        # The sub-3B regression is an embedded-head cost; a separate drafter
+        # (Gemma) is a cheap standalone model that wins below 3B, so exempt it.
+        _mtp_too_small = (
+            _mtp_size_b is not None and _mtp_size_b < _MTP_MIN_SIZE_B and not bool(mtp_draft_path)
+        )
 
         if user_owns_spec_type:
             # User --spec-type wins outright; suppress auto-emit to avoid a
@@ -3422,9 +3905,16 @@ class LlamaCppBackend:
                     "run `unsloth studio update`. Loading without "
                     "speculative decoding."
                 )
+                self._spec_fallback_reason = "binary_no_mtp"
                 return False
             draft_n_max = _resolved_draft_n_max()
             n_max_flag = caps.get("spec_draft_n_max_flag") or "--spec-draft-n-max"
+            # Separate-file drafter (Gemma): point llama-server at it. Baked-in
+            # heads (Qwen) pass no path -- llama-server reads them from the
+            # main GGUF.
+            if mtp_draft_path:
+                flags.extend(["--model-draft", mtp_draft_path])
+                logger.info(f"Using separate MTP drafter: {mtp_draft_path}")
             if chain_ngram:
                 ngram_knobs = _build_ngram_mod_flags(caps)
                 if ngram_knobs:
@@ -3483,33 +3973,40 @@ class LlamaCppBackend:
             _emit_ngram_mod()
             return flags
         if effective_mode == "mtp":
+            if not is_mtp_model:
+                # No head and no drafter: llama-server aborts on draft-mtp
+                # instead of no-op'ing, so default back.
+                logger.warning(
+                    "MTP requested but this GGUF has no MTP head or drafter; "
+                    "loading without speculative decoding."
+                )
+                flags.append("--spec-default")
+                self._speculative_type = "default"
+                return flags
             if _mtp_too_small:
                 logger.warning(
                     f"Forcing MTP on a {_mtp_size_b:.1f}B model; "
                     "the bench shows draft-mtp regresses below 3B. "
                     "Engaging anyway (user override)."
                 )
-            elif not is_mtp_model:
-                logger.warning(
-                    "Forcing MTP on a non-MTP GGUF; llama-server may "
-                    "fall back to spec-off if no nextn head is present. "
-                    "Engaging anyway (user override)."
-                )
             _emit_mtp(chain_ngram = False)
             return flags
         if effective_mode == "mtp+ngram":
+            if not is_mtp_model:
+                # No head/drafter: keep the ngram half (needs no head),
+                # drop the draft-mtp chain that would abort the server.
+                logger.warning(
+                    "MTP+Ngram requested but this GGUF has no MTP head or "
+                    "drafter; loading ngram-mod only."
+                )
+                _emit_ngram_mod()
+                return flags
             if _mtp_too_small:
                 logger.warning(
                     f"Forcing MTP+Ngram on a {_mtp_size_b:.1f}B model; "
                     "the bench shows the chain regresses below 3B. "
                     "Engaging anyway (user override)."
                 )
-            elif not is_mtp_model:
-                logger.warning(
-                    "Forcing MTP+Ngram on a non-MTP GGUF; llama-server "
-                    "may fall back to ngram-only if no nextn head is "
-                    "present. Engaging anyway (user override)."
-                )
             _emit_mtp(chain_ngram = True)
             return flags
 
@@ -3557,6 +4054,7 @@ class LlamaCppBackend:
         is_vision: bool,
         gguf_path: Optional[str] = None,
         spec_draft_n_max: Optional[int] = None,
+        mtp_draft_path: Optional[str] = None,
     ) -> bool:
         """True iff the live server already satisfies these load kwargs.
 
@@ -3615,6 +4113,22 @@ class LlamaCppBackend:
         if (self._chat_template_override or None) != (chat_template_override or None):
             return False
 
+        # A drafter appearing/disappearing next to a local GGUF changes the
+        # launch command (--model-draft) when the mode can use it; without
+        # this, adding mtp-*.gguf after a load is deduped away and MTP can't
+        # engage short of an unload. HF loads resolve the drafter inside
+        # load_model (gguf_path is None here), so only local paths compare;
+        # the route-level probe covers HF cache repos. No sub-3B gate: both
+        # sides come from the same config detection, so a sub-3B mismatch
+        # only happens when a drafter genuinely appeared (one benign reload,
+        # then the stored path converges).
+        if (
+            gguf_path is not None
+            and req_mode in ("auto", "mtp", "mtp+ngram")
+            and (mtp_draft_path or None) != (self._mtp_draft_path or None)
+        ):
+            return False
+
         # extra_args=None means "no opinion" (inherit handled at the route
         # layer); only an explicit list forces equality.
         if extra_args is not None:
@@ -3656,6 +4170,8 @@ class LlamaCppBackend:
             self._model_identifier = None
             self._gguf_path = None
             self._hf_repo = None
+            self._mtp_draft_path = None
+            self._spec_fallback_reason = None
             self._hf_variant = None
             self._is_vision = False
             self._is_audio = False
@@ -3917,6 +4433,23 @@ class LlamaCppBackend:
         """atexit handler to ensure llama-server is terminated."""
         self._kill_process()
 
+    @staticmethod
+    def _fit_off_retry_eligible(cmd: "list[str]", use_fit: bool) -> bool:
+        """Whether a llama-server startup crash may be retried with --fit off.
+
+        Only when Studio's own VRAM math placed the model (use_fit=False)
+        and nothing on the command line set the fit mode explicitly
+        (-fit / --fit, space- or equals-form). --fit-ctx / --fit-target /
+        -fitc / -fitt tune the fit step but do not select the mode, so
+        they do not block the retry.
+        """
+        if use_fit:
+            return False
+        for a in cmd:
+            if a in ("-fit", "--fit") or a.startswith(("-fit=", "--fit=")):
+                return False
+        return True
+
     def _wait_for_health(
         self,
         timeout: float = 120.0,
@@ -3933,9 +4466,17 @@ class LlamaCppBackend:
                 if self._stdout_thread is not None:
                     self._stdout_thread.join(timeout = 2)
                 output = "\n".join(self._stdout_lines[-50:])
+                # Keep the TAIL: crash details (abort reason, ROCm/CUDA error
+                # text) print last, after the long startup banner. Head
+                # truncation has cut off exactly the diagnostic line before.
+                _log_hint = (
+                    f" Full log: {self._llama_log_path}"
+                    if getattr(self, "_llama_log_path", None)
+                    else ""
+                )
                 logger.error(
                     f"llama-server exited with code {self._process.returncode}. "
-                    f"Output: {output[:2000]}"
+                    f"Output (tail): {output[-2000:]}{_log_hint}"
                 )
                 return False
 
@@ -3959,6 +4500,64 @@ class LlamaCppBackend:
         logger.error(f"llama-server health check timed out after {timeout}s")
         return False
 
+    @staticmethod
+    def _ctx_integrity_flags(
+        n_parallel: int, use_fit: bool, requested_ctx: int, effective_ctx: int, caps: dict
+    ) -> list[str]:
+        """Flags that keep the per-request window equal to the advertised ctx.
+
+        Explicit ``--parallel`` disables llama-server's auto-slots
+        ``--kv-unified`` default, silently splitting ``-c`` into per-slot
+        windows of ``-c / N``; restore the shared pool so one request can use
+        the full context. With ``--fit on``, ``--fit-ctx`` floors the fit step
+        at an explicitly requested ctx (default floor is 4096) so it offloads
+        or fails instead of silently shrinking the window.
+        """
+        flags: list[str] = []
+        if n_parallel > 1 and caps.get("supports_kv_unified"):
+            flags.append("--kv-unified")
+        if use_fit and requested_ctx > 0 and effective_ctx > 0 and caps.get("supports_fit_ctx"):
+            flags.extend(["--fit-ctx", str(effective_ctx)])
+        return flags
+
+    def _query_server_n_ctx(self) -> Optional[int]:
+        """Per-slot context llama-server actually allocated, from ``/props``.
+
+        The memory-fit step or ``--parallel`` slot split can leave this below
+        the requested ``-c``; requests are validated against this value.
+        """
+        url = f"http://127.0.0.1:{self._port}/props"
+        try:
+            resp = httpx.get(url, timeout = 5.0)
+            if resp.status_code != 200:
+                return None
+            settings = resp.json().get("default_generation_settings") or {}
+            n_ctx = settings.get("n_ctx")
+            return int(n_ctx) if n_ctx else None
+        except Exception:
+            return None
+
+    def _reconcile_effective_ctx_with_server(self) -> None:
+        """Adopt the server's real ``n_ctx`` when it is below Studio's value.
+
+        Keeps ``context_length`` (load response, status route, passthrough
+        ``max_tokens`` ceiling) honest; clients sized to the requested value
+        would otherwise hit ``exceed_context_size_error`` 400s early.
+        """
+        actual_n_ctx = self._query_server_n_ctx()
+        if not actual_n_ctx or actual_n_ctx <= 0:
+            return
+        if self._effective_context_length and actual_n_ctx < self._effective_context_length:
+            logger.warning(
+                "llama-server allocated a smaller per-request context than "
+                f"requested ({self._effective_context_length} -> {actual_n_ctx}; "
+                "memory fit or --parallel slot split); clients must treat "
+                f"{actual_n_ctx} as the real context window."
+            )
+            self._effective_context_length = actual_n_ctx
+        elif not self._effective_context_length:
+            self._effective_context_length = actual_n_ctx
+
     # ── Message building (OpenAI format) ──────────────────────────
 
     @staticmethod
diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py
index 66c7f10d42..1862c9c5de 100644
--- a/studio/backend/core/inference/llama_server_args.py
+++ b/studio/backend/core/inference/llama_server_args.py
@@ -132,7 +132,13 @@ _SPEC_FLAGS: frozenset[str] = frozenset(
         "--spec-ngram-size",
         "--draft-min",
         "--draft-max",
-        # MTP path (llama.cpp #22673).
+        # MTP path (llama.cpp #22673). --model-draft and aliases are
+        # Studio-managed since the separate-drafter support (Gemma 4): an
+        # inherited copy must not last-wins-override the auto-detected
+        # drafter. Explicit extras for the current load are never stripped.
+        "--model-draft",
+        "-md",
+        "--spec-draft-model",
         "--spec-draft-n-max",
         "--spec-draft-n-min",
         "--spec-draft-p-min",
diff --git a/studio/backend/core/inference/mcp_client.py b/studio/backend/core/inference/mcp_client.py
index 9a7fd1f074..a0d79bbdf4 100644
--- a/studio/backend/core/inference/mcp_client.py
+++ b/studio/backend/core/inference/mcp_client.py
@@ -25,16 +25,85 @@ def is_stdio(address: str) -> bool:
     return not address.strip().lower().startswith(("http://", "https://"))
 
 
+def _split_windows_command_line(address: str) -> list[str]:
+    """Parse a Windows command line using the same backslash/quote rules that
+    subprocess.list2cmdline() writes. This keeps trailing backslashes before a
+    closing quote from being doubled in the resulting argv."""
+    parts: list[str] = []
+    current: list[str] = []
+    in_quotes = False
+    backslashes = 0
+    arg_started = False
+    i = 0
+
+    while i < len(address):
+        ch = address[i]
+        if ch == "\\":
+            backslashes += 1
+            i += 1
+            continue
+        if ch == '"':
+            current.extend("\\" * (backslashes // 2))
+            if backslashes % 2:
+                current.append('"')
+            else:
+                in_quotes = not in_quotes
+            arg_started = True
+            backslashes = 0
+            i += 1
+            continue
+        if ch.isspace() and not in_quotes:
+            if backslashes:
+                current.extend("\\" * backslashes)
+                arg_started = True
+                backslashes = 0
+            if arg_started or current:
+                parts.append("".join(current))
+                current = []
+                arg_started = False
+            i += 1
+            while i < len(address) and address[i].isspace():
+                i += 1
+            continue
+        if backslashes:
+            current.extend("\\" * backslashes)
+            arg_started = True
+            backslashes = 0
+        current.append(ch)
+        arg_started = True
+        i += 1
+
+    if backslashes:
+        current.extend("\\" * backslashes)
+        arg_started = True
+    if in_quotes:
+        raise ValueError("No closing quotation")
+    if arg_started or current:
+        parts.append("".join(current))
+    return parts
+
+
 def parse_stdio_command(address: str) -> list[str]:
     """Split a stdio command line into argv. Shared by route validation and the
     transport so both agree on quoting (notably Windows backslash paths)."""
     posix = sys.platform != "win32"
-    parts = shlex.split(address, posix = posix)
-    if not posix:
-        # posix=False keeps backslash paths but also keeps surrounding quotes;
-        # strip a matched pair so argv reaches the subprocess clean.
-        parts = [p[1:-1] if len(p) >= 2 and p[0] == p[-1] and p[0] in "\"'" else p for p in parts]
-    return parts
+    if posix:
+        return shlex.split(address, posix = posix)
+    if address.lstrip().startswith("'"):
+        raise ValueError("Single-quoted executables are not supported on Windows")
+    return _split_windows_command_line(address)
+
+
+def join_stdio_command(parts: list[str]) -> str:
+    """Inverse of parse_stdio_command: join argv into a single command string
+    that parse_stdio_command() splits back into ``parts`` on this platform.
+    Config files (issue #5936) carry structured command + args; storage holds
+    one string in the url field. Windows uses list2cmdline so spaced/backslash
+    paths round-trip through the posix=False quote-strip; posix uses shlex."""
+    if sys.platform == "win32":
+        import subprocess
+        return subprocess.list2cmdline(parts)
+    return shlex.join(parts)
 
 
 def stdio_mcp_enabled() -> bool:
diff --git a/studio/backend/core/inference/mcp_config_import.py b/studio/backend/core/inference/mcp_config_import.py
new file mode 100644
index 0000000000..534ed84c1d
--- /dev/null
+++ b/studio/backend/core/inference/mcp_config_import.py
@@ -0,0 +1,169 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Parse a standard ``mcpServers`` JSON config (Claude Desktop / Cursor / Cline
+/ VS Code) into entries the existing MCP storage understands. See issue #5936.
+
+A stdio entry (``command`` + ``args`` + ``env``) is joined into the single
+command string the ``url`` field already stores; a remote entry (``url`` +
+``headers``) maps straight through. Parsing never raises on a single bad entry:
+it returns ``(entries, errors)`` so one malformed server can't sink the import.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Optional
+
+from core.inference.mcp_client import join_stdio_command
+
+_SCALAR = (str, int, float, bool)
+_UNSUPPORTED_STDIO_FIELDS = ("cwd", "envFile")
+_UNSUPPORTED_TIMEOUT_FIELDS = ("timeout", "timeoutMs", "timeoutSeconds")
+_HTTP_REMOTE_TYPES = ("http", "streamableHttp")
+
+
+@dataclass
+class ParsedMcpEntry:
+    display_name: str
+    url: str  # joined command (stdio) or http(s) url (remote)
+    headers: Optional[dict[str, str]]  # env vars (stdio) or http headers (remote)
+    is_stdio: bool
+    is_enabled: bool = True
+    use_oauth: bool = False
+
+
+def _coerce_str_dict(value: dict) -> dict[str, str]:
+    return {str(k): str(v) for k, v in value.items()}
+
+
+def _has_variable_reference(value: object) -> bool:
+    if isinstance(value, str):
+        return "${" in value
+    if isinstance(value, list):
+        return any(_has_variable_reference(item) for item in value)
+    if isinstance(value, dict):
+        return any(_has_variable_reference(item) for item in value.values())
+    return False
+
+
+def _has_null_value(value: object) -> bool:
+    return isinstance(value, dict) and any(item is None for item in value.values())
+
+
+def _enabled_from_spec(label: str, spec: dict) -> tuple[Optional[bool], Optional[str]]:
+    disabled = spec.get("disabled")
+    if disabled is None:
+        return True, None
+    if not isinstance(disabled, bool):
+        return None, f"{label}: 'disabled' must be true or false."
+    return not disabled, None
+
+
+def _parse_entry(name: str, spec: object) -> tuple[Optional[ParsedMcpEntry], Optional[str]]:
+    label = str(name).strip()
+    if not label:
+        return None, "Server entry has an empty name."
+    if not isinstance(spec, dict):
+        return None, f"{label}: entry must be an object."
+    if _has_variable_reference(spec):
+        return None, f"{label}: VS Code variable references are not supported by import."
+
+    is_enabled, error = _enabled_from_spec(label, spec)
+    if error:
+        return None, error
+
+    has_command = bool(spec.get("command"))
+    has_url = bool(spec.get("url"))
+    if has_command and has_url:
+        return None, f"{label}: entry has both 'command' and 'url'; use one."
+    if not has_command and not has_url:
+        return None, f"{label}: entry needs a 'command' (stdio) or 'url' (remote)."
+
+    if has_command:
+        command = spec["command"]
+        if not isinstance(command, str):
+            return None, f"{label}: 'command' must be a string."
+        entry_type = spec.get("type")
+        if entry_type is not None and entry_type != "stdio":
+            return None, f"{label}: stdio entry has unsupported type {entry_type!r}."
+        sandbox_enabled = spec.get("sandboxEnabled")
+        if sandbox_enabled is not None and not isinstance(sandbox_enabled, bool):
+            return None, f"{label}: 'sandboxEnabled' must be true or false."
+        if sandbox_enabled:
+            return None, f"{label}: sandboxed stdio servers cannot be preserved by import."
+        unsupported = [field for field in _UNSUPPORTED_STDIO_FIELDS if spec.get(field) is not None]
+        if unsupported:
+            return None, f"{label}: import cannot preserve {', '.join(unsupported)}."
+        if spec.get("oauth") is not None:
+            return None, f"{label}: 'oauth' is only supported for remote servers."
+        args = spec.get("args") or []
+        if not isinstance(args, list) or not all(isinstance(a, _SCALAR) for a in args):
+            return None, f"{label}: 'args' must be a list of strings."
+        env = spec.get("env")
+        if env is not None and not isinstance(env, dict):
+            return None, f"{label}: 'env' must be an object."
+        if _has_null_value(env):
+            return None, f"{label}: null environment values are not supported by import."
+        url = join_stdio_command([command, *(str(a) for a in args)])
+        headers = _coerce_str_dict(env) if env else None
+        return ParsedMcpEntry(label, url, headers, True, is_enabled = is_enabled), None
+
+    url = spec["url"]
+    if not isinstance(url, str):
+        return None, f"{label}: 'url' must be a string."
+    url = url.strip()
+    entry_type = spec.get("type")
+    if entry_type is not None and entry_type not in (*_HTTP_REMOTE_TYPES, "sse"):
+        return None, f"{label}: remote entry has unsupported type {entry_type!r}."
+    unsupported_timeout = [
+        field for field in _UNSUPPORTED_TIMEOUT_FIELDS if spec.get(field) is not None
+    ]
+    if unsupported_timeout:
+        return None, f"{label}: import cannot preserve {', '.join(unsupported_timeout)}."
+    url_infers_sse = url.rstrip("/").endswith("/sse")
+    if entry_type == "sse" and not url_infers_sse:
+        return None, f"{label}: explicit SSE transport cannot be preserved for this URL."
+    if entry_type in _HTTP_REMOTE_TYPES and url_infers_sse:
+        return None, f"{label}: explicit HTTP transport cannot be preserved for this URL."
+    oauth_raw = spec.get("oauth")
+    if oauth_raw is not None and not isinstance(oauth_raw, dict):
+        return None, f"{label}: 'oauth' must be an object."
+    headers_raw = spec.get("headers")
+    if headers_raw is not None and not isinstance(headers_raw, dict):
+        return None, f"{label}: 'headers' must be an object."
+    if _has_null_value(headers_raw):
+        return None, f"{label}: null header values are not supported by import."
+    headers = _coerce_str_dict(headers_raw) if headers_raw else None
+    return ParsedMcpEntry(
+        label,
+        url,
+        headers,
+        False,
+        is_enabled = is_enabled,
+        use_oauth = oauth_raw is not None,
+    ), None
+
+
+def parse_mcp_config(config: object) -> tuple[list[ParsedMcpEntry], list[str]]:
+    """Parse a Claude-Desktop/Cursor/Cline/VS Code config. Accepts the
+    ``mcpServers`` key (primary) or ``servers`` (VS Code alias). Returns
+    ``(entries, errors)``; a bad entry adds an error rather than raising."""
+    if not isinstance(config, dict):
+        return [], ["Config must be a JSON object."]
+    servers_key = "mcpServers" if "mcpServers" in config else "servers"
+    servers = config.get(servers_key)
+    if servers is None:
+        return [], ["Config has no 'mcpServers' (or 'servers') object."]
+    if not isinstance(servers, dict):
+        return [], [f"'{servers_key}' must be an object mapping name -> server."]
+
+    entries: list[ParsedMcpEntry] = []
+    errors: list[str] = []
+    for name, spec in servers.items():
+        entry, error = _parse_entry(name, spec)
+        if error:
+            errors.append(error)
+        elif entry:
+            entries.append(entry)
+    return entries, errors
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index fbbdb9a773..5c7799152f 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -7,6 +7,7 @@ instead of torch/transformers for model loading and generation.
 
 import threading
 from typing import Optional, Generator
+from core.inference.runtime_context import runtime_context_length
 from loggers import get_logger
 
 logger = get_logger(__name__)
@@ -175,6 +176,7 @@ class MLXInferenceBackend:
             "is_audio": False,
             "audio_type": None,
             "has_audio_input": False,
+            "context_length": runtime_context_length(self._model, max_seq_length),
         }
         # Capture chat_template_info so the worker IPC reply ships it back and
         # the route layer classifies capabilities like the other paths.
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index e394b342f0..4e4788ace9 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -17,6 +17,7 @@ Pattern follows core/training/training.py.
 import atexit
 import base64
 import os
+import signal
 import structlog
 from loggers import get_logger
 import multiprocessing as mp
@@ -239,6 +240,43 @@ class InferenceOrchestrator:
         """True if the subprocess is alive."""
         return self._proc is not None and self._proc.is_alive()
 
+    def _subprocess_crash_message(self, context: str) -> str:
+        """Return a user-facing crash message with the worker exit status."""
+        context_label = {
+            "wait": "loading the model",
+            "generation": "generating a response",
+            "audio generation": "generating audio",
+            "audio input generation": "processing audio input",
+        }.get(context, context)
+        message = f"The inference worker stopped unexpectedly while {context_label}."
+
+        if self._proc is None:
+            return f"{message} Details: process missing."
+
+        exitcode = self._proc.exitcode
+        pid = self._proc.pid
+        if exitcode is None:
+            return f"{message} Details: pid={pid}."
+
+        if exitcode < 0:
+            signum = -exitcode
+            try:
+                sig_name = signal.Signals(signum).name
+            except ValueError:
+                sig_name = f"SIG{signum}"
+
+            suffix = ""
+            if sig_name == "SIGKILL":
+                suffix = (
+                    " This usually means the system killed it under memory pressure. "
+                    "Try a smaller model, lower context length, or close other GPU-heavy apps."
+                )
+            return (
+                f"{message}{suffix} " f"Details: pid={pid}, signal={sig_name}, exitcode={exitcode}."
+            )
+
+        return f"{message} Details: pid={pid}, exitcode={exitcode}."
+
     # ------------------------------------------------------------------
     # Queue helpers
     # ------------------------------------------------------------------
@@ -286,7 +324,7 @@ class InferenceOrchestrator:
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    raise RuntimeError("Inference subprocess crashed during wait")
+                    raise RuntimeError(self._subprocess_crash_message("wait"))
                 continue
 
             rtype = resp.get("type", "")
@@ -510,7 +548,7 @@ class InferenceOrchestrator:
                 except queue.Empty:
                     # Timeout — check subprocess health
                     if not self._ensure_subprocess_alive():
-                        yield "Error: Inference subprocess crashed during generation"
+                        yield f"Error: {self._subprocess_crash_message('generation')}"
                         return
                     continue
 
@@ -689,6 +727,7 @@ class InferenceOrchestrator:
                         "is_audio": model_info.get("is_audio", False),
                         "audio_type": model_info.get("audio_type"),
                         "has_audio_input": model_info.get("has_audio_input", False),
+                        "context_length": model_info.get("context_length"),
                     }
                     # Mirror chat_template_info so routes can classify caps
                     # without re-entering the subprocess.
@@ -1028,7 +1067,7 @@ class InferenceOrchestrator:
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    yield "Error: Inference subprocess crashed during generation"
+                    yield f"Error: {self._subprocess_crash_message('generation')}"
                     return
                 continue
 
@@ -1125,7 +1164,7 @@ class InferenceOrchestrator:
 
             if resp is None:
                 if not self._ensure_subprocess_alive():
-                    raise RuntimeError("Inference subprocess crashed during audio generation")
+                    raise RuntimeError(self._subprocess_crash_message("audio generation"))
                 continue
 
             rtype = resp.get("type", "")
@@ -1247,7 +1286,7 @@ class InferenceOrchestrator:
 
                 if resp is None:
                     if not self._ensure_subprocess_alive():
-                        yield "Error: Inference subprocess crashed during audio input generation"
+                        yield ("Error: " + self._subprocess_crash_message("audio input generation"))
                         return
                     continue
 
diff --git a/studio/backend/core/inference/runtime_context.py b/studio/backend/core/inference/runtime_context.py
new file mode 100644
index 0000000000..8b881628be
--- /dev/null
+++ b/studio/backend/core/inference/runtime_context.py
@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Runtime context length helpers shared by inference backends."""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+
+def runtime_context_length(model: Any, fallback: Optional[int] = None) -> Optional[int]:
+    """Return the effective context length Unsloth attached to a loaded model."""
+    for value in (getattr(model, "max_seq_length", None), fallback):
+        if isinstance(value, bool):
+            continue
+        try:
+            value_int = int(value)
+        except (TypeError, ValueError):
+            continue
+        if value_int > 0:
+            return value_int
+    return None
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 4f5b985234..cc79654087 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -315,6 +315,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 "audio_type": getattr(mc, "audio_type", None),
                 "has_audio_input": getattr(mc, "has_audio_input", False),
             }
+            try:
+                _bm = getattr(backend, "models", {}) or {}
+                _entry = (
+                    _bm.get(mc.identifier)
+                    or _bm.get(getattr(backend, "active_model_name", None))
+                    or {}
+                )
+                _context_length = _entry.get("context_length")
+                if _context_length is not None:
+                    model_info["context_length"] = int(_context_length)
+            except Exception as _ctx_exc:
+                logger.warning("context_length forward failed: %s", _ctx_exc)
             # Forward chat_template_info so the parent can classify capabilities.
             try:
                 _bm = getattr(backend, "models", {}) or {}
@@ -881,6 +893,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                             name: {
                                 "is_vision": info.get("is_vision", False),
                                 "is_lora": info.get("is_lora", False),
+                                "context_length": info.get("context_length"),
                             }
                             for name, info in backend.models.items()
                         },
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 3d6ab2411c..57342b2453 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -58,7 +58,8 @@ from pathlib import Path
 from typing import Optional, Callable
 from dataclasses import dataclass
 import pandas as pd
-from datasets import Dataset, load_dataset
+from datasets import Dataset
+from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
 
 from core.inference.llama_cpp import _hf_offline_if_dns_dead
 from utils.models import is_vision_model, detect_audio_type
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index c5bc3dbad5..2764101d93 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -534,8 +534,19 @@ class TrainingBackend:
                 except (TypeError, ValueError):
                     logger.debug("Could not convert loss to float: %s", _raw_loss)
                     _safe_loss = None
-                if _safe_loss is not None and not math.isfinite(_safe_loss):
+                _loss_is_nonfinite = _safe_loss is not None and not math.isfinite(_safe_loss)
+                if _loss_is_nonfinite:
+                    # Drop the value rather than laundering it back to the last
+                    # finite loss; clients see loss=None at this step so the NaN
+                    # is not hidden behind a stale value. Training continues.
                     _safe_loss = None
+                    if not getattr(self._progress, "_nonfinite_loss_warned", False):
+                        self._progress._nonfinite_loss_warned = True
+                        logger.warning(
+                            "Training produced non-finite loss at step %s; "
+                            "loss field will report null until it recovers.",
+                            event.get("step", "?"),
+                        )
                 try:
                     _safe_lr = float(_raw_lr) if _raw_lr is not None else None
                 except (TypeError, ValueError):
@@ -545,6 +556,10 @@ class TrainingBackend:
                     _safe_lr = None
                 if _safe_loss is not None:
                     self._progress.loss = _safe_loss
+                elif _loss_is_nonfinite:
+                    # Clear stale finite loss so the API doesn't keep
+                    # reporting the last good value while NaN is happening.
+                    self._progress.loss = None
                 if _safe_lr is not None:
                     self._progress.learning_rate = _safe_lr
                 self._progress.total_steps = event.get("total_steps", self._progress.total_steps)
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index fc2b5a7343..3ed83024c8 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -26,6 +26,22 @@ import subprocess as _sp
 from pathlib import Path
 from typing import Any, Callable
 
+# ── WSL AMD Strix Halo (gfx1151): enable ROCDXG before any torch import ──────
+# Mirrors main.py. In WSL the AMD GPU is reached via the ROCDXG bridge
+# (librocdxg.so over /dev/dxg), which HSA loads only when HSA_ENABLE_DXG_
+# DETECTION=1 is set before torch touches the GPU. A worker spawned outside a
+# login shell misses the installer's persisted env and falls back to CPU.
+# Gated to no-op unless BOTH /dev/dxg and librocdxg.so exist, so native Linux
+# ROCm, NVIDIA, macOS and Windows are unaffected.
+if sys.platform.startswith("linux") and "HSA_ENABLE_DXG_DETECTION" not in os.environ:
+    try:
+        if os.path.exists("/dev/dxg") and any(
+            os.path.exists(_p + "/librocdxg.so") for _p in ("/opt/rocm/lib", "/opt/rocm/lib64")
+        ):
+            os.environ["HSA_ENABLE_DXG_DETECTION"] = "1"
+    except Exception:
+        pass
+
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
 from utils.wheel_utils import (
@@ -678,8 +694,11 @@ def _rocm_classify_unified_memory(props: Any) -> tuple[str, bool]:
       ``set_per_process_memory_fraction`` cap to leave OS headroom.
 
     Classification priority:
-    1. ``gcnArchName`` / variant spellings (stable, naming-independent).
-    2. Device-name substring match (last resort when all arch attrs absent;
+    1. ``props.is_integrated`` truthy (hipDeviceProp_t.integrated -- the
+       driver's own unified-memory answer; covers APUs beyond the hardcoded
+       arch set, e.g. gfx1103 Phoenix iGPUs). Only ever upgrades to unified.
+    2. ``gcnArchName`` / variant spellings (stable, naming-independent).
+    3. Device-name substring match (last resort when all arch attrs absent;
        AMD SDK / Radeon wheels may not populate them):
          - gfx1150 Strix Point: ``Radeon 890M``, ``Radeon 880M``
          - gfx1151 Strix Halo:  ``Radeon 8060S`` (Ryzen AI MAX+ 395),
@@ -692,6 +711,16 @@ def _rocm_classify_unified_memory(props: Any) -> tuple[str, bool]:
             gcn_arch = _v
             break
 
+    # Driver's own answer first: hipDeviceProp_t.integrated (exposed as
+    # props.is_integrated; same gate PR #5988's UMA safetensors fast-load
+    # uses). Strictly additive -- only a truthy value upgrades to unified;
+    # 0/absent falls through to the arch/name logic below, so a wheel that
+    # omits or zeroes the field can never downgrade the known APU set. This
+    # covers unified APUs outside the hardcoded arches (gfx1103 Phoenix
+    # iGPUs, future parts) with one universal signal.
+    if getattr(props, "is_integrated", 0):
+        return gcn_arch, True
+
     if gcn_arch:
         return gcn_arch, gcn_arch in {"gfx1150", "gfx1151"}
 
@@ -1069,7 +1098,74 @@ def _mlx_vlm_max_resized_size(width: int, height: int, target: int) -> tuple[int
     return new_w, new_h
 
 
-def _resize_mlx_vlm_image(image, resize):
+_MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE = {}
+
+
+def _mlx_vlm_resized_image_layout(processor = None) -> str | None:
+    """Return the numpy image layout expected after Studio-side VLM resizing."""
+    image_processor = getattr(processor, "image_processor", None)
+    if image_processor is None:
+        return None
+    cls = image_processor.__class__
+    key = (getattr(cls, "__module__", ""), getattr(cls, "__qualname__", cls.__name__))
+    if key in _MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE:
+        return _MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE[key]
+    copied_image_processor = _copy_mlx_vlm_image_processor(image_processor)
+    layout = (
+        _probe_mlx_vlm_numpy_image_layout(copied_image_processor)
+        if copied_image_processor is not None
+        else None
+    )
+    _MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE[key] = layout
+    return layout
+
+
+def _copy_mlx_vlm_image_processor(image_processor):
+    import copy
+    try:
+        return copy.deepcopy(image_processor)
+    except Exception:
+        try:
+            return copy.copy(image_processor)
+        except Exception:
+            return None
+
+
+def _probe_mlx_vlm_numpy_image_layout(image_processor) -> str | None:
+    try:
+        import numpy as np
+    except ImportError:
+        return None
+
+    def _accepts(candidate) -> bool:
+        try:
+            image_processor(images = [candidate])
+            return True
+        except TypeError:
+            try:
+                image_processor([candidate])
+                return True
+            except Exception:
+                return False
+        except Exception:
+            return False
+
+    # Use an asymmetric image so CHW-vs-HWC mistakes are visible to processors
+    # that skip conversion for 3D numpy arrays.
+    hwc = np.zeros((64, 96, 3), dtype = np.uint8)
+    chw = np.ascontiguousarray(hwc.transpose(2, 0, 1))
+    if _accepts(hwc):
+        return None
+    if _accepts(chw):
+        return "chw"
+    return None
+
+
+def _resize_mlx_vlm_image(
+    image,
+    resize,
+    image_layout = None,
+):
     if resize is None:
         return image
     try:
@@ -1087,16 +1183,27 @@ def _resize_mlx_vlm_image(image, resize):
     # On resize, hand mlx-vlm a writable RGB ndarray so its PIL-path
     # square-resize is skipped and HF processors don't warn on non-writable
     # views. resize=None above keeps the original PIL.
-    return np.array(image, copy = True)
+    array = np.array(image, copy = True)
+    if image_layout == "chw":
+        return np.ascontiguousarray(array.transpose(2, 0, 1))
+    return array
 
 
-def _resize_mlx_vlm_images(value, resize):
+def _resize_mlx_vlm_images(
+    value,
+    resize,
+    image_layout = None,
+):
     if isinstance(value, list):
-        return [_resize_mlx_vlm_image(image, resize) for image in value]
-    return _resize_mlx_vlm_image(value, resize)
+        return [_resize_mlx_vlm_image(image, resize, image_layout = image_layout) for image in value]
+    return _resize_mlx_vlm_image(value, resize, image_layout = image_layout)
 
 
-def _adapt_for_mlx_vlm(items, resize = None):
+def _adapt_for_mlx_vlm(
+    items,
+    resize = None,
+    image_layout = None,
+):
     """Adapt GPU-path VLM dataset output for mlx-vlm.
 
     The GPU path embeds PIL images in message content as
@@ -1116,7 +1223,13 @@ def _adapt_for_mlx_vlm(items, resize = None):
                     if isinstance(part, dict) and part.get("type") == "image":
                         img = part.get("image")
                         if img is not None:
-                            images.append(_resize_mlx_vlm_image(img, resize))
+                            images.append(
+                                _resize_mlx_vlm_image(
+                                    img,
+                                    resize,
+                                    image_layout = image_layout,
+                                )
+                            )
                         new_content.append({"type": "image"})
                     else:
                         new_content.append(part)
@@ -1127,9 +1240,17 @@ def _adapt_for_mlx_vlm(items, resize = None):
         if images:
             out["image"] = images[0] if len(images) == 1 else images
         elif "image" in item:
-            out["image"] = _resize_mlx_vlm_images(item["image"], resize)
+            out["image"] = _resize_mlx_vlm_images(
+                item["image"],
+                resize,
+                image_layout = image_layout,
+            )
         elif "images" in item:
-            out["images"] = _resize_mlx_vlm_images(item["images"], resize)
+            out["images"] = _resize_mlx_vlm_images(
+                item["images"],
+                resize,
+                image_layout = image_layout,
+            )
         adapted.append(out)
     return adapted
 
@@ -1256,7 +1377,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
             "(unsloth_zoo.mlx.loader / unsloth_zoo.mlx.trainer). Reinstall via "
             "install.sh on Apple Silicon."
         ) from e
-    from datasets import load_dataset
+    from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
 
     if mx.metal.is_available():
         info = mx.device_info()
@@ -1276,6 +1397,14 @@ def _run_mlx_training(event_queue, stop_queue, config):
         message = "LoftQ is not supported for MLX training yet."
         _send("error", error = message)
         raise NotImplementedError(message)
+    if config.get("is_embedding"):
+        message = "Embedding model training is not supported for MLX training yet."
+        _send("error", error = message)
+        raise NotImplementedError(message)
+    if config.get("training_type") == "Continued Pretraining":
+        message = "Continued Pretraining is not supported for MLX training yet."
+        _send("error", error = message)
+        raise NotImplementedError(message)
 
     optim_name = _normalize_mlx_studio_optimizer(config.get("optim", "adamw_8bit"))
     lr_scheduler_type = _normalize_mlx_studio_scheduler(config.get("lr_scheduler_type", "linear"))
@@ -1284,6 +1413,12 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # Force text-only for non-image datasets even on vision-capable models
     # (e.g. Qwen3.5-VL trained on plain alpaca text).
     _send("status", status_message = f"Loading {model_name}...")
+    # Pull through resume_from_checkpoint so MLXTrainer.train() can restore
+    # optimizer + step state and continue cleanly. Was previously dropped on
+    # the floor for the MLX path, so the Resume UI button silently restarted
+    # from step 0 (the CUDA path at lines 2729 / 3108 has been forwarding
+    # this all along).
+    resume_from_checkpoint = config.get("resume_from_checkpoint") or None
     is_dataset_image = bool(config.get("is_dataset_image", False))
     training_type = config.get("training_type", "LoRA/QLoRA")
     use_lora = training_type == "LoRA/QLoRA"
@@ -1317,7 +1452,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
             "status",
             status_message = f"MLX vision image resize: {vision_image_size} (max dimension)",
         )
-
     # ── 2. Apply LoRA / full FT ──
     # gradient_checkpointing stays a string ("mlx"/"unsloth"/"none"/etc.);
     # get_peft_model and MLXTrainer both accept and handle strings.
@@ -1426,6 +1560,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # Reuse the GPU format pipeline for VLM (auto-detects OCR/caption/llava/
     # sharegpt+images) and text (alpaca/sharegpt/chatml → "text" column).
     format_type = config.get("format_type", "")
+    custom_format_mapping = config.get("custom_format_mapping")
     try:
         from utils.datasets import format_and_template_dataset
         def _fmt_progress(status_message = "", **_kw):
@@ -1439,12 +1574,19 @@ def _run_mlx_training(event_queue, stop_queue, config):
                 tokenizer = tokenizer,
                 is_vlm = True,
                 dataset_name = hf_dataset or "local",
+                custom_format_mapping = custom_format_mapping,
                 progress_callback = _fmt_progress,
             )
             if vlm_info.get("success"):
+                vision_image_layout = (
+                    _mlx_vlm_resized_image_layout(tokenizer)
+                    if vision_image_size is not None
+                    else None
+                )
                 dataset = _adapt_for_mlx_vlm(
                     vlm_info["dataset"],
                     resize = vision_image_size,
+                    image_layout = vision_image_layout,
                 )
             else:
                 errors = vlm_info.get("errors", [])
@@ -1456,11 +1598,18 @@ def _run_mlx_training(event_queue, stop_queue, config):
                     tokenizer = tokenizer,
                     is_vlm = True,
                     dataset_name = hf_dataset or "local",
+                    custom_format_mapping = custom_format_mapping,
                 )
                 if ev_info.get("success"):
+                    vision_image_layout = (
+                        _mlx_vlm_resized_image_layout(tokenizer)
+                        if vision_image_size is not None
+                        else None
+                    )
                     eval_dataset = _adapt_for_mlx_vlm(
                         ev_info["dataset"],
                         resize = vision_image_size,
+                        image_layout = vision_image_layout,
                     )
 
         elif format_type:
@@ -1472,6 +1621,8 @@ def _run_mlx_training(event_queue, stop_queue, config):
                 is_vlm = False,
                 format_type = format_type,
                 dataset_name = hf_dataset or "local",
+                custom_format_mapping = custom_format_mapping,
+                progress_callback = _fmt_progress,
             )
             if info.get("success", True):
                 dataset = info.get("dataset", dataset)
@@ -1483,6 +1634,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
                     is_vlm = False,
                     format_type = format_type,
                     dataset_name = hf_dataset or "local",
+                    custom_format_mapping = custom_format_mapping,
                 )
                 if ev.get("success", True):
                     eval_dataset = ev.get("dataset", eval_dataset)
@@ -1730,7 +1882,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # ── 11. Run training ──
     gc.collect()
     mx.synchronize()
-    trainer.train()
+    trainer.train(resume_from_checkpoint = resume_from_checkpoint)
 
     # ── 12. Save and finalize ──
     if trainer.stop_requested and not _stop_save[0]:
@@ -1829,7 +1981,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                 }
             )
             return
-        # Activate correct transformers version (Gemma-4 needs 5.5.0, etc.)
+        # Activate correct transformers version (Gemma-4 needs a 5.x sidecar, etc.)
         # before any transformers/mlx-lm imports in _run_mlx_training.
         try:
             _activate_transformers_version(model_name)
@@ -2006,12 +2158,29 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                 os.environ["TORCHDYNAMO_DISABLE"] = "1"
                 logger.info("Windows ROCm: torch.compile (dynamo) disabled")
 
+            # bitsandbytes' import-time get_rocm_gpu_arch() probe runs
+            # `hipinfo.exe` from PATH; the AMD torch wheel ships it in the venv
+            # Scripts dir, which is on PATH only for activated venvs. Prepend
+            # it so the probe succeeds instead of logging a scary (harmless)
+            # "Could not detect ROCm GPU architecture" ERROR on every import.
+            # Normally inherited from main.py's env, but workers can also be
+            # spawned standalone (tests, CLI) -- keep the guard here too.
+            _scripts_dir = os.path.dirname(sys.executable)
+            if os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe")):
+                import shutil as _shutil
+                if not _shutil.which("hipinfo.exe"):
+                    os.environ["PATH"] = _scripts_dir + os.pathsep + os.environ.get("PATH", "")
+
             # BNB picks a rocm DLL from torch.version.hip, but AMD's Windows BNB
             # wheel may ship a DLL whose suffix doesn't match. Detect the actual
-            # DLL name and override; "72" is a safe fallback. Callers may
-            # pre-set the var to override.
-            if "BNB_ROCM_VERSION" not in os.environ:
+            # DLL name and override. Values seeded by the installer are
+            # redetectable defaults, while caller overrides remain authoritative.
+            if (
+                "BNB_ROCM_VERSION" not in os.environ
+                or os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") == "sitecustomize"
+            ):
                 _bnb_rocm_ver = None
+                _found_rocm_bnb = False
                 try:
                     import glob as _glob
                     import importlib.util as _ilu
@@ -2024,6 +2193,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                             for _dll in _glob.glob(
                                 os.path.join(_pkg_dir, "libbitsandbytes_rocm*.dll")
                             ):
+                                _found_rocm_bnb = True
                                 _m = _re.search(
                                     r"libbitsandbytes_rocm(\d+)\.dll",
                                     os.path.basename(_dll),
@@ -2035,14 +2205,20 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                             _bnb_rocm_ver = max(_all_vers, key = lambda v: int(v))
                 except Exception:
                     pass
-                _bnb_rocm_ver = _bnb_rocm_ver or "72"
-                os.environ["BNB_ROCM_VERSION"] = _bnb_rocm_ver
-                logger.info(
-                    "Windows ROCm: set BNB_ROCM_VERSION=%s "
-                    "(detected from installed BNB wheel; "
-                    "overrides torch.version.hip auto-detection)",
-                    _bnb_rocm_ver,
-                )
+                # Only when a ROCm bnb DLL actually exists (mirrors main.py):
+                # without one the seeded value and its marker stay untouched,
+                # so later import fixes can still redetect or opt out. DLL
+                # with unparsable name -> seeded value or "72".
+                if _found_rocm_bnb:
+                    _bnb_rocm_ver = _bnb_rocm_ver or os.environ.get("BNB_ROCM_VERSION") or "72"
+                    os.environ["BNB_ROCM_VERSION"] = _bnb_rocm_ver
+                    os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] = "detected"
+                    logger.info(
+                        "Windows ROCm: set BNB_ROCM_VERSION=%s "
+                        "(detected from installed BNB wheel; "
+                        "overrides torch.version.hip auto-detection)",
+                        _bnb_rocm_ver,
+                    )
 
             # Parse HIP version for the kernel-fix gate below, falling back to
             # the rocm version embedded in torch.__version__ when version.hip is
@@ -2202,10 +2378,25 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                 if _is_unified and not _gcn_arch:
                     logger.debug(
                         "ROCm OOM guard: gcnArchName absent -- inferred "
-                        "unified memory from device name %r; applying 0.80 cap",
+                        "unified memory from device name %r; applying unified cap",
                         _dev_name,
                     )
-                _mem_fraction = 0.80 if _is_unified else 0.90
+                # Unified hosts on native Windows: mem_get_info's total is the
+                # WDDM budget the driver grants HIP (BIOS carve + ~half of the
+                # remaining RAM) -- the OS share is already outside it, so the
+                # Linux 0.80 starve-protection double-taxes (48.49 GiB budget →
+                # 38.79 allowed) and blocks loads that fit in free memory.
+                # 1.0 removes the double-tax. Current AMD Windows wheels only
+                # enforce sub-1.0 fractions (measured on gfx1151: 0.5 caps,
+                # 1.0 still allocates past the budget via WDDM overcommit), so
+                # 1.0 behaves like torch's uncapped default, with WDDM
+                # arbitrating residency; on wheels that do enforce it, it caps
+                # at exactly the driver-granted budget. On Linux the total
+                # spans nearly all RAM, so keep the 0.80 OS headroom there.
+                if _is_unified:
+                    _mem_fraction = 1.0 if sys.platform == "win32" else 0.80
+                else:
+                    _mem_fraction = 0.90
                 _torch_mem.cuda.set_per_process_memory_fraction(_mem_fraction)
                 logger.info(
                     "ROCm OOM guard: set_per_process_memory_fraction(%.2f) — "
@@ -2215,6 +2406,28 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                     _dev_name,
                     _gcn_arch or "unknown arch",
                 )
+                # Unified Windows APUs: the WDDM budget is user-raisable, but
+                # nothing on the box says so -- users see "48 GB VRAM" on a
+                # 96 GB machine and assume a Studio bug. Say where the limit
+                # comes from and how to raise it.
+                if _is_unified and sys.platform == "win32":
+                    try:
+                        import psutil as _psutil
+
+                        _phys = _psutil.virtual_memory().total
+                        _granted = _torch_mem.cuda.mem_get_info(0)[1]
+                        if _granted < 0.75 * _phys:
+                            logger.info(
+                                "Windows grants the GPU %.1f GiB of %.1f GiB "
+                                "system RAM (driver/WDDM budget). To raise it: "
+                                "increase the BIOS UMA frame buffer size, or "
+                                "AMD Software > Performance > Tuning > "
+                                "Variable Graphics Memory.",
+                                _granted / 1024**3,
+                                _phys / 1024**3,
+                            )
+                    except Exception:
+                        pass
         except Exception as _oom_guard_err:
             logger.debug("Could not set GPU memory fraction: %s", _oom_guard_err)
 
@@ -2768,7 +2981,8 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
         )
         from sentence_transformers.losses import MultipleNegativesRankingLoss
         from sentence_transformers.training_args import BatchSamplers
-        from datasets import load_dataset, Dataset
+        from datasets import Dataset
+        from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
         from transformers import TrainerCallback
         from utils.paths import datasets_root, resolve_output_dir
     except ImportError as e:
diff --git a/studio/backend/hub/services/datasets/formatting.py b/studio/backend/hub/services/datasets/formatting.py
index 8b0ff39f63..1c78d80e21 100644
--- a/studio/backend/hub/services/datasets/formatting.py
+++ b/studio/backend/hub/services/datasets/formatting.py
@@ -223,7 +223,10 @@ def _load_processed_hf_preview_slice(
     if not _is_valid_repo_id(request.dataset_name):
         return None
     try:
-        from datasets import DownloadConfig, load_dataset
+        from datasets import DownloadConfig
+
+        # Non-streaming loads take the cached builder lock; use the EACCES-safe wrapper.
+        from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
     except Exception:
         return None
 
diff --git a/studio/backend/hub/services/datasets/local.py b/studio/backend/hub/services/datasets/local.py
index 2d2c7c3a0d..8d48c4f735 100644
--- a/studio/backend/hub/services/datasets/local.py
+++ b/studio/backend/hub/services/datasets/local.py
@@ -224,7 +224,8 @@ def _stream_file_preview_slice(path: Path, preview_size: int):
 
 
 def _load_local_preview_slice(*, dataset_path: Path, train_split: str, preview_size: int):
-    from datasets import load_dataset
+    # Non-streaming loads take the cached builder lock; use the EACCES-safe wrapper.
+    from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
 
     if dataset_path.is_dir():
         parquet_dir = (
diff --git a/studio/backend/hub/services/models/common.py b/studio/backend/hub/services/models/common.py
index 688d324a60..f381bffe9c 100644
--- a/studio/backend/hub/services/models/common.py
+++ b/studio/backend/hub/services/models/common.py
@@ -21,6 +21,7 @@ from hub.utils.gguf import (
     extract_quant_label,
     is_gguf_filename as _is_gguf_filename,
     is_mmproj_filename as _is_mmproj_filename,
+    is_mtp_drafter_path as _is_mtp_drafter_path,
 )
 from hub.utils.paths import is_valid_repo_id as _is_valid_repo_id
 
@@ -71,7 +72,7 @@ def _is_model_directory(d: Path) -> bool:
         if suffix == ".safetensors":
             return True
         if suffix == ".gguf":
-            return "mmproj" not in f.name.lower()
+            return "mmproj" not in f.name.lower() and not _is_mtp_drafter_path(f.name)
         if suffix == ".bin":
             name = f.name.lower()
             return (
@@ -276,7 +277,9 @@ def _classify_non_gguf_model_format(
 
 
 def _is_main_gguf_filename(name: str) -> bool:
-    return _is_gguf_filename(name) and not _is_mmproj_filename(name)
+    return (
+        _is_gguf_filename(name) and not _is_mmproj_filename(name) and not _is_mtp_drafter_path(name)
+    )
 
 
 def _iter_gguf_paths(root: Path):
diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index d4aed0d59f..ecc9f8426d 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -32,6 +32,7 @@ from hub.services.models.common import (
     _is_gguf_filename,
     _is_main_gguf_filename,
     _is_mmproj_filename,
+    _is_mtp_drafter_path,
 )
 
 logger = get_logger(__name__)
@@ -139,7 +140,10 @@ def _delete_gguf_variant_from_repos(
         if matched and not sibling_active and not _has_remaining_main_gguf(target_repo):
             companion_matches = _repo_file_matches(
                 target_repo,
-                lambda name: _is_gguf_filename(name) and _is_mmproj_filename(name),
+                # Companions: mmproj and the MTP drafter -- downloaded with
+                # every variant, so the last variant's delete reclaims them.
+                lambda name: _is_gguf_filename(name)
+                and (_is_mmproj_filename(name) or _is_mtp_drafter_path(name)),
             )
             for snap, _blob, name in companion_matches:
                 try:
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index efe6bdf6c4..0b4671c6d2 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -38,6 +38,7 @@ from hub.utils.paths import (
 )
 from hub.services.models.common import (
     _is_mmproj_filename,
+    _is_mtp_drafter_path,
     _iter_gguf_paths,
 )
 from hub.utils.gguf_plan import (
@@ -479,7 +480,7 @@ async def get_gguf_variants_response(
                         continue
                     key = rel.lower()
                     by_filename[key] = max(by_filename.get(key, 0), size)
-                    if _is_mmproj_filename(f.name):
+                    if _is_mmproj_filename(f.name) or _is_mtp_drafter_path(rel):
                         continue
                     q = extract_quant_label(rel).lower()
                     by_quant[q] = by_quant.get(q, 0) + size
@@ -593,7 +594,11 @@ async def get_gguf_variants_response(
                 requirement = requirements_by_quant.get(variant.quant.lower())
                 if requirement is None:
                     continue
-                if requirement.mmproj_hashes & incomplete_hashes and _filenames_cached(
+                # companion_hashes adds the MTP drafter (mmproj_hashes covers
+                # every mmproj precision in the repo, not just the planned one).
+                if (
+                    (requirement.mmproj_hashes | requirement.companion_hashes) & incomplete_hashes
+                ) and _filenames_cached(
                     requirement.main_filenames,
                     requirement.main_size_bytes,
                 ):
diff --git a/studio/backend/hub/utils/gguf.py b/studio/backend/hub/utils/gguf.py
index 4bd1961c33..c9b3ce97e3 100644
--- a/studio/backend/hub/utils/gguf.py
+++ b/studio/backend/hub/utils/gguf.py
@@ -83,6 +83,28 @@ def is_mmproj_filename(filename: str) -> bool:
     return "mmproj" in filename.lower()
 
 
+def is_mtp_drafter_path(path: str) -> bool:
+    """True for a separate-file MTP drafter (speculative head), a companion to
+    the main model rather than a selectable quant.
+
+    Covers the repo-root ``mtp-*.gguf`` (the Q8_0 copy unsloth ships for
+    llama.cpp ``-hf`` auto-discovery) and the ``MTP/`` subdir copies (Gemma 4).
+    Repos that bake the head into the main GGUF (Qwen) have no such file, so
+    this is False for them. Must be excluded from main-model selection
+    everywhere mmproj is.
+
+    CANONICAL COPY. Layering keeps two mirrors that must change in lockstep:
+    utils/models/model_config.py ``_is_mtp_drafter`` (utils cannot import
+    hub) and core/inference/llama_cpp.py ``_is_companion_gguf_path`` (core
+    avoids hub imports; bundles the mmproj check).
+    """
+    p = path.lower()
+    if not p.endswith(".gguf"):
+        return False
+    name = p.rsplit("/", 1)[-1]
+    return name.startswith("mtp-") or "/mtp/" in f"/{p}"
+
+
 def is_gguf_filename(filename: str) -> bool:
     return filename.lower().endswith(".gguf")
 
@@ -119,7 +141,9 @@ def iter_gguf_files(directory: Path, recursive: bool = False):
 
 def pick_best_gguf(filenames: list[str]) -> Optional[str]:
     gguf_files = [
-        name for name in filenames if is_gguf_filename(name) and not is_mmproj_filename(name)
+        name
+        for name in filenames
+        if is_gguf_filename(name) and not is_mmproj_filename(name) and not is_mtp_drafter_path(name)
     ]
     if not gguf_files:
         return None
@@ -270,6 +294,12 @@ def list_partial_gguf_variants_from_state(
             for expected in manifest.expected_files:
                 if not is_gguf_filename(expected.path):
                     continue
+                if is_mtp_drafter_path(expected.path):
+                    # Downloaded with every variant (like mmproj) but not a
+                    # selectable quant; count it so the shown download size
+                    # matches what is fetched.
+                    companion_bytes += max(0, int(expected.size or 0))
+                    continue
                 if is_mmproj_filename(expected.path):
                     has_vision = True
                     companion_bytes += max(0, int(expected.size or 0))
@@ -336,6 +366,8 @@ def list_gguf_variants(
         filename = getattr(sibling, "rfilename", None)
         if not isinstance(filename, str) or not is_gguf_filename(filename):
             continue
+        if is_mtp_drafter_path(filename):
+            continue
         if is_mmproj_filename(filename):
             has_vision = True
             continue
@@ -389,6 +421,8 @@ def list_local_gguf_variants(directory: str) -> tuple[list[GgufVariantInfo], boo
         except OSError:
             size = 0
         rel = file.relative_to(root).as_posix()
+        if is_mtp_drafter_path(rel):
+            continue
         quant = extract_quant_label(rel)
         quant_totals[quant] = quant_totals.get(quant, 0) + size
         quant_first_file.setdefault(quant, rel)
diff --git a/studio/backend/hub/utils/gguf_plan.py b/studio/backend/hub/utils/gguf_plan.py
index 03ec847f4c..60782fa109 100644
--- a/studio/backend/hub/utils/gguf_plan.py
+++ b/studio/backend/hub/utils/gguf_plan.py
@@ -7,7 +7,12 @@ from dataclasses import dataclass
 from typing import Optional, Sequence
 
 from hub.utils.download_manifest import ExpectedFile
-from hub.utils.gguf import extract_quant_label, is_gguf_filename, is_mmproj_filename
+from hub.utils.gguf import (
+    extract_quant_label,
+    is_gguf_filename,
+    is_mmproj_filename,
+    is_mtp_drafter_path,
+)
 
 
 @dataclass(frozen = True)
@@ -53,24 +58,30 @@ def expected_file_from_sibling(sibling) -> Optional[ExpectedFile]:
 
 
 def is_companion_gguf_path(path: str) -> bool:
-    return is_gguf_filename(path) and is_mmproj_filename(path)
+    """Companion (non-main) GGUF downloaded alongside a variant: the vision
+    mmproj or the separate MTP drafter (Gemma 4)."""
+    return is_gguf_filename(path) and (is_mmproj_filename(path) or is_mtp_drafter_path(path))
 
 
 def is_main_gguf_variant_path(path: str, variant: str) -> bool:
     return (
         is_gguf_filename(path)
         and not is_mmproj_filename(path)
+        and not is_mtp_drafter_path(path)
         and extract_quant_label(path).lower() == variant.lower()
     )
 
 
+def _gguf_rfilename(sibling) -> Optional[str]:
+    """The sibling's rfilename when it is a GGUF, else None."""
+    name = getattr(sibling, "rfilename", None)
+    if isinstance(name, str) and is_gguf_filename(name):
+        return name
+    return None
+
+
 def mmproj_siblings(siblings: Sequence) -> list:
-    return [
-        s
-        for s in siblings
-        if isinstance(getattr(s, "rfilename", None), str)
-        and is_companion_gguf_path(getattr(s, "rfilename"))
-    ]
+    return [s for s in siblings if (name := _gguf_rfilename(s)) and is_mmproj_filename(name)]
 
 
 def preferred_mmproj_sibling(siblings: Sequence) -> Optional[object]:
@@ -83,6 +94,25 @@ def preferred_mmproj_sibling(siblings: Sequence) -> Optional[object]:
     )
 
 
+def preferred_mtp_sibling(siblings: Sequence) -> Optional[object]:
+    """The separate MTP drafter to fetch with every variant: the repo-root
+    ``mtp-*.gguf`` copy unsloth ships for llama.cpp ``-hf`` auto-discovery
+    (Gemma 4). Same pick as the loader's drafter resolution (``mtp-`` basename
+    prefix, first in sort order) so download and load resolve the same file;
+    the higher-precision ``MTP/`` subdir copies are for explicit selection and
+    are not auto-fetched. None for repos with the head baked into the main
+    GGUF (Qwen)."""
+    candidates = sorted(
+        (
+            s
+            for s in siblings
+            if (name := _gguf_rfilename(s)) and name.lower().rsplit("/", 1)[-1].startswith("mtp-")
+        ),
+        key = lambda s: getattr(s, "rfilename"),
+    )
+    return candidates[0] if candidates else None
+
+
 def build_gguf_variant_plans(siblings: Sequence) -> dict[str, GgufVariantPlan]:
     main: dict[str, list] = {}
     all_mmproj = mmproj_siblings(siblings)
@@ -94,12 +124,20 @@ def build_gguf_variant_plans(siblings: Sequence) -> dict[str, GgufVariantPlan]:
     all_mmproj_hashes = frozenset(h for h in (sibling_sha256(s) for s in all_mmproj) if h)
     companion = preferred_mmproj_sibling(siblings)
     companion_expected = expected_file_from_sibling(companion) if companion is not None else None
+    mtp_sibling = preferred_mtp_sibling(siblings)
+    mtp_expected = expected_file_from_sibling(mtp_sibling) if mtp_sibling is not None else None
+    companions_expected = tuple(
+        file for file in (companion_expected, mtp_expected) if file is not None
+    )
 
     for sibling in siblings:
-        name = getattr(sibling, "rfilename", None)
-        if not isinstance(name, str) or not is_gguf_filename(name):
+        name = _gguf_rfilename(sibling)
+        if name is None:
             continue
-        if is_mmproj_filename(name):
+        # Companions are folded into every plan below; keep them out of the
+        # quant grouping so a drafter never lands in a variant's main files
+        # (the root mtp-*.gguf carries a quant label, e.g. Q8_0).
+        if is_mmproj_filename(name) or is_mtp_drafter_path(name):
             continue
         quant = extract_quant_label(name).lower()
         main.setdefault(quant, []).append(sibling)
@@ -111,11 +149,7 @@ def build_gguf_variant_plans(siblings: Sequence) -> dict[str, GgufVariantPlan]:
             for sibling in target_main_siblings
             if (file := expected_file_from_sibling(sibling)) is not None
         )
-        expected_files = (
-            (*main_expected, companion_expected)
-            if companion_expected is not None
-            else main_expected
-        )
+        expected_files = (*main_expected, *companions_expected)
         plans[quant] = plan_from_expected_files(
             quant,
             expected_files,
@@ -135,6 +169,9 @@ def plan_from_expected_files(
     expected = tuple(expected_files)
     main_files = tuple(file for file in expected if is_main_gguf_variant_path(file.path, variant))
     companion_files = tuple(file for file in expected if is_companion_gguf_path(file.path))
+    # Manifest-resume fallback for the mmproj fields below: companion_files
+    # also holds the MTP drafter, so keep an mmproj-only view.
+    mmproj_files = tuple(file for file in companion_files if is_mmproj_filename(file.path))
     main_hashes = frozenset(file.sha256 for file in main_files if file.sha256)
     companion_hashes = frozenset(file.sha256 for file in companion_files if file.sha256)
     required_hashes = frozenset(file.sha256 for file in expected if file.sha256)
@@ -149,9 +186,13 @@ def plan_from_expected_files(
         mmproj_filenames = (
             all_mmproj_filenames
             if all_mmproj_filenames is not None
-            else frozenset(file.path for file in companion_files)
+            else frozenset(file.path for file in mmproj_files)
+        ),
+        mmproj_hashes = (
+            all_mmproj_hashes
+            if all_mmproj_hashes is not None
+            else frozenset(file.sha256 for file in mmproj_files if file.sha256)
         ),
-        mmproj_hashes = (all_mmproj_hashes if all_mmproj_hashes is not None else companion_hashes),
         expected_files = expected,
         main_size_bytes = main_size,
         download_size_bytes = download_size,
diff --git a/studio/backend/hub/utils/inventory_scan.py b/studio/backend/hub/utils/inventory_scan.py
index a7281ad8b3..fe3aadfd38 100644
--- a/studio/backend/hub/utils/inventory_scan.py
+++ b/studio/backend/hub/utils/inventory_scan.py
@@ -25,7 +25,12 @@ from loggers import get_logger
 
 logger = get_logger(__name__)
 
-from hub.utils.gguf import extract_quant_label, is_gguf_filename, is_mmproj_filename
+from hub.utils.gguf import (
+    extract_quant_label,
+    is_gguf_filename,
+    is_mmproj_filename,
+    is_mtp_drafter_path,
+)
 from hub.utils.state_dir import RepoType
 
 from hub.utils.hf_cache_state import (
@@ -336,7 +341,7 @@ def _completed_gguf_variants(snapshot_dir: Optional[Path]) -> set[str]:
         except OSError:
             continue
         rel = path.relative_to(snapshot_dir).as_posix()
-        if not is_gguf_filename(rel) or is_mmproj_filename(rel):
+        if not is_gguf_filename(rel) or is_mmproj_filename(rel) or is_mtp_drafter_path(rel):
             continue
         quant = extract_quant_label(rel)
         split = _GGUF_SPLIT_RE.search(path.name)
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 35c955405e..2b34b11fef 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -63,17 +63,38 @@ if sys.platform == "win32":
     _add_rocm_dll_dirs()
     del _add_rocm_dll_dirs
 
+    # ── Windows AMD ROCm: make hipInfo.exe resolvable for subprocess probes ──
+    # bitsandbytes' get_rocm_gpu_arch() runs `hipinfo.exe` via PATH at import
+    # time; the AMD torch wheel ships it in the venv Scripts dir, which is on
+    # PATH only when the venv is activated -- Studio launches python directly.
+    # Without this, every bitsandbytes import logs a scary (but harmless)
+    # "Could not detect ROCm GPU architecture: [WinError 2]" ERROR + WARNING.
+    # Gated on the file existing: only AMD ROCm wheels ship hipInfo.exe, so
+    # NVIDIA/CPU hosts are untouched. os.add_dll_directory above does not help
+    # here -- subprocess PATH resolution ignores DLL search directories.
+    _scripts_dir = os.path.dirname(sys.executable)
+    if os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe")):
+        import shutil as _shutil
+        if not _shutil.which("hipinfo.exe"):
+            os.environ["PATH"] = _scripts_dir + os.pathsep + os.environ.get("PATH", "")
+        del _shutil
+    del _scripts_dir
+
     # ── Windows AMD ROCm: set BNB_ROCM_VERSION before any bitsandbytes import ─
     # bitsandbytes derives the rocm<ver>.dll name from torch.version.hip, but the
     # wheel ships rocm72.dll, so the server crashes ("Configured ROCm binary not
-    # found") without this. Detect the shipped DLL and fall back to "72" (mirrors
-    # worker.py). Gate on the rocm bnb DLL / HIP_PATH rather than torch.version.hip
-    # to avoid importing torch on every Windows host.
-    if "BNB_ROCM_VERSION" not in os.environ:
+    # found") without this. Detect the shipped DLL (mirrors worker.py); gate on
+    # the rocm bnb DLL rather than torch.version.hip to avoid importing torch on
+    # every Windows host.
+    # Values seeded by the installer's sitecustomize.py are redetectable
+    # defaults; explicit caller values remain authoritative.
+    if (
+        "BNB_ROCM_VERSION" not in os.environ
+        or os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") == "sitecustomize"
+    ):
         import glob as _glob
         import logging as _logging
 
-        _hip_env = bool(os.environ.get("HIP_PATH") or os.environ.get("ROCM_PATH"))
         _bnb_rocm_ver = None
         _found_rocm_bnb = False
         try:
@@ -96,18 +117,43 @@ if sys.platform == "win32":
                     _bnb_rocm_ver = max(_all_vers_main, key = lambda v: int(v))
         except Exception as _e:
             _logging.getLogger(__name__).warning(
-                "Windows ROCm: BNB DLL detection failed (%s); falling back to version '72'",
+                "Windows ROCm: BNB DLL detection failed (%s); leaving BNB_ROCM_VERSION as is",
                 _e,
             )
-        # rocm bnb DLL present, or HIP_PATH/ROCM_PATH set (DLL unparsable -> "72")
-        if _found_rocm_bnb or _hip_env:
-            _bnb_rocm_ver_final = _bnb_rocm_ver or "72"
+        # Only when a ROCm bnb DLL actually exists: HIP_PATH/ROCM_PATH alone
+        # (HIP SDK on a CUDA/CPU box) must not force a ROCm backend onto a
+        # non-ROCm bitsandbytes, which raises at import. DLL unparsable -> "72".
+        if _found_rocm_bnb:
+            _bnb_rocm_ver_final = _bnb_rocm_ver or os.environ.get("BNB_ROCM_VERSION") or "72"
             os.environ["BNB_ROCM_VERSION"] = _bnb_rocm_ver_final
+            os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] = "detected"
             _logging.getLogger(__name__).info(
                 "Windows ROCm: set BNB_ROCM_VERSION=%s (from installed BNB wheel)",
                 _bnb_rocm_ver_final,
             )
 
+# ── WSL AMD Strix Halo (gfx1151): enable ROCDXG before any torch import ──────
+# In WSL the AMD GPU is reached via the ROCDXG bridge (librocdxg.so over
+# /dev/dxg), which HSA loads only when HSA_ENABLE_DXG_DETECTION=1 is set BEFORE
+# torch touches the GPU. A worker launched outside a login shell (e.g.
+# `wsl.exe -d Ubuntu-24.04 python ...`) misses the installer's persisted env
+# and silently falls back to CPU. Set it here, gated to no-op unless BOTH
+# /dev/dxg AND librocdxg.so exist -- native Linux ROCm, NVIDIA, macOS and
+# Windows are unaffected.
+elif sys.platform.startswith("linux") and "HSA_ENABLE_DXG_DETECTION" not in os.environ:
+    try:
+        if os.path.exists("/dev/dxg") and any(
+            os.path.exists(os.path.join(_p, "librocdxg.so"))
+            for _p in ("/opt/rocm/lib", "/opt/rocm/lib64")
+        ):
+            os.environ["HSA_ENABLE_DXG_DETECTION"] = "1"
+            import logging as _logging
+            _logging.getLogger(__name__).info(
+                "WSL ROCm: set HSA_ENABLE_DXG_DETECTION=1 (librocdxg bridge present)"
+            )
+    except Exception:
+        pass
+
 # Put backend dir on sys.path so _platform_compat is importable when main.py
 # is launched directly (e.g. `uvicorn main:app`).
 _backend_dir = str(_Path(__file__).parent)
@@ -221,6 +267,7 @@ from routes import (
     training_history_router,
     training_router,
 )
+from routes.llama import router as llama_router
 from hub.routes import (
     inventory_router as hub_inventory_router,
     datasets_router as hub_datasets_router,
@@ -759,6 +806,7 @@ app.include_router(mcp_servers_router, prefix = "/api/mcp/servers", tags = ["mcp
 app.include_router(prompts_router, prefix = "/api/prompts", tags = ["prompts"])
 app.include_router(datasets_router, prefix = "/api/datasets", tags = ["datasets"])
 app.include_router(data_recipe_router, prefix = "/api/data-recipe", tags = ["data-recipe"])
+app.include_router(llama_router, prefix = "/api/llama", tags = ["llama"])
 app.include_router(export_router, prefix = "/api/export", tags = ["export"])
 app.include_router(rag_router, prefix = "/api/rag", tags = ["rag"])
 app.include_router(training_history_router, prefix = "/api/train", tags = ["training-history"])
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 9c12f8faac..b70202d6ff 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -72,7 +72,7 @@ class LoadRequest(BaseModel):
             "ngram-mod+draft-mtp chain on both platforms), 'off' (disabled). "
             "Legacy values 'default' (-> auto), 'draft-mtp' (-> mtp), "
             "'ngram-mod' (-> ngram), and 'ngram-simple' (kept as-is) are "
-            "still accepted. Ignored for non-GGUF and vision models."
+            "still accepted. Ignored for non-GGUF models."
         ),
     )
     spec_draft_n_max: Optional[int] = Field(
@@ -171,7 +171,7 @@ class LoadResponse(BaseModel):
         description = "Whether the model defaults require trust_remote_code to be enabled for loading.",
     )
     context_length: Optional[int] = Field(
-        None, description = "Model's native context length (from GGUF metadata)"
+        None, description = "Runtime context length in tokens for the loaded model"
     )
     max_context_length: Optional[int] = Field(
         None, description = "Maximum context length currently available on this hardware"
@@ -346,6 +346,17 @@ class InferenceStatusResponse(BaseModel):
             "False -> recommend `unsloth studio update`."
         ),
     )
+    spec_fallback_reason: Optional[str] = Field(
+        None,
+        description = (
+            "Why MTP was disabled on the loaded model despite being requested "
+            "(auto on an MTP model, or forced mtp / mtp+ngram). "
+            "'binary_no_mtp' / 'binary_outdated' -> a newer prebuilt would "
+            "re-enable it (show the update affordance); 'runtime_error' -> the "
+            "current build could not run it. None when MTP engaged or was not "
+            "requested."
+        ),
+    )
     llama_cpp_prebuilt_stale: bool = Field(
         False,
         description = (
@@ -533,8 +544,10 @@ class ChatMessage(BaseModel):
 
         if self.role == "tool":
             # tool_call_id resolution happens at ChatCompletionRequest scope.
-            if not self.content:
-                raise ValueError('role="tool" messages require non-empty "content".')
+            # OpenAI accepts empty tool results (commands with no output);
+            # normalize to "" instead of a 400 agentic clients treat as fatal.
+            if self.content is None or self.content == []:
+                self.content = ""
         elif self.role == "assistant":
             # Post-Stop sentinel: collapse content="" / [] to None.
             if (self.content == "" or self.content == []) and not self.tool_calls:
@@ -631,7 +644,8 @@ class ChatCompletionRequest(BaseModel):
         None, description = "[x-unsloth] Base64-encoded image for vision models"
     )
     audio_base64: Optional[str] = Field(
-        None, description = "[x-unsloth] Base64-encoded WAV for audio-input models (ASR)"
+        None,
+        description = "[x-unsloth] Base64-encoded audio (wav/mp3/ogg/flac/m4a) for audio-input models",
     )
     use_adapter: Optional[Union[bool, str]] = Field(
         None,
@@ -680,6 +694,16 @@ class ChatCompletionRequest(BaseModel):
         True,
         description = "[x-unsloth] Auto-detect and fix malformed tool calls from model output.",
     )
+    context_overflow: Optional[Literal["error", "truncate_middle"]] = Field(
+        None,
+        description = (
+            "[x-unsloth] Passthrough behavior when the prompt exceeds the real "
+            "context window. 'error' (default) returns a 400 with "
+            "code=context_length_exceeded. 'truncate_middle' drops middle "
+            "turn-groups (system prompt, first turn, and recent turns kept; "
+            "tool calls stay paired with their results) and retries."
+        ),
+    )
     max_tool_calls_per_message: Optional[int] = Field(
         25,
         ge = 0,
diff --git a/studio/backend/models/mcp_servers.py b/studio/backend/models/mcp_servers.py
index c696eb0faa..606c2423bf 100644
--- a/studio/backend/models/mcp_servers.py
+++ b/studio/backend/models/mcp_servers.py
@@ -44,3 +44,14 @@ class McpServerProbeResult(BaseModel):
     ok: bool
     tool_count: int = 0
     error: Optional[str] = None
+
+
+class McpServerImportRequest(BaseModel):
+    # A standard mcpServers JSON config (Claude Desktop / Cursor / Cline / VS Code).
+    config: dict
+
+
+class McpServerImportResult(BaseModel):
+    created: list[McpServerResponse] = Field(default_factory = list)
+    skipped: list[str] = Field(default_factory = list)  # display names skipped as duplicates
+    errors: list[str] = Field(default_factory = list)
diff --git a/studio/backend/routes/datasets.py b/studio/backend/routes/datasets.py
index 859f2958f9..46319ca2ba 100644
--- a/studio/backend/routes/datasets.py
+++ b/studio/backend/routes/datasets.py
@@ -412,7 +412,8 @@ def _build_local_dataset_items() -> list[LocalDatasetItem]:
 
 
 def _load_local_preview_slice(*, dataset_path: Path, train_split: str, preview_size: int):
-    from datasets import load_dataset
+    # Non-streaming loads take the cached builder lock; use the EACCES-safe wrapper.
+    from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
 
     if dataset_path.is_dir():
         parquet_dir = (
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
index abb89a13da..ef37daa158 100644
--- a/studio/backend/routes/export.py
+++ b/studio/backend/routes/export.py
@@ -78,8 +78,7 @@ async def load_checkpoint(
                 for _ in range(60):  # up to 30s
                     if not trn.is_training_active():
                         break
-                    import time
-                    time.sleep(0.5)
+                    await asyncio.sleep(0.5)
                 else:
                     logger.warning("Training subprocess did not exit within 30s, proceeding anyway")
         except Exception as e:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 9f0bb43c9a..3850d0cbb2 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -29,6 +29,16 @@ from utils.models import extract_model_size_b as _extract_model_size_b
 from utils.api_errors import openai_error_body, anthropic_error_body
 
 
+def _positive_int_or_none(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    return value_int if value_int > 0 else None
+
+
 def _install_httpcore_asyncgen_silencer() -> None:
     """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13.
 
@@ -242,6 +252,169 @@ def _openai_passthrough_error(status_code, text) -> "HTTPException":
     )
 
 
+_OVERFLOW_TRUNCATE_MAX_RETRIES = 3
+# Truncated-prompt share of the real window; the rest is generation headroom
+# so a near-full prompt cannot cut a tool call mid-JSON at the wall.
+_OVERFLOW_PROMPT_TARGET_FRACTION = 0.75
+
+
+def _overflow_truncation_requested(payload) -> bool:
+    """True when the request (or the UNSLOTH_CONTEXT_OVERFLOW server default,
+    for clients that cannot send custom fields) opted into truncation."""
+    requested = getattr(payload, "context_overflow", None)
+    if requested is not None:
+        return requested == "truncate_middle"
+    return os.environ.get("UNSLOTH_CONTEXT_OVERFLOW", "").strip().lower() == "truncate_middle"
+
+
+def _parse_overflow_counts(err_text: str):
+    """(n_prompt_tokens, n_ctx) from an exceed_context_size_error body, or
+    None. Tolerates \\" around keys (body may be a re-wrapped JSON string)."""
+    m_prompt = _re.search(r'n_prompt_tokens\\?"?\s*:\s*(\d+)', err_text)
+    m_ctx = _re.search(r'n_ctx\\?"?\s*:\s*(\d+)', err_text)
+    if m_prompt and m_ctx:
+        return int(m_prompt.group(1)), int(m_ctx.group(1))
+    return None
+
+
+def _estimate_message_tokens(msg: dict) -> int:
+    try:
+        return max(1, len(json.dumps(msg, ensure_ascii = False)) // 4)
+    except Exception:
+        return 1
+
+
+def _truncate_middle_messages(messages: list, keep_ratio: float):
+    """Drop whole turn-groups from the middle of an OpenAI message list.
+
+    Always kept: leading system message(s), the first group (task anchor),
+    and the trailing groups. A group is a user message, or an assistant
+    message plus its following tool results, so surviving tool_calls stay
+    paired with their results as chat templates require.
+    Returns (new_messages, dropped_message_count).
+    """
+    if not messages or keep_ratio >= 1.0:
+        return messages, 0
+
+    head: list = []
+    idx = 0
+    while idx < len(messages) and messages[idx].get("role") in ("system", "developer"):
+        head.append(messages[idx])
+        idx += 1
+
+    groups: list[list] = []
+    for msg in messages[idx:]:
+        role = msg.get("role")
+        if role == "tool" and groups:
+            groups[-1].append(msg)
+        elif role == "tool":
+            groups.append([msg])  # orphan tool result; treat as its own group
+        else:
+            groups.append([msg])
+
+    # Anchor group plus the last 3 groups stay.
+    protected_tail = min(3, max(1, len(groups) - 1))
+    if len(groups) <= 1 + protected_tail:
+        return messages, 0
+
+    total_est = sum(_estimate_message_tokens(m) for m in messages)
+    target_est = int(total_est * keep_ratio)
+
+    anchor = groups[0]
+    middle = groups[1:-protected_tail]
+    tail = groups[-protected_tail:]
+
+    current_est = total_est
+    kept_middle: list[list] = list(middle)
+    dropped = 0
+    # Drop oldest-first until the estimate fits the target.
+    while kept_middle and current_est > target_est:
+        victim = kept_middle.pop(0)
+        dropped += len(victim)
+        current_est -= sum(_estimate_message_tokens(m) for m in victim)
+
+    if dropped == 0:
+        return messages, 0
+
+    new_messages = head + anchor
+    for grp in kept_middle:
+        new_messages.extend(grp)
+    for grp in tail:
+        new_messages.extend(grp)
+    return new_messages, dropped
+
+
+_CLIP_MARKER = "\n[... truncated by context_overflow=truncate_middle ...]\n"
+# Generous head+tail first; cut harder if the estimate still misses the target.
+_CLIP_KEEP_CHARS = (1500, 400)
+
+
+def _clip_long_contents(messages: list, target_est: int) -> int:
+    """Clip oversized string contents middle-out until ``target_est`` is met.
+
+    Tool results first, then earlier user turns, the final message last.
+    Message count and roles never change, so tool pairing holds even when
+    group-dropping could not free enough. Returns messages clipped.
+    """
+
+    def _candidates():
+        tools = [m for m in messages if m.get("role") == "tool"]
+        users = [m for m in messages[:-1] if m.get("role") == "user"]
+        last = [messages[-1]] if messages else []
+        return tools + users + last
+
+    clipped = 0
+    for keep in _CLIP_KEEP_CHARS:
+        for msg in _candidates():
+            if sum(_estimate_message_tokens(m) for m in messages) <= target_est:
+                return clipped
+            content = msg.get("content")
+            if not isinstance(content, str) or len(content) <= 2 * keep + len(_CLIP_MARKER):
+                continue
+            msg["content"] = content[:keep] + _CLIP_MARKER + content[-keep:]
+            clipped += 1
+    return clipped
+
+
+def _apply_overflow_truncation(body: dict, err_text: str) -> bool:
+    """Shrink a passthrough body after an upstream context overflow: drop
+    middle turn-groups, clip still-oversized contents, clamp ``max_tokens``
+    to the generation headroom. Returns False when nothing could shrink."""
+    counts = _parse_overflow_counts(err_text)
+    messages = body.get("messages") or []
+    total_est = sum(_estimate_message_tokens(m) for m in messages)
+    if counts:
+        n_prompt, n_ctx = counts
+        keep_ratio = min(0.95, (_OVERFLOW_PROMPT_TARGET_FRACTION * n_ctx) / max(1, n_prompt))
+        # Scale the server-token target into char-estimate units.
+        target_est = int(total_est * keep_ratio)
+    else:
+        n_ctx = None
+        keep_ratio = 0.6  # no counts in the error; cut conservatively
+        target_est = int(total_est * keep_ratio)
+
+    new_messages, dropped = _truncate_middle_messages(messages, keep_ratio)
+    if dropped:
+        body["messages"] = new_messages
+    clipped = 0
+    if sum(_estimate_message_tokens(m) for m in body.get("messages") or []) > target_est:
+        clipped = _clip_long_contents(body.get("messages") or [], target_est)
+    if not dropped and not clipped:
+        return False
+    if n_ctx:
+        headroom = max(1024, int(n_ctx * (1.0 - _OVERFLOW_PROMPT_TARGET_FRACTION)))
+        cur_max = body.get("max_tokens")
+        body["max_tokens"] = min(cur_max, headroom) if cur_max else headroom
+    logger.warning(
+        "context_overflow=truncate_middle: dropped %d middle messages, clipped "
+        "%d contents (keep_ratio %.2f); retrying within the real window",
+        dropped,
+        clipped,
+        keep_ratio,
+    )
+    return True
+
+
 def _anthropic_stream_error_event(exc):
     """Anthropic in-band SSE ``error`` event for a mid-stream failure, or ``None``
     to fall through to a normal message_delta finish. Returns an event only for a
@@ -418,6 +591,7 @@ try:
         _DEFAULT_MAX_TOKENS_FLOOR,
         _DEFAULT_T_MAX_PREDICT_MS,
         _canonicalize_spec_mode,
+        _extra_args_set_spec_type,
         _hf_offline_if_dns_dead,
         detect_reasoning_flags,
     )
@@ -427,7 +601,10 @@ try:
     )
     from utils.models import ModelConfig
     from utils.inference import load_inference_config
-    from utils.models.model_config import load_model_defaults
+    from utils.models.model_config import (
+        detect_mtp_file,
+        load_model_defaults,
+    )
     from utils.native_path_leases import (
         NativePathLeaseError,
         display_label_for_native_path,
@@ -445,6 +622,7 @@ except ImportError:
         _DEFAULT_MAX_TOKENS_FLOOR,
         _DEFAULT_T_MAX_PREDICT_MS,
         _canonicalize_spec_mode,
+        _extra_args_set_spec_type,
         _hf_offline_if_dns_dead,
         detect_reasoning_flags,
     )
@@ -454,7 +632,10 @@ except ImportError:
     )
     from utils.models import ModelConfig
     from utils.inference import load_inference_config
-    from utils.models.model_config import load_model_defaults
+    from utils.models.model_config import (
+        detect_mtp_file,
+        load_model_defaults,
+    )
     from utils.native_path_leases import (
         NativePathLeaseError,
         display_label_for_native_path,
@@ -927,35 +1108,42 @@ def _strip_tool_xml_for_display(text: str, *, auto_heal_tool_calls: bool) -> str
 logger = get_logger(__name__)
 
 
-def _validate_native_mmproj_companion(mmproj_path: str | None, gguf_path: str | None) -> None:
-    if not mmproj_path or not gguf_path:
+def _validate_native_gguf_companion(
+    companion_path: str | None, gguf_path: str | None, label: str
+) -> None:
+    """Reject a companion GGUF (mmproj / MTP drafter) that a native-lease load
+    would otherwise hand to llama-server: must be a regular file (no symlink
+    escaping the leased directory) living next to the selected GGUF."""
+    if not companion_path or not gguf_path:
         return
     import stat as _stat_module
 
-    mm = Path(mmproj_path)
+    companion = Path(companion_path)
     gguf = Path(gguf_path)
     try:
-        mm_lstat = os.lstat(mm)
+        companion_lstat = os.lstat(companion)
     except OSError as exc:
         raise HTTPException(
             status_code = 400,
-            detail = "Native vision companion is no longer accessible.",
+            detail = f"Native {label} is no longer accessible.",
         ) from exc
-    if _stat_module.S_ISLNK(mm_lstat.st_mode) or not _stat_module.S_ISREG(mm_lstat.st_mode):
+    if _stat_module.S_ISLNK(companion_lstat.st_mode) or not _stat_module.S_ISREG(
+        companion_lstat.st_mode
+    ):
         raise HTTPException(
             status_code = 400,
-            detail = "Native vision companion must be a regular file.",
+            detail = f"Native {label} must be a regular file.",
         )
     try:
-        if mm.resolve(strict = True).parent != gguf.resolve(strict = True).parent:
+        if companion.resolve(strict = True).parent != gguf.resolve(strict = True).parent:
             raise HTTPException(
                 status_code = 400,
-                detail = "Native vision companion must live next to the selected GGUF.",
+                detail = f"Native {label} must live next to the selected GGUF.",
             )
     except OSError as exc:
         raise HTTPException(
             status_code = 400,
-            detail = "Native vision companion is no longer accessible.",
+            detail = f"Native {label} is no longer accessible.",
         ) from exc
 
 
@@ -980,13 +1168,11 @@ def _request_matches_loaded_settings(request: LoadRequest, llama_backend: LlamaC
         llama_backend.cache_type_kv
     ):
         return False
-    # Vision loads silently drop speculative decoding (llama_cpp.py gates spec
-    # on ``not is_vision``), so treat the request as ``off`` against the
-    # backend's ``None`` to avoid a redundant reload.
-    if llama_backend.is_vision:
-        req_mode = "off"
-    else:
-        req_mode = _canonicalize_spec_mode(request.speculative_type) or "auto"
+    # Spec decoding works on vision models too (MTP is mmproj-compatible,
+    # llama.cpp #22673; the old ``not is_vision`` gate is gone), so compare
+    # the real requested mode -- coercing vision to ``off`` here used to
+    # swallow every spec-mode change on a vision model as already_loaded.
+    req_mode = _canonicalize_spec_mode(request.speculative_type) or "auto"
     backend_mode = llama_backend.requested_spec_mode or "auto"
     if req_mode != backend_mode:
         return False
@@ -1008,6 +1194,33 @@ def _request_matches_loaded_settings(request: LoadRequest, llama_backend: LlamaC
     else:
         if list(request.llama_extra_args) != backend_extra:
             return False
+    # A separate drafter (Gemma's root mtp-*.gguf) appearing or disappearing
+    # next to the loaded weights changes the launch command (--model-draft),
+    # so a duplicate /load must reload rather than dedupe. Always compare the
+    # detected vs stored drafter when the mode can use one and the user does
+    # not own --spec-type: the resolved-path compare is cheap and handles all
+    # four cases (both None -> match; one None -> reload; equal -> match;
+    # different -> reload), including a drafter deleted out from under a
+    # running server. Runs last: it stats the filesystem, so every pure-memory
+    # comparison above short-circuits first. Resolve both sides since the
+    # stored launch path may be a snapshot symlink while detect_mtp_file
+    # returns the resolved blob.
+    if req_mode in ("auto", "mtp", "mtp+ngram") and llama_backend.gguf_path:
+        effective_extras = (
+            request.llama_extra_args
+            if request.llama_extra_args is not None
+            else llama_backend.extra_args
+        )
+        if not _extra_args_set_spec_type(effective_extras):
+            detected = detect_mtp_file(llama_backend.gguf_path)
+            stored = llama_backend.mtp_draft_path
+            try:
+                detected_resolved = Path(detected).resolve() if detected else None
+                stored_resolved = Path(stored).resolve() if stored else None
+            except OSError:
+                return False
+            if detected_resolved != stored_resolved:
+                return False
     return True
 
 
@@ -1187,6 +1400,7 @@ async def load_model(
                     reasoning_always_on = _sf_flags["reasoning_always_on"],
                     supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
                     supports_tools = _sf_flags["supports_tools"],
+                    context_length = _positive_int_or_none(_model_info.get("context_length")),
                     chat_template = _chat_template,
                 )
 
@@ -1313,12 +1527,26 @@ async def load_model(
                 )
             else:
                 # Local mode: llama-server loads via -m <path>
-                if native_grant_backed and config.gguf_mmproj_file:
-                    _validate_native_mmproj_companion(config.gguf_mmproj_file, config.gguf_file)
+                if native_grant_backed:
+                    if config.gguf_mmproj_file:
+                        _validate_native_gguf_companion(
+                            config.gguf_mmproj_file, config.gguf_file, "vision companion"
+                        )
+                    if config.gguf_mtp_file:
+                        # The drafter is optional (unlike mmproj for a vision
+                        # model): drop it rather than fail the load.
+                        try:
+                            _validate_native_gguf_companion(
+                                config.gguf_mtp_file, config.gguf_file, "MTP drafter"
+                            )
+                        except HTTPException as exc:
+                            logger.warning("Dropping MTP drafter for native load: %s", exc.detail)
+                            config.gguf_mtp_file = None
                 success = await asyncio.to_thread(
                     llama_backend.load_model,
                     gguf_path = config.gguf_file,
                     mmproj_path = config.gguf_mmproj_file,
+                    mtp_draft_path = config.gguf_mtp_file,
                     # Pass the resolved variant so _extra_args_source keys off
                     # the same string the inheritance check at the top of /load
                     # uses (#5401 followup).
@@ -1514,6 +1742,7 @@ async def load_model(
             reasoning_always_on = _sf_flags["reasoning_always_on"],
             supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
             supports_tools = _sf_flags["supports_tools"],
+            context_length = _positive_int_or_none(_model_info.get("context_length")),
             chat_template = _chat_template,
         )
 
@@ -1852,6 +2081,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
                 speculative_type = llama_backend.requested_spec_mode,
                 spec_draft_n_max = llama_backend.spec_draft_n_max,
                 llama_cpp_supports_mtp = _supports_mtp,
+                spec_fallback_reason = llama_backend.spec_fallback_reason,
                 llama_cpp_prebuilt_stale = _stale,
                 llama_cpp_installed_tag = _installed_tag,
                 llama_cpp_latest_tag = _latest_tag,
@@ -1901,6 +2131,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
             reasoning_always_on = _sf_flags["reasoning_always_on"],
             supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
             supports_tools = _sf_flags["supports_tools"],
+            context_length = _positive_int_or_none(model_info.get("context_length")),
             chat_template = chat_template,
             llama_cpp_supports_mtp = _supports_mtp,
             llama_cpp_prebuilt_stale = _stale,
@@ -2064,6 +2295,172 @@ def _decode_audio_base64(b64: str) -> np.ndarray:
     return waveform.squeeze(0).numpy()
 
 
+# Reject oversized audio before decoding. base64 inflates raw bytes by ~4/3, so
+# cap the encoded length to bound the upload. _MAX_AUDIO_SECONDS additionally
+# bounds the *decoded* length, since a small compressed file (opus/flac/etc.)
+# can expand to a far larger PCM array than the encoded-size cap implies.
+_MAX_AUDIO_RAW_BYTES = 25 * 1024 * 1024
+_MAX_AUDIO_B64_CHARS = _MAX_AUDIO_RAW_BYTES * 4 // 3
+_MAX_AUDIO_SECONDS = 30 * 60
+_WAV_HEADER_BYTES = 44
+_MIN_TRANSCODE_AUDIO_SAMPLE_RATE = 8000
+
+
+def _sniff_audio_container(raw: bytes) -> Optional[str]:
+    """Return 'wav' or 'mp3' if the bytes are a container llama-server accepts
+    directly (so we can forward them untouched), else None (needs transcoding)."""
+    if len(raw) >= 12 and raw[:4] == b"RIFF" and raw[8:12] == b"WAVE":
+        return "wav"
+    # mp3: ID3 tag, or an MPEG audio frame sync (no other accepted format leads
+    # with 0xFF, so the simple sync check doesn't collide).
+    if raw[:3] == b"ID3" or (len(raw) >= 2 and raw[0] == 0xFF and (raw[1] & 0xE0) == 0xE0):
+        return "mp3"
+    return None
+
+
+def _mono_f32_to_wav_bytes(arr: np.ndarray, sample_rate: int) -> bytes:
+    """Encode a mono float32 array as 16-bit PCM WAV bytes.
+
+    Torch-free (numpy + stdlib only) so it works on no-torch GGUF-only installs;
+    the shared audio_codecs helper pulls in torch at import time.
+    """
+    import io
+    import wave
+
+    arr = np.nan_to_num(np.asarray(arr, dtype = np.float32).flatten(), posinf = 0.0, neginf = 0.0)
+    if arr.size == 0:
+        raise ValueError("decoded audio is empty")
+    peak = float(np.abs(arr).max())
+    if peak > 1.0:
+        arr = arr / peak
+    pcm = (arr * 32767.0).astype(np.int16)
+
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as wf:
+        wf.setnchannels(1)
+        wf.setsampwidth(2)
+        wf.setframerate(int(sample_rate))
+        wf.writeframes(pcm.tobytes())
+    return buf.getvalue()
+
+
+def _resample_mono_linear(arr: np.ndarray, source_rate: int, target_rate: int) -> np.ndarray:
+    """Small numpy-only resampler for upload size limiting."""
+    if source_rate <= 0 or target_rate <= 0 or source_rate == target_rate:
+        return arr
+    duration = len(arr) / float(source_rate)
+    target_len = max(1, int(round(duration * target_rate)))
+    if target_len == len(arr):
+        return arr
+    source_x = np.linspace(0.0, duration, num = len(arr), endpoint = False)
+    target_x = np.linspace(0.0, duration, num = target_len, endpoint = False)
+    return np.interp(target_x, source_x, arr).astype(np.float32)
+
+
+def _fit_transcoded_audio_to_wav_cap(arr: np.ndarray, sample_rate: int) -> tuple[np.ndarray, int]:
+    """Downsample only when needed so transcoded WAV stays within the upload cap."""
+    if sample_rate <= 0:
+        raise ValueError("decoded audio has an invalid sample rate")
+    wav_bytes = _WAV_HEADER_BYTES + len(arr) * 2
+    if wav_bytes <= _MAX_AUDIO_RAW_BYTES:
+        return arr, sample_rate
+
+    duration = len(arr) / float(sample_rate)
+    max_samples = max(1, (_MAX_AUDIO_RAW_BYTES - _WAV_HEADER_BYTES) // 2)
+    target_rate = int(max_samples // duration)
+    if target_rate < _MIN_TRANSCODE_AUDIO_SAMPLE_RATE:
+        raise ValueError("decoded audio exceeds the transcoded WAV size limit")
+    target_rate = min(sample_rate, target_rate)
+    fitted = _resample_mono_linear(arr, sample_rate, target_rate)
+    if _WAV_HEADER_BYTES + len(fitted) * 2 > _MAX_AUDIO_RAW_BYTES:
+        raise ValueError("decoded audio exceeds the transcoded WAV size limit")
+    return fitted, target_rate
+
+
+def _decode_audio_mono(raw: bytes) -> tuple[np.ndarray, int]:
+    """Decode audio bytes to (mono float32 array, native sample_rate).
+
+    soundfile (libsndfile) reads wav/mp3/ogg/flac straight from memory. librosa
+    (ffmpeg-backed) additionally covers m4a/webm but needs a real path and is
+    absent on no-torch GGUF-only installs. Both imports are inside the fallback
+    so a missing decoder degrades to the next one (and finally a clear error)
+    rather than crashing.
+    """
+    import io
+
+    try:
+        import soundfile as sf
+        arr, sr = sf.read(io.BytesIO(raw), dtype = "float32")
+    except Exception:
+        try:
+            import librosa
+        except ModuleNotFoundError as e:
+            raise RuntimeError(
+                "this audio format needs librosa, which is not installed in "
+                "GGUF-only environments; use wav, mp3, ogg or flac"
+            ) from e
+        import os
+        import tempfile
+        from utils.paths import ensure_dir, tmp_root
+
+        with tempfile.NamedTemporaryFile(
+            suffix = ".audio",
+            delete = False,
+            dir = str(ensure_dir(tmp_root())),
+        ) as tmp:
+            tmp.write(raw)
+            tmp_path = tmp.name
+        try:
+            arr, sr = librosa.load(tmp_path, sr = None, mono = True)
+        finally:
+            os.unlink(tmp_path)
+    if arr.ndim > 1:
+        arr = arr.mean(axis = 1)
+    if sr > 0 and len(arr) > sr * _MAX_AUDIO_SECONDS:
+        raise ValueError(f"decoded audio exceeds the {_MAX_AUDIO_SECONDS // 60}-minute limit")
+    return arr, sr
+
+
+def _prepare_audio_for_llama(b64: str) -> tuple[str, str]:
+    """Return (base64, format) ready for llama-server's input_audio part.
+
+    llama-server's API only accepts wav/mp3, and decodes/resamples/down-mixes
+    them itself, so wav and mp3 uploads are forwarded untouched (no decode, no
+    PCM payload inflation). Other containers (m4a/ogg/webm/flac) are decoded to
+    a mono WAV. Blocking; call via a thread from async paths.
+    """
+    if b64.startswith("data:"):
+        b64 = b64.split(",", 1)[1] if "," in b64 else ""
+    raw = base64.b64decode(b64)
+    passthrough = _sniff_audio_container(raw)
+    if passthrough is not None:
+        return b64, passthrough
+
+    arr, sr = _decode_audio_mono(raw)
+    arr, sr = _fit_transcoded_audio_to_wav_cap(arr, sr)
+    return base64.b64encode(_mono_f32_to_wav_bytes(arr, sr)).decode("ascii"), "wav"
+
+
+def _inject_audio_part(messages: list[dict], audio_b64: str, audio_format: str) -> None:
+    """Append an input_audio part to the last user message, in place.
+
+    Audio rides in the message list like image_url parts do, so it flows through
+    both the plain and tool-calling generation paths.
+    """
+    part = {
+        "type": "input_audio",
+        "input_audio": {"data": audio_b64, "format": audio_format},
+    }
+    for msg in reversed(messages):
+        if msg.get("role") == "user":
+            content = msg.get("content")
+            if isinstance(content, list):
+                content.append(part)
+            else:
+                msg["content"] = [{"type": "text", "text": content or ""}, part]
+            return
+
+
 def _extract_content_parts(messages: list) -> tuple[str, list[dict], "Optional[str]"]:
     """
     Parse OpenAI-format messages into components the inference backend expects.
@@ -3021,9 +3418,12 @@ async def openai_chat_completions(
         if _wants_multiple_choices(payload):
             _raise_unsupported_n("GGUF tool or response_format passthrough")
         if payload.audio_base64:
+            # This path forwards the request verbatim, so the transcoded audio
+            # never gets injected. (The agentic tool loop below does support
+            # audio.)
             raise HTTPException(
                 status_code = 400,
-                detail = "Audio input is not supported for GGUF chat models yet.",
+                detail = "Audio input is not supported together with guided decoding or client-supplied tools yet.",
             )
 
         # Preserve the vision guard from the non-passthrough path below:
@@ -3074,11 +3474,34 @@ async def openai_chat_completions(
 
     # ── GGUF path: proxy to llama-server /v1/chat/completions ──
     if using_gguf:
+        # Forward uploaded audio as an input_audio part. wav/mp3 pass through
+        # untouched (llama-server decodes and resamples them via the mmproj
+        # audio encoder); other containers are transcoded to WAV here. The part
+        # is injected into the message list below so it rides through both the
+        # plain and tool-calling paths, exactly like image_url parts.
+        audio_b64 = None
+        audio_format = "wav"
         if payload.audio_base64:
-            raise HTTPException(
-                status_code = 400,
-                detail = "Audio input is not supported for GGUF chat models yet.",
-            )
+            if not getattr(llama_backend, "_has_audio_input", False):
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "Audio provided but current GGUF model does not support audio input.",
+                )
+            if len(payload.audio_base64) > _MAX_AUDIO_B64_CHARS:
+                raise HTTPException(
+                    status_code = 413,
+                    detail = "Audio file is too large (max ~25 MB).",
+                )
+            try:
+                audio_b64, audio_format = await asyncio.to_thread(
+                    _prepare_audio_for_llama, payload.audio_base64
+                )
+            except Exception as e:
+                logger.warning("Audio decode failed: %s", e, exc_info = True)
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "Could not decode the provided audio file.",
+                )
 
         gguf_messages, _ = _openai_messages_for_gguf_chat(
             payload,
@@ -3086,6 +3509,8 @@ async def openai_chat_completions(
         )
         gguf_messages = _set_or_prepend_system_message(gguf_messages, system_prompt)
         image_b64 = None
+        if audio_b64:
+            _inject_audio_part(gguf_messages, audio_b64, audio_format)
 
         cancel_event = threading.Event()
 
@@ -4207,26 +4632,45 @@ def _openai_model_objects() -> list[dict]:
     # Check GGUF backend
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded:
-        models.append(
-            {
-                "id": llama_backend.model_identifier,
-                "object": "model",
-                "created": _created,
-                "owned_by": "local",
-            }
-        )
+        entry = {
+            "id": llama_backend.model_identifier,
+            "object": "model",
+            "created": _created,
+            "owned_by": "local",
+        }
+        _ctx = _positive_int_or_none(getattr(llama_backend, "context_length", None))
+        if _ctx is not None:
+            entry["context_length"] = _ctx
+        _max_ctx = _positive_int_or_none(getattr(llama_backend, "max_context_length", None))
+        if _max_ctx is not None:
+            entry["max_context_length"] = _max_ctx
+        _native_ctx = _positive_int_or_none(getattr(llama_backend, "native_context_length", None))
+        if _native_ctx is not None:
+            entry["native_context_length"] = _native_ctx
+        models.append(entry)
 
     # Check Unsloth backend
     backend = get_inference_backend()
     if backend.active_model_name:
-        models.append(
-            {
-                "id": backend.active_model_name,
-                "object": "model",
-                "created": _created,
-                "owned_by": "local",
-            }
-        )
+        model_info = backend.models.get(backend.active_model_name, {})
+        entry = {
+            "id": backend.active_model_name,
+            "object": "model",
+            "created": _created,
+            "owned_by": "local",
+        }
+        _ctx = _positive_int_or_none(model_info.get("context_length"))
+        if _ctx is None:
+            for _candidate in (
+                getattr(backend, "context_length", None),
+                getattr(backend, "max_seq_length", None),
+            ):
+                _ctx = _positive_int_or_none(_candidate)
+                if _ctx is not None:
+                    break
+        if _ctx is not None:
+            entry["context_length"] = _ctx
+        models.append(entry)
 
     return models
 
@@ -4476,6 +4920,17 @@ def _responses_message_text(content: Union[str, list]) -> str:
     return "\n".join(parts)
 
 
+def _responses_tool_output_text(output: Union[str, list]) -> str:
+    """Return Chat Completions-safe content for a Responses tool result."""
+    if isinstance(output, str):
+        return output if output.strip() else "(no output)"
+
+    if output:
+        return json.dumps(output)
+
+    return "(no output)"
+
+
 def _normalise_responses_input(payload: ResponsesRequest) -> list[ChatMessage]:
     """Convert a ResponsesRequest's ``input`` into a Chat-format ``ChatMessage`` list.
 
@@ -4535,10 +4990,9 @@ def _normalise_responses_input(payload: ResponsesRequest) -> list[ChatMessage]:
 
         if isinstance(item, ResponsesFunctionCallOutputInputItem):
             # Chat Completions `role="tool"` requires string content; serialize
-            # a Responses content-array output.
-            output = item.output
-            if not isinstance(output, str):
-                output = json.dumps(output)
+            # a Responses content-array output and keep empty outputs from
+            # tripping the stricter ChatMessage role validator.
+            output = _responses_tool_output_text(item.output)
             messages.append(
                 ChatMessage(
                     role = "tool",
@@ -4783,7 +5237,9 @@ async def _responses_stream(
             detail = "Image provided but current GGUF model does not support vision.",
         )
 
-    body = _build_openai_passthrough_body(chat_req, backend_ctx = llama_backend.context_length)
+    body = _build_openai_passthrough_body(
+        chat_req, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+    )
     body["stream_options"] = {"include_usage": True}
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
 
@@ -6489,7 +6945,11 @@ def _extract_response_format(payload):
     return rf if isinstance(rf, dict) else None
 
 
-def _build_openai_passthrough_body(payload, backend_ctx = None) -> dict:
+def _build_openai_passthrough_body(
+    payload,
+    backend_ctx = None,
+    llama_backend = None,
+) -> dict:
     """Assemble the llama-server request body from a ChatCompletionRequest.
 
     Only known OpenAI / llama-server fields are forwarded, so Studio-specific
@@ -6500,12 +6960,19 @@ def _build_openai_passthrough_body(payload, backend_ctx = None) -> dict:
     system_prompt, _, _ = _extract_content_parts(payload.messages)
     messages = _set_or_prepend_system_message(messages, system_prompt)
     tool_choice = payload.tool_choice if payload.tool_choice is not None else "auto"
-    # When the caller asked for a specific reasoning mode, forward it via
-    # chat_template_kwargs so the Jinja template renders with (or without) the
-    # reasoning preamble.
-    tpl_kwargs = None
-    if payload.enable_thinking is not None:
-        tpl_kwargs = {"enable_thinking": bool(payload.enable_thinking)}
+    # Forward per-request reasoning fields (enable_thinking / reasoning_effort /
+    # preserve_thinking) via chat_template_kwargs so the Jinja template renders
+    # in the caller's mode, gated on the active template's capabilities exactly
+    # like the non-passthrough paths.
+    tpl_kwargs = (
+        llama_backend._request_reasoning_kwargs(
+            payload.enable_thinking,
+            payload.reasoning_effort,
+            payload.preserve_thinking,
+        )
+        if llama_backend is not None
+        else None
+    )
     return _build_passthrough_payload(
         messages,
         payload.tools,
@@ -6540,7 +7007,9 @@ async def _openai_passthrough_stream(
     the client sees a standard OpenAI response.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
-    body = _build_openai_passthrough_body(payload, backend_ctx = llama_backend.context_length)
+    body = _build_openai_passthrough_body(
+        payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+    )
 
     _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
     _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
@@ -6558,27 +7027,32 @@ async def _openai_passthrough_stream(
             limits = httpx.Limits(max_keepalive_connections = 0),
         )
         resp = None
-        try:
-            req = client.build_request("POST", target_url, json = body)
-            resp = await client.send(req, stream = True)
-        except httpx.RequestError as e:
-            # llama-server subprocess crashed / starting / unreachable.
-            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-            if resp is not None:
+        _truncate_budget = (
+            _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
+        )
+        while True:
+            try:
+                req = client.build_request("POST", target_url, json = body)
+                resp = await client.send(req, stream = True)
+            except httpx.RequestError as e:
+                # llama-server subprocess crashed / starting / unreachable.
+                logger.error("openai passthrough stream: upstream unreachable: %s", e)
+                if resp is not None:
+                    try:
+                        await resp.aclose()
+                    except Exception:
+                        pass
                 try:
-                    await resp.aclose()
+                    await client.aclose()
                 except Exception:
                     pass
-            try:
-                await client.aclose()
-            except Exception:
-                pass
-            raise HTTPException(
-                status_code = 502,
-                detail = _friendly_error(e),
-            )
+                raise HTTPException(
+                    status_code = 502,
+                    detail = _friendly_error(e),
+                )
 
-        if resp.status_code != 200:
+            if resp.status_code == 200:
+                break
             err_bytes = await resp.aread()
             err_text = err_bytes.decode("utf-8", errors = "replace")
             logger.error(
@@ -6591,6 +7065,14 @@ async def _openai_passthrough_stream(
                 await resp.aclose()
             except Exception:
                 pass
+            # Opt-in overflow policy: shrink and retry instead of a fatal 400.
+            if (
+                _truncate_budget > 0
+                and _classify_llama_generation_error(Exception(err_text))
+                and _apply_overflow_truncation(body, err_text)
+            ):
+                _truncate_budget -= 1
+                continue
             try:
                 await client.aclose()
             except Exception:
@@ -6685,22 +7167,37 @@ async def _openai_passthrough_non_streaming(llama_backend, payload, model_name):
     ``tool_calls``, and accurate ``usage`` token counts.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
-    body = _build_openai_passthrough_body(payload, backend_ctx = llama_backend.context_length)
+    body = _build_openai_passthrough_body(
+        payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+    )
 
-    try:
-        async with httpx.AsyncClient() as client:
-            resp = await client.post(target_url, json = body, timeout = 600)
-    except httpx.RequestError as e:
-        # llama-server subprocess crashed / starting / unreachable. Surface the
-        # same friendly message the sync chat path emits so operators don't see
-        # a bare 500 with no diagnostic.
-        logger.error("openai passthrough non-streaming: upstream unreachable: %s", e)
-        raise HTTPException(
-            status_code = 502,
-            detail = _friendly_error(e),
-        )
+    _truncate_budget = (
+        _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
+    )
+    while True:
+        try:
+            async with httpx.AsyncClient() as client:
+                resp = await client.post(target_url, json = body, timeout = 600)
+        except httpx.RequestError as e:
+            # llama-server subprocess crashed / starting / unreachable. Surface the
+            # same friendly message the sync chat path emits so operators don't see
+            # a bare 500 with no diagnostic.
+            logger.error("openai passthrough non-streaming: upstream unreachable: %s", e)
+            raise HTTPException(
+                status_code = 502,
+                detail = _friendly_error(e),
+            )
 
-    if resp.status_code != 200:
+        if resp.status_code == 200:
+            break
+        # Opt-in overflow policy: shrink and retry instead of a fatal 400.
+        if (
+            _truncate_budget > 0
+            and _classify_llama_generation_error(Exception(resp.text))
+            and _apply_overflow_truncation(body, resp.text)
+        ):
+            _truncate_budget -= 1
+            continue
         raise _openai_passthrough_error(resp.status_code, resp.text)
 
     # The guided-decoding fence wraps each choice's JSON content in a
diff --git a/studio/backend/routes/llama.py b/studio/backend/routes/llama.py
new file mode 100644
index 0000000000..3aae6f4209
--- /dev/null
+++ b/studio/backend/routes/llama.py
@@ -0,0 +1,85 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""llama.cpp prebuilt update endpoints.
+
+GET  /api/llama/update-status  -> is a newer prebuilt available + job state
+POST /api/llama/update         -> download + atomically swap to the latest
+
+Detection reuses utils.llama_cpp_freshness; the swap reuses
+install_llama_prebuilt.py via utils.llama_cpp_update. Both fail open so the UI
+never blocks on a missing marker / offline GitHub.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+from fastapi import APIRouter, Depends, Query
+from pydantic import BaseModel, Field
+
+from auth.authentication import get_current_subject
+from utils.llama_cpp_update import get_update_status, start_update
+
+router = APIRouter()
+
+
+class LlamaUpdateJob(BaseModel):
+    state: str = Field("idle", description = "idle | running | success | error")
+    message: str = ""
+    from_tag: Optional[str] = None
+    to_tag: Optional[str] = None
+    error: Optional[str] = None
+    progress: Optional[float] = Field(None, description = "0..1 while running, 1 on success.")
+    started_at: Optional[str] = None
+    finished_at: Optional[str] = None
+
+
+class LlamaUpdateStatusResponse(BaseModel):
+    supported: bool = Field(
+        False,
+        description = "True when the install came from an Unsloth prebuilt (has a marker).",
+    )
+    update_available: bool = Field(
+        False, description = "True when the latest release is genuinely newer than the install."
+    )
+    stale: bool = Field(
+        False, description = "Update available AND install older than the staleness threshold."
+    )
+    installed_tag: Optional[str] = None
+    latest_tag: Optional[str] = None
+    published_repo: Optional[str] = None
+    installed_at_utc: Optional[str] = None
+    age_days: Optional[int] = None
+    source_build: bool = Field(
+        False, description = "True when there is no marker (source build) but a prebuilt is offered."
+    )
+    job: LlamaUpdateJob = Field(default_factory = LlamaUpdateJob)
+
+
+class LlamaUpdateActionResponse(BaseModel):
+    started: bool
+    reason: Optional[str] = None
+    message: Optional[str] = None
+    job: LlamaUpdateJob = Field(default_factory = LlamaUpdateJob)
+
+
+@router.get("/update-status", response_model = LlamaUpdateStatusResponse)
+async def llama_update_status(
+    force_refresh: bool = Query(
+        False, description = "Bypass the 24h release cache for an explicit check."
+    ),
+    current_subject: str = Depends(get_current_subject),
+) -> LlamaUpdateStatusResponse:
+    # Off the event loop: detection may probe the host and read GitHub.
+    status = await asyncio.to_thread(get_update_status, force_refresh = force_refresh)
+    return LlamaUpdateStatusResponse(**status)
+
+
+@router.post("/update", response_model = LlamaUpdateActionResponse)
+async def llama_update(
+    current_subject: str = Depends(get_current_subject),
+) -> LlamaUpdateActionResponse:
+    action = await asyncio.to_thread(start_update)
+    return LlamaUpdateActionResponse(**action)
diff --git a/studio/backend/routes/mcp_servers.py b/studio/backend/routes/mcp_servers.py
index 098c64fe39..3001c6b7c9 100644
--- a/studio/backend/routes/mcp_servers.py
+++ b/studio/backend/routes/mcp_servers.py
@@ -18,8 +18,11 @@ from core.inference.mcp_client import (
     probe_timeout,
     stdio_mcp_enabled,
 )
+from core.inference.mcp_config_import import parse_mcp_config
 from models.mcp_servers import (
     McpServerCreate,
+    McpServerImportRequest,
+    McpServerImportResult,
     McpServerProbeResult,
     McpServerResponse,
     McpServerTestRequest,
@@ -240,6 +243,46 @@ async def refresh_mcp_server_tools(
     return McpServerProbeResult(ok = True, tool_count = len(tools))
 
 
+@router.post("/import", response_model = McpServerImportResult)
+async def import_mcp_servers(
+    payload: McpServerImportRequest, current_subject: str = Depends(get_current_subject)
+):
+    """Bulk-register servers from a standard mcpServers JSON config (issue
+    #5936). Each entry rides the existing create path: _validate_url applies
+    the same stdio gate (a stdio entry becomes a per-entry error when stdio is
+    off; http still imports), and entries whose url already exists are skipped
+    so re-importing the same file is idempotent. One bad entry never 400s the
+    whole batch -- failures are reported per entry."""
+    entries, errors = parse_mcp_config(payload.config)
+    created: list[McpServerResponse] = []
+    skipped: list[str] = []
+    seen_urls = {row["url"] for row in mcp_servers_db.list_servers()}
+
+    for entry in entries:
+        try:
+            url = _validate_url(entry.url)
+        except HTTPException as exc:
+            errors.append(f"{entry.display_name}: {exc.detail}")
+            continue
+        if url in seen_urls:
+            skipped.append(entry.display_name)
+            continue
+        headers = _normalize_headers(entry.headers)
+        server_id = uuid.uuid4().hex[:16]
+        mcp_servers_db.create_server(
+            id = server_id,
+            display_name = entry.display_name,
+            url = url,
+            headers_json = json.dumps(headers) if headers else None,
+            is_enabled = entry.is_enabled,
+            use_oauth = entry.use_oauth and not is_stdio(url),
+        )
+        seen_urls.add(url)
+        created.append(_row_to_response(mcp_servers_db.get_server(server_id)))
+
+    return McpServerImportResult(created = created, skipped = skipped, errors = errors)
+
+
 @router.post("/test", response_model = McpServerProbeResult)
 async def test_mcp_server(
     payload: McpServerTestRequest, current_subject: str = Depends(get_current_subject)
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 96403e21a5..d7687ffdee 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -665,10 +665,17 @@ async def stream_training_progress(
 
             # If not active, send final state and exit
             if not is_active:
-                if backend.step_history:
-                    final_step = backend.step_history[-1]
+                _live = (getattr(tp, "step", 0) or 0) if tp else 0
+                if backend.step_history or _live > 0:
+                    final_step = backend.step_history[-1] if backend.step_history else 0
                     final_loss = backend.loss_history[-1] if backend.loss_history else None
                     final_lr = backend.lr_history[-1] if backend.lr_history else None
+                    # Histories skip non-finite steps; report the live step with
+                    # loss=None instead of the last finite pair.
+                    if _live > final_step:
+                        final_step = _live
+                        final_loss = getattr(tp, "loss", None)
+                        final_lr = getattr(tp, "learning_rate", final_lr)
                     final_total_steps = getattr(tp, "total_steps", final_step) if tp else final_step
                     final_epoch = getattr(tp, "epoch", None) if tp else None
                     payload = build_progress(
@@ -697,11 +704,18 @@ async def stream_training_progress(
 
         while backend.is_training_active():
             try:
-                if backend.step_history:
-                    current_step = backend.step_history[-1]
+                tp_inner = getattr(getattr(backend, "trainer", None), "training_progress", None)
+                live_step = (getattr(tp_inner, "step", 0) or 0) if tp_inner else 0
+                if backend.step_history or live_step > 0:
+                    current_step = backend.step_history[-1] if backend.step_history else 0
                     current_loss = backend.loss_history[-1] if backend.loss_history else None
                     current_lr = backend.lr_history[-1] if backend.lr_history else None
-                    tp_inner = getattr(getattr(backend, "trainer", None), "training_progress", None)
+                    # Histories skip non-finite steps; follow the live progress
+                    # step and report its loss (None until it recovers).
+                    if live_step > current_step:
+                        current_step = live_step
+                        current_loss = getattr(tp_inner, "loss", None)
+                        current_lr = getattr(tp_inner, "learning_rate", current_lr)
                     current_total_steps = (
                         getattr(tp_inner, "total_steps", current_step) if tp_inner else current_step
                     )
@@ -798,6 +812,13 @@ async def stream_training_progress(
         final_loss = backend.loss_history[-1] if backend.loss_history else None
         final_lr = backend.lr_history[-1] if backend.lr_history else None
         final_tp = getattr(getattr(backend, "trainer", None), "training_progress", None)
+        # If the run ended on a non-finite stretch, report the live step with
+        # loss=None instead of rolling back to the last finite pair.
+        _final_live_step = (getattr(final_tp, "step", 0) or 0) if final_tp else 0
+        if _final_live_step > (final_step if final_step is not None else -1):
+            final_step = _final_live_step
+            final_loss = getattr(final_tp, "loss", None)
+            final_lr = getattr(final_tp, "learning_rate", final_lr)
         final_total_steps = getattr(final_tp, "total_steps", final_step) if final_tp else final_step
         final_epoch = getattr(final_tp, "epoch", None) if final_tp else None
         final_payload = build_progress(
diff --git a/studio/backend/run.py b/studio/backend/run.py
index c992a4bec2..a883154c4a 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -8,6 +8,7 @@ Self-contained; can be moved to any directory.
 
 import os
 import sys
+import time
 from pathlib import Path
 from typing import Optional
 
@@ -231,6 +232,9 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
     public internet. Synchronous so output lands between the banner URLs and the
     stop hint. Bounded at ~15s; failures swallowed (verifier failing != Studio
     failing). Only meaningful for a wildcard bind."""
+    global _public_reachable
+    # Reset to "unknown" each run; set True/False only when the probe decides.
+    _public_reachable = None
     import ipaddress
     import json
     import time
@@ -323,12 +327,14 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
 
         print("", flush = True)
         if ok_nodes:
+            _public_reachable = True
             print(
                 f"{ok_c}  Reachability check: {url}/ is reachable from the "
                 f"public internet ({ok_nodes}/{total} probe nodes connected).{reset}",
                 flush = True,
             )
         elif err_nodes:
+            _public_reachable = False
             print(
                 f"{err_c}  Reachability check: {url}/ is NOT reachable from "
                 f"the public internet ({err_nodes}/{total} probe nodes failed).{reset}",
@@ -413,9 +419,31 @@ def _emit_startup_output(host: str, port: int, display_host: str) -> None:
         print_studio_stop_hint()
     elif wildcard_bind:
         _verify_global_reachability(display_host, port)
+        _print_cloudflare_line()
         print_studio_stop_hint()
 
 
+def _print_cloudflare_line() -> None:
+    """Print the Cloudflare quick-tunnel URL for 0.0.0.0 binds, if one is up.
+
+    Reads the module-level URL set by ``run_server``. Prints nothing when the
+    tunnel is disabled or failed -- failures are silently ignored. When the public
+    reachability probe just failed (``_public_reachable is False``) but the tunnel
+    is up, reword to point the user at the Cloudflare link as the way in.
+    """
+    if not _cloudflare_url:
+        return
+    from startup_banner import stdout_supports_color
+
+    accent = "\033[38;5;150;1m"
+    reset = "\033[0m"
+    if _public_reachable is False:
+        line = f"  Use the secure link access via Cloudflare instead: {_cloudflare_url}"
+    else:
+        line = f"  Secure link access via Cloudflare: {_cloudflare_url}"
+    print(f"{accent}{line}{reset}" if stdout_supports_color() else line)
+
+
 def _get_pid_on_port(port: int) -> "tuple[int, str] | None":
     """Return (pid, process_name) listening on *port*, or None.
 
@@ -583,6 +611,13 @@ def _graceful_shutdown(server = None):
     except Exception as e:
         logger.warning("Error shutting down llama-server: %s", e)
 
+    # 6. Stop the Cloudflare tunnel (if started).
+    try:
+        from cloudflare_tunnel import stop_studio_tunnel
+        stop_studio_tunnel()
+    except Exception as e:
+        logger.warning("Error stopping Cloudflare tunnel: %s", e)
+
     logger.info("All subprocesses cleaned up")
 
 
@@ -593,6 +628,16 @@ _server = None
 # Shutdown event -- wakes the main loop on signal.
 _shutdown_event = None
 
+# trycloudflare.com URL for 0.0.0.0 binds (set by run_server, read by the banner);
+# None when there is no tunnel (loopback, disabled, or a silently-ignored failure).
+_cloudflare_url = None
+
+# Public reachability from the last _verify_global_reachability run, read by the
+# Cloudflare banner line. True when the public ip:port probe confirmed reachable,
+# False when it confirmed NOT reachable, None when the probe did not run or could
+# not decide (timeout, blocked, private address).
+_public_reachable = None
+
 
 _DEFAULT_FRONTEND_PATH = Path(__file__).resolve().parent.parent / "frontend" / "dist"
 
@@ -673,6 +718,94 @@ def _resolve_frontend_path(frontend_path: Path) -> tuple[Optional[Path], list[Pa
     return None, attempted
 
 
+class _TeeStream:
+    """Mirror writes to the original stream and a session log file.
+
+    Console behavior is unchanged (writes/returns delegate to the original
+    stream; Tauri's structured-stdout protocol and isatty probes see exactly
+    what they saw before). The file copy is best-effort: a full disk or a
+    closed handle must never break the console."""
+
+    def __init__(self, stream, log_fh):
+        self._stream = stream
+        self._log_fh = log_fh
+
+    def write(self, data):
+        try:
+            self._log_fh.write(data)
+        except Exception:
+            pass
+        return self._stream.write(data)
+
+    def flush(self):
+        try:
+            self._log_fh.flush()
+        except Exception:
+            pass
+        try:
+            self._stream.flush()
+        except Exception:
+            pass
+
+    def __getattr__(self, name):
+        return getattr(self._stream, name)
+
+
+def _setup_server_disk_logging():
+    """Tee stdout/stderr to ~/.unsloth/studio/logs/server/ and aim
+    faulthandler at the same file so hard crashes (access violations /
+    SIGSEGV in the GPU runtime) leave a stack trace on disk.
+
+    Also exports PYTHONFAULTHANDLER=1 so child Python processes (training
+    workers) dump native-crash stacks to their captured stderr. Keeps the
+    newest 20 session logs. Opt out with UNSLOTH_STUDIO_NO_FILE_LOG=1.
+    Returns the log path, or None when disabled/unavailable.
+    """
+    if os.environ.get("UNSLOTH_STUDIO_NO_FILE_LOG") == "1":
+        return None
+    try:
+        from utils.paths import studio_root
+        log_dir = Path(studio_root()) / "logs" / "server"
+    except Exception:
+        home = (
+            os.environ.get("UNSLOTH_STUDIO_HOME")
+            or os.environ.get("STUDIO_HOME")
+            or os.path.join(os.path.expanduser("~"), ".unsloth", "studio")
+        )
+        log_dir = Path(home) / "logs" / "server"
+    try:
+        log_dir.mkdir(parents = True, exist_ok = True)
+        stamp = time.strftime("%Y%m%d-%H%M%S")
+        log_path = log_dir / f"server-{stamp}-pid{os.getpid()}.log"
+        # Line-buffered so the tail survives a hard kill; errors="replace"
+        # so a console encoding quirk can never take the server down.
+        log_fh = open(log_path, "w", encoding = "utf-8", errors = "replace", buffering = 1)
+    except Exception:
+        return None
+
+    import faulthandler
+
+    try:
+        faulthandler.enable(file = log_fh, all_threads = True)
+    except Exception:
+        pass
+    # Children (training workers) inherit: their native-crash stacks land on
+    # the stderr the server already captures.
+    os.environ.setdefault("PYTHONFAULTHANDLER", "1")
+
+    sys.stdout = _TeeStream(sys.stdout, log_fh)
+    sys.stderr = _TeeStream(sys.stderr, log_fh)
+
+    # Best-effort retention: keep the newest 20 session logs.
+    try:
+        logs = sorted(log_dir.glob("server-*.log"), key = lambda p: p.stat().st_mtime)
+        for old in logs[:-20]:
+            old.unlink(missing_ok = True)
+    except Exception:
+        pass
+    return log_path
+
+
 def run_server(
     host: str = "127.0.0.1",
     port: int = 8888,
@@ -680,6 +813,7 @@ def run_server(
     silent: bool = False,
     api_only: bool = False,
     llama_parallel_slots: int = 1,
+    cloudflare: bool = True,
 ):
     """
     Start the FastAPI server.
@@ -705,6 +839,16 @@ def run_server(
         except Exception:
             pass
 
+    # Persist a session log + native-crash stacks BEFORE importing main, so
+    # even import-time failures leave evidence on disk. Field report: Studio
+    # "terminates without a warning" -- a native crash in the GPU runtime
+    # kills the process with no Python traceback, and a desktop-shortcut
+    # console closes before anything can be read. Console-only logging made
+    # that undiagnosable.
+    _session_log = _setup_server_disk_logging()
+    if _session_log is not None and not silent:
+        print(f"Session log: {_session_log}")
+
     # Set env var BEFORE importing main so CORS middleware picks it up.
     if api_only:
         os.environ["UNSLOTH_API_ONLY"] = "1"
@@ -874,6 +1018,21 @@ def run_server(
     if api_only:
         print(f"TAURI_PORT={port}", flush = True)
 
+    # Free trycloudflare.com tunnel for 0.0.0.0 binds (the raw ip:port is often
+    # unreachable). Started pre-banner and even when silent so the CLI banner can
+    # read app.state.cloudflare_url; torn down by _graceful_shutdown.
+    global _cloudflare_url
+    _cloudflare_url = None
+    app.state.cloudflare_url = None
+    _cloudflare_enabled = cloudflare and host == "0.0.0.0" and not api_only and not _IS_COLAB
+    if _cloudflare_enabled:
+        try:  # best-effort: any failure must not block startup
+            from cloudflare_tunnel import start_studio_tunnel
+            _cloudflare_url = start_studio_tunnel(port)
+            app.state.cloudflare_url = _cloudflare_url
+        except Exception as e:
+            logger.debug("Cloudflare tunnel skipped: %s", e)
+
     if not silent:
         _emit_startup_output(host, port, display_host)
 
@@ -912,6 +1071,13 @@ if __name__ == "__main__":
         action = "store_true",
         help = "API server only, no frontend (for Tauri)",
     )
+    parser.add_argument(
+        "--cloudflare",
+        action = argparse.BooleanOptionalAction,
+        default = True,
+        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 "
+        "(default on; --no-cloudflare to disable)",
+    )
     # Mirror unsloth_cli/commands/studio.py's _PARALLEL_*. Default 1 is for direct
     # backend launches; `unsloth studio run` always passes its own value (4).
     _PARALLEL_MIN = 1
@@ -938,6 +1104,7 @@ if __name__ == "__main__":
         silent = args.silent,
         api_only = args.api_only,
         llama_parallel_slots = args.parallel,
+        cloudflare = args.cloudflare,
     )
     if args.frontend is not None:
         kwargs["frontend_path"] = Path(args.frontend)
diff --git a/studio/backend/tests/test_cloudflare_tunnel.py b/studio/backend/tests/test_cloudflare_tunnel.py
new file mode 100644
index 0000000000..873547631d
--- /dev/null
+++ b/studio/backend/tests/test_cloudflare_tunnel.py
@@ -0,0 +1,479 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the Cloudflare quick-tunnel helper and run.py wiring.
+
+cloudflare_tunnel.py is stdlib-only (storage_roots is imported lazily), so it
+loads via spec_from_file_location without the studio venv. run.py defaults are
+checked by AST so we never import its heavy deps (uvicorn/structlog).
+"""
+
+import ast
+import importlib.util
+import io
+import sys
+import tarfile
+import types
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+_CT_PY = _BACKEND / "cloudflare_tunnel.py"
+_RUN_PY = _BACKEND / "run.py"
+
+
+def _load_ct():
+    spec = importlib.util.spec_from_file_location("cloudflare_tunnel", _CT_PY)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+ct = _load_ct()
+
+
+# ── URL parsing ──────────────────────────────────────────────────────
+
+
+def test_url_regex_extracts_and_ignores_noise():
+    blob = (
+        "2026-06-11T10:00:00Z INF Thank you for trying Cloudflare Tunnel.\n"
+        "2026-06-11T10:00:01Z INF Requesting new quick Tunnel on trycloudflare.com...\n"
+        "2026-06-11T10:00:01Z INF |  https://setting-democracy-gathering.trycloudflare.com  |\n"
+        "2026-06-11T10:00:02Z INF Registered tunnel connection https://not-the-url.example.com\n"
+    )
+    m = ct._URL_RE.search(blob)
+    assert m is not None
+    assert m.group(0) == "https://setting-democracy-gathering.trycloudflare.com"
+
+
+def test_url_regex_no_match_on_unrelated():
+    assert ct._URL_RE.search("INF connecting to https://api.cloudflare.com/v4") is None
+
+
+# ── asset mapping ────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "system,machine,expected",
+    [
+        ("Linux", "x86_64", ("cloudflared-linux-amd64", False)),
+        ("Linux", "aarch64", ("cloudflared-linux-arm64", False)),
+        ("Darwin", "arm64", ("cloudflared-darwin-arm64.tgz", True)),
+        ("Darwin", "x86_64", ("cloudflared-darwin-amd64.tgz", True)),
+        ("Windows", "AMD64", ("cloudflared-windows-amd64.exe", False)),
+        ("Windows", "x86", ("cloudflared-windows-386.exe", False)),
+        ("Linux", "mips", None),
+        ("Plan9", "x86_64", None),
+    ],
+)
+def test_asset_name(monkeypatch, system, machine, expected):
+    monkeypatch.setattr(ct.platform, "system", lambda: system)
+    monkeypatch.setattr(ct.platform, "machine", lambda: machine)
+    assert ct._asset_name() == expected
+
+
+# ── binary discovery ─────────────────────────────────────────────────
+
+
+def test_find_cloudflared_prefers_path(monkeypatch):
+    monkeypatch.setattr(ct.shutil, "which", lambda name: "/usr/local/bin/cloudflared")
+    assert ct.find_cloudflared() == "/usr/local/bin/cloudflared"
+
+
+def test_find_cloudflared_falls_back_to_cache(monkeypatch, tmp_path):
+    cached = tmp_path / "cloudflared"
+    cached.write_text("#!/bin/sh\n")
+    cached.chmod(0o755)
+    monkeypatch.setattr(ct.shutil, "which", lambda name: None)
+    monkeypatch.setattr(ct, "_cache_path", lambda: cached)
+    assert ct.find_cloudflared() == str(cached)
+
+
+def test_find_cloudflared_none_when_missing(monkeypatch, tmp_path):
+    monkeypatch.setattr(ct.shutil, "which", lambda name: None)
+    monkeypatch.setattr(ct, "_cache_path", lambda: tmp_path / "absent")
+    assert ct.find_cloudflared() is None
+
+
+# ── ensure / download ────────────────────────────────────────────────
+
+
+def test_ensure_downloads_and_chmods_when_missing(monkeypatch, tmp_path):
+    cached = tmp_path / "cloudflared"
+    monkeypatch.setattr(ct, "find_cloudflared", lambda: None)
+    monkeypatch.setattr(ct, "_asset_name", lambda: ("cloudflared-linux-amd64", False))
+    monkeypatch.setattr(ct, "_cache_path", lambda: cached)
+
+    def fake_download(url, dest):
+        assert url.endswith("/cloudflared-linux-amd64")
+        dest.write_bytes(b"ELF-ish")
+        return True
+
+    monkeypatch.setattr(ct, "_download", fake_download)
+    monkeypatch.setattr(ct.sys, "platform", "linux")
+    path = ct.ensure_cloudflared()
+    assert path == str(cached)
+    assert cached.exists()
+    assert cached.stat().st_mode & 0o111  # executable bit set
+
+
+def test_ensure_returns_none_on_download_failure(monkeypatch, tmp_path):
+    monkeypatch.setattr(ct, "find_cloudflared", lambda: None)
+    monkeypatch.setattr(ct, "_asset_name", lambda: ("cloudflared-linux-amd64", False))
+    monkeypatch.setattr(ct, "_cache_path", lambda: tmp_path / "cloudflared")
+    monkeypatch.setattr(ct, "_download", lambda url, dest: False)
+    assert ct.ensure_cloudflared() is None
+
+
+def test_ensure_returns_none_for_unsupported_arch(monkeypatch, tmp_path):
+    monkeypatch.setattr(ct, "find_cloudflared", lambda: None)
+    monkeypatch.setattr(ct, "_asset_name", lambda: None)
+    monkeypatch.setattr(ct, "_cache_path", lambda: tmp_path / "cloudflared")
+    assert ct.ensure_cloudflared() is None
+
+
+def test_download_sets_user_agent(monkeypatch, tmp_path):
+    import urllib.request
+
+    captured = {}
+
+    class _Resp:
+        _sent = False
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+        def read(self, n = -1):
+            if self._sent:
+                return b""
+            self._sent = True
+            return b"payload"
+
+    def fake_urlopen(req, timeout = None):
+        captured["ua"] = req.get_header("User-agent")
+        return _Resp()
+
+    monkeypatch.setattr(urllib.request, "urlopen", fake_urlopen)
+    dest = tmp_path / "cloudflared"
+    assert ct._download("https://github.com/cloudflare/cloudflared/x", dest) is True
+    assert captured["ua"] == "unsloth-studio"  # GitHub CDN 403s the default UA
+    assert dest.read_bytes() == b"payload"
+
+
+# ── cross-platform: Windows (.exe), macOS (.tgz) ─────────────────────
+
+
+def test_cache_path_uses_exe_on_windows(monkeypatch, tmp_path):
+    import types
+
+    fake_sr = types.ModuleType("utils.paths.storage_roots")
+    fake_sr.studio_bin_root = lambda: tmp_path
+    monkeypatch.setitem(sys.modules, "utils.paths.storage_roots", fake_sr)
+    monkeypatch.setattr(ct.sys, "platform", "win32")
+    assert ct._cache_path() == tmp_path / "cloudflared.exe"
+
+
+def test_ensure_windows_downloads_exe(monkeypatch, tmp_path):
+    cached = tmp_path / "cloudflared.exe"
+    monkeypatch.setattr(ct, "find_cloudflared", lambda: None)
+    monkeypatch.setattr(ct, "_asset_name", lambda: ("cloudflared-windows-amd64.exe", False))
+    monkeypatch.setattr(ct, "_cache_path", lambda: cached)
+    monkeypatch.setattr(ct.sys, "platform", "win32")
+
+    def fake_download(url, dest):
+        assert url.endswith("/cloudflared-windows-amd64.exe")
+        dest.write_bytes(b"MZ")  # PE header magic
+        return True
+
+    monkeypatch.setattr(ct, "_download", fake_download)
+    # chmod is skipped on Windows; would raise on a path that does not exist yet.
+    monkeypatch.setattr(ct.os, "chmod", lambda *a, **k: pytest.fail("chmod called on win32"))
+    assert ct.ensure_cloudflared() == str(cached)
+    assert cached.read_bytes() == b"MZ"
+
+
+def test_ensure_macos_extracts_tgz_and_chmods(monkeypatch, tmp_path):
+    cached = tmp_path / "cloudflared"
+    monkeypatch.setattr(ct, "find_cloudflared", lambda: None)
+    monkeypatch.setattr(ct, "_asset_name", lambda: ("cloudflared-darwin-arm64.tgz", True))
+    monkeypatch.setattr(ct, "_cache_path", lambda: cached)
+    monkeypatch.setattr(ct.sys, "platform", "darwin")
+
+    def fake_download(url, dest):
+        # dest is cached.with_suffix(".tgz"); write a real archive there.
+        assert url.endswith("/cloudflared-darwin-arm64.tgz")
+        with tarfile.open(dest, "w:gz") as tar:
+            data = b"mach-o"
+            info = tarfile.TarInfo(name = "cloudflared")
+            info.size = len(data)
+            tar.addfile(info, io.BytesIO(data))
+        return True
+
+    monkeypatch.setattr(ct, "_download", fake_download)
+    path = ct.ensure_cloudflared()
+    assert path == str(cached)
+    assert cached.read_bytes() == b"mach-o"
+    assert cached.stat().st_mode & 0o111  # chmod applied on posix
+    assert not cached.with_suffix(".tgz").exists()  # temp archive cleaned up
+
+
+# ── .tgz extraction (darwin) ─────────────────────────────────────────
+
+
+def _make_tgz(
+    tmp_path,
+    member_name,
+    data = b"bin",
+):
+    tgz = tmp_path / "cf.tgz"
+    with tarfile.open(tgz, "w:gz") as tar:
+        info = tarfile.TarInfo(name = member_name)
+        info.size = len(data)
+        tar.addfile(info, io.BytesIO(data))
+    return tgz
+
+
+def test_tgz_extraction_extracts_clean_member(tmp_path):
+    tgz = _make_tgz(tmp_path, "cloudflared")
+    dest = tmp_path / "out"
+    assert ct._extract_tgz_member(tgz, dest) is True
+    assert dest.read_bytes() == b"bin"
+
+
+def test_tgz_extraction_rejects_traversal(tmp_path):
+    tgz = _make_tgz(tmp_path, "../cloudflared")
+    dest = tmp_path / "out"
+    assert ct._extract_tgz_member(tgz, dest) is False
+    assert not dest.exists()
+
+
+def test_tgz_extraction_missing_member(tmp_path):
+    tgz = _make_tgz(tmp_path, "README")
+    dest = tmp_path / "out"
+    assert ct._extract_tgz_member(tgz, dest) is False
+
+
+# ── tunnel lifecycle ─────────────────────────────────────────────────
+
+
+class _FakePopen:
+    def __init__(self):
+        self.terminated = False
+        self.killed = False
+        self._alive = True
+
+    def poll(self):
+        return None if self._alive else 0
+
+    def terminate(self):
+        self.terminated = True
+        self._alive = False
+
+    def wait(self, timeout = None):
+        if self._alive:
+            raise ct.subprocess.TimeoutExpired(cmd = "cloudflared", timeout = timeout)
+        return 0
+
+    def kill(self):
+        self.killed = True
+        self._alive = False
+
+
+def test_stop_terminates_process():
+    t = ct.CloudflareTunnel(8080, "/bin/cloudflared")
+    fake = _FakePopen()
+    t._proc = fake
+    t.stop()
+    assert fake.terminated is True
+    assert t._proc is None
+    # second stop is a no-op (idempotent)
+    t.stop()
+
+
+def test_wait_for_url_times_out_without_blocking():
+    t = ct.CloudflareTunnel(8080, "/bin/cloudflared")
+    assert t.wait_for_url(timeout = 0.05) is None
+
+
+def test_start_studio_tunnel_no_binary(monkeypatch):
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: None)
+    assert ct.start_studio_tunnel(8080) is None
+
+
+def test_start_studio_tunnel_registers_before_wait(monkeypatch):
+    # The tunnel must be visible to stop_studio_tunnel() during the URL wait,
+    # else a shutdown in that window orphans cloudflared.
+    seen = {}
+
+    class _Stub:
+        def __init__(self, port, binary):
+            self.url = None
+
+        def start(self):
+            pass
+
+        def wait_for_url(self, timeout):
+            seen["active_during_wait"] = ct._active_tunnel is self
+            self.url = "https://x.trycloudflare.com"
+            return self.url
+
+        def stop(self):
+            seen["stopped"] = True
+
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: "/bin/cloudflared")
+    monkeypatch.setattr(ct, "CloudflareTunnel", _Stub)
+    try:
+        assert ct.start_studio_tunnel(8080) == "https://x.trycloudflare.com"
+        assert seen["active_during_wait"] is True
+    finally:
+        ct.stop_studio_tunnel()
+
+
+def test_start_studio_tunnel_clears_and_stops_on_no_url(monkeypatch):
+    seen = {}
+
+    class _Stub:
+        def __init__(self, port, binary):
+            self.url = None
+
+        def start(self):
+            pass
+
+        def wait_for_url(self, timeout):
+            return None
+
+        def stop(self):
+            seen["stopped"] = True
+
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: "/bin/cloudflared")
+    monkeypatch.setattr(ct, "CloudflareTunnel", _Stub)
+    assert ct.start_studio_tunnel(8080) is None
+    assert seen.get("stopped") is True
+    assert ct._active_tunnel is None
+
+
+def test_start_studio_tunnel_returns_url(monkeypatch):
+    class _StubTunnel:
+        def __init__(self, port, binary):
+            self.url = None
+
+        def start(self):
+            self.url = "https://stub-xyz.trycloudflare.com"
+
+        def wait_for_url(self, timeout):
+            return self.url
+
+        def stop(self):
+            pass
+
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: "/bin/cloudflared")
+    monkeypatch.setattr(ct, "CloudflareTunnel", _StubTunnel)
+    try:
+        assert ct.start_studio_tunnel(8080) == "https://stub-xyz.trycloudflare.com"
+    finally:
+        ct.stop_studio_tunnel()
+
+
+# ── run.py source-level pins (AST / source, no heavy import) ─────────
+
+
+def _func_param_defaults(source, func_name):
+    tree = ast.parse(source)
+    for node in ast.walk(tree):
+        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)) and node.name == func_name:
+            args = node.args.args
+            defaults = node.args.defaults
+            offset = len(args) - len(defaults)
+            out = {}
+            for i, d in enumerate(defaults):
+                if isinstance(d, ast.Constant):
+                    out[args[offset + i].arg] = d.value
+            return out
+    return {}
+
+
+def _argparse_default(source, option):
+    tree = ast.parse(source)
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute):
+            if node.func.attr == "add_argument" and node.args:
+                a0 = node.args[0]
+                if isinstance(a0, ast.Constant) and a0.value == option:
+                    for kw in node.keywords:
+                        if kw.arg == "default" and isinstance(kw.value, ast.Constant):
+                            return kw.value.value
+    return None
+
+
+def test_run_server_cloudflare_default_true():
+    defaults = _func_param_defaults(_RUN_PY.read_text(), "run_server")
+    assert defaults.get("cloudflare") is True
+
+
+def test_argparse_cloudflare_default_true():
+    assert _argparse_default(_RUN_PY.read_text(), "--cloudflare") is True
+
+
+def test_run_server_gates_tunnel_on_wildcard():
+    # Guard against accidentally widening the trigger beyond 0.0.0.0.
+    source = _RUN_PY.read_text()
+    assert "_cloudflare_enabled" in source
+    assert 'host == "0.0.0.0"' in source
+
+
+def _run_print_cloudflare_line(monkeypatch, *, cloudflare_url, public_reachable):
+    """Exec the real _print_cloudflare_line source in isolation (run.py has heavy
+    deps), with the two module globals injected and startup_banner stubbed."""
+    src = _RUN_PY.read_text()
+    tree = ast.parse(src)
+    func_src = next(
+        ast.get_source_segment(src, n)
+        for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef) and n.name == "_print_cloudflare_line"
+    )
+    stub = types.ModuleType("startup_banner")
+    stub.stdout_supports_color = lambda: False
+    monkeypatch.setitem(sys.modules, "startup_banner", stub)
+    captured: list[str] = []
+    ns = {
+        "_cloudflare_url": cloudflare_url,
+        "_public_reachable": public_reachable,
+        "print": lambda *a, **k: captured.append(" ".join(str(x) for x in a)),
+    }
+    exec(compile(func_src, "<print_cloudflare_line>", "exec"), ns)
+    ns["_print_cloudflare_line"]()
+    return "\n".join(captured)
+
+
+def test_cloudflare_line_reworded_when_public_unreachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = False
+    )
+    assert "Use the secure link access via Cloudflare instead: https://x.trycloudflare.com" in out
+
+
+def test_cloudflare_line_default_wording_when_reachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = True
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Use the secure link" not in out
+
+
+def test_cloudflare_line_default_wording_when_unknown(monkeypatch):
+    # Probe did not run / could not decide -> keep the existing wording.
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = None
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Use the secure link" not in out
+
+
+def test_cloudflare_line_prints_nothing_without_tunnel(monkeypatch):
+    out = _run_print_cloudflare_line(monkeypatch, cloudflare_url = None, public_reachable = False)
+    assert out == ""
diff --git a/studio/backend/tests/test_context_overflow_truncation.py b/studio/backend/tests/test_context_overflow_truncation.py
new file mode 100644
index 0000000000..4f4c240934
--- /dev/null
+++ b/studio/backend/tests/test_context_overflow_truncation.py
@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the opt-in ``context_overflow="truncate_middle"`` passthrough policy.
+
+On ``exceed_context_size_error`` the passthrough drops middle turn-groups and
+retries inside the real window instead of surfacing a fatal 400. Truncation
+keeps the system prompt, the first turn, and recent turns, and never orphans
+a tool result from its tool_calls turn. Also covers ``/v1/models`` exposing
+the real post-readback context window.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from routes.inference import (
+    _apply_overflow_truncation,
+    _clip_long_contents,
+    _CLIP_MARKER,
+    _estimate_message_tokens,
+    _openai_model_objects,
+    _overflow_truncation_requested,
+    _parse_overflow_counts,
+    _truncate_middle_messages,
+)
+import routes.inference as routes_mod
+
+
+# Nick's actual error body from the Discord report logs.
+_NICK_ERROR = (
+    '{"detail":"llama-server error: {\\"error\\":{\\"code\\":400,'
+    '\\"message\\":\\"request (70494 tokens) exceeds the available context size '
+    '(67584 tokens), try increasing it\\",\\"type\\":\\"exceed_context_size_error\\",'
+    '\\"n_prompt_tokens\\":70494,\\"n_ctx\\":67584}}"}'
+)
+
+
+def _tool_turn(i: int, result_chars: int = 400) -> list[dict]:
+    """An assistant tool_calls turn paired with its tool result."""
+    return [
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "id": f"call_{i}",
+                    "type": "function",
+                    "function": {"name": "read", "arguments": f'{{"filePath":"/f{i}"}}'},
+                }
+            ],
+        },
+        {"role": "tool", "tool_call_id": f"call_{i}", "content": "x" * result_chars},
+    ]
+
+
+def _conversation(n_tool_turns: int = 12) -> list[dict]:
+    msgs = [
+        {"role": "system", "content": "You are an agent." * 20},
+        {"role": "user", "content": "Do the big task." * 20},
+    ]
+    for i in range(n_tool_turns):
+        msgs.extend(_tool_turn(i))
+    msgs.append({"role": "assistant", "content": "halfway summary"})
+    msgs.append({"role": "user", "content": "keep going"})
+    return msgs
+
+
+# ---------------------------------------------------------------------------
+# _parse_overflow_counts
+# ---------------------------------------------------------------------------
+
+
+def test_parse_overflow_counts_nick_error():
+    assert _parse_overflow_counts(_NICK_ERROR) == (70494, 67584)
+
+
+def test_parse_overflow_counts_missing_fields():
+    assert _parse_overflow_counts('{"error":"something else"}') is None
+
+
+# ---------------------------------------------------------------------------
+# _truncate_middle_messages
+# ---------------------------------------------------------------------------
+
+
+def test_truncation_drops_middle_keeps_anchors():
+    msgs = _conversation()
+    new, dropped = _truncate_middle_messages(msgs, keep_ratio = 0.5)
+    assert dropped > 0
+    assert len(new) == len(msgs) - dropped
+    # System prompt and task anchor survive.
+    assert new[0]["role"] == "system"
+    assert new[1] == msgs[1]
+    # The most recent turns survive verbatim.
+    assert new[-1] == msgs[-1]
+    assert new[-2] == msgs[-2]
+
+
+def test_truncation_never_orphans_tool_results():
+    msgs = _conversation()
+    new, dropped = _truncate_middle_messages(msgs, keep_ratio = 0.4)
+    assert dropped > 0
+    surviving_call_ids = {
+        tc["id"] for m in new if m.get("role") == "assistant" for tc in (m.get("tool_calls") or [])
+    }
+    for m in new:
+        if m.get("role") == "tool":
+            assert m["tool_call_id"] in surviving_call_ids
+
+
+def test_truncation_reduces_estimated_size_toward_target():
+    msgs = _conversation()
+    total = sum(_estimate_message_tokens(m) for m in msgs)
+    new, dropped = _truncate_middle_messages(msgs, keep_ratio = 0.5)
+    new_total = sum(_estimate_message_tokens(m) for m in new)
+    assert dropped > 0
+    assert new_total < total
+    # Should land at or below the requested share, modulo one whole group.
+    biggest_group = max(
+        _estimate_message_tokens(a) + _estimate_message_tokens(b)
+        for a, b in zip(msgs[2:-2:2], msgs[3:-2:2])
+    )
+    assert new_total <= int(total * 0.5) + biggest_group
+
+
+def test_truncation_noop_when_keep_ratio_full():
+    msgs = _conversation()
+    new, dropped = _truncate_middle_messages(msgs, keep_ratio = 1.0)
+    assert dropped == 0
+    assert new == msgs
+
+
+def test_truncation_noop_when_only_protected_turns_remain():
+    msgs = [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "task"},
+        *_tool_turn(0),
+        {"role": "user", "content": "latest"},
+    ]
+    new, dropped = _truncate_middle_messages(msgs, keep_ratio = 0.1)
+    assert dropped == 0
+    assert new == msgs
+
+
+# ---------------------------------------------------------------------------
+# _apply_overflow_truncation
+# ---------------------------------------------------------------------------
+
+
+def test_apply_overflow_truncation_mutates_body_and_clamps_max_tokens():
+    body = {"messages": _conversation(), "max_tokens": 32000}
+    assert _apply_overflow_truncation(body, _NICK_ERROR) is True
+    assert len(body["messages"]) < len(_conversation())
+    # Generation headroom: max_tokens clamped to the non-prompt share of n_ctx.
+    assert body["max_tokens"] <= max(1024, int(67584 * 0.25))
+
+
+def test_apply_overflow_truncation_returns_false_when_nothing_droppable():
+    body = {
+        "messages": [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "task"},
+            {"role": "user", "content": "latest"},
+        ],
+        "max_tokens": 32000,
+    }
+    assert _apply_overflow_truncation(body, _NICK_ERROR) is False
+
+
+def test_apply_overflow_truncation_clips_giant_protected_tool_results():
+    """One giant burst (few turn-groups, all protected) must still shrink:
+    stage 2 clips oversized tool contents instead of giving up."""
+    msgs = [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "task"},
+        *_tool_turn(0, result_chars = 60000),
+        *_tool_turn(1, result_chars = 60000),
+    ]
+    body = {"messages": msgs, "max_tokens": 32000}
+    n_before = len(msgs)
+    assert _apply_overflow_truncation(body, _NICK_ERROR) is True
+    # No message disappeared (pairing intact), but contents were clipped.
+    assert len(body["messages"]) == n_before
+    clipped = [m for m in body["messages"] if _CLIP_MARKER in str(m.get("content"))]
+    assert clipped, "expected at least one clipped tool result"
+    surviving_call_ids = {
+        tc["id"]
+        for m in body["messages"]
+        if m.get("role") == "assistant"
+        for tc in (m.get("tool_calls") or [])
+    }
+    for m in body["messages"]:
+        if m.get("role") == "tool":
+            assert m["tool_call_id"] in surviving_call_ids
+
+
+def test_clip_long_contents_reaches_target_and_keeps_structure():
+    msgs = [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "task"},
+        *_tool_turn(0, result_chars = 40000),
+        {"role": "user", "content": "latest question"},
+    ]
+    total = sum(_estimate_message_tokens(m) for m in msgs)
+    clipped = _clip_long_contents(msgs, target_est = total // 4)
+    assert clipped >= 1
+    assert sum(_estimate_message_tokens(m) for m in msgs) <= total // 4
+    # Roles and count unchanged; the short final user message untouched.
+    assert [m["role"] for m in msgs] == ["system", "user", "assistant", "tool", "user"]
+    assert msgs[-1]["content"] == "latest question"
+
+
+def test_overflow_truncation_requested_reads_field(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_CONTEXT_OVERFLOW", raising = False)
+
+    class _P:
+        context_overflow = "truncate_middle"
+
+    class _Q:
+        context_overflow = None
+
+    assert _overflow_truncation_requested(_P()) is True
+    assert _overflow_truncation_requested(_Q()) is False
+    assert _overflow_truncation_requested(object()) is False
+
+
+def test_overflow_truncation_server_default_env(monkeypatch):
+    """UNSLOTH_CONTEXT_OVERFLOW enables the policy for clients that cannot
+    send custom body fields; an explicit per-request 'error' still wins."""
+
+    class _Unset:
+        context_overflow = None
+
+    class _ExplicitError:
+        context_overflow = "error"
+
+    monkeypatch.setenv("UNSLOTH_CONTEXT_OVERFLOW", "truncate_middle")
+    assert _overflow_truncation_requested(_Unset()) is True
+    assert _overflow_truncation_requested(_ExplicitError()) is False
+    monkeypatch.setenv("UNSLOTH_CONTEXT_OVERFLOW", "error")
+    assert _overflow_truncation_requested(_Unset()) is False
+
+
+# ---------------------------------------------------------------------------
+# /v1/models context metadata
+# ---------------------------------------------------------------------------
+
+
+class _FakeLlamaBackend:
+    is_loaded = True
+    model_identifier = "unsloth/Qwen3.6-27B-GGUF"
+    context_length = 67584
+    max_context_length = 262144
+
+
+class _FakeEmptyBackend:
+    active_model_name = None
+
+
+def test_v1_models_exposes_real_context_window(monkeypatch):
+    monkeypatch.setattr(routes_mod, "get_llama_cpp_backend", lambda: _FakeLlamaBackend())
+    monkeypatch.setattr(routes_mod, "get_inference_backend", lambda: _FakeEmptyBackend())
+    models = _openai_model_objects()
+    assert len(models) == 1
+    entry = models[0]
+    assert entry["id"] == "unsloth/Qwen3.6-27B-GGUF"
+    # The REAL (post /props readback) window, not the requested one.
+    assert entry["context_length"] == 67584
+    assert entry["max_context_length"] == 262144
diff --git a/studio/backend/tests/test_desktop_auth.py b/studio/backend/tests/test_desktop_auth.py
index 50bd0e89d7..b4b39ba1a9 100644
--- a/studio/backend/tests/test_desktop_auth.py
+++ b/studio/backend/tests/test_desktop_auth.py
@@ -419,6 +419,8 @@ def test_health_response_reports_desktop_capability_fields(monkeypatch):
     routes_module.__path__ = []
     settings_module = ModuleType("routes.settings")
     settings_module.router = APIRouter()
+    llama_module = ModuleType("routes.llama")
+    llama_module.router = APIRouter()
     prompts_module = ModuleType("routes.prompts")
     prompts_module.router = APIRouter()
 
@@ -440,9 +442,11 @@ def test_health_response_reports_desktop_capability_fields(monkeypatch):
     }.items():
         setattr(routes_module, name, router)
     routes_module.settings = settings_module
+    routes_module.llama = llama_module
 
     monkeypatch.setitem(sys.modules, "routes", routes_module)
     monkeypatch.setitem(sys.modules, "routes.settings", settings_module)
+    monkeypatch.setitem(sys.modules, "routes.llama", llama_module)
     monkeypatch.setitem(sys.modules, "routes.prompts", prompts_module)
 
     import studio.backend.main as backend_main
diff --git a/studio/backend/tests/test_external_provider_proxy_env.py b/studio/backend/tests/test_external_provider_proxy_env.py
new file mode 100644
index 0000000000..f17b655908
--- /dev/null
+++ b/studio/backend/tests/test_external_provider_proxy_env.py
@@ -0,0 +1,80 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from pathlib import Path
+import importlib.util
+import sys
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_EXTERNAL_PROVIDER_PATH = (
+    Path(__file__).resolve().parent.parent / "core/inference/external_provider.py"
+)
+
+
+def _load_external_provider_module():
+    spec = importlib.util.spec_from_file_location(
+        "external_provider_under_test",
+        _EXTERNAL_PROVIDER_PATH,
+    )
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_shared_http_client_ignores_unsupported_proxy_scheme(monkeypatch):
+    ep_mod = _load_external_provider_module()
+    calls = []
+
+    class FakeAsyncClient:
+        def __init__(self, **kwargs):
+            calls.append(kwargs)
+            if kwargs.get("trust_env") is not False:
+                raise ValueError("Unknown scheme for proxy URL URL('socks4://127.0.0.1:12345')")
+
+    monkeypatch.setattr(ep_mod.httpx, "AsyncClient", FakeAsyncClient)
+
+    client = ep_mod._create_shared_http_client()
+
+    assert isinstance(client, FakeAsyncClient)
+    assert calls == [{}, {"trust_env": False}]
+
+
+def test_shared_http_client_ignores_missing_socksio(monkeypatch):
+    ep_mod = _load_external_provider_module()
+    calls = []
+
+    class FakeAsyncClient:
+        def __init__(self, **kwargs):
+            calls.append(kwargs)
+            if kwargs.get("trust_env") is not False:
+                raise ImportError(
+                    "Using SOCKS proxy, but the 'socksio' package is not installed. "
+                    "Make sure to install httpx using `pip install httpx[socks]`."
+                )
+
+    monkeypatch.setattr(ep_mod.httpx, "AsyncClient", FakeAsyncClient)
+
+    client = ep_mod._create_shared_http_client()
+
+    assert isinstance(client, FakeAsyncClient)
+    assert calls == [{}, {"trust_env": False}]
+
+
+def test_shared_http_client_reraises_other_value_errors(monkeypatch):
+    ep_mod = _load_external_provider_module()
+
+    class FakeAsyncClient:
+        def __init__(self, **kwargs):
+            raise ValueError("different httpx setup error")
+
+    monkeypatch.setattr(ep_mod.httpx, "AsyncClient", FakeAsyncClient)
+
+    try:
+        ep_mod._create_shared_http_client()
+    except ValueError as exc:
+        assert str(exc) == "different httpx setup error"
+    else:
+        raise AssertionError("expected ValueError")
diff --git a/studio/backend/tests/test_inference_orchestrator_crash_message.py b/studio/backend/tests/test_inference_orchestrator_crash_message.py
new file mode 100644
index 0000000000..be1a673e62
--- /dev/null
+++ b/studio/backend/tests/test_inference_orchestrator_crash_message.py
@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from pathlib import Path
+from types import SimpleNamespace
+import importlib.util
+import sys
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+
+def test_subprocess_crash_message_includes_signal_and_oom_hint():
+    spec = importlib.util.spec_from_file_location(
+        "inference_orchestrator_under_test",
+        Path(__file__).resolve().parent.parent / "core/inference/orchestrator.py",
+    )
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+
+    orchestrator = module.InferenceOrchestrator.__new__(module.InferenceOrchestrator)
+    orchestrator._proc = SimpleNamespace(pid = 1234, exitcode = -9)
+
+    msg = orchestrator._subprocess_crash_message("wait")
+
+    assert msg.startswith("The inference worker stopped unexpectedly while loading the model.")
+    assert "memory pressure" in msg
+    assert "smaller model" in msg
+    assert "Details:" in msg
+    assert "signal=SIGKILL" in msg
+    assert "exitcode=-9" in msg
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
new file mode 100644
index 0000000000..ede5629664
--- /dev/null
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -0,0 +1,173 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""install_llama_prebuilt.py: host->repo mapping and the --resolve-prebuilt mode.
+
+These back the in-app update for source-build (markerless) installs: the backend
+asks the installer whether an official prebuilt exists for this host without
+downloading. Network and host detection are stubbed; no GPU or internet needed.
+"""
+
+from __future__ import annotations
+
+import importlib
+import json
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_studio = Path(__file__).resolve().parent.parent.parent
+if str(_studio) not in sys.path:
+    sys.path.insert(0, str(_studio))
+
+ilp = importlib.import_module("install_llama_prebuilt")
+
+if not hasattr(ilp, "published_repo_for_host") or not hasattr(
+    ilp, "resolve_simple_install_release_plans"
+):
+    pytest.skip("PR symbols not present - check branch", allow_module_level = True)
+
+FORK = ilp.DEFAULT_PUBLISHED_REPO  # unslothai/llama.cpp
+UPSTREAM = ilp.UPSTREAM_REPO  # ggml-org/llama.cpp
+
+
+def _host(**kw):
+    base = dict(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = False,
+        is_macos = False,
+        is_x86_64 = False,
+        is_arm64 = False,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+        has_rocm = False,
+        rocm_gfx_target = None,
+        macos_version = None,
+    )
+    base.update(kw)
+    return ilp.HostInfo(**base)
+
+
+def test_published_repo_for_host():
+    # CPU-only Linux (x64 and arm64) -> ggml-org upstream.
+    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True)) == UPSTREAM
+    assert (
+        ilp.published_repo_for_host(_host(is_linux = True, is_arm64 = True, machine = "aarch64"))
+        == UPSTREAM
+    )
+    # GPU Linux -> fork.
+    assert (
+        ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_usable_nvidia = True))
+        == FORK
+    )
+    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_rocm = True)) == FORK
+    # CPU-only Windows -> ggml-org (setup.ps1: the fork ships no win-cpu bundle).
+    assert (
+        ilp.published_repo_for_host(_host(system = "Windows", is_windows = True, is_x86_64 = True))
+        == UPSTREAM
+    )
+    # GPU Windows -> fork.
+    assert (
+        ilp.published_repo_for_host(
+            _host(system = "Windows", is_windows = True, is_x86_64 = True, has_usable_nvidia = True)
+        )
+        == FORK
+    )
+    # macOS -> fork regardless of GPU (ggml-org macOS bundles need too-new macOS).
+    assert (
+        ilp.published_repo_for_host(
+            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
+        )
+        == FORK
+    )
+    # Linux with AMD tooling but no probed GPU -> fork (setup.sh routes on tooling).
+    assert (
+        ilp.published_repo_for_host(
+            _host(is_linux = True, is_x86_64 = True), linux_amd_tooling_present = True
+        )
+        == FORK
+    )
+    # The tooling hint is Linux-only: Windows CPU stays on ggml-org.
+    assert (
+        ilp.published_repo_for_host(
+            _host(system = "Windows", is_windows = True, is_x86_64 = True),
+            linux_amd_tooling_present = True,
+        )
+        == UPSTREAM
+    )
+
+
+def _run_resolve(monkeypatch, capsys, plans_or_exc):
+    monkeypatch.setattr(
+        ilp,
+        "detect_host",
+        lambda: _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64"),
+    )
+
+    def _resolver(tag, host, repo, published_release_tag):
+        if isinstance(plans_or_exc, Exception):
+            raise plans_or_exc
+        return ("b9585", plans_or_exc)
+
+    monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        ["install_llama_prebuilt.py", "--resolve-prebuilt", "latest", "--output-format", "json"],
+    )
+    rc = ilp.main()
+    assert rc == ilp.EXIT_SUCCESS
+    return json.loads(capsys.readouterr().out.strip().splitlines()[-1])
+
+
+def test_resolve_prebuilt_available(monkeypatch, capsys):
+    plan = SimpleNamespace(
+        release_tag = "b9585",
+        llama_tag = "b9585",
+        attempts = [
+            SimpleNamespace(name = "llama-b9585-bin-macos-arm64.tar.gz", install_kind = "macos-arm64")
+        ],
+    )
+    out = _run_resolve(monkeypatch, capsys, [plan])
+    assert out["prebuilt_available"] is True
+    assert out["repo"] == FORK
+    assert out["release_tag"] == "b9585"
+    assert out["asset"] == "llama-b9585-bin-macos-arm64.tar.gz"
+    assert out["install_kind"] == "macos-arm64"
+
+
+def test_resolve_prebuilt_unavailable(monkeypatch, capsys):
+    out = _run_resolve(monkeypatch, capsys, ilp.PrebuiltFallback("no macOS asset"))
+    assert out["prebuilt_available"] is False
+    assert out["repo"] == FORK
+
+
+def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
+    # CPU-probed Linux host but rocminfo on PATH: the dispatch must route to the
+    # fork so a HIP source build is not offered an upstream CPU prebuilt.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    monkeypatch.setattr(ilp.shutil, "which", lambda tool: tool == "rocminfo")
+    seen = {}
+
+    def _resolver(tag, host, repo, published_release_tag):
+        seen["repo"] = repo
+        raise ilp.PrebuiltFallback("no asset")
+
+    monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        ["install_llama_prebuilt.py", "--resolve-prebuilt", "latest", "--output-format", "json"],
+    )
+    assert ilp.main() == ilp.EXIT_SUCCESS
+    out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
+    assert seen["repo"] == FORK
+    assert out["repo"] == FORK
diff --git a/studio/backend/tests/test_lemonade_llamacpp_rocm_bins_mock.py b/studio/backend/tests/test_lemonade_llamacpp_rocm_bins_mock.py
index 3443f53a12..7c6c514b8f 100644
--- a/studio/backend/tests/test_lemonade_llamacpp_rocm_bins_mock.py
+++ b/studio/backend/tests/test_lemonade_llamacpp_rocm_bins_mock.py
@@ -1,10 +1,11 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Validates that the installer resolves lemonade ROCm prebuilt assets.
+"""Validates that the installer correctly resolves lemonade ROCm prebuilt assets.
 
-Uses a faked HostInfo so no AMD GPU is needed. The lemonade GitHub API calls
-are stubbed so the suite runs offline and isn't subject to rate limits.
+Uses a faked HostInfo so no AMD GPU is needed. Network calls to the lemonade
+GitHub API are stubbed out so the suite runs without internet access and is
+not subject to rate limits.
 """
 
 from __future__ import annotations
@@ -31,14 +32,19 @@ if resolve_lemonade_rocm_choice is None or _LEMONADE_GFX_FAMILIES is None:
 
 @pytest.fixture(autouse = True)
 def _clear_lemonade_release_cache():
-    """Prevent cross-test pollution of the lemonade release lru_cache when
-    tests vary the fetch_json mock return value."""
+    """Prevent cross-test pollution of the lemonade release lru_cache and
+    selection-log dedup set when tests vary the fetch_json mock return value."""
     _cache = getattr(_mod, "_fetch_lemonade_release_cached", None)
+    _logged: set | None = getattr(_mod, "_lemonade_selection_logged", None)
     if _cache is not None and hasattr(_cache, "cache_clear"):
         _cache.cache_clear()
+    if _logged is not None:
+        _logged.clear()
     yield
     if _cache is not None and hasattr(_cache, "cache_clear"):
         _cache.cache_clear()
+    if _logged is not None:
+        _logged.clear()
 
 
 _STUB_TAG = "b1262"
@@ -89,7 +95,9 @@ def _lookup_family(gfx: str) -> str | None:
     return None
 
 
+# ---------------------------------------------------------------------------
 # GPU family mapping
+# ---------------------------------------------------------------------------
 
 
 @pytest.mark.parametrize(
@@ -111,7 +119,9 @@ def test_unknown_gpu_not_in_families():
     assert _lookup_family("gfx999") is None
 
 
+# ---------------------------------------------------------------------------
 # Asset resolution - hits real lemonade GitHub API
+# ---------------------------------------------------------------------------
 
 
 @pytest.mark.parametrize(
@@ -141,59 +151,71 @@ def test_unknown_gpu_falls_through_to_upstream():
     assert result is None
 
 
-# Simple-policy dispatcher must plan a lemonade ROCm attempt for AMD-only hosts.
-# This is the path setup.sh invokes (via --simple-policy), so the lemonade
-# integration is useless if it isn't wired in here.
+# ---------------------------------------------------------------------------
+# The Linux attempt builder must plan a lemonade ROCm attempt for AMD-only hosts.
+# This is the path setup.sh actually invokes (fork hosts now select from the
+# manifest), so the lemonade integration is useless if it isn't wired in here.
+# ---------------------------------------------------------------------------
 
-direct_linux_release_plan = getattr(_mod, "direct_linux_release_plan", None)
+_linux_published_attempts = getattr(_mod, "_linux_published_attempts", None)
 direct_upstream_release_plan = getattr(_mod, "direct_upstream_release_plan", None)
 
+PublishedLlamaArtifact = _mod.PublishedLlamaArtifact
+PublishedReleaseBundle = _mod.PublishedReleaseBundle
 
-def _stub_unsloth_release(release_tag: str = "b9022") -> dict:
-    # Minimal payload parse_direct_linux_release_bundle accepts. It needs at
-    # least one `app-{label}-linux-x64*.tar.gz` asset to recognise the bundle;
-    # we ship a bare CPU one so the planner has a baseline non-ROCm fallback.
-    asset_name = f"app-{release_tag}-linux-x64.tar.gz"
-    return {
-        "tag_name": release_tag,
-        "name": release_tag,
-        "assets": [
-            {
-                "name": asset_name,
-                "browser_download_url": f"https://example.invalid/{asset_name}",
-            },
-        ],
-    }
+
+def _rocm_bundle(gfx_family: str, mapped_targets: list[str]) -> "PublishedReleaseBundle":
+    """A fork manifest bundle exposing a per-gfx linux-rocm artifact, so
+    published_rocm_choice_for_host can match the host before the lemonade
+    fallback is appended."""
+    asset_name = f"app-b9457-linux-x64-rocm-{gfx_family}.tar.gz"
+    artifact = PublishedLlamaArtifact(
+        asset_name = asset_name,
+        install_kind = "linux-rocm",
+        runtime_line = None,
+        coverage_class = None,
+        supported_sms = [],
+        min_sm = None,
+        max_sm = None,
+        bundle_profile = None,
+        rank = 1000,
+        gfx_target = gfx_family,
+        mapped_targets = mapped_targets,
+    )
+    return PublishedReleaseBundle(
+        repo = "unslothai/llama.cpp",
+        release_tag = "v1.0",
+        upstream_tag = "b9457",
+        assets = {asset_name: f"https://example.invalid/{asset_name}"},
+        artifacts = [artifact],
+    )
 
 
 @pytest.mark.skipif(
-    direct_linux_release_plan is None,
-    reason = "simple-policy dispatcher not present on this branch",
+    _linux_published_attempts is None,
+    reason = "Linux attempt builder not present on this branch",
 )
-def test_simple_policy_plans_lemonade_for_rocm_host():
+def test_linux_attempts_include_fork_rocm_and_lemonade_for_rocm_host():
     host = _make_rocm_host("gfx1151")
+    bundle = _rocm_bundle("gfx1151", ["gfx1151"])
     with patch.object(_mod, "fetch_json", return_value = _stub_lemonade_release()):
-        plan = direct_linux_release_plan(
-            _stub_unsloth_release(),
-            host,
-            "unslothai/llama.cpp",
-            "latest",
-        )
-    assert plan is not None, "ROCm host should not be skipped by simple-policy planner"
-    kinds = [a.install_kind for a in plan.attempts]
-    assert (
-        "linux-rocm" in kinds
-    ), f"simple-policy planner did not include a lemonade ROCm attempt; got {kinds}"
-    rocm_attempt = next(a for a in plan.attempts if a.install_kind == "linux-rocm")
-    assert rocm_attempt.source_label == "lemonade"
-    assert "gfx1151" in rocm_attempt.name
+        attempts = _linux_published_attempts(host, bundle, "latest")
+    kinds = [a.install_kind for a in attempts]
+    assert "linux-rocm" in kinds, f"builder did not include any linux-rocm attempt; got {kinds}"
+    sources = {a.source_label for a in attempts if a.install_kind == "linux-rocm"}
+    # The fork's own per-gfx bundle is preferred, with the lemonade prebuilt as
+    # the fallback -- both must be present for a covered ROCm host.
+    assert "published" in sources, f"fork ROCm bundle missing; got {sources}"
+    assert "lemonade" in sources, f"lemonade ROCm fallback missing; got {sources}"
+    lemonade_attempt = next(a for a in attempts if a.source_label == "lemonade")
+    assert "gfx1151" in lemonade_attempt.name
 
 
 @pytest.mark.skipif(
     direct_upstream_release_plan is None,
-    reason = "simple-policy dispatcher not present on this branch",
+    reason = "direct release planners not present on this branch",
 )
-def test_simple_policy_plans_lemonade_for_windows_hip_host():
+def test_direct_upstream_plan_includes_lemonade_for_windows_hip_host():
     host = _make_rocm_host("gfx1151", windows = True)
     release = {
         "tag_name": "b9022",
@@ -204,16 +226,14 @@ def test_simple_policy_plans_lemonade_for_windows_hip_host():
         plan = direct_upstream_release_plan(release, host, "ggml-org/llama.cpp", "latest")
     assert plan is not None, "Windows ROCm host should plan a lemonade HIP attempt"
     kinds = [a.install_kind for a in plan.attempts]
-    assert (
-        "windows-hip" in kinds
-    ), f"simple-policy planner did not include a lemonade HIP attempt; got {kinds}"
+    assert "windows-hip" in kinds, f"planner did not include a lemonade HIP attempt; got {kinds}"
 
 
 @pytest.mark.skipif(
     direct_upstream_release_plan is None,
-    reason = "simple-policy dispatcher not present on this branch",
+    reason = "direct release planners not present on this branch",
 )
-def test_simple_policy_windows_hip_falls_back_to_upstream_when_lemonade_unavailable():
+def test_windows_hip_falls_back_to_upstream_when_lemonade_unavailable():
     """If lemonade returns None (e.g. gfx999 or transient API failure), the planner
     must still include the upstream HIP asset rather than silently downgrading to CPU."""
     host = _make_rocm_host("gfx999", windows = True)
@@ -247,8 +267,9 @@ def test_lemonade_release_api_url_pinned_tag():
 
 
 def test_lemonade_release_api_url_encodes_tag():
-    """Slashes / hashes in the tag must be URL-encoded so the URL can't be
-    reshaped (defence in depth -- tags should already be sanitised upstream)."""
+    """Unexpected slashes / hashes in the tag must be URL-encoded so the URL
+    cannot be reshaped (defence in depth -- tags should already be sanitised
+    upstream)."""
     url = _mod._lemonade_release_api_for("b1260/../latest")
     assert "/releases/tags/b1260%2F..%2Flatest" in url
     assert "//latest" not in url.split("/releases/tags/", 1)[1]
@@ -263,9 +284,9 @@ def test_lemonade_resolver_skipped_by_opt_out_env(monkeypatch):
 
 
 def test_lemonade_resolver_rejects_non_github_url(monkeypatch):
-    """If the GitHub API response contained an off-host download URL, the
-    resolver must refuse it (lemonade assets aren't in the approved-hash
-    manifest)."""
+    """If the GitHub API response somehow contained an off-host download URL,
+    the resolver must refuse to use it (lemonade assets are not in the
+    approved-hash manifest)."""
     bad_release = {
         "tag_name": _STUB_TAG,
         "assets": [
@@ -289,7 +310,7 @@ def test_lemonade_resolver_rejects_http_scheme():
 
 
 def test_lemonade_resolver_accepts_github_cdn():
-    # Real GitHub release CDN URLs carry the /github-production-release-asset- prefix
+    # Real GitHub release CDN URLs carry the /github-production-release-asset- prefix.
     assert _mod._is_trusted_github_release_url(
         "https://objects.githubusercontent.com/github-production-release-asset-abc123/456/789?token=x",
         "lemonade-sdk/llamacpp-rocm",
@@ -297,7 +318,7 @@ def test_lemonade_resolver_accepts_github_cdn():
 
 
 def test_lemonade_resolver_rejects_arbitrary_cdn_path():
-    # A CDN URL without the release-asset path prefix must be rejected
+    # A CDN URL without the release-asset path prefix must be rejected.
     assert not _mod._is_trusted_github_release_url(
         "https://objects.githubusercontent.com/abc/def",
         "lemonade-sdk/llamacpp-rocm",
@@ -339,7 +360,7 @@ def test_lemonade_runtime_patterns_include_hip_runtime():
 
     Lemonade ZIPs carry transitive deps (libamd_comgr, libLLVM, libclang-cpp,
     ...) whose names change across ROCm releases. A broad ``lib*.so*`` glob
-    avoids enumerating every transitive dependency by name.
+    avoids having to enumerate every transitive dependency by name.
     """
     from install_llama_prebuilt import runtime_patterns_for_choice, AssetChoice
 
@@ -353,7 +374,7 @@ def test_lemonade_runtime_patterns_include_hip_runtime():
     )
     pats = runtime_patterns_for_choice(choice)
     # The broad glob must be present so every .so in the lemonade bundle
-    # (including future transitive deps) gets overlaid.
+    # (including transitive deps added in future ROCm releases) gets overlaid.
     assert "lib*.so*" in pats, f"'lib*.so*' missing from linux-rocm patterns: {pats}"
 
 
@@ -365,9 +386,9 @@ _pick_rocm_gfx_target = getattr(_mod, "_pick_rocm_gfx_target", None)
     reason = "_pick_rocm_gfx_target not present on this branch",
 )
 def test_pick_rocm_gfx_target_honors_cuda_visible_devices(monkeypatch):
-    """AMD HIP honours CUDA_VISIBLE_DEVICES like HIP_VISIBLE_DEVICES; on a
-    gfx1151 + gfx1100 mixed host, CUDA_VISIBLE_DEVICES=1 must select gfx1100."""
-    # Two GPUs; rocminfo reports each token twice (as in real tool output).
+    """AMD HIP honours CUDA_VISIBLE_DEVICES identically to HIP_VISIBLE_DEVICES;
+    on a gfx1151 + gfx1100 mixed host, CUDA_VISIBLE_DEVICES=1 must select gfx1100."""
+    # Two GPUs; rocminfo reports each token twice (as in the real tool output).
     probe_out = "gfx1151\ngfx1151\ngfx1100\ngfx1100"
     monkeypatch.delenv("HIP_VISIBLE_DEVICES", raising = False)
     monkeypatch.delenv("ROCR_VISIBLE_DEVICES", raising = False)
@@ -396,7 +417,7 @@ def test_pick_rocm_gfx_target_same_arch_multi_gpu(monkeypatch):
     """Regression: [gfx1100, gfx1100, gfx1151] with HIP_VISIBLE_DEVICES=2 must
     return gfx1151, not fall back to GPU 0 due to dict.fromkeys collapsing the
     two gfx1100 entries into one and making index 2 out of range."""
-    # rocminfo output for 3 GPUs (2x gfx1100 dGPU + 1x gfx1151 APU).
+    # Simulate rocminfo output for 3 GPUs (2x gfx1100 dGPU + 1x gfx1151 APU).
     # Each GPU gets its own Agent section with a few token mentions.
     probe_out = (
         "***\nAgent 1\n***\n  gfx1100 some info\n  gfx1100\n"
@@ -407,3 +428,96 @@ def test_pick_rocm_gfx_target_same_arch_multi_gpu(monkeypatch):
     monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising = False)
     monkeypatch.setenv("HIP_VISIBLE_DEVICES", "2")
     assert _pick_rocm_gfx_target(probe_out) == "gfx1151"
+
+
+# ---------------------------------------------------------------------------
+# Fork release scan: Windows ROCm resolves lemonade by the requested tag
+# ---------------------------------------------------------------------------
+
+_resolve_release_asset_choice = getattr(_mod, "resolve_release_asset_choice", None)
+_ApprovedReleaseChecksums = getattr(_mod, "ApprovedReleaseChecksums", None)
+
+
+@pytest.mark.skipif(
+    _resolve_release_asset_choice is None or _ApprovedReleaseChecksums is None,
+    reason = "fork release planner not present on this branch",
+)
+def test_fork_scan_windows_rocm_resolves_lemonade_by_requested_tag():
+    """The fork release scan pins llama_tag to per-release upstream tags
+    (b9457, ...) that lemonade's own tag series never contains, so the
+    lemonade lookup must use the requested tag ("latest") instead. Pinning
+    lemonade to the per-release tag 404s on every scanned release and a
+    Windows ROCm host ends in a rate-limited fatal instead of the lemonade
+    prebuilt."""
+    host = _make_rocm_host("gfx1151", windows = True)
+    # No windows-rocm artifact in the bundle, matching current fork releases.
+    bundle = _rocm_bundle("gfx1151", ["gfx1151"])
+    checksums = _ApprovedReleaseChecksums(
+        repo = "unslothai/llama.cpp",
+        release_tag = "v1.0",
+        upstream_tag = "b9457",
+        artifacts = {},
+    )
+    seen_urls: list[str] = []
+
+    def _fake_fetch(api_url, *args, **kwargs):
+        seen_urls.append(api_url)
+        if "lemonade-sdk" in api_url:
+            if api_url.endswith("/releases/latest"):
+                return _stub_lemonade_release()
+            raise RuntimeError(f"unexpected pinned lemonade fetch: {api_url}")
+        # ggml-org asset listing for the upstream HIP/CPU filename fallbacks.
+        return {"tag_name": "b9457", "assets": []}
+
+    with patch.object(_mod, "fetch_json", side_effect = _fake_fetch):
+        attempts = _resolve_release_asset_choice(
+            host,
+            "b9457",  # concrete per-release upstream tag from the scan loop
+            bundle,
+            checksums,
+            requested_tag = "latest",
+        )
+
+    lemonade = [a for a in attempts if a.source_label == "lemonade"]
+    assert lemonade, f"lemonade attempt missing for Windows ROCm host; got {attempts}"
+    assert "gfx1151" in lemonade[0].name
+    assert any(
+        u.endswith("/releases/latest") for u in seen_urls
+    ), f"lemonade was never resolved via /releases/latest; fetches: {seen_urls}"
+    assert not any(
+        "lemonade-sdk" in u and "/releases/tags/" in u for u in seen_urls
+    ), f"lemonade lookup was pinned to the fork release tag: {seen_urls}"
+
+
+@pytest.mark.skipif(
+    direct_upstream_release_plan is None,
+    reason = "direct release planners not present on this branch",
+)
+def test_direct_upstream_plan_includes_lemonade_for_linux_rocm_host():
+    """A Linux ROCm host on the ggml-org direct path (e.g. a --published-repo
+    override) must plan lemonade before the CPU tarball, mirroring the Windows
+    branch. The lemonade planning previously lived in the removed
+    --simple-policy dispatcher, so without this leg such hosts silently
+    install the CPU build."""
+    host = _make_rocm_host("gfx1151")
+    release = {
+        "tag_name": "b9022",
+        "name": "b9022",
+        "assets": [
+            {
+                "name": "llama-b9022-bin-ubuntu-x64.tar.gz",
+                "browser_download_url": (
+                    "https://github.com/ggml-org/llama.cpp/releases/download/"
+                    "b9022/llama-b9022-bin-ubuntu-x64.tar.gz"
+                ),
+            }
+        ],
+    }
+    with patch.object(_mod, "fetch_json", return_value = _stub_lemonade_release()):
+        plan = direct_upstream_release_plan(release, host, "ggml-org/llama.cpp", "latest")
+    assert plan is not None, "Linux ROCm host should produce a direct plan"
+    kinds = [a.install_kind for a in plan.attempts]
+    sources = [a.source_label for a in plan.attempts]
+    assert "linux-rocm" in kinds, f"lemonade ROCm attempt missing; got {kinds}"
+    assert sources[0] == "lemonade", f"lemonade must be the first attempt; got {sources}"
+    assert "gfx1151" in plan.attempts[0].name
diff --git a/studio/backend/tests/test_llama_cpp_freshness.py b/studio/backend/tests/test_llama_cpp_freshness.py
index cb17e0d5e7..f8e4619ded 100644
--- a/studio/backend/tests/test_llama_cpp_freshness.py
+++ b/studio/backend/tests/test_llama_cpp_freshness.py
@@ -21,12 +21,25 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
+
+class _NoopLogger:
+    """structlog-style logger: every method swallows positional + kwargs.
+
+    A stdlib logging.Logger rejects structlog's keyword fields (e.g.
+    ``logger.warning(msg, error=...)``), which leaked into the update module's
+    error path and failed only when this file's stub loaded first.
+    """
+
+    def __getattr__(self, _name):
+        return lambda *a, **k: None
+
+
 _loggers_stub = _types.ModuleType("loggers")
-_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+_loggers_stub.get_logger = lambda *a, **k: _NoopLogger()
 sys.modules.setdefault("loggers", _loggers_stub)
 
 _structlog_stub = _types.ModuleType("structlog")
-_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+_structlog_stub.get_logger = lambda *a, **k: _NoopLogger()
 sys.modules.setdefault("structlog", _structlog_stub)
 
 import pytest
@@ -51,6 +64,11 @@ def _write_marker(install_dir: Path, **overrides) -> Path:
         .replace("+00:00", "Z"),
     }
     payload.update(overrides)
+    # The installer always writes `tag` and `release_tag` from the same release
+    # (a normalized base vs the full release tag), so keep the pair consistent
+    # when a test overrides only `tag`.
+    if "tag" in overrides and "release_tag" not in overrides:
+        payload["release_tag"] = overrides["tag"]
     install_dir.mkdir(parents = True, exist_ok = True)
     (install_dir / "UNSLOTH_PREBUILT_INFO.json").write_text(json.dumps(payload))
     return install_dir / "UNSLOTH_PREBUILT_INFO.json"
@@ -303,3 +321,115 @@ def test_format_stale_warning_singular_day():
     msg = fr.format_stale_warning({"installed_tag": "b9190", "latest_tag": "b9300", "age_days": 1})
     assert "1 day" in msg
     assert "1 days" not in msg
+
+
+# parse_base_build / is_behind.
+
+
+def test_parse_base_build():
+    assert fr.parse_base_build("b9596") == 9596
+    assert fr.parse_base_build(" b9596 ") == 9596
+    assert fr.parse_base_build("b9596-mix-e6f2453") == 9596  # mix suffix doesn't defeat it
+    assert fr.parse_base_build("9596") is None
+    assert fr.parse_base_build("master-abc") is None
+    assert fr.parse_base_build("") is None
+    assert fr.parse_base_build(None) is None
+
+
+@pytest.mark.parametrize(
+    "installed, latest, expected",
+    [
+        (
+            "b9596-mix-e6f2453",
+            "b9596-mix-e6f2453",
+            False,
+        ),  # already on the mix latest -> not behind
+        ("b9596", "b9594", False),  # latest is an older build -> downgrade guard
+        ("b9596", "b9594-mix-xxx", False),  # older mix latest -> still guarded
+        ("b9500", "b9596-mix-e6f2453", True),  # newer base -> behind
+        ("b9596-mix-aaa", "b9596-mix-bbb", True),  # new mix at same base -> behind
+        ("b9596", "b9596-mix-bbb", True),  # clean -> mix at same base -> behind
+        ("b9596-mix-aaa", "b9596", False),  # bare base never supersedes a mix install
+        ("b9596", "b9596", False),  # identical -> not behind
+        (" b9596 ", "b9596", False),  # whitespace-only diff -> not behind
+        ("master-abc", "master-def", True),  # non-bNNNN both -> plain inequality
+        ("master-abc", "master-abc", False),
+        (None, "b9596", False),
+        ("b9596", None, False),
+    ],
+)
+def test_is_behind(installed, latest, expected):
+    assert fr.is_behind(installed, latest) is expected
+
+
+def test_check_prebuilt_freshness_not_behind_on_mix_latest(monkeypatch, tmp_path):
+    # Installed the mix latest: marker base tag b9596, full release_tag with sha,
+    # GitHub latest is that same full tag. Must not report behind (sticky bug).
+    install_dir = tmp_path / "llama.cpp"
+    _write_marker(install_dir, tag = "b9596", release_tag = "b9596-mix-e6f2453")
+    bin_path = _fake_binary(install_dir, layout = "root")
+    monkeypatch.setattr(
+        fr, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["behind"] is False
+    assert info["stale"] is False
+
+
+def test_check_prebuilt_freshness_downgrade_guard(monkeypatch, tmp_path):
+    # A lagging latest (older build than installed) must never read as behind/stale.
+    install_dir = tmp_path / "llama.cpp"
+    _write_marker(
+        install_dir,
+        tag = "b9585",
+        installed_at_utc = (datetime.now(tz = timezone.utc) - timedelta(days = 30))
+        .isoformat()
+        .replace("+00:00", "Z"),
+    )
+    bin_path = _fake_binary(install_dir, layout = "root")
+    monkeypatch.setattr(fr, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["behind"] is False
+    assert info["stale"] is False
+
+
+def test_fetch_latest_release_tag_uses_publish_time(monkeypatch):
+    # Resolves newest by published_at (like the installer), skips drafts/prereleases,
+    # and does NOT just take GitHub's first/`/releases/latest` item.
+    import urllib.request
+
+    class _Resp:
+        def __init__(self, payload):
+            self._p = json.dumps(payload).encode()
+
+        def read(self):
+            return self._p
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    payload = [
+        {
+            "tag_name": "b9518",
+            "draft": False,
+            "prerelease": False,
+            "published_at": "2026-06-04T21:11:19Z",
+        },
+        {
+            "tag_name": "b9596-mix-e6f2453",
+            "draft": False,
+            "prerelease": False,
+            "published_at": "2026-06-11T22:50:41Z",
+        },
+        {
+            "tag_name": "b9999-draft",
+            "draft": True,
+            "prerelease": False,
+            "published_at": "2026-06-12T00:00:00Z",
+        },
+    ]
+    monkeypatch.setattr(urllib.request, "urlopen", lambda req, timeout = 5.0: _Resp(payload))
+    assert fr._fetch_latest_release_tag("unslothai/llama.cpp") == "b9596-mix-e6f2453"
diff --git a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
new file mode 100644
index 0000000000..6ef94545fe
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
@@ -0,0 +1,187 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the llama-server mmproj text-only fallback.
+
+A GGUF vision model is launched with ``--mmproj <projector>``. When the
+installed llama.cpp prebuilt is older than the model's projector format,
+llama-server aborts at startup with ``clip.cpp:NNNN: Unknown projector
+type`` (exit -6). load_model now retries once WITHOUT ``--mmproj`` so the
+base model still loads text-only, warns the user to update llama.cpp, and
+marks the session non-vision. These tests pin the two decision helpers:
+``_is_projector_incompatibility`` (when to retry) and ``_strip_mmproj_args``
+(how the retry argv is built). Unrelated failures must NOT trigger a retry.
+"""
+
+from __future__ import annotations
+
+import sys
+import types as _types
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Match the stubbing pattern in sibling tests so the module imports in a
+# lightweight env without fastapi.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+_structlog_stub = _types.ModuleType("structlog")
+_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("structlog")
+sys.modules.setdefault("structlog", _structlog_stub)
+if not hasattr(sys.modules["structlog"], "get_logger"):
+    sys.modules["structlog"].get_logger = _structlog_stub.get_logger
+
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+_detect = LlamaCppBackend._is_projector_incompatibility
+_strip = LlamaCppBackend._strip_mmproj_args
+
+# Real abort captured loading gemma-4 on a 3-day-old prebuilt (build b9496).
+_GEMMA4_OLD_LLAMACPP_OUT = (
+    "srv    load_model: loading model 'gemma-4-E2B-it-UD-Q4_K_XL.gguf'\n"
+    "/build_work/src/llama.cpp-b9496/tools/mtmd/clip.cpp:4391: "
+    "Unknown projector type\n"
+    "libggml-base.so.0(ggml_abort+0x152)\n"
+    "libmtmd.so.0(clip_n_mmproj_embd)\n"
+)
+# Unrelated failures that must keep their own handling (no projector retry).
+_OOM_OUT = (
+    "ggml_backend_cuda_buffer_type_alloc_buffer: allocating 12000.00 MiB on "
+    "device 0: cudaMalloc failed: out of memory"
+)
+_BAD_ARCH_OUT = "llama_model_load: error loading model: unknown model architecture: 'qwen_image'"
+_PORT_OUT = "srv start: failed to bind: address already in use"
+_MISSING_OUT = "error: failed to open GGUF file: no such file or directory"
+# A healthy startup log that merely mentions the projector must not match.
+_HEALTHY_VISION_OUT = (
+    "Using mmproj for vision: /cache/mmproj-F16.gguf\n"
+    "clip_model_loader: loaded meta data with 20 key-value pairs\n"
+    "srv  update_slots: all slots are idle"
+)
+
+
+class TestProjectorIncompatibilityDetector:
+    def test_gemma4_on_old_llamacpp_triggers_retry(self):
+        # Headline case: a 3-day-old llama.cpp aborts on Gemma-4's projector.
+        assert _detect(_GEMMA4_OLD_LLAMACPP_OUT) is True
+
+    @pytest.mark.parametrize(
+        "out",
+        [
+            "clip.cpp:4391: Unknown projector type",
+            "error: unsupported projector type for this model",
+            "llama_mmproj: unsupported mmproj file version",
+            "clip.cpp: projector type 'gemma4' is not supported",
+        ],
+    )
+    def test_projector_format_errors_match(self, out):
+        assert _detect(out) is True
+
+    def test_case_insensitive(self):
+        assert _detect("UNKNOWN PROJECTOR TYPE") is True
+
+    @pytest.mark.parametrize(
+        "out",
+        [
+            _OOM_OUT,
+            _BAD_ARCH_OUT,
+            _PORT_OUT,
+            _MISSING_OUT,
+            _HEALTHY_VISION_OUT,
+            "",
+            # bare multimodal words without a failure term must not match
+            "loading clip model",
+            "mmproj file resolved from cache",
+        ],
+    )
+    def test_unrelated_failures_do_not_retry(self, out):
+        assert _detect(out) is False
+
+
+# A realistic vision launch argv (mirrors the live "Starting llama-server"
+# command), projector pair at the end.
+_VISION_CMD = [
+    "/home/u/.unsloth/llama.cpp/build/bin/llama-server",
+    "-m",
+    "/cache/gemma-4-E2B-it-UD-Q4_K_XL.gguf",
+    "--port",
+    "55473",
+    "-c",
+    "131072",
+    "--parallel",
+    "1",
+    "--flash-attn",
+    "on",
+    "--no-context-shift",
+    "-ngl",
+    "-1",
+    "--threads",
+    "-1",
+    "--jinja",
+    "--spec-default",
+    "--mmproj",
+    "/cache/mmproj-F16.gguf",
+]
+
+
+class TestStripMmprojArgs:
+    def test_removes_mmproj_pair(self):
+        stripped = _strip(_VISION_CMD)
+        assert "--mmproj" not in stripped
+        assert "/cache/mmproj-F16.gguf" not in stripped
+
+    def test_preserves_every_text_flag(self):
+        stripped = _strip(_VISION_CMD)
+        for flag in (
+            "-m",
+            "/cache/gemma-4-E2B-it-UD-Q4_K_XL.gguf",
+            "--port",
+            "55473",
+            "-c",
+            "131072",
+            "-ngl",
+            "-1",
+            "--jinja",
+            "--spec-default",
+            "--flash-attn",
+            "on",
+        ):
+            assert flag in stripped
+        # Exactly the two projector tokens are dropped.
+        assert len(stripped) == len(_VISION_CMD) - 2
+
+    def test_strips_mmproj_in_the_middle(self):
+        cmd = ["llama-server", "--mmproj", "/p/mm.gguf", "-c", "4096", "--jinja"]
+        assert _strip(cmd) == ["llama-server", "-c", "4096", "--jinja"]
+
+    def test_noop_when_no_mmproj(self):
+        cmd = ["llama-server", "-m", "/p/model.gguf", "-c", "4096", "--jinja"]
+        assert _strip(cmd) == cmd
+
+    def test_returns_new_list(self):
+        cmd = ["llama-server", "--mmproj", "/p/mm.gguf"]
+        out = _strip(cmd)
+        assert out is not cmd
+        assert cmd[-1] == "/p/mm.gguf"  # input untouched
+
+
+class TestRetryContract:
+    """The two helpers compose into the load_model retry decision."""
+
+    def test_gemma4_failure_yields_valid_text_only_command(self):
+        # Old-llama.cpp projector abort -> retry, and the retry argv is a
+        # valid text-only launch (model + context kept, projector gone).
+        assert _detect(_GEMMA4_OLD_LLAMACPP_OUT) is True
+        retry_cmd = _strip(_VISION_CMD)
+        assert "--mmproj" not in retry_cmd
+        assert "-m" in retry_cmd and "--jinja" in retry_cmd
+
+    def test_oom_does_not_retry_text_only(self):
+        # An OOM with --mmproj present must NOT be treated as a projector
+        # problem: load_model errors out instead of dropping vision.
+        assert _detect(_OOM_OUT) is False
diff --git a/studio/backend/tests/test_llama_cpp_mtp_detection.py b/studio/backend/tests/test_llama_cpp_mtp_detection.py
index 784ab1b259..b00cd7169e 100644
--- a/studio/backend/tests/test_llama_cpp_mtp_detection.py
+++ b/studio/backend/tests/test_llama_cpp_mtp_detection.py
@@ -1056,8 +1056,8 @@ _SUB_3B_MTP_MODEL = "unsloth/Qwen3.5-0.8B-MTP-GGUF"
         ("mtp", False, _MTP_MODEL, "draft-mtp", "3", False),
         # ── mtp forced on sub-3B: engage anyway ──
         ("mtp", True, _SUB_3B_MTP_MODEL, "draft-mtp", "2", False),
-        # ── mtp forced on non-MTP: engage anyway ──
-        ("mtp", True, _NON_MTP_MODEL, "draft-mtp", "2", False),
+        # ── mtp forced on non-MTP: default back (no head/drafter) ──
+        ("mtp", True, _NON_MTP_MODEL, None, None, False),
         # ── ngram forced: ngram-mod alone on BOTH platforms ──
         ("ngram", True, _MTP_MODEL, "ngram-mod", None, True),
         ("ngram", False, _MTP_MODEL, "ngram-mod", None, True),
@@ -1066,6 +1066,8 @@ _SUB_3B_MTP_MODEL = "unsloth/Qwen3.5-0.8B-MTP-GGUF"
         ("mtp+ngram", True, _MTP_MODEL, "ngram-mod,draft-mtp", "2", True),
         ("mtp+ngram", False, _MTP_MODEL, "ngram-mod,draft-mtp", "3", True),
         ("mtp+ngram", True, _SUB_3B_MTP_MODEL, "ngram-mod,draft-mtp", "2", True),
+        # ── mtp+ngram forced on non-MTP: keep ngram, drop draft-mtp ──
+        ("mtp+ngram", True, _NON_MTP_MODEL, "ngram-mod", None, True),
         # ── off: nothing emitted ──
         ("off", True, _MTP_MODEL, None, None, False),
         ("off", False, _MTP_MODEL, None, None, False),
@@ -1178,3 +1180,281 @@ def test_build_speculative_flags_mtp_token_missing_logs_and_skips(monkeypatch):
     # choice is still reflected in _requested_spec_mode.
     assert backend.requested_spec_mode == "mtp"
     assert backend.speculative_type is None
+
+
+def test_forced_mtp_on_non_mtp_model_defaults_back(monkeypatch):
+    # Forcing MTP on a model with no head/drafter must NOT emit draft-mtp:
+    # llama-server aborts on it ("failed to measure MTP context memory")
+    # rather than no-op'ing. Default back to --spec-default instead.
+    backend = _resolver_backend(monkeypatch)
+    flags = backend._build_speculative_flags(
+        speculative_type = "mtp",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _NON_MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    assert "--spec-type" not in flags
+    assert "--spec-default" in flags
+    assert backend.speculative_type == "default"
+    assert backend.requested_spec_mode == "mtp"
+
+
+def test_forced_mtp_ngram_on_non_mtp_model_keeps_ngram(monkeypatch):
+    # mtp+ngram on a non-MTP model drops the doomed draft-mtp chain but keeps
+    # the ngram half, which needs no head.
+    backend = _resolver_backend(monkeypatch)
+    flags = backend._build_speculative_flags(
+        speculative_type = "mtp+ngram",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _NON_MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    parsed = _flags_dict(flags)
+    assert parsed.get("--spec-type") == "ngram-mod"
+    assert backend.speculative_type == "ngram-mod"
+    assert backend.requested_spec_mode == "mtp+ngram"
+
+
+# ── Full named-repo resolver matrix (the shipping Studio families) ─────
+#
+# Locks auto / off / forced-mtp routing for every Qwen3.5 (MTP + plain) and
+# gemma-4 (regular + QAT) GGUF repo, including the giant MoEs that stay
+# resolver-only (122B-A10B / 397B-A17B). Expectations are derived from the
+# same signals load_model uses -- _extract_model_size_b (active>effective>
+# total, so E2B->2, A3B->3, A10B->10, A17B->17), _is_mtp_model_name, and the
+# separate-drafter flag -- so each row mirrors what the loader emits on a
+# B200 (GPU default, n=2). gemma carries no -MTP marker; its MTP comes from
+# the root mtp-*.gguf drafter, modelled here by passing mtp_draft_path.
+#
+# auto_spec: "draft-mtp" = head/drafter engaged (>=3B MTP, or any size with a
+# separate drafter); "ngram-mod" = embedded sub-3B drop (zero-VRAM); None =
+# non-MTP -> llama-server --spec-default.
+
+_GEMMA_DRAFTER = "/snap/mtp-gemma-4-it.gguf"  # stand-in separate drafter
+
+_REAL_REPO_MATRIX = [
+    # repo, drafter, auto_spec, auto_ngram_knobs
+    ("unsloth/Qwen3.5-0.8B-MTP-GGUF", None, "ngram-mod", True),
+    ("unsloth/Qwen3.5-2B-MTP-GGUF", None, "ngram-mod", True),
+    ("unsloth/Qwen3.5-4B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-9B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-27B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-35B-A3B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-122B-A10B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-397B-A17B-MTP-GGUF", None, "draft-mtp", False),
+    ("unsloth/Qwen3.5-0.8B-GGUF", None, None, False),
+    ("unsloth/Qwen3.5-2B-GGUF", None, None, False),
+    ("unsloth/Qwen3.5-4B-GGUF", None, None, False),
+    ("unsloth/Qwen3.5-9B-GGUF", None, None, False),
+    # E2B is 2B but ships a separate drafter -> exempt from the sub-3B drop.
+    ("unsloth/gemma-4-E2B-it-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-E4B-it-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-12b-it-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-26B-A4B-it-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-31B-it-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-E2B-it-qat-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-E4B-it-qat-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-12b-it-qat-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-26B-A4B-it-qat-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+    ("unsloth/gemma-4-31B-it-qat-GGUF", _GEMMA_DRAFTER, "draft-mtp", False),
+]
+
+
+def _resolve_real(monkeypatch, repo, drafter, mode):
+    backend = _resolver_backend(monkeypatch)
+    flags = backend._build_speculative_flags(
+        speculative_type = mode,
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = repo,
+        model_path = None,
+        gpus = True,  # B200 default
+        binary = "/fake/llama-server",
+        mtp_draft_path = drafter,
+    )
+    return backend, flags, _flags_dict(flags)
+
+
+@pytest.mark.parametrize(
+    "repo, drafter, auto_spec, auto_ngram_knobs",
+    _REAL_REPO_MATRIX,
+    ids = [r[0].split("/")[-1] for r in _REAL_REPO_MATRIX],
+)
+def test_real_repo_auto_routing(monkeypatch, repo, drafter, auto_spec, auto_ngram_knobs):
+    # Auto is the default mode the dropdown ships with.
+    backend, flags, parsed = _resolve_real(monkeypatch, repo, drafter, "auto")
+    if auto_spec is None:
+        # Non-MTP: no draft-mtp, hand off to llama-server's own default.
+        assert "--spec-type" not in parsed
+        assert "--spec-default" in flags
+        assert backend.speculative_type == "default"
+    elif auto_spec == "draft-mtp":
+        assert parsed.get("--spec-type") == "draft-mtp"
+        assert parsed.get("--spec-draft-n-max") == "2"
+        assert backend.speculative_type == "draft-mtp"
+        # gemma ships a separate drafter; Qwen bakes the head into the GGUF.
+        assert (
+            (parsed.get("--model-draft") == drafter) if drafter else ("--model-draft" not in parsed)
+        )
+    else:  # ngram-mod (sub-3B MTP drop)
+        assert parsed.get("--spec-type") == "ngram-mod"
+        assert "--model-draft" not in parsed  # draft head dropped
+        assert backend.speculative_type == "ngram-mod"
+    if auto_ngram_knobs:
+        assert "--spec-ngram-mod-n-match" in parsed
+    assert backend.requested_spec_mode == "auto"
+
+
+@pytest.mark.parametrize(
+    "repo, drafter",
+    [(r[0], r[1]) for r in _REAL_REPO_MATRIX],
+    ids = [r[0].split("/")[-1] for r in _REAL_REPO_MATRIX],
+)
+def test_real_repo_off_emits_nothing(monkeypatch, repo, drafter):
+    # Off must suppress speculative decoding for every family.
+    backend, flags, _ = _resolve_real(monkeypatch, repo, drafter, "off")
+    assert flags == []
+    assert backend.speculative_type is None
+    assert backend.requested_spec_mode == "off"
+
+
+@pytest.mark.parametrize(
+    "repo, drafter",
+    [(r[0], r[1]) for r in _REAL_REPO_MATRIX],
+    ids = [r[0].split("/")[-1] for r in _REAL_REPO_MATRIX],
+)
+def test_real_repo_forced_mtp_never_aborts(monkeypatch, repo, drafter):
+    # Forcing MTP on the dropdown: real MTP models (name marker or separate
+    # drafter) engage draft-mtp even below 3B; non-MTP models default back to
+    # --spec-default instead of emitting a draft-mtp llama-server will abort on.
+    backend, flags, parsed = _resolve_real(monkeypatch, repo, drafter, "mtp")
+    is_real_mtp = _is_mtp_model_name(repo) or bool(drafter)
+    if is_real_mtp:
+        assert parsed.get("--spec-type") == "draft-mtp"
+        assert backend.speculative_type == "draft-mtp"
+        assert (
+            (parsed.get("--model-draft") == drafter) if drafter else ("--model-draft" not in parsed)
+        )
+    else:
+        assert "--spec-type" not in parsed
+        assert "--spec-default" in flags
+        assert backend.speculative_type == "default"
+    assert backend.requested_spec_mode == "mtp"
+
+
+# ── Sub-3B separate-drafter exemption (Gemma) ─────────────────────────
+#
+# The sub-3B MTP drop is an embedded-head cost (Qwen). A separate drafter
+# (Gemma's root mtp-*.gguf) is a cheap standalone model that wins below 3B
+# (B200 Q4_K_XL: gemma-4-E2B draft-mtp n=2 = 1.21x vs OFF), so it is exempt.
+
+
+def test_sub3b_gemma_separate_drafter_engages_mtp(monkeypatch):
+    backend = _resolver_backend(monkeypatch)
+    flags = backend._build_speculative_flags(
+        speculative_type = "auto",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = "unsloth/gemma-4-E2B-it-GGUF",  # 2B
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+        mtp_draft_path = "/snap/mtp-gemma-4-E2B-it.gguf",  # separate drafter
+    )
+    parsed = _flags_dict(flags)
+    assert parsed.get("--spec-type") == "draft-mtp"
+    assert parsed.get("--model-draft") == "/snap/mtp-gemma-4-E2B-it.gguf"
+    assert "--spec-ngram-mod-n-match" not in parsed
+    assert backend.speculative_type == "draft-mtp"
+
+
+def test_sub3b_qwen_embedded_head_still_drops_to_ngram(monkeypatch):
+    backend = _resolver_backend(monkeypatch)
+    flags = backend._build_speculative_flags(
+        speculative_type = "auto",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = "unsloth/Qwen3.5-2B-MTP-GGUF",  # 2B, embedded head
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+        mtp_draft_path = None,  # no separate drafter
+    )
+    parsed = _flags_dict(flags)
+    assert parsed.get("--spec-type") == "ngram-mod"
+    assert "--model-draft" not in parsed
+    assert backend.speculative_type == "ngram-mod"
+
+
+def test_auto_mode_drops_mtp_exempts_separate_drafter():
+    from core.inference.llama_cpp import _auto_mode_drops_mtp
+
+    assert _auto_mode_drops_mtp("auto", 2.0) is True
+    assert _auto_mode_drops_mtp("auto", 2.0, has_separate_drafter = True) is False
+    assert _auto_mode_drops_mtp("auto", 4.0) is False
+    assert _auto_mode_drops_mtp("mtp", 2.0) is False  # forced engages regardless
+
+
+# ── spec_fallback_reason (drives the "update llama.cpp" UI hint) ───────
+
+
+def test_spec_fallback_reason_set_when_binary_lacks_mtp(monkeypatch):
+    # Outdated llama-server with no mtp token: a forced MTP request can't emit
+    # draft-mtp, so record the reason for the UI update affordance.
+    backend = _resolver_backend(monkeypatch, mtp_token = None)
+    backend._build_speculative_flags(
+        speculative_type = "mtp",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    assert backend.spec_fallback_reason == "binary_no_mtp"
+
+
+def test_spec_fallback_reason_none_when_mtp_engages(monkeypatch):
+    backend = _resolver_backend(monkeypatch)
+    backend._build_speculative_flags(
+        speculative_type = "auto",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    assert backend.speculative_type == "draft-mtp"
+    assert backend.spec_fallback_reason is None
+
+
+def test_spec_fallback_reason_reset_on_off(monkeypatch):
+    # A subsequent off load must clear a stale reason.
+    backend = _resolver_backend(monkeypatch, mtp_token = None)
+    backend._build_speculative_flags(
+        speculative_type = "mtp",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    assert backend.spec_fallback_reason == "binary_no_mtp"
+    backend._build_speculative_flags(
+        speculative_type = "off",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+    assert backend.spec_fallback_reason is None
diff --git a/studio/backend/tests/test_llama_cpp_props_readback.py b/studio/backend/tests/test_llama_cpp_props_readback.py
new file mode 100644
index 0000000000..d87c05f2c6
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_props_readback.py
@@ -0,0 +1,254 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the post-launch /props context readback.
+
+llama-server's memory-fit step or --parallel slot split can allocate less
+context than the requested -c while Studio keeps advertising the requested
+value; clients sized to it then die on exceed_context_size_error 400s.
+``_reconcile_effective_ctx_with_server`` must adopt the server's real
+``default_generation_settings.n_ctx`` whenever it is smaller.
+
+Stubbed httpx; no subprocess, GPU, or network. Cross-platform.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import types as _types
+from pathlib import Path
+
+import pytest
+
+# ---------------------------------------------------------------------------
+# Stub heavy/unavailable deps before importing the module under test.
+# Mirrors test_llama_cpp_context_fit.py.
+# ---------------------------------------------------------------------------
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Prefer the real modules so importing this file first cannot poison later
+# test modules with stubs; only stub what the environment genuinely lacks.
+try:
+    import loggers  # noqa: F401
+except ImportError:
+    _loggers_stub = _types.ModuleType("loggers")
+    _loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+    sys.modules.setdefault("loggers", _loggers_stub)
+
+try:
+    import structlog  # noqa: F401
+except ImportError:
+    sys.modules.setdefault("structlog", _types.ModuleType("structlog"))
+
+try:
+    import httpx  # noqa: F401
+except ImportError:
+    _httpx_stub = _types.ModuleType("httpx")
+    for _exc_name in (
+        "ConnectError",
+        "TimeoutException",
+        "ReadTimeout",
+        "ReadError",
+        "RemoteProtocolError",
+        "CloseError",
+        "WriteError",
+        "HTTPError",
+    ):
+        setattr(_httpx_stub, _exc_name, type(_exc_name, (Exception,), {}))
+
+    class _FakeTimeout:
+        def __init__(self, *a, **kw):
+            pass
+
+    _httpx_stub.Timeout = _FakeTimeout
+    _httpx_stub.Client = type(
+        "Client",
+        (),
+        {
+            "__init__": lambda self, **kw: None,
+            "__enter__": lambda self: self,
+            "__exit__": lambda self, *a: None,
+        },
+    )
+    _httpx_stub.get = lambda *a, **kw: (_ for _ in ()).throw(RuntimeError("unstubbed httpx.get"))
+    sys.modules.setdefault("httpx", _httpx_stub)
+
+from core.inference.llama_cpp import LlamaCppBackend
+import core.inference.llama_cpp as llama_cpp_mod
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+class _FakeResponse:
+    def __init__(
+        self,
+        status_code = 200,
+        body = None,
+    ):
+        self.status_code = status_code
+        self._body = body or {}
+
+    def json(self):
+        return self._body
+
+
+def _make_backend(effective_ctx = 98304, port = 51234):
+    inst = LlamaCppBackend.__new__(LlamaCppBackend)
+    inst._port = port
+    inst._effective_context_length = effective_ctx
+    inst._context_length = 262144
+    return inst
+
+
+def _stub_props(
+    monkeypatch,
+    status_code = 200,
+    body = None,
+    exc = None,
+):
+    def fake_get(url, timeout = None):
+        assert url.endswith("/props")
+        if exc is not None:
+            raise exc
+        return _FakeResponse(status_code, body)
+
+    monkeypatch.setattr(llama_cpp_mod.httpx, "get", fake_get, raising = False)
+
+
+# ---------------------------------------------------------------------------
+# _query_server_n_ctx parsing
+# ---------------------------------------------------------------------------
+
+
+def test_query_n_ctx_reads_default_generation_settings(monkeypatch):
+    _stub_props(
+        monkeypatch,
+        body = {"default_generation_settings": {"n_ctx": 67584}},
+    )
+    assert _make_backend()._query_server_n_ctx() == 67584
+
+
+def test_query_n_ctx_non_200_returns_none(monkeypatch):
+    _stub_props(monkeypatch, status_code = 503)
+    assert _make_backend()._query_server_n_ctx() is None
+
+
+def test_query_n_ctx_missing_key_returns_none(monkeypatch):
+    _stub_props(monkeypatch, body = {"default_generation_settings": {}})
+    assert _make_backend()._query_server_n_ctx() is None
+
+
+def test_query_n_ctx_swallows_transport_errors(monkeypatch):
+    _stub_props(monkeypatch, exc = RuntimeError("connection refused"))
+    assert _make_backend()._query_server_n_ctx() is None
+
+
+# ---------------------------------------------------------------------------
+# _reconcile_effective_ctx_with_server decisions
+# ---------------------------------------------------------------------------
+
+
+def test_fit_shrunk_ctx_overwrites_advertised_value(monkeypatch):
+    """The Nick repro: requested/advertised 98304, server really at 67584."""
+    inst = _make_backend(effective_ctx = 98304)
+    _stub_props(
+        monkeypatch,
+        body = {"default_generation_settings": {"n_ctx": 67584}},
+    )
+    inst._reconcile_effective_ctx_with_server()
+    assert inst._effective_context_length == 67584
+    assert inst.context_length == 67584
+
+
+def test_matching_ctx_is_left_alone(monkeypatch):
+    inst = _make_backend(effective_ctx = 98304)
+    _stub_props(
+        monkeypatch,
+        body = {"default_generation_settings": {"n_ctx": 98304}},
+    )
+    inst._reconcile_effective_ctx_with_server()
+    assert inst._effective_context_length == 98304
+
+
+def test_larger_server_ctx_does_not_inflate_advertised_value(monkeypatch):
+    """Never advertise more than the user asked for, even if the server could."""
+    inst = _make_backend(effective_ctx = 32768)
+    _stub_props(
+        monkeypatch,
+        body = {"default_generation_settings": {"n_ctx": 65536}},
+    )
+    inst._reconcile_effective_ctx_with_server()
+    assert inst._effective_context_length == 32768
+
+
+def test_unset_effective_ctx_adopts_server_value(monkeypatch):
+    inst = _make_backend(effective_ctx = None)
+    inst._context_length = None
+    _stub_props(
+        monkeypatch,
+        body = {"default_generation_settings": {"n_ctx": 40960}},
+    )
+    inst._reconcile_effective_ctx_with_server()
+    assert inst._effective_context_length == 40960
+
+
+def test_props_failure_keeps_studio_value(monkeypatch):
+    """A flaky /props must never wipe the computed context."""
+    inst = _make_backend(effective_ctx = 98304)
+    _stub_props(monkeypatch, exc = RuntimeError("boom"))
+    inst._reconcile_effective_ctx_with_server()
+    assert inst._effective_context_length == 98304
+
+
+# ---------------------------------------------------------------------------
+# _ctx_integrity_flags: keep the per-request window equal to the advertised ctx
+# ---------------------------------------------------------------------------
+
+_CAPS_ALL = {"supports_kv_unified": True, "supports_fit_ctx": True}
+_CAPS_NONE = {"supports_kv_unified": False, "supports_fit_ctx": False}
+
+
+def test_kv_unified_added_for_multi_slot():
+    """Explicit --parallel N disables llama-server's auto-slots kv-unified
+    default, splitting -c into per-slot windows of -c/N; Studio must restore
+    the shared pool so one request can use the full advertised context."""
+    flags = LlamaCppBackend._ctx_integrity_flags(4, False, 98304, 98304, _CAPS_ALL)
+    assert "--kv-unified" in flags
+
+
+def test_kv_unified_skipped_for_single_slot_or_old_build():
+    assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags(
+        1, False, 98304, 98304, _CAPS_ALL
+    )
+    assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags(
+        4, False, 98304, 98304, _CAPS_NONE
+    )
+
+
+def test_fit_ctx_floors_explicit_request_under_fit():
+    flags = LlamaCppBackend._ctx_integrity_flags(1, True, 98304, 98304, _CAPS_ALL)
+    assert flags[flags.index("--fit-ctx") + 1] == "98304"
+
+
+def test_fit_ctx_skipped_without_fit_or_explicit_ctx_or_support():
+    assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(
+        1, False, 98304, 98304, _CAPS_ALL
+    )
+    assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(1, True, 0, 262144, _CAPS_ALL)
+    assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(
+        1, True, 98304, 98304, _CAPS_NONE
+    )
+
+
+def test_probe_missing_binary_reports_new_capabilities_false():
+    info = LlamaCppBackend.probe_server_capabilities(binary = "/nonexistent/llama-server")
+    assert info["found"] is False
+    assert info["supports_kv_unified"] is False
+    assert info["supports_fit_ctx"] is False
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
new file mode 100644
index 0000000000..3b8f511a7c
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -0,0 +1,843 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Hermetic tests for the in-app llama.cpp update orchestration.
+
+No network, no real install: the GitHub release lookup and the installer
+subprocess are both monkeypatched. Verifies detection (update_available) and
+the apply flow (job lifecycle, installer invocation, post-swap re-read).
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import time
+from pathlib import Path
+from types import ModuleType
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import utils.llama_cpp_freshness as freshness  # noqa: E402
+import utils.llama_cpp_update as upd  # noqa: E402
+
+MARKER = "UNSLOTH_PREBUILT_INFO.json"
+
+
+class _FakeInstallerPopen:
+    """Stands in for the streamed installer process in _run_update."""
+
+    def __init__(
+        self,
+        cmd,
+        *,
+        returncode = 0,
+        lines = None,
+        on_start = None,
+        captured_kwargs = None,
+        **kwargs,
+    ):
+        if captured_kwargs is not None:
+            captured_kwargs.update(kwargs)
+        if on_start is not None:
+            on_start(list(cmd))
+        self.returncode = returncode
+        self.stdout = iter(lines or [])
+
+    def wait(self):
+        return self.returncode
+
+    def kill(self):
+        pass
+
+
+def _patch_installer_popen(
+    monkeypatch,
+    *,
+    returncode = 0,
+    lines = None,
+    on_start = None,
+    captured_kwargs = None,
+):
+    monkeypatch.setattr(
+        upd.subprocess,
+        "Popen",
+        lambda cmd, **kw: _FakeInstallerPopen(
+            cmd,
+            returncode = returncode,
+            lines = lines,
+            on_start = on_start,
+            captured_kwargs = captured_kwargs,
+            **kw,
+        ),
+    )
+
+
+def _write_install(
+    dir_: Path,
+    tag: str,
+    repo: str = "unslothai/llama.cpp",
+    asset: str | None = None,
+    release_tag: str | None = None,
+) -> str:
+    """Create a fake prebuilt install tree and return the llama-server path.
+
+    ``asset`` is the bundle filename recorded in the marker; omit it to model an
+    older marker that predates asset-based ROCm forwarding (backward compat).
+    ``release_tag`` is the full release tag (e.g. a ``b9596-mix-<sha>`` mix
+    build); defaults to ``tag`` for a plain prebuilt."""
+    bin_dir = dir_ / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    binary = bin_dir / "llama-server"
+    binary.write_text("#!/bin/sh\necho stub\n")
+    marker = {
+        "tag": tag,
+        "release_tag": release_tag or tag,
+        "published_repo": repo,
+        "installed_at_utc": "2020-01-01T00:00:00Z",
+        "bundle_profile": "cuda13-newer",
+        "runtime_line": "cuda13",
+    }
+    if asset is not None:
+        marker["asset"] = asset
+    (dir_ / MARKER).write_text(json.dumps(marker))
+    return str(binary)
+
+
+@pytest.fixture(autouse = True)
+def _clean_state(monkeypatch, tmp_path):
+    freshness.reset_caches()
+    upd._reset_job_for_tests()
+    upd._resolve_memo.clear()
+    # Isolate the freshness disk cache so the suite never writes the real
+    # ~/.unsloth cache (the default when storage_roots can't be imported).
+    monkeypatch.setattr(freshness, "_cache_dir", lambda: tmp_path / ".freshness_cache")
+    # Deterministic markerless paths: no host-pinned binary, no custom dir.
+    monkeypatch.delenv("LLAMA_SERVER_PATH", raising = False)
+    monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
+    # Never hit the network in these tests.
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: None)
+    yield
+    freshness.reset_caches()
+    upd._reset_job_for_tests()
+    upd._resolve_memo.clear()
+
+
+def _no_prebuilt(monkeypatch):
+    """Stub the host prebuilt probe to 'none available' (no source-build offer)."""
+    monkeypatch.setattr(upd, "_resolve_prebuilt_for_host", lambda *, force_refresh = False: None)
+
+
+def _prebuilt(
+    monkeypatch,
+    *,
+    repo = "unslothai/llama.cpp",
+    release_tag = "b9585",
+    llama_tag = None,
+    asset = None,
+):
+    """Stub the host prebuilt probe to report an available prebuilt."""
+    payload = {
+        "prebuilt_available": True,
+        "repo": repo,
+        "release_tag": release_tag,
+        "llama_tag": llama_tag or release_tag,
+        "asset": asset or f"llama-{release_tag}-bin-macos-arm64.tar.gz",
+        "install_kind": "macos-arm64",
+    }
+    monkeypatch.setattr(upd, "_resolve_prebuilt_for_host", lambda *, force_refresh = False: payload)
+
+
+def test_status_no_marker_no_prebuilt(monkeypatch, tmp_path):
+    # No marker AND no prebuilt available for the host -> unsupported (the genuine
+    # source-build-with-nothing-to-offer case).
+    binary = tmp_path / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")  # no marker file alongside
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    _no_prebuilt(monkeypatch)
+    st = upd.get_update_status()
+    assert st["supported"] is False
+    assert st["update_available"] is False
+    assert st["installed_tag"] is None
+
+
+def test_status_source_build_offers_prebuilt(monkeypatch, tmp_path):
+    # Markerless source build with a prebuilt now available for the host: surface
+    # the update. Unknown installed version (source build) is treated as behind.
+    binary = tmp_path / "llama.cpp" / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    _prebuilt(monkeypatch, release_tag = "b9585")
+    monkeypatch.setattr(upd, "_installed_build_number", lambda b: None)
+    st = upd.get_update_status()
+    assert st["supported"] is True
+    assert st["update_available"] is True
+    assert st["source_build"] is True
+    assert st["latest_tag"] == "b9585"
+    assert st["published_repo"] == "unslothai/llama.cpp"
+
+
+def test_status_source_build_compares_llama_tag(monkeypatch, tmp_path):
+    # release_tag may be a fork wrapper (v1.0); compare/display the upstream
+    # llama_tag (b9457) so a source build is not wrongly judged newer.
+    binary = tmp_path / "llama.cpp" / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    _prebuilt(monkeypatch, release_tag = "v1.0", llama_tag = "b9457")
+    monkeypatch.setattr(upd, "_installed_build_number", lambda b: 9000)
+    st = upd.get_update_status()
+    assert st["latest_tag"] == "b9457"  # not the wrapper tag
+    assert st["update_available"] is True  # 9000 < 9457
+
+
+def test_status_source_build_pinned_binary_not_offered(monkeypatch, tmp_path):
+    # LLAMA_SERVER_PATH pins a custom binary outside any llama.cpp dir; an apply
+    # could not take effect, so the button must not surface.
+    binary = tmp_path / "custom" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setenv("LLAMA_SERVER_PATH", str(binary))
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    _prebuilt(monkeypatch)
+    st = upd.get_update_status()
+    assert st["supported"] is False
+    assert st["update_available"] is False
+
+
+def test_llama_install_root_pinned_returns_none(monkeypatch, tmp_path):
+    binary = tmp_path / "custom" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setenv("LLAMA_SERVER_PATH", str(binary))
+    assert upd._llama_install_root(str(binary)) is None
+
+
+def test_status_source_build_suppressed_when_newer(monkeypatch, tmp_path):
+    # A source build already newer than the latest prebuilt is not nagged.
+    binary = tmp_path / "llama.cpp" / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    _prebuilt(monkeypatch, release_tag = "b9518")
+    monkeypatch.setattr(upd, "_installed_build_number", lambda b: 9600)
+    st = upd.get_update_status()
+    assert st["supported"] is True
+    assert st["update_available"] is False
+    assert st["installed_tag"] == "b9600"
+
+
+def test_status_source_build_skips_probe_while_job_runs(monkeypatch, tmp_path):
+    # While the updater swaps the tree, status polls must not exec the binary
+    # being replaced (on Windows that exec can fail the installer's os.replace);
+    # the 3s poller only consumes job progress.
+    binary = tmp_path / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    probes = {"resolve": 0, "version": 0}
+
+    def _count_resolve(*, force_refresh = False):
+        probes["resolve"] += 1
+        return None
+
+    def _count_version(b):
+        probes["version"] += 1
+        return None
+
+    monkeypatch.setattr(upd, "_resolve_prebuilt_for_host", _count_resolve)
+    monkeypatch.setattr(upd, "_installed_build_number", _count_version)
+    with upd._job_lock:
+        upd._job["state"] = upd._JOB_RUNNING
+    st = upd.get_update_status()
+    assert st["job"]["state"] == "running"
+    assert probes == {"resolve": 0, "version": 0}
+
+
+def test_status_update_available(monkeypatch, tmp_path):
+    binary = _write_install(tmp_path, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    st = upd.get_update_status(force_refresh = True)
+    assert st["supported"] is True
+    assert st["installed_tag"] == "b9493"
+    assert st["latest_tag"] == "b9518"
+    assert st["update_available"] is True
+
+
+def test_status_up_to_date(monkeypatch, tmp_path):
+    binary = _write_install(tmp_path, "b9518")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    st = upd.get_update_status(force_refresh = True)
+    assert st["installed_tag"] == "b9518"
+    assert st["latest_tag"] == "b9518"
+    assert st["update_available"] is False
+
+
+def test_start_update_no_marker_no_prebuilt_refuses(monkeypatch, tmp_path):
+    binary = tmp_path / "llama-server"
+    binary.write_text("stub")  # no marker
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    _no_prebuilt(monkeypatch)
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "no_prebuilt_available"
+
+
+def test_start_update_source_build_installs_prebuilt(monkeypatch, tmp_path):
+    # Markerless install + available prebuilt: install in place into the resolved
+    # root, with the asset-derived ROCm forwarding and the resolved repo.
+    install_dir = tmp_path / "llama.cpp"
+    binary = install_dir / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")  # no marker
+    monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    _prebuilt(
+        monkeypatch, repo = "unslothai/llama.cpp", asset = "app-b9585-linux-x64-rocm-gfx110X.tar.gz"
+    )
+
+    captured = {}
+
+    class _Proc:
+        returncode = 0
+        stdout = "installed"
+        stderr = ""
+
+    def _fake_run(cmd, **kwargs):
+        cmd = list(cmd)
+        assert "--version" in cmd  # only status polls still use run()
+        return _Proc()
+
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        _write_install(install_dir, "b9585")  # installer writes the marker
+
+    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
+
+    res = upd.start_update()
+    assert res["started"] is True, res
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        if upd.get_update_status()["job"]["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    cmd = captured["cmd"]
+    assert "--install-dir" in cmd and str(install_dir) in cmd
+    assert "--published-repo" in cmd and "unslothai/llama.cpp" in cmd
+    assert "--llama-tag" in cmd and "latest" in cmd
+    assert cmd[cmd.index("--rocm-gfx") + 1] == "gfx110x"
+    assert "--simple-policy" not in cmd and "--cpu-fallback" not in cmd
+
+
+def test_start_update_happy_path(monkeypatch, tmp_path):
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    captured = {}
+
+    class _Proc:
+        returncode = 0
+        stdout = "installed"
+        stderr = ""
+
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        # Simulate the installer writing a new marker with the latest tag.
+        _write_install(install_dir, "b9518")
+
+    popen_kwargs: dict = {}
+    _patch_installer_popen(
+        monkeypatch,
+        lines = [
+            "[llama-prebuilt] resolving release\n",
+            "Downloading llama.zip:  35.0% (12.0 MiB/35.0 MiB) at 9.0 MiB/s\n",
+            "Downloading llama.zip:  80.0% (28.0 MiB/35.0 MiB) at 9.0 MiB/s\n",
+        ],
+        on_start = _on_start,
+        captured_kwargs = popen_kwargs,
+    )
+
+    res = upd.start_update()
+    assert res["started"] is True
+    assert res["job"]["from_tag"] == "b9493"
+    assert res["job"]["progress"] == 0.0
+
+    # Wait for the background worker.
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+    assert job["to_tag"] == "b9518"
+    # Installer was invoked with the resolved install dir + latest + repo.
+    assert "--install-dir" in captured["cmd"]
+    assert str(install_dir) in captured["cmd"]
+    assert "--llama-tag" in captured["cmd"] and "latest" in captured["cmd"]
+    assert "unslothai/llama.cpp" in captured["cmd"]
+    # Progress lines were parsed and success pins progress at 1.0.
+    assert job["progress"] == 1.0
+    # The worker asks the installer for fine-grained progress milestones.
+    assert popen_kwargs["env"]["UNSLOTH_PROGRESS_PERCENT_STEP"] == "5"
+
+
+def test_start_update_installer_failure_reports_error(monkeypatch, tmp_path):
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    _patch_installer_popen(monkeypatch, returncode = 2, lines = ["boom: network error\n"])
+
+    res = upd.start_update()
+    assert res["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "error"
+    assert "boom" in (job["error"] or "")
+
+
+# --- installer-argument construction (mirrors the post-#5963 setup scripts) ---
+
+
+def test_rocm_install_args_lemonade_gfx():
+    # Lemonade HIP app bundle: gfx family lives in the asset name.
+    assert upd._rocm_install_args("app-b9585-linux-x64-rocm-gfx110X.tar.gz") == [
+        "--rocm-gfx",
+        "gfx110x",
+    ]
+    assert upd._rocm_install_args("app-b9585-windows-x64-rocm-gfx1150.zip") == [
+        "--rocm-gfx",
+        "gfx1150",
+    ]
+
+
+def test_rocm_install_args_fork_version_bundle():
+    # Fork ROCm bundles encode a ROCm version, not a gfx -> forward --has-rocm.
+    assert upd._rocm_install_args("llama-b9334-bin-ubuntu-rocm-6.4-x64.tar.gz") == ["--has-rocm"]
+
+
+def test_rocm_install_args_windows_hip():
+    assert upd._rocm_install_args("llama-b9334-bin-win-hip-radeon-x64.zip") == ["--has-rocm"]
+
+
+def test_rocm_install_args_non_rocm_and_missing():
+    assert upd._rocm_install_args("llama-b9334-bin-ubuntu-x64.tar.gz") == []
+    assert upd._rocm_install_args("app-b9585-linux-x64-cuda13.tar.gz") == []
+    assert upd._rocm_install_args(None) == []
+
+
+def _capture_install_cmd(
+    monkeypatch,
+    tmp_path,
+    *,
+    tag = "b9493",
+    repo = "unslothai/llama.cpp",
+    asset = None,
+    latest = "b9518",
+) -> list:
+    """Run start_update() with the installer subprocess stubbed; return the argv."""
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, tag, repo = repo, asset = asset)
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: latest)
+
+    captured = {}
+
+    class _Proc:
+        returncode = 0
+        stdout = "installed"
+        stderr = ""
+
+    def _fake_run(cmd, **kwargs):
+        cmd = list(cmd)
+        assert "--version" in cmd  # only status polls still use run()
+        return _Proc()
+
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        _write_install(install_dir, latest, repo = repo, asset = asset)
+
+    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
+
+    res = upd.start_update()
+    assert res["started"] is True, res
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        if upd.get_update_status()["job"]["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    return captured.get("cmd", [])
+
+
+def test_install_cmd_rocm_marker_forwards_gfx(monkeypatch, tmp_path):
+    cmd = _capture_install_cmd(
+        monkeypatch, tmp_path, asset = "app-b9585-linux-x64-rocm-gfx110X.tar.gz"
+    )
+    assert "--rocm-gfx" in cmd
+    assert cmd[cmd.index("--rocm-gfx") + 1] == "gfx110x"
+    assert "--has-rocm" not in cmd
+    assert "--cpu-fallback" not in cmd
+    assert "--simple-policy" not in cmd
+    assert "--published-repo" in cmd and "unslothai/llama.cpp" in cmd
+
+
+def test_install_cmd_fork_rocm_marker_forwards_has_rocm(monkeypatch, tmp_path):
+    cmd = _capture_install_cmd(
+        monkeypatch, tmp_path, asset = "llama-b9334-bin-ubuntu-rocm-6.4-x64.tar.gz"
+    )
+    assert "--has-rocm" in cmd
+    assert "--rocm-gfx" not in cmd
+
+
+def test_install_cmd_ggml_cpu_marker_has_no_cpu_fallback(monkeypatch, tmp_path):
+    # CPU installs come from ggml-org. Re-running into the same install-dir/repo
+    # reproduces the same CPU bundle; --cpu-fallback (which force-drops GPU
+    # detection) is reserved for setup.sh's arm64 rescue and must not appear here.
+    cmd = _capture_install_cmd(
+        monkeypatch,
+        tmp_path,
+        repo = "ggml-org/llama.cpp",
+        asset = "llama-b9334-bin-ubuntu-x64.tar.gz",
+    )
+    assert "--cpu-fallback" not in cmd
+    assert "--rocm-gfx" not in cmd
+    assert "--has-rocm" not in cmd
+    assert "--simple-policy" not in cmd
+    assert "--published-repo" in cmd and "ggml-org/llama.cpp" in cmd
+
+
+def test_install_cmd_cuda_marker_minimal_and_backward_compatible(monkeypatch, tmp_path):
+    # Marker without an asset field (older install): no ROCm flags, no crash, and
+    # never the obsolete --simple-policy that #5963 removed from setup.
+    cmd = _capture_install_cmd(monkeypatch, tmp_path, asset = None)
+    assert "--simple-policy" not in cmd
+    assert "--rocm-gfx" not in cmd
+    assert "--has-rocm" not in cmd
+    assert "--cpu-fallback" not in cmd
+
+
+# --- refusal + maintenance-state coordination ---
+
+
+def test_start_update_already_running_refuses(monkeypatch, tmp_path):
+    binary = _write_install(tmp_path / "llama.cpp", "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    with upd._job_lock:
+        upd._job.update(state = upd._JOB_RUNNING)
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "already_running"
+
+
+def test_start_update_installer_missing_refuses(monkeypatch, tmp_path):
+    binary = _write_install(tmp_path / "llama.cpp", "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: None)
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "installer_missing"
+
+
+class _FakeBackend:
+    """Minimal stand-in for LlamaCppBackend's update-coordination surface."""
+
+    def __init__(self):
+        import threading
+
+        self._serial_load_lock = threading.Lock()
+        self._llama_update_in_progress = False
+        self.is_active = True
+        self.unloaded = False
+
+    def unload_model(self):
+        self.unloaded = True
+
+
+def _inject_backend(monkeypatch, backend):
+    routes_pkg = ModuleType("routes")
+    routes_pkg.__path__ = []
+    inference_mod = ModuleType("routes.inference")
+    inference_mod.get_llama_cpp_backend = lambda: backend
+    monkeypatch.setitem(sys.modules, "routes", routes_pkg)
+    monkeypatch.setitem(sys.modules, "routes.inference", inference_mod)
+
+
+def test_update_sets_maintenance_flag_and_unloads(monkeypatch, tmp_path):
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    backend = _FakeBackend()
+    _inject_backend(monkeypatch, backend)
+
+    seen = {}
+
+    def _on_start(cmd):
+        # The maintenance flag must be set while the installer runs.
+        seen["flag_during_install"] = backend._llama_update_in_progress
+        _write_install(install_dir, "b9518")
+
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
+
+    res = upd.start_update()
+    assert res["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        if upd.get_update_status()["job"]["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+
+    assert backend.unloaded is True
+    assert seen.get("flag_during_install") is True
+    # Cleared in the finally so model loads work again after the swap.
+    assert backend._llama_update_in_progress is False
+
+
+def test_update_clears_maintenance_flag_on_installer_failure(monkeypatch, tmp_path):
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    backend = _FakeBackend()
+    _inject_backend(monkeypatch, backend)
+
+    _patch_installer_popen(monkeypatch, returncode = 1, lines = ["boom\n"])
+
+    res = upd.start_update()
+    assert res["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        if upd.get_update_status()["job"]["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert upd.get_update_status()["job"]["state"] == "error"
+    assert backend._llama_update_in_progress is False
+
+
+def test_update_fails_open_when_backend_unavailable(monkeypatch, tmp_path):
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    def _raise():
+        raise RuntimeError("no backend")
+
+    inference_mod = ModuleType("routes.inference")
+    inference_mod.get_llama_cpp_backend = lambda: _raise()
+    routes_pkg = ModuleType("routes")
+    routes_pkg.__path__ = []
+    monkeypatch.setitem(sys.modules, "routes", routes_pkg)
+    monkeypatch.setitem(sys.modules, "routes.inference", inference_mod)
+
+    _patch_installer_popen(monkeypatch, on_start = lambda cmd: _write_install(install_dir, "b9518"))
+
+    res = upd.start_update()
+    assert res["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+
+
+# --- markerless helper units ---
+
+
+def test_resolve_prebuilt_parses_and_caches(monkeypatch, tmp_path):
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    calls = {"n": 0}
+
+    class _Proc:
+        returncode = 0
+        # stderr noise plus the JSON line on stdout (installer logs to stderr).
+        stdout = (
+            '{"prebuilt_available": true, "repo": "unslothai/llama.cpp", "release_tag": "b9585"}'
+        )
+        stderr = "[llama-prebuilt] some log\n"
+
+    def _fake_run(cmd, **kwargs):
+        calls["n"] += 1
+        assert "--resolve-prebuilt" in cmd
+        return _Proc()
+
+    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    res = upd._resolve_prebuilt_for_host()
+    assert res["prebuilt_available"] is True and res["release_tag"] == "b9585"
+    # Second call is memoized (no second subprocess).
+    upd._resolve_prebuilt_for_host()
+    assert calls["n"] == 1
+
+
+def test_resolve_prebuilt_fails_open(monkeypatch, tmp_path):
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+
+    def _boom(cmd, **kwargs):
+        raise OSError("subprocess failed")
+
+    monkeypatch.setattr(upd.subprocess, "run", _boom)
+    assert upd._resolve_prebuilt_for_host() is None
+    # Failures are not cached: a later success is observed.
+
+    class _Proc:
+        returncode = 0
+        stdout = '{"prebuilt_available": false}'
+        stderr = ""
+
+    monkeypatch.setattr(upd.subprocess, "run", lambda cmd, **kw: _Proc())
+    assert upd._resolve_prebuilt_for_host() == {"prebuilt_available": False}
+
+
+def test_installed_build_number(monkeypatch):
+    def _ver(text):
+        class _Proc:
+            returncode = 0
+            stdout = ""
+            stderr = text
+
+        monkeypatch.setattr(upd.subprocess, "run", lambda cmd, **kw: _Proc())
+        return upd._installed_build_number("/bin/llama-server")
+
+    assert _ver("version: 9585 (abc1234)\nbuilt with clang\n") == 9585
+    assert _ver("version: 1 (deadbee)\n") is None  # source build without tags
+    assert _ver("no version here") is None
+    assert upd._installed_build_number(None) is None
+
+
+def test_llama_install_root_finds_llama_cpp_ancestor(monkeypatch, tmp_path):
+    root = tmp_path / "llama.cpp"
+    binary = root / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
+    assert upd._llama_install_root(str(binary)) == root
+
+
+def test_llama_install_root_unmanaged_path_returns_none(monkeypatch, tmp_path):
+    # A binary on PATH (no marker, no env pin, no llama.cpp ancestor) is foreign:
+    # installing elsewhere would not replace it, so report no manageable root.
+    binary = tmp_path / "usr" / "local" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
+    assert upd._llama_install_root(str(binary)) is None
+
+
+def test_llama_install_root_unsloth_env_dir(monkeypatch, tmp_path):
+    # UNSLOTH_LLAMA_CPP_PATH dir holding the active binary is the managed root.
+    root = tmp_path / "vendor" / "llama"
+    binary = root / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setenv("UNSLOTH_LLAMA_CPP_PATH", str(root))
+    assert upd._llama_install_root(str(binary)) == root
+
+
+def test_llama_install_root_ignores_inactive_env_root(monkeypatch, tmp_path):
+    # UNSLOTH_LLAMA_CPP_PATH set but the active binary is not under it: do not
+    # target the stale env root, resolve from the binary's own llama.cpp tree.
+    inactive = tmp_path / "custom-empty"
+    inactive.mkdir()
+    active = tmp_path / "llama.cpp"
+    binary = active / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setenv("UNSLOTH_LLAMA_CPP_PATH", str(inactive))
+    assert upd._llama_install_root(str(binary)) == active
+
+
+def test_llama_install_root_refuses_pinned_checkout_under_llama_cpp(monkeypatch, tmp_path):
+    # The LLAMA_SERVER_PATH pin guard must run before the ancestor scan, or a
+    # user's own llama.cpp checkout could be handed to the installer.
+    root = tmp_path / "my-project" / "llama.cpp"
+    binary = root / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    monkeypatch.setenv("LLAMA_SERVER_PATH", str(binary))
+    monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
+    assert upd._llama_install_root(str(binary)) is None
+
+
+def test_start_update_source_build_refuses_when_newer(monkeypatch, tmp_path):
+    # A direct POST on a source build already newer than the prebuilt must not
+    # downgrade it; start_update mirrors the detection suppression.
+    install_dir = tmp_path / "llama.cpp"
+    binary = install_dir / "build" / "bin" / "llama-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")  # no marker
+    monkeypatch.setattr(upd, "_find_binary", lambda: str(binary))
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    _prebuilt(monkeypatch, release_tag = "b9518")
+    monkeypatch.setattr(upd, "_installed_build_number", lambda b: 9600)
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "up_to_date"
+
+
+# --- mix-tag detection + apply guard (the reported banner bug) ---
+
+
+def test_status_not_offered_on_mix_latest(monkeypatch, tmp_path):
+    # Installed the mix latest; GitHub latest is that same full tag -> no banner.
+    binary = _write_install(tmp_path / "llama.cpp", "b9596", release_tag = "b9596-mix-e6f2453")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    st = upd.get_update_status()
+    assert st["update_available"] is False
+    assert st["installed_tag"] == "b9596"
+    assert st["latest_tag"] == "b9596-mix-e6f2453"
+
+
+def test_status_not_offered_when_latest_lags(monkeypatch, tmp_path):
+    # A lagging latest (older build than installed) must never be offered.
+    binary = _write_install(tmp_path / "llama.cpp", "b9585")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    st = upd.get_update_status()
+    assert st["update_available"] is False
+
+
+def test_start_update_marked_refuses_when_not_behind(monkeypatch, tmp_path):
+    # A direct POST / stale banner must not reinstall when already on the latest.
+    binary = _write_install(tmp_path / "llama.cpp", "b9596", release_tag = "b9596-mix-e6f2453")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "up_to_date"
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_health.py b/studio/backend/tests/test_llama_cpp_wait_for_health.py
index 33f9e9d803..1ba6c9f7b5 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_health.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_health.py
@@ -145,3 +145,111 @@ class TestWaitForHealthResilience:
         monkeypatch.setattr(httpx, "get", should_not_be_called)
         assert b._wait_for_health(timeout = 5.0, interval = 0.01) is False
         assert called["n"] == 0
+
+
+class TestCrashLogTail:
+    """The "exited with code X" log must keep the TAIL of the output.
+
+    Crash diagnostics (abort reason, ROCm/CUDA error text) print last,
+    after the long startup banner; head truncation has cut off exactly
+    the diagnostic line in field reports (gfx1151 fit-step abort)."""
+
+    @staticmethod
+    def _capture_error_logs(monkeypatch) -> list:
+        """Capture module-logger .error() messages directly -- immune to
+        whatever logging/structlog config sibling test modules installed."""
+        import core.inference.llama_cpp as _llama_mod
+
+        records: list = []
+        fake_logger = mock.Mock()
+        fake_logger.error = mock.Mock(side_effect = lambda msg, *a, **k: records.append(msg))
+        monkeypatch.setattr(_llama_mod, "logger", fake_logger)
+        return records
+
+    def test_crash_log_keeps_tail_not_head(self, monkeypatch):
+        records = self._capture_error_logs(monkeypatch)
+        b = _make_backend()
+        b._process.poll.return_value = 1
+        b._process.returncode = 1
+        # >2000 chars of banner, diagnostic on the final line.
+        banner = [f"load_model: tensor blk.{i} buffer ROCm0" for i in range(80)]
+        diagnostic = "ggml-cuda.cu:103: ROCm error: out of memory"
+        b._stdout_lines = banner + [diagnostic]
+
+        assert b._wait_for_health(timeout = 1.0, interval = 0.01) is False
+
+        crash_logs = [m for m in records if "exited with code" in m]
+        assert crash_logs, "crash must produce an exited-with-code log"
+        assert diagnostic in crash_logs[-1]
+        assert "Output (tail)" in crash_logs[-1]
+        # The head of the banner must be the part sacrificed to truncation.
+        assert "blk.0 buffer" not in crash_logs[-1]
+
+    def test_crash_log_mentions_log_file_when_present(self, monkeypatch):
+        records = self._capture_error_logs(monkeypatch)
+        b = _make_backend()
+        b._process.poll.return_value = 1
+        b._process.returncode = 1
+        b._stdout_lines = ["boom"]
+        b._llama_log_path = Path("C:/logs/llama-123-port-1234.log")
+
+        assert b._wait_for_health(timeout = 1.0, interval = 0.01) is False
+
+        crash_logs = [m for m in records if "exited with code" in m]
+        assert crash_logs and "llama-123-port-1234.log" in crash_logs[-1]
+
+
+class TestRetryLogFilenameUnique:
+    """The --fit off retry can respawn within the same epoch second; the log
+    filename must carry the attempt index or the second open ("w") truncates
+    the crash log the retry warning just referenced (found by simulation:
+    frozen time.time -> single file, crash evidence gone)."""
+
+    def test_log_name_includes_attempt_index(self):
+        src = (
+            Path(__file__).resolve().parent.parent / "core" / "inference" / "llama_cpp.py"
+        ).read_text(encoding = "utf-8")
+        assert "-try{_spawn_attempt}.log" in src
+
+
+class TestFitOffRetryEligible:
+    """Gate for the one-shot --fit off startup-crash retry.
+
+    Retry only when Studio's own VRAM math placed the model and nothing
+    on the command line chose the fit mode explicitly."""
+
+    def test_eligible_for_plain_ngl_launch(self):
+        cmd = ["llama-server", "-m", "x.gguf", "-ngl", "-1", "--jinja"]
+        assert LlamaCppBackend._fit_off_retry_eligible(cmd, use_fit = False) is True
+
+    def test_not_eligible_when_use_fit(self):
+        cmd = ["llama-server", "-m", "x.gguf", "--fit", "on"]
+        assert LlamaCppBackend._fit_off_retry_eligible(cmd, use_fit = True) is False
+
+    @pytest.mark.parametrize(
+        "fit_args",
+        [
+            ["--fit", "on"],
+            ["--fit", "off"],
+            ["-fit", "off"],
+            ["--fit=on"],
+            ["-fit=off"],
+        ],
+    )
+    def test_not_eligible_with_explicit_fit_flag(self, fit_args):
+        cmd = ["llama-server", "-m", "x.gguf", *fit_args]
+        assert LlamaCppBackend._fit_off_retry_eligible(cmd, use_fit = False) is False
+
+    @pytest.mark.parametrize(
+        "tuning_args",
+        [
+            ["--fit-ctx", "8192"],
+            ["--fit-target", "1024"],
+            ["-fitc", "4096"],
+            ["-fitt", "512"],
+            ["--fit-ctx=8192"],
+        ],
+    )
+    def test_fit_tuning_flags_do_not_block_retry(self, tuning_args):
+        cmd = ["llama-server", "-m", "x.gguf", *tuning_args]
+        assert LlamaCppBackend._fit_off_retry_eligible(cmd, use_fit = False) is True
diff --git a/studio/backend/tests/test_llama_route.py b/studio/backend/tests/test_llama_route.py
new file mode 100644
index 0000000000..bf0c4b731f
--- /dev/null
+++ b/studio/backend/tests/test_llama_route.py
@@ -0,0 +1,111 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""routes/llama.py: the source_build field is exposed and the handlers run the
+(now subprocess-touching) detection off the event loop via a worker thread.
+
+The route file is loaded standalone with a stubbed auth dependency so the test
+does not pull the whole routes package (matplotlib-heavy training router) and
+works in a minimal env.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import importlib.util
+import sys
+import threading
+import types
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+pytest.importorskip("fastapi")
+
+
+def _load_route():
+    # Prefer the real auth module; stub it only in minimal envs where its
+    # deps are absent. Stubs are popped after the load so they never leak
+    # into sys.modules for the rest of the suite.
+    stubbed = []
+    try:
+        import auth.authentication  # noqa: F401
+    except Exception:
+        auth_pkg = types.ModuleType("auth")
+        auth_pkg.__path__ = []
+        auth_mod = types.ModuleType("auth.authentication")
+        auth_mod.get_current_subject = lambda: "test"
+        for name, stub in (("auth", auth_pkg), ("auth.authentication", auth_mod)):
+            if name not in sys.modules:
+                sys.modules[name] = stub
+                stubbed.append(name)
+    try:
+        spec = importlib.util.spec_from_file_location(
+            "llama_route_under_test", str(_BACKEND / "routes" / "llama.py")
+        )
+        mod = importlib.util.module_from_spec(spec)
+        sys.modules["llama_route_under_test"] = mod  # so pydantic resolves forward refs
+        spec.loader.exec_module(mod)
+        return mod
+    finally:
+        for name in stubbed:
+            sys.modules.pop(name, None)
+
+
+rl = _load_route()
+
+
+def test_status_response_exposes_source_build():
+    payload = {
+        "supported": True,
+        "update_available": True,
+        "stale": False,
+        "installed_tag": None,
+        "latest_tag": "b9585",
+        "published_repo": "unslothai/llama.cpp",
+        "installed_at_utc": None,
+        "age_days": None,
+        "source_build": True,
+        "job": {"state": "idle"},
+    }
+    model = rl.LlamaUpdateStatusResponse(**payload)
+    assert model.model_dump()["source_build"] is True
+    # Extra/unknown keys must not crash the response model.
+    rl.LlamaUpdateStatusResponse(**{**payload, "unexpected": 1})
+
+
+def test_status_handler_runs_off_event_loop(monkeypatch):
+    seen = {}
+
+    def fake_status(force_refresh = False):
+        seen["thread"] = threading.current_thread()
+        return {
+            "supported": True,
+            "update_available": True,
+            "source_build": True,
+            "latest_tag": "b9585",
+            "job": {"state": "idle"},
+        }
+
+    monkeypatch.setattr(rl, "get_update_status", fake_status)
+    out = asyncio.run(rl.llama_update_status(force_refresh = False, current_subject = "t"))
+    assert out.source_build is True
+    # Detection ran in a worker thread, not the event-loop thread.
+    assert seen["thread"] is not threading.main_thread()
+
+
+def test_update_handler_runs_off_event_loop(monkeypatch):
+    seen = {}
+
+    def fake_start():
+        seen["thread"] = threading.current_thread()
+        return {"started": True, "reason": None, "job": {"state": "running"}}
+
+    monkeypatch.setattr(rl, "start_update", fake_start)
+    out = asyncio.run(rl.llama_update(current_subject = "t"))
+    assert out.started is True
+    assert seen["thread"] is not threading.main_thread()
diff --git a/studio/backend/tests/test_llama_server_args.py b/studio/backend/tests/test_llama_server_args.py
index 52e3848157..09775707ee 100644
--- a/studio/backend/tests/test_llama_server_args.py
+++ b/studio/backend/tests/test_llama_server_args.py
@@ -507,3 +507,28 @@ def test_strip_shadowing_flags_defaults_strip_everything():
         ["-c", "4096", "--cache-type-k", "q8_0", "--spec-default", "--jinja"]
     )
     assert out == []
+
+
+def test_strip_shadowing_flags_drops_model_draft_with_spec():
+    # --model-draft (and aliases) are Studio-managed since the separate
+    # MTP drafter support: an inherited copy must not last-wins-override
+    # the auto-detected drafter.
+    out = strip_shadowing_flags(
+        ["--model-draft", "/old/mtp.gguf", "-md", "/old2.gguf", "--top-k", "20"],
+        strip_context = False,
+        strip_cache = False,
+        strip_spec = True,
+        strip_template = False,
+    )
+    assert out == ["--top-k", "20"]
+
+
+def test_strip_shadowing_flags_keeps_model_draft_without_spec():
+    out = strip_shadowing_flags(
+        ["--model-draft", "/custom/mtp.gguf"],
+        strip_context = True,
+        strip_cache = False,
+        strip_spec = False,
+        strip_template = False,
+    )
+    assert out == ["--model-draft", "/custom/mtp.gguf"]
diff --git a/studio/backend/tests/test_mcp_config_import.py b/studio/backend/tests/test_mcp_config_import.py
new file mode 100644
index 0000000000..4082733490
--- /dev/null
+++ b/studio/backend/tests/test_mcp_config_import.py
@@ -0,0 +1,341 @@
+"""Tests for MCP config-file import (issue #5936).
+
+Covers the round-trip-safe command join/split inverse (join_stdio_command ↔
+parse_stdio_command, on both posix and win32 using the issue's Windows
+fixtures), the pure config parser (parse_mcp_config), and the POST /import
+route (stdio gate on/off, url dedup, one bad entry not sinking the batch).
+
+Run from studio/backend:  python -m pytest tests/test_mcp_config_import.py -q
+"""
+
+import sys
+
+import pytest
+
+from core.inference import mcp_client
+from core.inference.mcp_config_import import parse_mcp_config
+from storage import mcp_servers_db
+
+
+def _reset_db(tmp_path, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(mcp_servers_db, "_schema_ready", False)
+
+
+def _enable(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", "1")
+
+
+def _disable(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", raising = False)
+
+
+# ── 1. join_stdio_command ↔ parse_stdio_command round-trip ──────────
+
+
+@pytest.mark.parametrize(
+    "parts",
+    [
+        ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
+        ["python", "-m", "mod", "--name", "a b"],
+        ["uvx", "some-server", "--flag"],
+        ["/usr/local/bin/my-server"],
+        ["mcp-server-sqlite"],
+    ],
+)
+def test_join_parse_roundtrip_posix(monkeypatch, parts):
+    monkeypatch.setattr(sys, "platform", "linux")
+    joined = mcp_client.join_stdio_command(parts)
+    assert mcp_client.parse_stdio_command(joined) == parts
+
+
+@pytest.mark.parametrize(
+    "parts",
+    [
+        # Issue #5936's literal Windows examples: absolute .exe with a path, and
+        # backslash drive/dir args must survive the join→split round-trip intact.
+        [
+            "C:\\Users\\user\\Documents\\Office-Word-MCP-Server\\.venv\\Scripts\\python.exe",
+            "C:\\Users\\user\\Documents\\Office-Word-MCP-Server\\word_mcp_server.py",
+        ],
+        [
+            "node",
+            "C:\\Users\\user\\Documents\\DesktopCommanderMCP\\dist\\index.js",
+            "--no-onboarding",
+        ],
+        [
+            "node",
+            "C:\\Users\\user\\AppData\\Roaming\\npm\\node_modules\\@modelcontextprotocol\\server-filesystem\\dist\\index.js",
+            "D:\\",
+            "O:\\",
+        ],
+        # A command path with spaces is the case that actually needs quoting.
+        ["C:\\Program Files\\node\\node.exe", "server.js"],
+        ["C:\\Program Files\\Foo\\", "server.js"],
+        ["C:\\Program Files\\Foo\\", '{"foo":"bar"}'],
+        ["'C:\\Program Files\\node\\node.exe'", "server.js"],
+        ["node", "O'Reilly"],
+        ["node", "C:\\Users\\O'Reilly\\server.js"],
+        ["node", "'draft'"],
+        ["node", "'open", "close'"],
+        ["node", ""],
+    ],
+)
+def test_join_parse_roundtrip_win32(monkeypatch, parts):
+    monkeypatch.setattr(sys, "platform", "win32")
+    joined = mcp_client.join_stdio_command(parts)
+    assert mcp_client.parse_stdio_command(joined) == parts
+
+
+def test_parse_rejects_manual_single_quoted_windows_executable(monkeypatch):
+    monkeypatch.setattr(sys, "platform", "win32")
+    command = "'C:\\Program Files\\node\\node.exe' server.js"
+    with pytest.raises(ValueError):
+        mcp_client.parse_stdio_command(command)
+
+
+def test_parse_windows_apostrophes_as_literals(monkeypatch):
+    monkeypatch.setattr(sys, "platform", "win32")
+    assert mcp_client.parse_stdio_command("node O'Reilly") == ["node", "O'Reilly"]
+    assert mcp_client.parse_stdio_command("node C:\\Users\\O'Reilly\\server.js") == [
+        "node",
+        "C:\\Users\\O'Reilly\\server.js",
+    ]
+    assert mcp_client.parse_stdio_command("node 'draft'") == ["node", "'draft'"]
+    assert mcp_client.parse_stdio_command("node 'open close'") == ["node", "'open", "close'"]
+
+
+def test_parse_rejects_unterminated_windows_double_quote(monkeypatch):
+    monkeypatch.setattr(sys, "platform", "win32")
+    with pytest.raises(ValueError):
+        mcp_client.parse_stdio_command('node "C:\\path with spaces')
+
+
+# ── 2. parse_mcp_config ─────────────────────────────────────────────
+
+
+def test_parse_stdio_entry():
+    cfg = {
+        "mcpServers": {
+            "fs": {
+                "command": "npx",
+                "args": ["-y", "server", "/tmp"],
+                "env": {"K": "v"},
+            }
+        }
+    }
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert len(entries) == 1
+    entry = entries[0]
+    assert entry.display_name == "fs"
+    assert entry.is_stdio is True
+    assert entry.headers == {"K": "v"}
+    assert mcp_client.parse_stdio_command(entry.url) == ["npx", "-y", "server", "/tmp"]
+
+
+def test_parse_remote_entry():
+    cfg = {
+        "mcpServers": {
+            "remote": {
+                "url": "https://example.com/mcp",
+                "headers": {"Authorization": "Bearer x"},
+            }
+        }
+    }
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert entries[0].url == "https://example.com/mcp"
+    assert entries[0].is_stdio is False
+    assert entries[0].headers == {"Authorization": "Bearer x"}
+
+
+def test_parse_preserves_disabled_and_oauth():
+    cfg = {
+        "servers": {
+            "remote": {
+                "type": "http",
+                "url": "https://example.com/mcp",
+                "oauth": {"clientId": "client"},
+                "disabled": True,
+            }
+        }
+    }
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert entries[0].is_enabled is False
+    assert entries[0].use_oauth is True
+
+
+def test_parse_accepts_cline_streamable_http_alias():
+    cfg = {
+        "mcpServers": {
+            "remote": {
+                "type": "streamableHttp",
+                "url": "https://example.com/mcp",
+            }
+        }
+    }
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert entries[0].url == "https://example.com/mcp"
+    assert entries[0].is_stdio is False
+
+
+@pytest.mark.parametrize(
+    "server",
+    [
+        {"command": "node", "args": ["server.js"], "cwd": "/tmp/server"},
+        {"command": "node", "args": ["server.js"], "envFile": ".env"},
+        {"command": "node", "args": ["server.js"], "env": {"API_KEY": "${input:api-key}"}},
+        {"command": "node", "args": ["${workspaceFolder}/server.js"]},
+        {"command": "node", "args": ["server.js"], "env": {"HTTP_PROXY": None}},
+        {"command": "node", "args": ["server.js"], "sandboxEnabled": True},
+        {"url": "https://example.com/mcp", "headers": {"Authorization": "Bearer ${input:token}"}},
+        {"url": "https://example.com/mcp", "headers": {"Authorization": None}},
+        {"type": "http", "url": "https://example.com/sse"},
+        {"type": "http", "url": "https://example.com/sse "},
+        {"type": "streamableHttp", "url": "https://example.com/sse"},
+        {"url": "https://example.com/mcp", "timeout": 120},
+        {"url": "https://example.com/mcp", "timeoutMs": 120000},
+        {"url": "https://example.com/mcp", "timeoutSeconds": 120},
+        {"type": "sse", "url": "https://example.com/custom"},
+    ],
+)
+def test_parse_rejects_unrepresentable_imports(server):
+    entries, errors = parse_mcp_config({"servers": {"bad": server}})
+    assert entries == []
+    assert len(errors) == 1
+
+
+def test_servers_alias_key():
+    # VS Code uses "servers" instead of "mcpServers".
+    cfg = {"servers": {"fs": {"command": "node", "args": ["x.js"]}}}
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert len(entries) == 1
+
+
+def test_env_and_args_values_coerced_to_str():
+    cfg = {"mcpServers": {"fs": {"command": "node", "args": [8080], "env": {"PORT": 8080}}}}
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert entries[0].headers == {"PORT": "8080"}
+    assert mcp_client.parse_stdio_command(entries[0].url) == ["node", "8080"]
+
+
+def test_args_optional():
+    cfg = {"mcpServers": {"sqlite": {"command": "mcp-server-sqlite"}}}
+    entries, errors = parse_mcp_config(cfg)
+    assert errors == []
+    assert entries[0].url == "mcp-server-sqlite"
+    assert entries[0].headers is None
+
+
+def test_bad_entry_does_not_sink_batch():
+    cfg = {
+        "mcpServers": {
+            "good": {"command": "node", "args": ["x.js"]},
+            "both": {"command": "node", "url": "https://x/mcp"},
+            "neither": {"name": "oops"},
+            "bad_args": {"command": "node", "args": "x.js"},
+            "bad_env": {"command": "node", "env": ["NOT", "A", "DICT"]},
+        }
+    }
+    entries, errors = parse_mcp_config(cfg)
+    assert {e.display_name for e in entries} == {"good"}
+    assert len(errors) == 4
+
+
+def test_not_a_dict():
+    entries, errors = parse_mcp_config([])
+    assert entries == []
+    assert len(errors) == 1
+
+
+def test_missing_servers_key():
+    entries, errors = parse_mcp_config({"foo": {}})
+    assert entries == []
+    assert len(errors) == 1
+
+
+def test_servers_alias_error_names_actual_key():
+    entries, errors = parse_mcp_config({"servers": []})
+    assert entries == []
+    assert errors == ["'servers' must be an object mapping name -> server."]
+
+
+# ── 3. POST /import route ───────────────────────────────────────────
+
+
+def test_import_route_creates_and_dedups(tmp_path, monkeypatch):
+    import asyncio
+
+    from models.mcp_servers import McpServerImportRequest
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    cfg = {
+        "mcpServers": {
+            "fs": {
+                "command": "npx",
+                "args": ["-y", "server", "/tmp"],
+                "env": {"API_KEY": "sk"},
+            },
+            "remote": {"url": "https://example.com/mcp"},
+            "oauth": {
+                "type": "http",
+                "url": "https://auth.example.com/mcp",
+                "oauth": {"clientId": "client"},
+            },
+            "disabled": {
+                "url": "https://disabled.example.com/mcp",
+                "disabled": True,
+            },
+        }
+    }
+    res = asyncio.run(
+        routes_mcp.import_mcp_servers(McpServerImportRequest(config = cfg), current_subject = "u")
+    )
+    assert res.errors == []
+    assert res.skipped == []
+    assert {c.display_name for c in res.created} == {"fs", "remote", "oauth", "disabled"}
+    fs = next(c for c in res.created if c.display_name == "fs")
+    assert fs.headers == {"API_KEY": "sk"}
+    assert fs.use_oauth is False
+    assert fs.is_enabled is True
+    oauth = next(c for c in res.created if c.display_name == "oauth")
+    assert oauth.use_oauth is True
+    disabled = next(c for c in res.created if c.display_name == "disabled")
+    assert disabled.is_enabled is False
+
+    # Re-importing the same config skips both by url.
+    res2 = asyncio.run(
+        routes_mcp.import_mcp_servers(McpServerImportRequest(config = cfg), current_subject = "u")
+    )
+    assert res2.created == []
+    assert set(res2.skipped) == {"fs", "remote", "oauth", "disabled"}
+
+
+def test_import_route_gates_stdio_when_disabled(tmp_path, monkeypatch):
+    import asyncio
+
+    from models.mcp_servers import McpServerImportRequest
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    _disable(monkeypatch)
+    cfg = {
+        "mcpServers": {
+            "fs": {"command": "npx", "args": ["server"]},
+            "remote": {"url": "https://example.com/mcp"},
+        }
+    }
+    res = asyncio.run(
+        routes_mcp.import_mcp_servers(McpServerImportRequest(config = cfg), current_subject = "u")
+    )
+    # Remote still imports; the stdio entry is rejected per-entry (gate off).
+    assert {c.display_name for c in res.created} == {"remote"}
+    assert any("fs" in err for err in res.errors)
+    assert len(mcp_servers_db.list_servers()) == 1
diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index 811971ec36..4402031467 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -63,6 +63,10 @@ _worker = _load_worker_module()
 _normalize_mlx_studio_optimizer = _worker._normalize_mlx_studio_optimizer
 _normalize_mlx_studio_scheduler = _worker._normalize_mlx_studio_scheduler
 _mlx_vlm_max_resized_size = _worker._mlx_vlm_max_resized_size
+_mlx_vlm_resized_image_layout = _worker._mlx_vlm_resized_image_layout
+_copy_mlx_vlm_image_processor = _worker._copy_mlx_vlm_image_processor
+_resize_mlx_vlm_image = _worker._resize_mlx_vlm_image
+_adapt_for_mlx_vlm = _worker._adapt_for_mlx_vlm
 
 
 def test_mlx_studio_optimizer_aliases_are_explicit():
@@ -90,3 +94,123 @@ def test_mlx_vlm_resize_uses_max_dimension_like_torch_trainer():
     # Half-pixel cases must match the Torch collator (not banker's round).
     assert _mlx_vlm_max_resized_size(333, 1000, 500) == (167, 500)
     assert _mlx_vlm_max_resized_size(1000, 333, 500) == (500, 167)
+
+
+def test_mlx_vlm_resize_keeps_default_numpy_layout_hwc():
+    Image = pytest.importorskip("PIL.Image")
+    image = Image.new("RGB", (320, 200), color = (10, 20, 30))
+
+    resized = _resize_mlx_vlm_image(image, 128)
+
+    assert resized.shape == (80, 128, 3)
+    assert resized.flags.c_contiguous
+
+
+def test_mlx_vlm_resize_uses_requested_chw_numpy_layout():
+    Image = pytest.importorskip("PIL.Image")
+    image = Image.new("RGB", (320, 200), color = (10, 20, 30))
+
+    resized = _resize_mlx_vlm_image(image, 128, image_layout = "chw")
+
+    assert resized.shape == (3, 80, 128)
+    assert resized.flags.c_contiguous
+
+
+def test_mlx_vlm_resized_image_layout_probes_processor_contract():
+    class ChwOnlyImageProcessor:
+        def __call__(self, images = None):
+            image = images[0]
+            if image.shape[0] == 3:
+                return {"pixel_values": image}
+            raise ValueError("expected CHW")
+
+    class HwcImageProcessor:
+        def __call__(self, images = None):
+            image = images[0]
+            if image.shape[-1] == 3:
+                return {"pixel_values": image}
+            raise ValueError("expected HWC")
+
+    assert (
+        _mlx_vlm_resized_image_layout(
+            types.SimpleNamespace(image_processor = ChwOnlyImageProcessor())
+        )
+        == "chw"
+    )
+    assert (
+        _mlx_vlm_resized_image_layout(types.SimpleNamespace(image_processor = HwcImageProcessor()))
+        is None
+    )
+
+
+def test_mlx_vlm_layout_probe_copies_image_processor():
+    class StatefulImageProcessor:
+        def __init__(self):
+            self.calls = 0
+
+        def __call__(self, images = None):
+            self.calls += 1
+            image = images[0]
+            if image.shape[0] == 3:
+                return {"pixel_values": image}
+            raise ValueError("expected CHW")
+
+    image_processor = StatefulImageProcessor()
+
+    layout = _mlx_vlm_resized_image_layout(types.SimpleNamespace(image_processor = image_processor))
+
+    assert layout == "chw"
+    assert image_processor.calls == 0
+
+
+def test_mlx_vlm_image_processor_copy_refuses_uncopyable_processors():
+    class UncopyableImageProcessor:
+        def __copy__(self):
+            raise RuntimeError("no copy")
+
+        def __deepcopy__(self, _memo):
+            raise RuntimeError("no deepcopy")
+
+    image_processor = UncopyableImageProcessor()
+
+    assert _copy_mlx_vlm_image_processor(image_processor) is None
+
+
+def test_mlx_vlm_layout_probe_skips_uncopyable_processors():
+    class UncopyableImageProcessor:
+        def __copy__(self):
+            raise RuntimeError("no copy")
+
+        def __deepcopy__(self, _memo):
+            raise RuntimeError("no deepcopy")
+
+        def __call__(self, images = None):
+            raise AssertionError("live processor should not be probed")
+
+    assert (
+        _mlx_vlm_resized_image_layout(
+            types.SimpleNamespace(image_processor = UncopyableImageProcessor())
+        )
+        is None
+    )
+
+
+def test_mlx_vlm_adapter_applies_chw_layout_to_message_images():
+    Image = pytest.importorskip("PIL.Image")
+    image = Image.new("RGB", (320, 200), color = (10, 20, 30))
+    item = {
+        "messages": [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "image": image},
+                    {"type": "text", "text": "Describe it."},
+                ],
+            }
+        ]
+    }
+
+    adapted = _adapt_for_mlx_vlm([item], resize = 128, image_layout = "chw")
+
+    assert adapted[0]["image"].shape == (3, 80, 128)
+    assert adapted[0]["messages"][0]["content"][0] == {"type": "image"}
diff --git a/studio/backend/tests/test_mtp_drafter_companion.py b/studio/backend/tests/test_mtp_drafter_companion.py
new file mode 100644
index 0000000000..d5e8d13652
--- /dev/null
+++ b/studio/backend/tests/test_mtp_drafter_companion.py
@@ -0,0 +1,276 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Separate-file MTP drafter (Gemma 4) contracts.
+
+Pins: the drafter-path predicate and its two layering mirrors, Gemma
+effective-size extraction, companion classification in variant plans
+(including resume from pre-fix manifests where the drafter leaked into a
+quant's main files), and local drafter detection / self-pairing rejection.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from hub.utils.download_manifest import ExpectedFile
+from hub.utils.gguf import is_mtp_drafter_path
+from hub.utils.gguf_plan import build_gguf_variant_plans, plan_from_expected_files
+from utils.models.model_config import (
+    _is_mtp_drafter,
+    detect_gguf_model,
+    detect_mtp_file,
+    extract_model_size_b,
+)
+
+
+# ── Predicate + layering mirrors ─────────────────────────────────────
+
+DRAFTER_CASES = [
+    ("mtp-gemma-4-12b-it.gguf", True),
+    ("MTP/gemma-4-12b-it-Q8_0-MTP.gguf", True),
+    ("foo/MTP/bar.gguf", True),
+    ("gemma-4-12b-it-Q8_0.gguf", False),
+    # Baked-in Qwen MTP repos: the head is inside the main GGUF, the file
+    # IS the model -- must never be classified as a companion.
+    ("Qwen3.6-27B-MTP-Q4_K_M.gguf", False),
+    ("prompt-mtp-test.gguf", False),
+    ("smtp/model.gguf", False),
+    ("mtp-readme.txt", False),
+]
+
+
+@pytest.mark.parametrize("path,expected", DRAFTER_CASES)
+def test_drafter_predicate_and_mirrors_agree(path, expected):
+    from core.inference.llama_cpp import _is_companion_gguf_path
+
+    assert is_mtp_drafter_path(path) is expected
+    assert _is_mtp_drafter(path) is expected
+    # The core mirror bundles mmproj; none of these inputs are mmproj, so
+    # it must agree with the canonical predicate.
+    assert _is_companion_gguf_path(path) is expected
+
+
+# ── Gemma effective-size extraction ──────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "model_id,size_b",
+    [
+        ("unsloth/gemma-4-E2B-it-GGUF", 2.0),
+        ("unsloth/gemma-4-E4B-it", 4.0),
+        ("unsloth/gemma-3n-E4B-it", 4.0),
+        # MoE active params beat effective and total notation.
+        ("unsloth/Qwen3.5-35B-A3B", 3.0),
+        ("unsloth/gemma-4-12b-it-GGUF", 12.0),
+        ("unsloth/Qwen3.5-9B-MTP-GGUF", 9.0),
+        ("no-size-here", None),
+    ],
+)
+def test_extract_model_size_b(model_id, size_b):
+    assert extract_model_size_b(model_id) == size_b
+
+
+# ── Variant plan companion classification ────────────────────────────
+
+
+def _sib(name: str, size: int, sha: str):
+    return SimpleNamespace(rfilename = name, size = size, lfs = {"sha256": sha})
+
+
+GEMMA_SIBLINGS = [
+    _sib("gemma-4-12b-it-Q4_K_M.gguf", 4_000, "main-q4"),
+    _sib("gemma-4-12b-it-Q8_0.gguf", 8_000, "main-q8"),
+    _sib("mtp-gemma-4-12b-it.gguf", 100, "drafter"),
+    _sib("MTP/gemma-4-12b-it-Q8_0-MTP.gguf", 100, "mtp-sub-q8"),
+    _sib("MTP/gemma-4-12b-it-BF16-MTP.gguf", 200, "mtp-sub-bf16"),
+    _sib("mmproj-F16.gguf", 500, "mmproj"),
+]
+
+
+def test_variant_plans_carry_drafter_as_companion():
+    plans = build_gguf_variant_plans(GEMMA_SIBLINGS)
+
+    # No phantom quants from the drafter's Q8_0 label or the MTP/ copies.
+    assert set(plans) == {"q4_k_m", "q8_0"}
+    for plan in plans.values():
+        assert "mtp-gemma-4-12b-it.gguf" in plan.target_filenames
+        assert not any("MTP/" in name for name in plan.target_filenames)
+        assert "drafter" in plan.companion_hashes
+        assert "drafter" not in plan.main_hashes
+        assert plan.mmproj_filenames == frozenset({"mmproj-F16.gguf"})
+
+    q4 = plans["q4_k_m"]
+    assert q4.main_filenames == frozenset({"gemma-4-12b-it-Q4_K_M.gguf"})
+    assert q4.main_size_bytes == 4_000
+    # Download size = main + mmproj + drafter.
+    assert q4.download_size_bytes == 4_600
+
+
+def test_baked_in_repo_plans_unchanged():
+    plans = build_gguf_variant_plans([_sib("Qwen3.6-27B-MTP-Q4_K_M.gguf", 4_000, "q4")])
+    assert plans["q4_k_m"].target_filenames == ("Qwen3.6-27B-MTP-Q4_K_M.gguf",)
+
+
+def test_old_manifest_resume_reclassifies_drafter():
+    # Pre-fix manifests could leak the drafter into a quant's expected
+    # files; resume must classify it as a companion, not a main shard.
+    old = [
+        ExpectedFile(path = "gemma-4-12b-it-Q8_0.gguf", size = 8_000, sha256 = "main-q8"),
+        ExpectedFile(path = "mtp-gemma-4-12b-it.gguf", size = 100, sha256 = "drafter"),
+    ]
+    plan = plan_from_expected_files("Q8_0", old)
+    assert plan.main_hashes == frozenset({"main-q8"})
+    assert plan.companion_hashes == frozenset({"drafter"})
+    assert plan.mmproj_filenames == frozenset()
+
+
+# ── Local detection / self-pairing ───────────────────────────────────
+
+
+def test_detect_mtp_file_finds_root_sibling(tmp_path):
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-model.gguf").write_bytes(b"x")
+    (tmp_path / "MTP").mkdir()
+    (tmp_path / "MTP" / "model-Q8_0-MTP.gguf").write_bytes(b"x")
+
+    found = detect_mtp_file(str(tmp_path / "model-Q4_K_M.gguf"))
+    assert found is not None
+    assert found.endswith("mtp-model.gguf")
+
+
+def test_detect_mtp_file_none_without_sibling(tmp_path):
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"x")
+    assert detect_mtp_file(str(tmp_path / "model-Q4_K_M.gguf")) is None
+
+
+def test_detect_gguf_model_rejects_drafter_file(tmp_path):
+    drafter = tmp_path / "mtp-model.gguf"
+    drafter.write_bytes(b"x")
+    assert detect_gguf_model(str(drafter)) is None
+
+
+def test_detect_gguf_model_dir_skips_companions(tmp_path):
+    main = tmp_path / "model-Q4_K_M.gguf"
+    main.write_bytes(b"xxxx")
+    # Companions are larger so a size-sorted pick would wrongly win.
+    (tmp_path / "mtp-model.gguf").write_bytes(b"x" * 64)
+    (tmp_path / "mmproj-F16.gguf").write_bytes(b"x" * 128)
+
+    assert detect_gguf_model(str(tmp_path)) == str(main.resolve())
+
+
+def test_detect_mtp_file_pairs_by_weight_name(tmp_path):
+    # Multi-model folder: each weight must get its own drafter, never the
+    # first-sorted foreign one.
+    (tmp_path / "gemma-4-12b-it-Q4_K_M.gguf").write_bytes(b"x")
+    (tmp_path / "gemma-4-31B-it-Q4_K_M.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-gemma-4-12b-it.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-gemma-4-31B-it.gguf").write_bytes(b"x")
+
+    found = detect_mtp_file(str(tmp_path / "gemma-4-31B-it-Q4_K_M.gguf"))
+    assert found is not None and found.endswith("mtp-gemma-4-31B-it.gguf")
+
+
+def test_detect_mtp_file_skips_foreign_drafter(tmp_path):
+    (tmp_path / "qwen3-8b-Q4_K_M.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-gemma-4-12b-it.gguf").write_bytes(b"x")
+    assert detect_mtp_file(str(tmp_path / "qwen3-8b-Q4_K_M.gguf")) is None
+
+
+def test_detect_mtp_file_qat_prefix_layout(tmp_path):
+    # unsloth's qat repo: drafter stem omits the -qat suffix but prefixes
+    # the weight name (mtp-gemma-4-12B-it.gguf / gemma-4-12B-it-qat-Q4_0.gguf).
+    (tmp_path / "gemma-4-12B-it-qat-Q4_0.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-gemma-4-12B-it.gguf").write_bytes(b"x")
+    found = detect_mtp_file(str(tmp_path / "gemma-4-12B-it-qat-Q4_0.gguf"))
+    assert found is not None and found.endswith("mtp-gemma-4-12B-it.gguf")
+
+
+def test_detect_mtp_file_search_root(tmp_path):
+    # Weight in a quant subdir, drafter at the granted directory root.
+    sub = tmp_path / "Q4_K_M"
+    sub.mkdir()
+    (sub / "gemma-4-12b-it-Q4_K_M.gguf").write_bytes(b"x")
+    (tmp_path / "mtp-gemma-4-12b-it.gguf").write_bytes(b"x")
+    found = detect_mtp_file(str(sub / "gemma-4-12b-it-Q4_K_M.gguf"), search_root = str(tmp_path))
+    assert found is not None and found.endswith("mtp-gemma-4-12b-it.gguf")
+
+
+# ── Reload dedup includes the drafter ────────────────────────────────
+
+
+def _loaded_backend(weight, drafter_path):
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    b = LlamaCppBackend()
+    # Shape matches atexit cleanup expectations (terminate/wait/kill).
+    b._process = SimpleNamespace(
+        poll = lambda: None,
+        terminate = lambda: None,
+        wait = lambda timeout = None: 0,
+        kill = lambda: None,
+    )
+    b._healthy = True
+    b._model_identifier = "local-gemma"
+    b._gguf_path = str(weight)
+    b._hf_variant = None
+    b._requested_n_ctx = 4096
+    b._cache_type_kv = None
+    b._requested_spec_mode = "auto"
+    b._speculative_type = "draft-mtp" if drafter_path else "default"
+    b._spec_draft_n_max = None
+    b._chat_template_override = None
+    b._extra_args = None
+    b._mtp_draft_path = drafter_path
+    return b
+
+
+def _target_state_kwargs(weight, mtp_draft_path):
+    return dict(
+        model_identifier = "local-gemma",
+        hf_variant = None,
+        n_ctx = 4096,
+        cache_type_kv = None,
+        speculative_type = "auto",
+        spec_draft_n_max = None,
+        chat_template_override = None,
+        extra_args = None,
+        is_vision = False,
+        gguf_path = str(weight),
+        mtp_draft_path = mtp_draft_path,
+    )
+
+
+def test_already_in_target_state_bounces_on_new_drafter(tmp_path):
+    weight = tmp_path / "gemma-4-12b-it-Q4_K_M.gguf"
+    weight.write_bytes(b"x")
+    drafter = tmp_path / "mtp-gemma-4-12b-it.gguf"
+    drafter.write_bytes(b"x")
+
+    # Loaded without a drafter; one now exists on disk -> must reload.
+    b = _loaded_backend(weight, None)
+    assert not b._already_in_target_state(**_target_state_kwargs(weight, str(drafter)))
+    # Same drafter as launched -> still deduped.
+    b = _loaded_backend(weight, str(drafter))
+    assert b._already_in_target_state(**_target_state_kwargs(weight, str(drafter)))
+
+
+def test_detect_gguf_model_rejects_mtp_subdir_copy(tmp_path):
+    # Direct selection of an MTP/ copy: the basename alone has no mtp-
+    # prefix, so rejection relies on the parent dir name.
+    sub = tmp_path / "MTP"
+    sub.mkdir()
+    copy = sub / "gemma-4-12b-it-BF16-MTP.gguf"
+    copy.write_bytes(b"x")
+    assert detect_gguf_model(str(copy)) is None
+    # Selecting the MTP dir itself must not surface the copies as models.
+    assert detect_gguf_model(str(sub)) is None
diff --git a/studio/backend/tests/test_native_context_length.py b/studio/backend/tests/test_native_context_length.py
index 0290bb1308..de1ca0649e 100644
--- a/studio/backend/tests/test_native_context_length.py
+++ b/studio/backend/tests/test_native_context_length.py
@@ -271,6 +271,7 @@ class TestPydanticModels:
     def test_load_response_has_field(self):
         """Field exists in LoadResponse.model_fields."""
         assert "native_context_length" in LoadResponse.model_fields
+        assert "context_length" in LoadResponse.model_fields
 
     def test_load_response_defaults_none(self):
         """Omitting native_context_length defaults to None."""
@@ -319,6 +320,7 @@ class TestPydanticModels:
     def test_status_response_has_field(self):
         """Field exists in InferenceStatusResponse.model_fields."""
         assert "native_context_length" in InferenceStatusResponse.model_fields
+        assert "context_length" in InferenceStatusResponse.model_fields
 
     def test_status_response_has_chat_template_field(self):
         """Status includes chat_template so the UI can rehydrate after refresh."""
@@ -347,6 +349,18 @@ class TestPydanticModels:
         roundtripped = LoadResponse.model_validate_json(resp.model_dump_json())
         assert roundtripped.native_context_length == 131072
 
+    def test_context_length_roundtrip(self):
+        """Runtime context_length serializes for non-GGUF/hub models."""
+        resp = LoadResponse(
+            status = "loaded",
+            model = "test",
+            display_name = "Test",
+            inference = {},
+            context_length = 8192,
+        )
+        roundtripped = LoadResponse.model_validate_json(resp.model_dump_json())
+        assert roundtripped.context_length == 8192
+
 
 # =====================================================================
 # D. TestRouteCompleteness -- source-level verification
@@ -408,6 +422,16 @@ class TestRouteCompleteness:
                 "native_context_length" not in block
             ), f"Non-GGUF LoadResponse should not set native_context_length:\n{block[:200]}"
 
+    def test_non_gguf_load_responses_set_runtime_context_length(self):
+        """Non-GGUF LoadResponse blocks report runtime context_length."""
+        blocks = self._find_construction_blocks("LoadResponse")
+        non_gguf = [b for b in blocks if "is_gguf = True" not in b and "is_gguf=True" not in b]
+        assert non_gguf, "Expected at least one non-GGUF LoadResponse block"
+        for block in non_gguf:
+            assert (
+                "context_length" in block
+            ), f"Non-GGUF LoadResponse should set context_length:\n{block[:200]}"
+
     def test_status_path(self):
         """InferenceStatusResponse construction with llama_backend has the field."""
         blocks = self._find_construction_blocks("InferenceStatusResponse")
@@ -420,6 +444,21 @@ class TestRouteCompleteness:
             found
         ), "No InferenceStatusResponse block with llama_backend has native_context_length"
 
+    def test_non_gguf_status_path_reports_runtime_context_length(self):
+        """Non-GGUF InferenceStatusResponse reports context_length from model_info."""
+        blocks = self._find_construction_blocks("InferenceStatusResponse")
+        found = False
+        for block in blocks:
+            if "is_gguf = False" in block and "context_length" in block:
+                found = True
+                break
+        assert found, "No non-GGUF InferenceStatusResponse block with context_length"
+
+    def test_openai_models_listing_reports_context_length(self):
+        """/v1/models includes context_length when the backend knows it."""
+        assert 'entry["context_length"]' in self._source
+        assert 'model_info.get("context_length")' in self._source
+
 
 # =====================================================================
 # E. TestEdgeCases
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index db4c07cdc8..1d994b46c0 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -166,10 +166,11 @@ class TestChatMessageToolRoles:
         with pytest.raises(ValidationError):
             ChatMessage(role = "user", content = [])
 
-    def test_tool_empty_content_rejected(self):
-        with pytest.raises(ValidationError) as exc_info:
-            ChatMessage(role = "tool", tool_call_id = "call_1", content = "")
-        assert "content" in str(exc_info.value)
+    def test_tool_empty_content_accepted(self):
+        # Empty tool output (mkdir, git add, ...) is routine in agentic loops;
+        # OpenAI and llama-server both accept it, so Studio must not 400.
+        msg = ChatMessage(role = "tool", tool_call_id = "call_1", content = "")
+        assert msg.content == ""
 
     def test_assistant_without_content_or_tool_calls_tolerated(self):
         # Stop-button leaves an empty assistant turn; tolerate for replay.
@@ -654,6 +655,100 @@ class TestBuildPassthroughPayloadToolChoice:
         ]
 
 
+# =====================================================================
+# Passthrough reasoning kwargs — enable_thinking / reasoning_effort /
+# preserve_thinking must reach llama-server via chat_template_kwargs,
+# gated on template capabilities like the non-passthrough paths.
+# =====================================================================
+
+
+def _reasoning_backend(
+    supports_reasoning = True,
+    reasoning_style = "enable_thinking",
+    reasoning_always_on = False,
+    supports_preserve_thinking = False,
+):
+    """Bare LlamaCppBackend with just the reasoning capability flags set,
+    so _build_openai_passthrough_body exercises the real
+    _request_reasoning_kwargs gating."""
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    backend = LlamaCppBackend.__new__(LlamaCppBackend)
+    backend._supports_reasoning = supports_reasoning
+    backend._reasoning_style = reasoning_style
+    backend._reasoning_always_on = reasoning_always_on
+    backend._supports_preserve_thinking = supports_preserve_thinking
+    return backend
+
+
+class TestPassthroughReasoningKwargs:
+    def _payload(self, **fields):
+        return ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "hi"}],
+            **fields,
+        )
+
+    def test_enable_thinking_forwarded(self):
+        body = _build_openai_passthrough_body(
+            self._payload(enable_thinking = False),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(),
+        )
+        assert body["chat_template_kwargs"] == {"enable_thinking": False}
+
+    def test_preserve_thinking_forwarded_when_template_supports_it(self):
+        body = _build_openai_passthrough_body(
+            self._payload(enable_thinking = True, preserve_thinking = True),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(supports_preserve_thinking = True),
+        )
+        assert body["chat_template_kwargs"] == {
+            "enable_thinking": True,
+            "preserve_thinking": True,
+        }
+
+    def test_preserve_thinking_dropped_when_template_lacks_it(self):
+        body = _build_openai_passthrough_body(
+            self._payload(preserve_thinking = True),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(supports_preserve_thinking = False),
+        )
+        assert "chat_template_kwargs" not in body
+
+    def test_reasoning_effort_forwarded_for_effort_style_models(self):
+        body = _build_openai_passthrough_body(
+            self._payload(reasoning_effort = "high"),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(reasoning_style = "reasoning_effort"),
+        )
+        assert body["chat_template_kwargs"] == {"reasoning_effort": "high"}
+
+    def test_enable_thinking_maps_to_effort_for_effort_style_models(self):
+        body = _build_openai_passthrough_body(
+            self._payload(enable_thinking = False),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(reasoning_style = "reasoning_effort"),
+        )
+        assert body["chat_template_kwargs"] == {"reasoning_effort": "low"}
+
+    def test_always_on_reasoning_skips_thinking_kwargs(self):
+        body = _build_openai_passthrough_body(
+            self._payload(enable_thinking = False),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(reasoning_always_on = True),
+        )
+        assert "chat_template_kwargs" not in body
+
+    def test_no_reasoning_fields_omits_chat_template_kwargs(self):
+        body = _build_openai_passthrough_body(
+            self._payload(),
+            backend_ctx = 4096,
+            llama_backend = _reasoning_backend(supports_preserve_thinking = True),
+        )
+        assert "chat_template_kwargs" not in body
+
+
 # =====================================================================
 # OpenAI API compatibility helpers — verified spec edge cases
 # =====================================================================
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index f0f2714214..69fb0a78c2 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -58,6 +58,7 @@ from routes.inference import (
     _build_chat_request,
     _chat_tool_calls_to_responses_output,
     _normalise_responses_input,
+    _responses_tool_output_text,
     _responses_stream,
     _translate_responses_tool_choice_to_chat,
     _translate_responses_tools_to_chat,
@@ -393,6 +394,100 @@ class TestNormaliseResponsesInputWithTools:
         # Content is serialised so llama-server sees a string.
         assert json.loads(msgs[0].content) == [{"type": "output_text", "text": "ok"}]
 
+    def test_empty_function_call_output_gets_no_output_sentinel(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": "",
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].role == "tool"
+        assert msgs[0].tool_call_id == "call_1"
+        assert msgs[0].content == "(no output)"
+        ChatMessage(**msgs[0].model_dump(exclude_none = True))
+
+    def test_whitespace_function_call_output_gets_no_output_sentinel(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": "   \n\t",
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].content == "(no output)"
+
+    def test_empty_content_array_output_gets_no_output_sentinel(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": [],
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].content == "(no output)"
+
+    def test_image_content_array_tool_output_is_serialised(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": [
+                        {
+                            "type": "image",
+                            "source": {
+                                "type": "base64",
+                                "media_type": "image/png",
+                                "data": "iVBORw0KGgo=",
+                            },
+                        }
+                    ],
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].role == "tool"
+        assert json.loads(msgs[0].content)[0]["type"] == "image"
+
+    def test_image_payload_outside_output_gets_no_output_sentinel(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": "",
+                    "content": [
+                        {
+                            "type": "image",
+                            "source": {
+                                "type": "base64",
+                                "media_type": "image/png",
+                                "data": "iVBORw0KGgo=",
+                            },
+                        }
+                    ],
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].role == "tool"
+        assert msgs[0].tool_call_id == "call_1"
+        assert msgs[0].content == "(no output)"
+
+    def test_tool_output_serializer_preserves_non_empty_text(self):
+        assert _responses_tool_output_text("done") == "done"
+        assert _responses_tool_output_text("  done  ") == "  done  "
+
 
 # =====================================================================
 # Response mapping — tool_calls → function_call output items
@@ -533,6 +628,10 @@ class TestResponsesStreamAdapter:
                 is_vision = False,
                 context_length = 4096,
                 base_url = "http://llama.test",
+                # Non-reasoning template: the real backend returns None here.
+                _request_reasoning_kwargs = (
+                    lambda enable_thinking = None, reasoning_effort = None, preserve_thinking = None: None
+                ),
             ),
         )
 
@@ -806,3 +905,17 @@ class TestTranslatedMessagesValidate:
             # Building a fresh ChatMessage from the dump round-trips the
             # role-shape validator — the passthrough's key invariant.
             ChatMessage(**m.model_dump(exclude_none = True))
+
+    def test_empty_tool_output_round_trips_through_chat_message_validator(self):
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_empty",
+                    "output": "",
+                },
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        for m in msgs:
+            ChatMessage(**m.model_dump(exclude_none = True))
diff --git a/studio/backend/tests/test_rocm_oom_guard.py b/studio/backend/tests/test_rocm_oom_guard.py
index 21edabd142..6e70c7cde4 100644
--- a/studio/backend/tests/test_rocm_oom_guard.py
+++ b/studio/backend/tests/test_rocm_oom_guard.py
@@ -13,6 +13,7 @@ wrong headroom factor on a 128 GiB unified-memory pool.
 
 from __future__ import annotations
 
+from pathlib import Path
 from types import SimpleNamespace
 
 import pytest
@@ -28,6 +29,46 @@ def _props(**kwargs) -> SimpleNamespace:
     return SimpleNamespace(**kwargs)
 
 
+# ── Path 0: props.is_integrated (driver's own unified-memory answer) ─────────
+
+
+class TestIsIntegratedSignal:
+    """hipDeviceProp_t.integrated wins when truthy; 0/absent never downgrades.
+
+    Same universal gate PR #5988's UMA safetensors fast-load uses -- keeps
+    Studio's two unified-memory consumers on one signal."""
+
+    def test_integrated_upgrades_unknown_apu(self) -> None:
+        # gfx1103 Phoenix iGPU: outside the hardcoded arch set, but the
+        # driver says integrated -> unified.
+        props = _props(gcnArchName = "gfx1103", name = "Radeon 780M", is_integrated = 1)
+        gcn, is_unified = _rocm_classify_unified_memory(props)
+        assert gcn == "gfx1103"
+        assert is_unified is True
+
+    def test_integrated_wins_without_any_arch(self) -> None:
+        props = _props(name = "Some Future APU", is_integrated = 1)
+        gcn, is_unified = _rocm_classify_unified_memory(props)
+        assert gcn == ""
+        assert is_unified is True
+
+    def test_zero_does_not_downgrade_known_apu(self) -> None:
+        # A wheel that zeroes the field must not flip Strix Halo to discrete.
+        props = _props(gcnArchName = "gfx1151", name = "x", is_integrated = 0)
+        gcn, is_unified = _rocm_classify_unified_memory(props)
+        assert is_unified is True
+
+    def test_absent_keeps_existing_behavior(self) -> None:
+        props = _props(gcnArchName = "gfx1201", name = "RX 9070 XT")
+        gcn, is_unified = _rocm_classify_unified_memory(props)
+        assert is_unified is False
+
+    def test_discrete_with_zero_stays_discrete(self) -> None:
+        props = _props(gcnArchName = "gfx1100", name = "RX 7900 XTX", is_integrated = 0)
+        gcn, is_unified = _rocm_classify_unified_memory(props)
+        assert is_unified is False
+
+
 # ── Path 1: canonical gcnArchName ────────────────────────────────────────────
 
 
@@ -168,3 +209,35 @@ class TestDeviceNameFallback:
         gcn, is_unified = _rocm_classify_unified_memory(props)
         assert gcn == ""
         assert is_unified is False
+
+
+# ── Fraction selection (source-pinned) ───────────────────────────────────────
+
+
+_WORKER_PY = Path(__file__).resolve().parents[1] / "core" / "training" / "worker.py"
+
+
+class TestMemFractionSelection:
+    """Pin the per-platform fraction policy in worker.py section 1g.
+
+    On native Windows, torch.cuda.mem_get_info's total is the WDDM budget
+    the driver grants HIP -- the OS share of RAM is already outside it, so
+    a 0.80 cap double-taxes (field report: 48.49 GiB budget -> '38.79 GiB
+    allowed' OOM denying a 47.29 GiB load that fit in free memory). 1.0
+    removes the double-tax; current AMD Windows wheels enforce only
+    sub-1.0 fractions, so it behaves like torch's uncapped default with
+    WDDM arbitrating residency (measured on gfx1151)."""
+
+    def test_unified_win32_uses_budget_exact_fraction(self) -> None:
+        source = _WORKER_PY.read_text(encoding = "utf-8")
+        assert '1.0 if sys.platform == "win32" else 0.80' in source
+
+    def test_discrete_keeps_090(self) -> None:
+        source = _WORKER_PY.read_text(encoding = "utf-8")
+        assert "_mem_fraction = 0.90" in source
+
+    def test_win32_unified_logs_vgm_hint(self) -> None:
+        """Users must learn the WDDM budget is raisable (BIOS UMA / AMD
+        Software Variable Graphics Memory) instead of assuming a bug."""
+        source = _WORKER_PY.read_text(encoding = "utf-8")
+        assert "Variable Graphics Memory" in source
diff --git a/studio/backend/tests/test_server_disk_logging.py b/studio/backend/tests/test_server_disk_logging.py
new file mode 100644
index 0000000000..05d03d869c
--- /dev/null
+++ b/studio/backend/tests/test_server_disk_logging.py
@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the server session log + native-crash capture in run.py.
+
+Field regression: Studio "terminates without a warning" -- a native crash in
+the GPU runtime kills the process with no Python traceback, and a desktop-
+shortcut console closes before anything can be read. The server must tee its
+console output to disk and aim faulthandler at the same file so even hard
+crashes leave evidence.
+"""
+
+from __future__ import annotations
+
+import io
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import run as run_mod  # noqa: E402
+
+
+class TestTeeStream:
+    def test_writes_reach_both_and_return_original(self):
+        console, log = io.StringIO(), io.StringIO()
+        tee = run_mod._TeeStream(console, log)
+        n = tee.write("hello")
+        assert console.getvalue() == "hello" == log.getvalue()
+        assert n == 5  # delegate's return value, console contract unchanged
+
+    def test_log_failure_never_breaks_console(self):
+        class Broken:
+            def write(self, data):
+                raise OSError("disk full")
+
+            def flush(self):
+                raise OSError("disk full")
+
+        console = io.StringIO()
+        tee = run_mod._TeeStream(console, Broken())
+        assert tee.write("still works") == len("still works")
+        tee.flush()  # must not raise
+        assert console.getvalue() == "still works"
+
+    def test_attribute_proxy(self):
+        console, log = io.StringIO(), io.StringIO()
+        tee = run_mod._TeeStream(console, log)
+        # isatty / encoding probes must see the original stream's answers.
+        assert tee.isatty() == console.isatty()
+
+
+class TestSetupServerDiskLogging:
+    def test_opt_out_env(self, monkeypatch):
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_FILE_LOG", "1")
+        assert run_mod._setup_server_disk_logging() is None
+
+    def test_creates_log_and_enables_faulthandler(self, monkeypatch, tmp_path):
+        import faulthandler
+
+        monkeypatch.delenv("UNSLOTH_STUDIO_NO_FILE_LOG", raising = False)
+        monkeypatch.delenv("PYTHONFAULTHANDLER", raising = False)
+        # Both resolution paths (utils.paths.studio_root and the env
+        # fallback) honor UNSLOTH_STUDIO_HOME, so this redirects the log dir.
+        monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+        orig_out, orig_err = sys.stdout, sys.stderr
+        was_enabled = faulthandler.is_enabled()
+        try:
+            log_path = run_mod._setup_server_disk_logging()
+            assert log_path is not None
+            assert Path(log_path).is_file()
+            assert "logs" in str(log_path)
+            # faulthandler armed at the file; children inherit the env switch.
+            assert faulthandler.is_enabled()
+            import os
+
+            assert os.environ.get("PYTHONFAULTHANDLER") == "1"
+            print("tee-capture-marker")
+            sys.stdout.flush()
+            assert "tee-capture-marker" in Path(log_path).read_text(
+                encoding = "utf-8", errors = "replace"
+            )
+        finally:
+            sys.stdout, sys.stderr = orig_out, orig_err
+            if not was_enabled:
+                faulthandler.disable()
+
+    def test_run_server_wires_logging_before_main_import(self):
+        src = (Path(_BACKEND_DIR) / "run.py").read_text(encoding = "utf-8")
+        call_idx = src.index("_setup_server_disk_logging()", src.index("def run_server"))
+        main_import_idx = src.index("from main import app", src.index("def run_server"))
+        assert call_idx < main_import_idx, (
+            "disk logging must be armed before importing main so import-time "
+            "failures leave evidence on disk"
+        )
diff --git a/studio/backend/tests/test_tool_message_empty_content.py b/studio/backend/tests/test_tool_message_empty_content.py
new file mode 100644
index 0000000000..d63b16ce80
--- /dev/null
+++ b/studio/backend/tests/test_tool_message_empty_content.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Empty ``role="tool"`` content must be accepted on the OpenAI-compat surface.
+
+Agentic clients send ``content: ""`` when a command produced no output;
+OpenAI and llama-server both accept it. Studio used to 400, which standard
+clients treat as non-retryable and kill the session. The validator must
+normalize empty/missing tool content to ``""`` instead of raising.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from models.inference import ChatMessage
+
+
+def test_tool_message_empty_string_content_is_accepted():
+    msg = ChatMessage(role = "tool", content = "", tool_call_id = "call_1")
+    assert msg.content == ""
+
+
+def test_tool_message_none_content_normalizes_to_empty_string():
+    msg = ChatMessage(role = "tool", content = None, tool_call_id = "call_1")
+    assert msg.content == ""
+
+
+def test_tool_message_empty_list_content_normalizes_to_empty_string():
+    msg = ChatMessage(role = "tool", content = [], tool_call_id = "call_1")
+    assert msg.content == ""
+
+
+def test_tool_message_real_content_is_preserved():
+    msg = ChatMessage(role = "tool", content = "ok", tool_call_id = "call_1")
+    assert msg.content == "ok"
+
+
+def test_user_message_still_requires_content():
+    with pytest.raises(ValueError):
+        ChatMessage(role = "user", content = None)
+
+
+def test_assistant_empty_content_still_collapses_to_none():
+    msg = ChatMessage(role = "assistant", content = "")
+    assert msg.content is None
diff --git a/studio/backend/tests/test_training_nan_loss_handling.py b/studio/backend/tests/test_training_nan_loss_handling.py
new file mode 100644
index 0000000000..a2dc78bee2
--- /dev/null
+++ b/studio/backend/tests/test_training_nan_loss_handling.py
@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Pin Studio's behavior when a training event reports non-finite (NaN/Inf) loss.
+
+The training event handler used to filter NaN/Inf to None silently while
+leaving the previous finite loss in progress.loss — so the API kept reporting
+the stale value as if everything were fine. We now drop the stale value:
+clients see loss=None at the affected step and a one-shot warning is logged.
+Training continues; the run is not marked failed.
+"""
+
+from __future__ import annotations
+
+import math
+import os
+import sys
+
+import pytest
+
+_BACKEND = os.path.join(os.path.dirname(__file__), "..")
+if _BACKEND not in sys.path:
+    sys.path.insert(0, _BACKEND)
+
+from core.training.training import TrainingBackend
+
+
+def _make_backend() -> TrainingBackend:
+    return TrainingBackend()
+
+
+def _progress_event(
+    step: int,
+    loss: float,
+    lr: float = 1e-4,
+) -> dict:
+    return {
+        "type": "progress",
+        "step": step,
+        "loss": loss,
+        "learning_rate": lr,
+        "epoch": 0.0,
+        "total_steps": 100,
+    }
+
+
+class TestNonfiniteLossSoftHandling:
+    def test_finite_loss_updates_progress_normally(self):
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = 0.97))
+        assert b._progress.loss == pytest.approx(0.97)
+        assert b._progress.error is None
+        assert b._should_stop is False
+        assert getattr(b._progress, "_nonfinite_loss_warned", False) is False
+
+    def test_nan_loss_clears_progress_loss(self):
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = 0.97))
+        assert b._progress.loss == pytest.approx(0.97)
+        b._handle_event(_progress_event(step = 2, loss = float("nan")))
+        # Stale finite loss must NOT leak through
+        assert b._progress.loss is None
+        # Run is not marked failed
+        assert b._progress.error is None
+        assert b._should_stop is False
+        # Warning flag is set so we don't re-log on every subsequent NaN step
+        assert b._progress._nonfinite_loss_warned is True
+
+    def test_inf_loss_clears_progress_loss(self):
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = float("inf")))
+        assert b._progress.loss is None
+        assert b._progress.error is None
+        assert b._should_stop is False
+        assert b._progress._nonfinite_loss_warned is True
+
+    def test_negative_inf_loss_clears_progress_loss(self):
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = float("-inf")))
+        assert b._progress.loss is None
+        assert b._progress.error is None
+        assert b._should_stop is False
+        assert b._progress._nonfinite_loss_warned is True
+
+    def test_repeated_nan_only_warns_once(self):
+        """Subsequent NaN events must not re-fire the warning flag setter.
+        The flag should already be True after the first NaN."""
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = 0.97))
+        b._handle_event(_progress_event(step = 2, loss = float("nan")))
+        assert b._progress._nonfinite_loss_warned is True
+        # Further NaN steps don't change anything we care about
+        b._handle_event(_progress_event(step = 3, loss = float("nan")))
+        b._handle_event(_progress_event(step = 4, loss = float("nan")))
+        assert b._progress._nonfinite_loss_warned is True
+        assert b._progress.loss is None
+        assert b._progress.error is None
+        assert b._should_stop is False
+
+    def test_recovery_updates_loss_when_finite_again(self):
+        """If a NaN step is followed by a finite step, progress.loss must
+        reflect the new finite value (not stay stuck at None)."""
+        b = _make_backend()
+        b._handle_event(_progress_event(step = 1, loss = 0.97))
+        b._handle_event(_progress_event(step = 2, loss = float("nan")))
+        assert b._progress.loss is None
+        b._handle_event(_progress_event(step = 3, loss = 0.85))
+        assert b._progress.loss == pytest.approx(0.85)
+        # Warning flag stays set (we don't reset it on recovery)
+        assert b._progress._nonfinite_loss_warned is True
diff --git a/studio/backend/tests/test_training_progress_stream_nan.py b/studio/backend/tests/test_training_progress_stream_nan.py
new file mode 100644
index 0000000000..899527a04d
--- /dev/null
+++ b/studio/backend/tests/test_training_progress_stream_nan.py
@@ -0,0 +1,129 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The SSE progress stream must follow the live progress step during
+non-finite-loss stretches (loss reported as null) instead of replaying the
+last finite step/loss pair from the metric histories, which skip NaN steps."""
+
+import asyncio
+import json
+import sys
+import types
+
+import pytest
+
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *args, **kwargs: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger,
+        get_logger = lambda *args, **kwargs: _DummyLogger(),
+    )
+
+import routes.training as rt
+
+
+class _Progress:
+    def __init__(self):
+        self.step = 5
+        self.total_steps = 10
+        self.loss = None  # cleared by the NaN honesty fix in core training
+        self.learning_rate = 8e-5
+        self.epoch = 0.1
+        self.grad_norm = None
+        self.num_tokens = None
+        self.eval_loss = None
+        self.elapsed_seconds = None
+        self.eta_seconds = None
+
+
+class _FakeBackend:
+    """Finite history stops at step 2; live progress is at step 5 with NaN
+    (loss=None). Active for a few polls, then done."""
+
+    def __init__(self, active_polls = 2):
+        self.current_job_id = "job-1"
+        self.step_history = [1, 2]
+        self.loss_history = [2.0, 1.5]
+        self.lr_history = [1e-4, 9e-5]
+        self.eval_enabled = False
+        self._active_calls = 0
+        self._active_polls = active_polls
+        self.trainer = types.SimpleNamespace(training_progress = _Progress())
+
+    def is_training_active(self):
+        self._active_calls += 1
+        return self._active_calls <= self._active_polls
+
+
+class _FakeRequest:
+    headers = {}
+
+
+def _collect_events(response, timeout = 15):
+    async def _drain():
+        chunks = []
+        async for chunk in response.body_iterator:
+            chunks.append(chunk)
+        return "".join(c.decode() if isinstance(c, bytes) else c for c in chunks)
+
+    return asyncio.run(asyncio.wait_for(_drain(), timeout))
+
+
+def _progress_payloads(raw):
+    payloads = []
+    for block in raw.split("\n\n"):
+        lines = block.strip().splitlines()
+        data = next((l[6:] for l in lines if l.startswith("data: ")), None)
+        if data:
+            payloads.append(json.loads(data))
+    return payloads
+
+
+def test_stream_reports_live_step_with_null_loss_during_nan(monkeypatch):
+    backend = _FakeBackend(active_polls = 2)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    response = asyncio.run(rt.stream_training_progress(_FakeRequest(), current_subject = "tester"))
+    raw = _collect_events(response)
+    payloads = _progress_payloads(raw)
+    assert payloads, f"no SSE payloads parsed from: {raw!r}"
+
+    live = [p for p in payloads if p.get("step") == 5]
+    assert live, (
+        "stream never advanced to the live progress step during the NaN "
+        f"stretch; steps seen: {[p.get('step') for p in payloads]}"
+    )
+    assert live[0]["loss"] is None
+    # The stale finite pair must not be re-emitted as the latest progress.
+    stale = [p for p in payloads if p.get("step") == 2 and p.get("loss") == 1.5]
+    assert not stale
+
+
+def test_inactive_stream_completes_with_live_step_and_null_loss(monkeypatch):
+    # Fresh connection after the run already ended during a NaN stretch: the
+    # immediate complete event must not replay the stale finite pair either.
+    backend = _FakeBackend(active_polls = 0)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    response = asyncio.run(rt.stream_training_progress(_FakeRequest(), current_subject = "tester"))
+    payloads = _progress_payloads(_collect_events(response))
+    final = payloads[-1]
+    assert final["step"] == 5
+    assert final["loss"] is None
+
+
+def test_stream_uses_finite_history_when_progress_in_sync(monkeypatch):
+    backend = _FakeBackend(active_polls = 2)
+    # Live progress agrees with the history tail: normal finite behavior.
+    backend.trainer.training_progress.step = 2
+    backend.trainer.training_progress.loss = 1.5
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    response = asyncio.run(rt.stream_training_progress(_FakeRequest(), current_subject = "tester"))
+    payloads = _progress_payloads(_collect_events(response))
+    finite = [p for p in payloads if p.get("step") == 2]
+    assert finite and finite[0]["loss"] == 1.5
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 915bc4b13b..7c497ba1b1 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -30,8 +30,11 @@ sys.modules.setdefault("loggers", _loggers_stub)
 from utils.transformers_version import (
     _resolve_base_model,
     _check_tokenizer_config_needs_v5,
+    _check_config_needs_510,
     _check_config_needs_550,
+    _config_json_cache,
     _tokenizer_class_cache,
+    _config_needs_510_cache,
     _config_needs_550_cache,
     needs_transformers_5,
     get_transformers_tier,
@@ -200,6 +203,7 @@ class TestCheckConfigNeeds550:
     """Tests for _check_config_needs_550() local config.json checks."""
 
     def setup_method(self):
+        _config_json_cache.clear()
         _config_needs_550_cache.clear()
 
     def test_gemma4_architecture(self, tmp_path: Path):
@@ -253,6 +257,106 @@ class TestCheckConfigNeeds550:
             mock_urlopen.assert_not_called()
 
 
+# ---------------------------------------------------------------------------
+# _check_config_needs_510 — config.json architecture/model_type check
+# ---------------------------------------------------------------------------
+
+
+class TestCheckConfigNeeds510:
+    """Tests for _check_config_needs_510() local config.json checks."""
+
+    def setup_method(self):
+        _config_json_cache.clear()
+        _config_needs_510_cache.clear()
+
+    def test_gemma4_unified_architecture(self, tmp_path: Path):
+        """config.json with Gemma4UnifiedForConditionalGeneration should return True."""
+        cfg = {
+            "architectures": ["Gemma4UnifiedForConditionalGeneration"],
+            "model_type": "gemma4_unified",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_unified_model_type_only(self, tmp_path: Path):
+        """config.json with model_type=gemma4_unified should return True."""
+        cfg = {"model_type": "gemma4_unified"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_unified_assistant_architecture(self, tmp_path: Path):
+        """Assistant Gemma 4 Unified configs should return True."""
+        cfg = {
+            "architectures": ["Gemma4UnifiedAssistantForCausalLM"],
+            "model_type": "gemma4_unified_assistant",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_unified_assistant_model_type_only(self, tmp_path: Path):
+        """Assistant Gemma 4 Unified model_type should return True."""
+        cfg = {"model_type": "gemma4_unified_assistant"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_assistant_architecture(self, tmp_path: Path):
+        """Assistant Gemma 4 configs should return True."""
+        cfg = {
+            "architectures": ["Gemma4AssistantForCausalLM"],
+            "model_type": "gemma4_assistant",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_assistant_model_type_only(self, tmp_path: Path):
+        """Assistant Gemma 4 model_type should return True."""
+        cfg = {"model_type": "gemma4_assistant"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is True
+
+    def test_gemma4_non_unified_returns_false(self, tmp_path: Path):
+        """Older Gemma 4 config should stay on the 550 tier."""
+        cfg = {
+            "architectures": ["Gemma4ForConditionalGeneration"],
+            "model_type": "gemma4",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_510(str(tmp_path)) is False
+
+    def test_no_config_json(self, tmp_path: Path):
+        """Missing config.json should return False (fail-open)."""
+        # Patch network call to avoid real fetch
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            mock_urlopen.side_effect = Exception("no network")
+            assert _check_config_needs_510(str(tmp_path)) is False
+
+    def test_result_is_cached(self, tmp_path: Path):
+        """Subsequent calls should use the cache."""
+        cfg = {"architectures": ["Gemma4UnifiedForConditionalGeneration"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        key = str(tmp_path)
+        _check_config_needs_510(key)
+        assert key in _config_needs_510_cache
+        assert _config_needs_510_cache[key] is True
+
+    def test_local_file_skips_network(self, tmp_path: Path):
+        """When local config.json exists, no network request should be made."""
+        cfg = {"architectures": ["LlamaForCausalLM"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            _check_config_needs_510(str(tmp_path))
+            mock_urlopen.assert_not_called()
+
+
 # ---------------------------------------------------------------------------
 # get_transformers_tier — tier detection
 # ---------------------------------------------------------------------------
@@ -263,11 +367,19 @@ class TestGetTransformersTier:
 
     def setup_method(self):
         _tokenizer_class_cache.clear()
+        _config_json_cache.clear()
+        _config_needs_510_cache.clear()
         _config_needs_550_cache.clear()
 
     def test_gemma4_substring_returns_550(self):
         assert get_transformers_tier("google/gemma-4-E2B-it") == "550"
 
+    def test_gemma4_12b_substring_returns_510(self):
+        assert get_transformers_tier("unsloth/gemma-4-12b-it") == "510"
+
+    def test_gemma4_assistant_substring_returns_510(self):
+        assert get_transformers_tier("google/gemma-4-E2B-it-assistant") == "510"
+
     def test_gemma4_alt_substring_returns_550(self):
         assert get_transformers_tier("unsloth/gemma4-E4B-it") == "550"
 
@@ -281,17 +393,92 @@ class TestGetTransformersTier:
 
         assert get_transformers_tier(str(tmp_path)) == "550"
 
+    def test_gemma4_unified_config_json_returns_510(self, tmp_path: Path):
+        """Local checkpoint with Gemma4 Unified architecture → 510."""
+        cfg = {
+            "architectures": ["Gemma4UnifiedForConditionalGeneration"],
+            "model_type": "gemma4_unified",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert get_transformers_tier(str(tmp_path)) == "510"
+
+    def test_gemma4_assistant_config_json_returns_510(self, tmp_path: Path):
+        """Local checkpoint with Gemma4 Assistant architecture → 510."""
+        cfg = {
+            "architectures": ["Gemma4AssistantForCausalLM"],
+            "model_type": "gemma4_assistant",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert get_transformers_tier(str(tmp_path)) == "510"
+
+    def test_local_config_json_short_circuits_path_substrings(self, tmp_path: Path):
+        """Local config.json should prevent false matches from parent directory names."""
+        model_dir = tmp_path / "gemma-4-12b-experiment" / "llama-checkpoint"
+        model_dir.mkdir(parents = True)
+        (model_dir / "config.json").write_text(
+            json.dumps(
+                {
+                    "architectures": ["LlamaForCausalLM"],
+                    "model_type": "llama",
+                }
+            )
+        )
+        (model_dir / "tokenizer_config.json").write_text(
+            json.dumps({"tokenizer_class": "LlamaTokenizerFast"})
+        )
+
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            assert get_transformers_tier(str(model_dir)) == "default"
+            mock_urlopen.assert_not_called()
+
+    def test_remote_config_json_is_fetched_once_for_config_tiers(self):
+        """510 and 550 slow-path checks should share one config.json fetch."""
+
+        class _Response:
+            def __enter__(self):
+                return self
+
+            def __exit__(self, exc_type, exc, tb):
+                return False
+
+            def read(self):
+                return json.dumps(
+                    {
+                        "architectures": ["Gemma4ForConditionalGeneration"],
+                        "model_type": "gemma4",
+                    }
+                ).encode()
+
+        with patch("urllib.request.urlopen", return_value = _Response()) as mock_urlopen:
+            assert get_transformers_tier("org/no-fast-substring-model") == "550"
+
+        assert mock_urlopen.call_count == 1
+
     def test_qwen35_returns_530(self):
-        with patch(
-            "utils.transformers_version._check_config_needs_550",
-            return_value = False,
+        with (
+            patch(
+                "utils.transformers_version._check_config_needs_550",
+                return_value = False,
+            ),
+            patch(
+                "utils.transformers_version._check_config_needs_510",
+                return_value = False,
+            ),
         ):
             assert get_transformers_tier("Qwen/Qwen3.5-9B") == "530"
 
     def test_ministral_returns_530(self):
-        with patch(
-            "utils.transformers_version._check_config_needs_550",
-            return_value = False,
+        with (
+            patch(
+                "utils.transformers_version._check_config_needs_550",
+                return_value = False,
+            ),
+            patch(
+                "utils.transformers_version._check_config_needs_510",
+                return_value = False,
+            ),
         ):
             assert get_transformers_tier("mistralai/Ministral-3-8B-Instruct-2512") == "530"
 
@@ -301,6 +488,10 @@ class TestGetTransformersTier:
                 "utils.transformers_version._check_config_needs_550",
                 return_value = False,
             ),
+            patch(
+                "utils.transformers_version._check_config_needs_510",
+                return_value = False,
+            ),
             patch(
                 "utils.transformers_version._check_tokenizer_config_needs_v5",
                 return_value = False,
@@ -309,15 +500,22 @@ class TestGetTransformersTier:
             assert get_transformers_tier("meta-llama/Llama-3-8B") == "default"
 
     def test_550_checked_before_530(self):
-        """5.5.0 is checked first — a model matching both gets 550."""
+        """5.5.0 is checked before 5.3.0 - a model matching both gets 550."""
         assert get_transformers_tier("gemma-4-model") == "550"
 
     def test_needs_transformers_5_compat(self):
-        """needs_transformers_5 should return True for both 530 and 550 models."""
+        """needs_transformers_5 should return True for 510, 530, and 550 models."""
+        assert needs_transformers_5("unsloth/gemma-4-12b-it") is True
         assert needs_transformers_5("google/gemma-4-E2B-it") is True
-        with patch(
-            "utils.transformers_version._check_config_needs_550",
-            return_value = False,
+        with (
+            patch(
+                "utils.transformers_version._check_config_needs_550",
+                return_value = False,
+            ),
+            patch(
+                "utils.transformers_version._check_config_needs_510",
+                return_value = False,
+            ),
         ):
             assert needs_transformers_5("Qwen/Qwen3.5-9B") is True
         with (
@@ -325,6 +523,10 @@ class TestGetTransformersTier:
                 "utils.transformers_version._check_config_needs_550",
                 return_value = False,
             ),
+            patch(
+                "utils.transformers_version._check_config_needs_510",
+                return_value = False,
+            ),
             patch(
                 "utils.transformers_version._check_tokenizer_config_needs_v5",
                 return_value = False,
diff --git a/studio/backend/utils/datasets/cache_safe.py b/studio/backend/utils/datasets/cache_safe.py
new file mode 100644
index 0000000000..e629210f33
--- /dev/null
+++ b/studio/backend/utils/datasets/cache_safe.py
@@ -0,0 +1,51 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Permission-safe wrapper around datasets.load_dataset.
+
+A shared HF datasets cache can contain subtrees owned by another user (for
+example populated by an earlier root-run job). datasets then raises
+"[Errno 13] Permission denied: ..._builder.lock" while locking the cached
+builder, killing the training run even though the dataset itself is fine.
+Retry such loads in a Studio-owned cache so the run proceeds; the worst case
+is one rebuild of the dataset in the fallback location.
+"""
+
+import logging
+import os
+
+from utils.paths.storage_roots import cache_root
+
+logger = logging.getLogger(__name__)
+
+
+def studio_datasets_cache() -> str:
+    path = cache_root() / "hf-datasets"
+    path.mkdir(parents = True, exist_ok = True)
+    return str(path)
+
+
+def load_dataset_cache_safe(*args, **kwargs):
+    """datasets.load_dataset, retried in a Studio-owned cache on EACCES."""
+    from datasets import load_dataset
+    try:
+        return load_dataset(*args, **kwargs)
+    except PermissionError as error:
+        fallback = studio_datasets_cache()
+        logger.warning(
+            "HF datasets cache is not writable (%s); rebuilding in %s",
+            error,
+            fallback,
+        )
+        kwargs["cache_dir"] = fallback
+        # Nested builders consult the env var while the load runs; restore it
+        # after so other datasets keep trying the shared cache first.
+        old_env = os.environ.get("HF_DATASETS_CACHE")
+        os.environ["HF_DATASETS_CACHE"] = fallback
+        try:
+            return load_dataset(*args, **kwargs)
+        finally:
+            if old_env is None:
+                os.environ.pop("HF_DATASETS_CACHE", None)
+            else:
+                os.environ["HF_DATASETS_CACHE"] = old_env
diff --git a/studio/backend/utils/hardware/amd.py b/studio/backend/utils/hardware/amd.py
index 39aafe0489..27dfe187cc 100644
--- a/studio/backend/utils/hardware/amd.py
+++ b/studio/backend/utils/hardware/amd.py
@@ -12,6 +12,7 @@ import math
 import os
 import platform
 import re
+import shutil
 import subprocess
 import sys
 from typing import Any, Optional
@@ -33,24 +34,85 @@ _amd_smi_consecutive_failures = 0
 _amd_smi_disabled = False
 
 
+def _hip_sdk_present() -> bool:
+    """True if a HIP SDK is detectable (hipinfo on PATH or under HIP_PATH/
+    ROCM_PATH), meaning amd-smi has a working runtime and runs un-elevated."""
+    if shutil.which("hipinfo"):
+        return True
+    for var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
+        root = os.environ.get(var)
+        if root and os.path.exists(os.path.join(root, "bin", "hipinfo.exe")):
+            return True
+    return False
+
+
+def _amd_smi_allowed() -> bool:
+    """Whether it is safe to spawn amd-smi here.
+
+    On Windows without a working HIP runtime, amd-smi elevates a child at
+    runtime -- popping a UAC/DiskPart prompt that RunAsInvoker can't suppress
+    (its manifest is asInvoker). So only call it on Windows with a HIP SDK
+    present or UNSLOTH_ENABLE_AMD_SMI=1. Linux amd-smi never elevates.
+    """
+    if platform.system() != "Windows":
+        return True
+    flag = os.environ.get("UNSLOTH_ENABLE_AMD_SMI", "").strip().lower()
+    if flag in ("1", "true", "yes", "on"):
+        return True
+    if flag in ("0", "false", "no", "off"):
+        return False
+    return _hip_sdk_present()
+
+
 def _run_amd_smi(*args: str, timeout: int = _AMD_SMI_DEFAULT_TIMEOUT) -> Optional[Any]:
     """Run amd-smi with the given args and return parsed JSON, or None."""
     global _amd_smi_consecutive_failures, _amd_smi_disabled
     if _amd_smi_disabled:
         return None
+    if not _amd_smi_allowed():
+        # Permanently skip amd-smi on Windows w/o a HIP SDK: every call would
+        # pop a UAC/DiskPart prompt (see _amd_smi_allowed). VRAM polling is then
+        # unavailable, but that beats the prompt. Opt back in with
+        # UNSLOTH_ENABLE_AMD_SMI=1.
+        if not _amd_smi_disabled:
+            logger.info(
+                "amd-smi disabled on Windows (no HIP SDK detected) to avoid a "
+                "UAC/DiskPart elevation prompt; GPU VRAM polling unavailable. "
+                "Set UNSLOTH_ENABLE_AMD_SMI=1 to force amd-smi."
+            )
+            _amd_smi_disabled = True
+        return None
+    if shutil.which("amd-smi") is None:
+        # amd-smi does not exist on Windows (neither Adrenalin nor the HIP SDK
+        # ship a CLI) and can be absent on minimal Linux installs. Disable the
+        # poller in one step instead of burning the 3-strike circuit breaker
+        # on guaranteed FileNotFoundError spawns. Studio's VRAM display falls
+        # back to torch mem_get_info.
+        if not _amd_smi_disabled:
+            logger.info(
+                "amd-smi not found on PATH; GPU utilization polling via "
+                "amd-smi unavailable (VRAM falls back to torch mem_get_info)."
+            )
+            _amd_smi_disabled = True
+        return None
+    _amd_env = child_env_without_native_path_secret()
+    if platform.system() == "Windows":
+        # RunAsInvoker belt-and-suspenders for any manifest-elevating helper;
+        # the real guard is _amd_smi_allowed() above. Mirrors install scripts.
+        _amd_env = {**_amd_env, "__COMPAT_LAYER": "RunAsInvoker"}
     try:
         result = subprocess.run(
             ["amd-smi", *args, "--json"],
             capture_output = True,
             text = True,
             timeout = timeout,
-            env = child_env_without_native_path_secret(),
+            env = _amd_env,
             **windows_hidden_subprocess_kwargs(),
         )
     except (OSError, subprocess.TimeoutExpired) as e:
         if isinstance(e, FileNotFoundError):
-            # amd-smi ships with Adrenalin, not the HIP SDK; absence is expected
-            # on HIP SDK-only Windows setups.
+            # Raced a PATH change after the which() check above; absence is
+            # expected on Windows (no AMD product ships an amd-smi CLI there).
             logger.debug("amd-smi not found (not in PATH): %s", e)
         else:
             logger.warning("amd-smi query failed: %s", e)
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 7292fa185f..893d0364f7 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -1818,7 +1818,13 @@ def apply_gpu_ids(gpu_ids) -> None:
             )
     if _is_rocm:
         os.environ["HIP_VISIBLE_DEVICES"] = value
-        os.environ["ROCR_VISIBLE_DEVICES"] = value
+        # ROCR_VISIBLE_DEVICES operates at the HSA agent level and uses
+        # different indexing semantics to HIP_VISIBLE_DEVICES. Setting it
+        # to a physical GPU index breaks multi-GPU ROCm systems where the
+        # parent already set ROCR_VISIBLE_DEVICES (e.g. "0,1"): narrowing
+        # to "1" causes torch.cuda.is_available() to return False in the
+        # worker subprocess. HIP_VISIBLE_DEVICES is sufficient for GPU
+        # selection on ROCm -- leave ROCR_VISIBLE_DEVICES inherited.
     _visible_gpu_count = None
     if _is_rocm:
         logger.info("Applied gpu_ids: CUDA_VISIBLE_DEVICES='%s' (rocm)", value)
diff --git a/studio/backend/utils/llama_cpp_freshness.py b/studio/backend/utils/llama_cpp_freshness.py
index 3e5066ca2d..f5fd745334 100644
--- a/studio/backend/utils/llama_cpp_freshness.py
+++ b/studio/backend/utils/llama_cpp_freshness.py
@@ -13,6 +13,7 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import time
 from datetime import datetime, timezone
 from pathlib import Path
@@ -104,11 +105,18 @@ def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
 
 
 def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
-    """GitHub API call. None on any failure (offline, rate-limited, etc)."""
+    """Newest published release tag for `repo`, by publish time.
+
+    Resolves "latest" the way install_llama_prebuilt.py does (newest
+    non-draft/non-prerelease by ``published_at``), NOT via GitHub's
+    ``/releases/latest`` pointer. That pointer sorts by commit date and can lag
+    behind the build the installer actually installs, so detection and apply
+    disagreed -- the cause of the downgrade/sticky banner. None on any failure
+    (offline, rate-limited, etc)."""
     import urllib.error
     import urllib.request
 
-    url = f"https://api.github.com/repos/{repo}/releases/latest"
+    url = f"https://api.github.com/repos/{repo}/releases?per_page=30"
     headers = {
         "Accept": "application/vnd.github+json",
         "User-Agent": "unsloth-studio-freshness-check",
@@ -128,8 +136,21 @@ def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
     ) as exc:
         logger.debug("freshness fetch failed", repo = repo, error = str(exc))
         return None
-    tag = data.get("tag_name")
-    return tag if isinstance(tag, str) and tag else None
+    if not isinstance(data, list):
+        return None
+    published = [
+        r
+        for r in data
+        if isinstance(r, dict)
+        and not r.get("draft")
+        and not r.get("prerelease")
+        and isinstance(r.get("tag_name"), str)
+        and r.get("tag_name")
+    ]
+    if not published:
+        return None
+    newest = max(published, key = lambda r: r.get("published_at") or "")
+    return newest["tag_name"]
 
 
 def latest_published_release(repo: str, *, force_refresh: bool = False) -> Optional[str]:
@@ -172,19 +193,58 @@ def _parse_installed_at(value: object) -> Optional[datetime]:
     return dt
 
 
+def parse_base_build(tag: object) -> Optional[int]:
+    """Numeric base build from a release tag. Handles both a plain ``bNNNN`` and
+    a mix-build tag like ``b9596-mix-<sha>`` (anchored at the start, so the mix
+    suffix doesn't defeat it). None for anything not starting with ``bNNNN``."""
+    if not isinstance(tag, str):
+        return None
+    m = re.match(r"b(\d+)", tag.strip())
+    return int(m.group(1)) if m else None
+
+
+def is_behind(installed: Optional[str], latest: Optional[str]) -> bool:
+    """Whether `installed` is genuinely behind `latest`, comparing the FULL
+    release identity (so a mix build can legitimately be the latest) with a
+    base-build guard so a lagging GitHub /releases/latest can never read as an
+    update or a downgrade.
+
+    - identical tags -> not behind (clears the sticky banner post-update)
+    - higher base build on `latest` -> behind; lower -> NOT behind (downgrade guard)
+    - same base build: a different/new mix -> behind, but a bare ``bNNNN`` never
+      supersedes a mix build (extra PRs) at that base -> not behind
+    - non-bNNNN tags -> behind (plain inequality, since they already differ)
+    """
+    if not installed or not latest:
+        return False
+    installed, latest = installed.strip(), latest.strip()
+    if installed == latest:
+        return False
+    ib, lb = parse_base_build(installed), parse_base_build(latest)
+    if ib is None or lb is None:
+        return True
+    if lb != ib:
+        return lb > ib
+    # Same base build, different tags: offer a mix (latest carries a suffix), but
+    # never offer a bare base over a mix install at the same base.
+    return latest != f"b{lb}"
+
+
 def check_prebuilt_freshness(
     binary_path: Optional[str],
     *,
     threshold_days: int = STALENESS_THRESHOLD_DAYS,
     now: Optional[datetime] = None,
 ) -> dict:
-    """Returns {has_marker, stale, installed_tag, latest_tag,
+    """Returns {has_marker, stale, behind, installed_tag, latest_tag,
     installed_at_utc, age_days, published_repo, threshold_days}.
-    stale = True iff installed != latest AND age >= threshold.
-    Fails open on missing data (stale stays False)."""
+    behind = installed genuinely older than latest (see is_behind).
+    stale = behind AND age >= threshold.
+    Fails open on missing data (behind/stale stay False)."""
     out: dict = {
         "has_marker": False,
         "stale": False,
+        "behind": False,
         "installed_tag": None,
         "latest_tag": None,
         "installed_at_utc": None,
@@ -196,16 +256,25 @@ def check_prebuilt_freshness(
     if not marker:
         return out
     out["has_marker"] = True
+    # Display prefers the normalized base ("tag"); comparison below prefers the
+    # full "release_tag" -- deliberately opposite fallbacks.
     out["installed_tag"] = marker.get("tag") or marker.get("release_tag")
     out["installed_at_utc"] = marker.get("installed_at_utc")
     out["published_repo"] = marker.get("published_repo")
 
+    # The marker records both a normalized base tag ("tag", e.g. b9596) and the
+    # full release tag ("release_tag", e.g. b9596-mix-<sha>). Compare against the
+    # FULL identity, since GitHub /releases/latest returns the full tag_name --
+    # comparing the normalized base against the full latest is what produced the
+    # permanent "downgrade" banner on every mix release.
+    installed_full = marker.get("release_tag") or marker.get("tag")
     repo = out["published_repo"]
-    if not repo or not out["installed_tag"]:
+    if not repo or not installed_full:
         return out
     latest = latest_published_release(repo)
     out["latest_tag"] = latest
-    if not latest or latest == out["installed_tag"]:
+    out["behind"] = is_behind(installed_full, latest)
+    if not out["behind"]:
         return out
 
     installed_at = _parse_installed_at(out["installed_at_utc"])
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
new file mode 100644
index 0000000000..654ade6cd4
--- /dev/null
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -0,0 +1,560 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""In-app llama.cpp prebuilt update.
+
+Builds on utils.llama_cpp_freshness (which detects whether a newer prebuilt
+release exists) and adds the *apply* half: run install_llama_prebuilt.py to
+download the newest bundle for this host and atomically swap it in place, so
+the next model load uses it.
+
+Design notes:
+- Detection is delegated to check_prebuilt_freshness(). We surface an
+  ``update_available`` flag (installed_tag != latest_tag) which is laxer than
+  freshness' ``stale`` (which additionally requires the install to be >= 3 days
+  old). The UI shows the "Update llama.cpp" affordance on update_available.
+- The install is slow (download + extract + validate), so it runs on a daemon
+  thread; callers poll get_update_status() for the job state.
+- Everything fails open: a missing marker / offline GitHub / source build just
+  reports update_available=False and never blocks the app.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import sys
+import threading
+import time
+from pathlib import Path
+from typing import Optional
+
+import structlog
+
+from utils.llama_cpp_freshness import (
+    _INSTALL_MARKER_NAME,
+    check_prebuilt_freshness,
+    latest_published_release,
+    read_install_marker,
+    reset_caches,
+)
+
+logger = structlog.get_logger(__name__)
+
+DEFAULT_PUBLISHED_REPO = "unslothai/llama.cpp"
+_INSTALL_TIMEOUT_SECONDS = 1800  # 30 min ceiling for download + build/validate
+
+# Background job state. Single in-flight update at a time, guarded by _job_lock.
+_JOB_IDLE = "idle"
+_JOB_RUNNING = "running"
+_JOB_SUCCESS = "success"
+_JOB_ERROR = "error"
+
+_job_lock = threading.Lock()
+_job: dict = {
+    "state": _JOB_IDLE,
+    "message": "",
+    "from_tag": None,
+    "to_tag": None,
+    "error": None,
+    "progress": None,
+    "started_at": None,
+    "finished_at": None,
+}
+
+# Matches the installer's download progress lines, e.g.
+# "Downloading x.zip:  35.0% (12.3 MiB/35.1 MiB) at 8.2 MiB/s".
+_PROGRESS_LINE_RE = re.compile(r"(\d+(?:\.\d+)?)%\s*\(")
+# The download dominates the update; extract/validate fill the last slice.
+_DOWNLOAD_PROGRESS_CEILING = 0.95
+
+
+def _utcnow() -> str:
+    return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+
+
+def _find_binary() -> Optional[str]:
+    """Locate the active llama-server binary via the inference backend's own
+    resolver, so update targets exactly what Studio runs. Lazy import keeps the
+    heavy inference module off this module's import path."""
+    try:
+        from core.inference.llama_cpp import LlamaCppBackend
+        return LlamaCppBackend._find_llama_server_binary()
+    except Exception as exc:  # pragma: no cover - defensive
+        logger.debug("llama update: binary discovery failed", error = str(exc))
+        return None
+
+
+def _install_dir_for(binary_path: Optional[str]) -> Optional[Path]:
+    """The directory holding UNSLOTH_PREBUILT_INFO.json -- i.e. the install root
+    install_llama_prebuilt.py wrote and the one we re-install into. Walks up from
+    the binary the same way read_install_marker() does."""
+    if not binary_path:
+        return None
+    p = Path(binary_path)
+    for parent in p.parents[:5]:
+        if (parent / _INSTALL_MARKER_NAME).is_file():
+            return parent
+    return None
+
+
+def _installer_script() -> Optional[Path]:
+    """Locate install_llama_prebuilt.py. Honours UNSLOTH_LLAMA_INSTALLER, then
+    searches up from this file for both ``<root>/install_llama_prebuilt.py`` and
+    ``<root>/studio/install_llama_prebuilt.py`` so it works in the dev tree and
+    in an installed Studio layout."""
+    env = os.environ.get("UNSLOTH_LLAMA_INSTALLER")
+    if env and Path(env).is_file():
+        return Path(env)
+    here = Path(__file__).resolve()
+    for up in here.parents:
+        for cand in (up / "install_llama_prebuilt.py", up / "studio" / "install_llama_prebuilt.py"):
+            if cand.is_file():
+                return cand
+    return None
+
+
+# Markerless (source-build) installs have no UNSLOTH_PREBUILT_INFO.json, so we
+# ask the installer whether an official prebuilt now exists for this host. Memo
+# is 24h; only successful answers are cached so a network blip retries.
+_RESOLVE_TTL_SECONDS = 24 * 60 * 60
+_resolve_memo: dict = {}
+
+
+def _resolve_prebuilt_for_host(*, force_refresh: bool = False) -> Optional[dict]:
+    """Run install_llama_prebuilt.py --resolve-prebuilt (no download) and return
+    {prebuilt_available, repo, release_tag, llama_tag, asset, install_kind} or
+    None. Fail-open: any error -> None so a source build never blocks the app."""
+    now = time.time()
+    if not force_refresh and _resolve_memo:
+        if now - _resolve_memo.get("at", 0.0) < _RESOLVE_TTL_SECONDS:
+            return _resolve_memo.get("value")
+    script = _installer_script()
+    if script is None:
+        return None
+    value: Optional[dict] = None
+    try:
+        proc = subprocess.run(
+            [
+                sys.executable,
+                str(script),
+                "--resolve-prebuilt",
+                "latest",
+                "--output-format",
+                "json",
+            ],
+            capture_output = True,
+            text = True,
+            timeout = 60,
+        )
+        out = (proc.stdout or "").strip()
+        if proc.returncode == 0 and out:
+            parsed = json.loads(out.splitlines()[-1])
+            if isinstance(parsed, dict):
+                value = parsed
+    except Exception as exc:  # pragma: no cover - subprocess/json defensive
+        logger.debug("llama update: resolve-prebuilt failed", error = str(exc))
+        value = None
+    if value is not None:  # cache real answers; let failures retry next poll
+        _resolve_memo.update(at = now, value = value)
+    return value
+
+
+def _installed_build_number(binary: Optional[str]) -> Optional[int]:
+    """Best-effort build number from ``llama-server --version`` (e.g.
+    'version: 9585 (abc)'). None when unparseable or <= 1: a source build with
+    no git tags reports 'version: 1', which we treat as unknown (offer update)."""
+    if not binary:
+        return None
+    try:
+        proc = subprocess.run([binary, "--version"], capture_output = True, text = True, timeout = 20)
+    except Exception:  # pragma: no cover - defensive
+        return None
+    m = re.search(r"version:\s*(\d+)", (proc.stderr or "") + (proc.stdout or ""))
+    if not m:
+        return None
+    n = int(m.group(1))
+    return n if n > 1 else None
+
+
+def _is_under(path: Path, root: Path) -> bool:
+    try:
+        p, r = path.resolve(), root.resolve()
+    except (OSError, ValueError):
+        p, r = path, root
+    return p == r or r in p.parents
+
+
+def _llama_install_root(binary: Optional[str]) -> Optional[Path]:
+    """The Studio-managed llama.cpp root the active binary lives under, or None
+    when the binary is unmanaged. Installing anywhere the active binary is not
+    would not replace what _find_llama_server_binary runs (which prefers a pinned
+    LLAMA_SERVER_PATH, then UNSLOTH_LLAMA_CPP_PATH, then a llama.cpp tree), so we
+    refuse rather than silently install into an inactive or foreign tree."""
+    marked = _install_dir_for(binary)
+    if marked is not None:
+        return marked
+    if not binary:
+        return None
+    # LLAMA_SERVER_PATH is an explicit user pin that always wins in discovery;
+    # never auto-replace its tree (even a user's own llama.cpp checkout).
+    if os.environ.get("LLAMA_SERVER_PATH"):
+        return None
+    p = Path(binary)
+    env = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
+    if env and _is_under(p, Path(env)):
+        return Path(env)
+    for parent in p.parents:
+        if parent.name == "llama.cpp":
+            return parent
+    # PATH / system / custom install: not a managed tree, so do not offer.
+    return None
+
+
+def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]:
+    """Update status for a markerless (source-build) install: offer the official
+    prebuilt when one exists for this host and is newer than the installed
+    binary. None -> caller falls through to the no-marker default (unsupported)."""
+    res = _resolve_prebuilt_for_host(force_refresh = force_refresh)
+    if not res or not res.get("prebuilt_available"):
+        return None
+    # llama_tag is the upstream build (bNNNN, what --version reports); release_tag
+    # can be a fork wrapper tag, so compare/display against llama_tag.
+    latest = res.get("llama_tag") or res.get("release_tag")
+    if not latest:
+        return None
+    # No resolvable install root (e.g. a pinned LLAMA_SERVER_PATH we cannot
+    # manage) means an apply would not take effect, so do not offer.
+    if _llama_install_root(binary) is None:
+        return None
+    installed_build = _installed_build_number(binary)
+    m = re.search(r"(\d+)", latest)
+    latest_build = int(m.group(1)) if m else None
+    # Suppress only when the source build is reliably newer/equal; unknown
+    # version (the involuntary source-build case) is treated as behind.
+    update_available = (
+        installed_build is None or latest_build is None or installed_build < latest_build
+    )
+    with _job_lock:
+        job = dict(_job)
+    return {
+        "supported": True,
+        "update_available": update_available,
+        "stale": False,
+        "installed_tag": (f"b{installed_build}" if installed_build else None),
+        "latest_tag": latest,
+        "published_repo": res.get("repo"),
+        "installed_at_utc": None,
+        "age_days": None,
+        "source_build": True,
+        "job": job,
+    }
+
+
+def get_update_status(*, force_refresh: bool = False) -> dict:
+    """Report whether a newer prebuilt exists plus the current job state.
+
+    force_refresh bypasses the 24h release cache for an explicit "check now".
+    """
+    binary = _find_binary()
+    marker = read_install_marker(binary)
+
+    with _job_lock:
+        job_running = _job["state"] == _JOB_RUNNING
+
+    # No marker = source build / custom path. Offer the official prebuilt if one
+    # now exists for this host (this is why macOS source builds showed no button).
+    # Skipped while the updater swaps the tree: each 3s poll would exec the
+    # half-replaced binary (on Windows that exec can make the installer's
+    # os.replace fail) and the poller only consumes job progress.
+    if marker is None and binary is not None and not job_running:
+        src = _source_build_status(binary, force_refresh = force_refresh)
+        if src is not None:
+            return src
+
+    repo = (marker or {}).get("published_repo") or DEFAULT_PUBLISHED_REPO
+
+    if force_refresh and repo:
+        # Prime the cache so the freshness read below sees the newest tag.
+        try:
+            latest_published_release(repo, force_refresh = True)
+        except Exception as exc:  # pragma: no cover - network defensive
+            logger.debug("llama update: force refresh failed", error = str(exc))
+
+    freshness = check_prebuilt_freshness(binary)
+    installed = freshness.get("installed_tag")
+    latest = freshness.get("latest_tag")
+    # `behind` compares the full release identity with a base-build guard, so a
+    # lagging /releases/latest or a mix-tagged latest can't show a false update
+    # (see llama_cpp_freshness.is_behind).
+    update_available = bool(freshness.get("has_marker") and freshness.get("behind"))
+
+    with _job_lock:
+        job = dict(_job)
+
+    return {
+        "supported": bool(freshness.get("has_marker")),
+        "update_available": update_available,
+        "stale": bool(freshness.get("stale")),
+        "installed_tag": installed,
+        "latest_tag": latest,
+        "published_repo": freshness.get("published_repo") or repo,
+        "installed_at_utc": freshness.get("installed_at_utc"),
+        "age_days": freshness.get("age_days"),
+        "source_build": False,
+        "job": job,
+    }
+
+
+def _rocm_install_args(asset: Optional[str]) -> list[str]:
+    """Forward --rocm-gfx/--has-rocm from the marker asset, mirroring setup.sh.
+    The installer probe can miss the gfx arch on amd-smi-only hosts; lemonade
+    bundles carry the family in the name (rocm-gfx110X), fork bundles only rocm/hip."""
+    if not asset:
+        return []
+    low = asset.lower()
+    if "rocm" not in low and "hip" not in low:
+        return []
+    gfx = re.search(r"-gfx[0-9a-z]+", low)
+    if gfx:
+        # _normalize_forwarded_gfx accepts the family form (gfx110x -> gfx110X).
+        return ["--rocm-gfx", gfx.group(0).lstrip("-")]
+    return ["--has-rocm"]
+
+
+def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path) -> None:
+    """Worker: put the backend into a maintenance state, run the installer for
+    the latest prebuilt, then refresh caches so the next load uses the new build."""
+    backend = None
+    model_was_active = False
+    try:
+        # Maintenance state so no load starts a server from the half-swapped binary
+        # (and the old binary is freed for the swap). Fails open without a backend.
+        try:
+            from routes.inference import get_llama_cpp_backend
+            backend = get_llama_cpp_backend()
+        except Exception as exc:
+            logger.debug(
+                "llama update: backend unavailable, skipping load coordination", error = str(exc)
+            )
+            backend = None
+
+        if backend is not None:
+            try:
+                with backend._serial_load_lock:
+                    backend._llama_update_in_progress = True
+                    # is_active covers the loading/unhealthy window is_loaded misses
+                    # (a live process also locks the exe on Windows during the swap).
+                    if getattr(backend, "is_active", False):
+                        model_was_active = True
+                        backend.unload_model()
+            except Exception as exc:
+                logger.debug("llama update: load coordination failed", error = str(exc))
+
+        cmd = [
+            sys.executable,
+            str(script),
+            "--install-dir",
+            str(install_dir),
+            "--llama-tag",
+            "latest",
+            "--published-repo",
+            repo,
+        ]
+        cmd.extend(_rocm_install_args(asset))
+        logger.info("llama update: installing", cmd = " ".join(cmd))
+        # Stream the installer output so download percent lines feed
+        # job["progress"]; finer milestones via UNSLOTH_PROGRESS_PERCENT_STEP.
+        env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
+        proc = subprocess.Popen(
+            cmd,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
+            text = True,
+            env = env,
+        )
+        timed_out = threading.Event()
+
+        def _kill_on_timeout() -> None:
+            timed_out.set()
+            proc.kill()
+
+        watchdog = threading.Timer(_INSTALL_TIMEOUT_SECONDS, _kill_on_timeout)
+        watchdog.daemon = True
+        watchdog.start()
+        tail_lines: list[str] = []
+        try:
+            assert proc.stdout is not None
+            for line in proc.stdout:
+                tail_lines.append(line)
+                if len(tail_lines) > 80:
+                    del tail_lines[0]
+                m = _PROGRESS_LINE_RE.search(line)
+                if m is None:
+                    continue
+                fraction = min(float(m.group(1)) / 100.0, 1.0) * _DOWNLOAD_PROGRESS_CEILING
+                with _job_lock:
+                    _job["progress"] = max(_job.get("progress") or 0.0, fraction)
+            returncode = proc.wait()
+        finally:
+            watchdog.cancel()
+        if timed_out.is_set():
+            raise RuntimeError(f"installer timed out after {_INSTALL_TIMEOUT_SECONDS}s")
+        if returncode != 0:
+            tail = "".join(tail_lines).strip()[-1500:]
+            raise RuntimeError(f"installer exited {returncode}: {tail or 'no output'}")
+
+        # New UNSLOTH_PREBUILT_INFO.json is on disk; drop in-memory caches and
+        # re-prime the 24h disk freshness cache with the true newest, so the
+        # banner can't linger on a stale same-base value after the swap.
+        reset_caches()
+        try:
+            latest_published_release(repo, force_refresh = True)
+        except Exception as exc:  # pragma: no cover - network defensive
+            logger.debug("llama update: post-install freshness refresh failed", error = str(exc))
+        new_marker = read_install_marker(_find_binary())
+        new_tag = (new_marker or {}).get("tag") or (new_marker or {}).get("release_tag")
+
+        with _job_lock:
+            _job.update(
+                state = _JOB_SUCCESS,
+                message = (
+                    f"Updated llama.cpp to {new_tag}."
+                    + (" Reload your model to use it." if model_was_active else "")
+                ),
+                to_tag = new_tag,
+                error = None,
+                progress = 1.0,
+                finished_at = _utcnow(),
+            )
+        logger.info("llama update: success", to_tag = new_tag)
+    except Exception as exc:
+        logger.warning("llama update: failed", error = str(exc))
+        with _job_lock:
+            _job.update(
+                state = _JOB_ERROR,
+                message = "llama.cpp update failed.",
+                error = str(exc),
+                finished_at = _utcnow(),
+            )
+    finally:
+        # Lift the maintenance state so model loads work again, success or not.
+        if backend is not None:
+            try:
+                backend._llama_update_in_progress = False
+            except Exception:  # pragma: no cover - defensive
+                pass
+
+
+def start_update() -> dict:
+    """Kick off a background update. Idempotent: a second call while one is
+    running returns the in-flight job rather than starting another."""
+    binary = _find_binary()
+    marker = read_install_marker(binary)
+    script = _installer_script()
+    if script is None:
+        return {
+            "started": False,
+            "reason": "installer_missing",
+            "message": "install_llama_prebuilt.py could not be located.",
+            "job": get_update_status()["job"],
+        }
+
+    # A job already in flight wins over any freshness re-check below (and skips
+    # its network call). The final lock block re-checks to close the TOCTOU.
+    with _job_lock:
+        if _job["state"] == _JOB_RUNNING:
+            return {"started": False, "reason": "already_running", "job": dict(_job)}
+
+    if marker:
+        # Mirror the detection guard: a direct POST or a stale banner must not
+        # start an install when the latest is not actually newer (force a fresh
+        # check so a stale 24h cache can't wrongly block a real update either).
+        status = get_update_status(force_refresh = True)
+        if not status.get("update_available"):
+            return {
+                "started": False,
+                "reason": "up_to_date",
+                "message": "The installed llama.cpp build is already at the latest prebuilt.",
+                "job": status["job"],
+            }
+        install_dir = _install_dir_for(binary)
+        repo = marker.get("published_repo") or DEFAULT_PUBLISHED_REPO
+        from_tag = marker.get("tag") or marker.get("release_tag")
+        asset = marker.get("asset")
+    else:
+        # Source build / custom path: only proceed when the same detection logic
+        # would offer the update (prebuilt exists, install is behind, root is
+        # manageable), so a direct POST cannot downgrade a newer source build.
+        src = _source_build_status(binary, force_refresh = True) if binary else None
+        if src is None:
+            return {
+                "started": False,
+                "reason": "no_prebuilt_available",
+                "message": (
+                    "No official llama.cpp prebuilt is available for this host, "
+                    "so the source build cannot be swapped automatically."
+                ),
+                "job": get_update_status()["job"],
+            }
+        if not src.get("update_available"):
+            return {
+                "started": False,
+                "reason": "up_to_date",
+                "message": "The installed llama.cpp build is already at or newer than the latest prebuilt.",
+                "job": get_update_status()["job"],
+            }
+        res = _resolve_prebuilt_for_host()
+        install_dir = _llama_install_root(binary)
+        repo = (res or {}).get("repo") or DEFAULT_PUBLISHED_REPO
+        from_tag = None
+        asset = (res or {}).get("asset")
+
+    if install_dir is None:
+        return {
+            "started": False,
+            "reason": "no_install_dir",
+            "message": "Could not determine the llama.cpp install directory.",
+            "job": get_update_status()["job"],
+        }
+
+    with _job_lock:
+        if _job["state"] == _JOB_RUNNING:
+            return {"started": False, "reason": "already_running", "job": dict(_job)}
+        _job.update(
+            state = _JOB_RUNNING,
+            message = "Downloading and installing the latest llama.cpp prebuilt...",
+            from_tag = from_tag,
+            to_tag = None,
+            error = None,
+            progress = 0.0,
+            started_at = _utcnow(),
+            finished_at = None,
+        )
+        job_snapshot = dict(_job)
+
+    thread = threading.Thread(
+        target = _run_update,
+        args = (install_dir, repo, asset, script),
+        name = "llama-cpp-update",
+        daemon = True,
+    )
+    thread.start()
+    return {"started": True, "reason": None, "job": job_snapshot}
+
+
+def _reset_job_for_tests() -> None:
+    """Test-only: return the job tracker to idle."""
+    with _job_lock:
+        _job.update(
+            state = _JOB_IDLE,
+            message = "",
+            from_tag = None,
+            to_tag = None,
+            error = None,
+            progress = None,
+            started_at = None,
+            finished_at = None,
+        )
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 0df7a1a477..f61c210cf6 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -58,25 +58,28 @@ import re as _re
 _MODEL_SIZE_RE = _re.compile(r"(?:^|[-_/])(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
 # MoE active-parameter pattern: "A3B", "A3.5B", etc.
 _ACTIVE_SIZE_RE = _re.compile(r"(?:^|[-_/])a(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
+# Gemma 3n/4 effective-parameter pattern: "E2B", "E4B" -- the runtime
+# footprint (MatFormer + per-layer embeddings), which is the size that
+# matters for size-gated policies like sub-3B speculative-decoding fallback.
+_EFFECTIVE_SIZE_RE = _re.compile(r"(?:^|[-_/])e(\d+\.?\d*)\s*([bm])(?:$|[-_/])", _re.IGNORECASE)
 
 
 def extract_model_size_b(model_id: str) -> float | None:
     """Extract model size in billions from a model identifier.
 
     Prefers MoE active-parameter notation (e.g. ``A3B`` in
-    ``Qwen3.5-35B-A3B``) over total params. Handles ``B`` (billions)
-    and ``M`` (millions) suffixes.
+    ``Qwen3.5-35B-A3B``), then Gemma effective-parameter notation
+    (e.g. ``E2B``), over total params. Handles ``B`` (billions) and
+    ``M`` (millions) suffixes.
     """
     mid = (model_id or "").lower()
-    active = _ACTIVE_SIZE_RE.search(mid)
-    if active:
-        val = float(active.group(1))
-        return val / 1000.0 if active.group(2).lower() == "m" else val
-    size = _MODEL_SIZE_RE.search(mid)
-    if not size:
-        return None
-    val = float(size.group(1))
-    return val / 1000.0 if size.group(2).lower() == "m" else val
+    # First match wins, in priority order: active > effective > total.
+    for pattern in (_ACTIVE_SIZE_RE, _EFFECTIVE_SIZE_RE, _MODEL_SIZE_RE):
+        m = pattern.search(mid)
+        if m:
+            val = float(m.group(1))
+            return val / 1000.0 if m.group(2).lower() == "m" else val
+    return None
 
 
 # Maps equivalent model names to their canonical YAML config file.
@@ -510,7 +513,7 @@ _VLM_MODEL_TYPES = {
 _AUDIO_ONLY_MODEL_TYPES = {"csm", "whisper"}
 
 # Pre-computed .venv_t5 paths and backend dir for subprocess version switching.
-# Vision check uses 5.5.0 (newest, recognizes all architectures).
+# Vision check uses the Gemma 4 5.5 sidecar for existing Gemma 4 architectures.
 from utils.paths.storage_roots import studio_root as _studio_root  # noqa: E402
 
 _VENV_T5_DIR = str(_studio_root() / ".venv_t5_550")
@@ -930,6 +933,24 @@ def _is_mmproj(filename: str) -> bool:
     return "mmproj" in filename.lower()
 
 
+def _is_mtp_drafter(path: str) -> bool:
+    """True for a separate-file MTP drafter (speculative head), a companion
+    to the main model rather than a selectable quant: the repo-root
+    ``mtp-*.gguf`` or the ``MTP/`` subdir copies (Gemma 4).
+
+    Mirrors hub.utils.gguf.is_mtp_drafter_path (utils cannot import hub).
+    Must be excluded everywhere mmproj is, or the drafter leaks into variant
+    menus (a phantom quant) and quant-matched file lookups -- e.g. a ``Q8_0``
+    request must not resolve to ``MTP/...-Q8_0-MTP.gguf``, which sorts ahead
+    of the real weight.
+    """
+    p = path.lower()
+    if not p.endswith(".gguf"):
+        return False
+    name = p.rsplit("/", 1)[-1]
+    return name.startswith("mtp-") or "/mtp/" in f"/{p}"
+
+
 # Family tokens for #5347's filename fallback. Lowercase; order irrelevant.
 _MODEL_FAMILY_TOKENS: tuple[str, ...] = (
     "qwen",
@@ -1137,6 +1158,48 @@ def detect_mmproj_file(path: str, search_root: Optional[str] = None) -> Optional
     return str(best[1])
 
 
+def detect_mtp_file(path: str, search_root: Optional[str] = None) -> Optional[str]:
+    """Find the separate MTP drafter (``mtp-*.gguf``) for a local GGUF model.
+
+    The drafter that pairs with the main weights sits at the repo/snapshot
+    root (Gemma 4); the weight itself may be at the root or in a quant subdir,
+    so scan the weight's directory and ``search_root``. Matches by the
+    ``mtp-`` filename prefix unsloth uses for ``-hf`` auto-discovery -- the
+    same signal as the HF download path. Repos that bake the head into the
+    main GGUF (Qwen) have no such sibling, so this returns None.
+
+    Pairs by name so a multi-model folder can't attach a foreign drafter:
+    unsloth names the drafter ``mtp-<model>.gguf`` where ``<model>`` prefixes
+    the weight filename across all Gemma 4 repos (e.g.
+    ``mtp-gemma-4-12B-it.gguf`` next to ``gemma-4-12B-it-qat-Q4_0.gguf``).
+    An unmatched drafter is skipped (fail-safe: no MTP).
+    """
+    p = Path(path)
+    weight_name = p.name.lower() if p.suffix.lower() == ".gguf" else None
+    start_dir = p.parent if p.is_file() else p
+    dirs = [start_dir]
+    if search_root is not None:
+        dirs.append(Path(search_root))
+    for d in dirs:
+        try:
+            entries = sorted(d.iterdir())
+        except OSError:
+            continue
+        for f in entries:
+            name = f.name.lower()
+            if not (name.startswith("mtp-") and name.endswith(".gguf")):
+                continue
+            stem = name[len("mtp-") : -len(".gguf")]
+            if not stem or (weight_name is not None and not weight_name.startswith(stem)):
+                continue
+            try:
+                if f.is_file():
+                    return str(f.resolve())
+            except OSError:
+                continue
+    return None
+
+
 def detect_gguf_model(path: str) -> Optional[str]:
     """Check if a local path is or contains a GGUF model file.
 
@@ -1148,7 +1211,12 @@ def detect_gguf_model(path: str) -> Optional[str]:
 
     # Case 1: direct .gguf file
     if p.suffix.lower() == ".gguf":
-        if _is_mmproj(p.name):
+        # Companions are not models: rejecting a drafter here also keeps
+        # detect_mtp_file from pairing the same file with itself
+        # (-m drafter --model-draft drafter). Include the immediate parent
+        # dir so the MTP/ subdir copies are caught -- the basename alone
+        # (...-MTP.gguf) doesn't match the predicate's mtp- prefix.
+        if _is_mmproj(p.name) or _is_mtp_drafter(f"{p.parent.name}/{p.name}"):
             return None
         # Extension is authoritative: don't gate on is_file()/exists(), which
         # can fail in the Windows lock window after llama-server is killed.
@@ -1160,10 +1228,14 @@ def detect_gguf_model(path: str) -> Optional[str]:
             return str(p.absolute())  # absolute() keeps symlink names readable
         # Directory named "*.gguf": fall through to the dir scan below.
 
-    # Case 2: directory containing .gguf files (skip mmproj)
+    # Case 2: directory containing .gguf files (skip mmproj / MTP drafter)
     if p.is_dir():
         gguf_files = sorted(
-            (f for f in _iter_gguf_files(p) if not _is_mmproj(f.name)),
+            (
+                f
+                for f in _iter_gguf_files(p)
+                if not _is_mmproj(f.name) and not _is_mtp_drafter(f"{f.parent.name}/{f.name}")
+            ),
             key = lambda f: f.stat().st_size,
             reverse = True,
         )
@@ -1393,6 +1465,9 @@ def list_gguf_variants(
         if "mmproj" in fname.lower():
             has_vision = True
             continue
+        # MTP drafters are speculative-decoding companions, not quants.
+        if _is_mtp_drafter(fname):
+            continue
 
         quant = _extract_quant_label(fname)
         quant_totals[quant] = quant_totals.get(quant, 0) + size
@@ -1468,6 +1543,8 @@ def list_local_gguf_variants(directory: str) -> tuple[list[GgufVariantInfo], boo
         # Use the relative path so ``BF16/foo.gguf`` and ``Q4_K_M/foo.gguf``
         # get distinct quant labels instead of collapsing on basename.
         rel = f.relative_to(p).as_posix()
+        if _is_mtp_drafter(rel):
+            continue
         quant = _extract_quant_label(rel)
         quant_totals[quant] = quant_totals.get(quant, 0) + size
         if quant not in quant_first_file:
@@ -1504,7 +1581,9 @@ def _find_local_gguf_by_variant(directory: str, variant: str) -> Optional[str]:
     matches = sorted(
         f
         for f in _iter_gguf_files(p, recursive = True)
-        if not _is_mmproj(f.name) and _extract_quant_label(f.relative_to(p).as_posix()) == variant
+        if not _is_mmproj(f.name)
+        and not _is_mtp_drafter(f.relative_to(p).as_posix())
+        and _extract_quant_label(f.relative_to(p).as_posix()) == variant
     )
     if matches:
         return str(matches[0].resolve())
@@ -1519,9 +1598,9 @@ def _detect_gguf_from_hf_cache(repo_id: str) -> Optional[str]:
     """
     for snap in _iter_hf_cache_snapshots(repo_id):
         rel_files = [
-            f.relative_to(snap).as_posix()
+            rel
             for f in _iter_gguf_files(snap, recursive = True)
-            if not _is_mmproj(f.name)
+            if not _is_mtp_drafter(rel := f.relative_to(snap).as_posix()) and not _is_mmproj(f.name)
         ]
         if rel_files:
             return _pick_best_gguf(rel_files)
@@ -2068,6 +2147,7 @@ class ModelConfig:
     has_audio_input: bool = False  # Accepts audio input (ASR/speech understanding)
     gguf_file: Optional[str] = None  # Full path to the .gguf file (local mode)
     gguf_mmproj_file: Optional[str] = None  # Full path to the mmproj .gguf file (vision projection)
+    gguf_mtp_file: Optional[str] = None  # Full path to the separate MTP drafter (local mode)
     gguf_hf_repo: Optional[str] = (
         None  # HF repo ID for -hf mode (e.g. "unsloth/gemma-3-4b-it-GGUF")
     )
@@ -2207,6 +2287,11 @@ class ModelConfig:
                 elif base_is_vision:
                     logger.warning(f"Base model is vision but no mmproj file found in {gguf_dir}")
 
+                # Separate MTP drafter sibling (Gemma 4), mirroring mmproj.
+                mtp_file = detect_mtp_file(gguf_file, search_root = path)
+                if mtp_file:
+                    logger.info(f"Detected MTP drafter: {mtp_file}")
+
                 return cls(
                     identifier = identifier,
                     display_name = display_name,
@@ -2218,6 +2303,7 @@ class ModelConfig:
                     is_gguf = True,
                     gguf_file = gguf_file,
                     gguf_mmproj_file = mmproj_file,
+                    gguf_mtp_file = mtp_file,
                 )
         else:
             # Does the HF repo contain GGUF files?
diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py
index 10597a689b..2d4f5ac243 100644
--- a/studio/backend/utils/paths/storage_roots.py
+++ b/studio/backend/utils/paths/storage_roots.py
@@ -60,6 +60,11 @@ def cache_root() -> Path:
     return studio_root() / "cache"
 
 
+def studio_bin_root() -> Path:
+    """Dir for Studio-managed executables (the `unsloth` shim, downloaded tools like cloudflared)."""
+    return studio_root() / "bin"
+
+
 def assets_root() -> Path:
     return studio_root() / "assets"
 
diff --git a/studio/backend/utils/studio_version.py b/studio/backend/utils/studio_version.py
index 98c48fe45c..9c18070fbb 100644
--- a/studio/backend/utils/studio_version.py
+++ b/studio/backend/utils/studio_version.py
@@ -15,6 +15,7 @@ _DEV_VERSION = "dev"
 _GIT_TIMEOUT_SECONDS = 1.0
 _STUDIO_TAG_RE = re.compile(r"^v\d+\.\d+\.\d+(?:-[0-9A-Za-z.][0-9A-Za-z.-]*)?$")
 _GIT_DESCRIBE_SUFFIX_RE = re.compile(r"-\d+-g[0-9A-Fa-f]+(?:-dirty)?$")
+_GIT_BRANCH_RE = re.compile(r"^[0-9A-Za-z._/-]+$")
 _MAX_VERSION_LENGTH = 64
 
 
@@ -71,6 +72,35 @@ def _exact_git_studio_tag(repo_root: Path) -> str | None:
     return tag if is_valid_studio_release_version(tag) else None
 
 
+def _git_branch(repo_root: Path) -> str | None:
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--abbrev-ref", "HEAD"],
+            cwd = repo_root,
+            check = False,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            text = True,
+            timeout = _GIT_TIMEOUT_SECONDS,
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        return None
+
+    if result.returncode != 0:
+        return None
+
+    branch = result.stdout.strip()
+    # "HEAD" means detached, e.g. a tag or commit checkout.
+    if (
+        not branch
+        or branch == "HEAD"
+        or len(branch) > _MAX_VERSION_LENGTH
+        or _GIT_BRANCH_RE.fullmatch(branch) is None
+    ):
+        return None
+    return branch
+
+
 def get_studio_version(repo_root: Path | None = None) -> str:
     """Return the installed Studio release tag for display, or ``dev``.
 
@@ -81,7 +111,10 @@ def get_studio_version(repo_root: Path | None = None) -> str:
 
     if _is_source_checkout(resolved_repo_root):
         git_tag = _exact_git_studio_tag(resolved_repo_root)
-        return git_tag if git_tag is not None else _DEV_VERSION
+        if git_tag is not None:
+            return git_tag
+        branch = _git_branch(resolved_repo_root)
+        return f"GitHub {branch}" if branch is not None else _DEV_VERSION
 
     stamped_version = _studio_release_build.STUDIO_RELEASE_VERSION
     if is_valid_studio_release_version(stamped_version):
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index bf8d14b7cb..6e1571d5ce 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -3,12 +3,27 @@
 
 """Automatic transformers version switching.
 
-Some newer architectures need transformers>=5.3.0 (.venv_t5_530/); Gemma 4
-needs >=5.5.0 (.venv_t5_550/). Everything else uses the default 4.57.x. A
-custom-named LoRA adapter's base model is resolved from adapter_config.json.
+Some newer model architectures (Ministral-3, GLM-4.7-Flash, Qwen3-30B-A3B MoE,
+tiny_qwen3_moe) require transformers>=5.3.0, while Gemma 4 models require a
+newer 5.x sidecar.  Everything else needs the default 4.57.x that ships with
+Unsloth.
 
-Training/inference run in subprocesses that activate the right version via
-sys.path; export (in-process) uses ensure_transformers_version() for the swap.
+Two separate target directories are maintained:
+  - .venv_t5_530/  — transformers 5.3.0 (Ministral-3, GLM, Qwen3 MoE, etc.)
+  - .venv_t5_550/  — transformers 5.5.0 (Gemma 4)
+  - .venv_t5_510/  — transformers 5.10.2 (Gemma 4 Unified / 12B)
+
+When loading a LoRA adapter with a custom name, we resolve the base model from
+``adapter_config.json`` and check *that* against the model list.
+
+Strategy:
+  Training and inference run in subprocesses that activate the correct version
+  via sys.path (prepending the appropriate .venv_t5_*/ directory). See:
+    - core/training/worker.py
+    - core/inference/worker.py
+
+  For export (still in-process), ensure_transformers_version() does a lightweight
+  sys.path swap using the same directories pre-installed by setup.sh.
 """
 
 import importlib
@@ -53,13 +68,32 @@ TRANSFORMERS_5_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "lfm2.5-vl-450m",  # LiquidAI/LFM2.5-VL-450M
 )
 
-# Lowercase substrings for models that require transformers 5.5.0 (checked first).
+# Lowercase substrings for models that require transformers 5.10.x (checked first).
+TRANSFORMERS_510_MODEL_SUBSTRINGS: tuple[str, ...] = (
+    "gemma-4-12b",  # Gemma 4 Unified 12B
+    "gemma4-12b",
+)
+
+# Lowercase substrings for models that require the Gemma 4 transformers 5.5 sidecar.
 TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "gemma-4",  # Gemma-4 (E2B-it, E4B-it, 31B-it, 26B-A4B-it)
     "gemma4",  # Gemma-4 alternate naming
     "qwen3.6",
 )
 
+# Architecture classes / model_type values that require transformers 5.10.x.
+# Checked via config.json (local or HuggingFace).
+_TRANSFORMERS_510_ARCHITECTURES: set[str] = {
+    "Gemma4UnifiedForConditionalGeneration",
+    "Gemma4AssistantForCausalLM",
+    "Gemma4UnifiedAssistantForCausalLM",
+}
+_TRANSFORMERS_510_MODEL_TYPES: set[str] = {
+    "gemma4_unified",
+    "gemma4_assistant",
+    "gemma4_unified_assistant",
+}
+
 # Architecture classes / model_type values that require transformers 5.5.0.
 # Checked via config.json (local or HuggingFace).
 _TRANSFORMERS_550_ARCHITECTURES: set[str] = {
@@ -78,21 +112,26 @@ _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
 _tokenizer_class_cache: dict[str, bool] = {}
 
 # Cache for dynamic config.json lookups (architecture/model_type checks).
+_config_json_cache: dict[str, dict | None] = {}
+_config_needs_510_cache: dict[str, bool] = {}
 _config_needs_550_cache: dict[str, bool] = {}
 
 # Versions
+TRANSFORMERS_510_VERSION = "5.10.2"
 TRANSFORMERS_550_VERSION = "5.5.0"
 TRANSFORMERS_530_VERSION = "5.3.0"
 TRANSFORMERS_DEFAULT_VERSION = "4.57.6"
-# Backwards-compat alias — points to 5.5.0 (the highest 5.x tier).
-# Consumers should prefer TRANSFORMERS_530_VERSION / TRANSFORMERS_550_VERSION.
-TRANSFORMERS_5_VERSION = TRANSFORMERS_550_VERSION
+# Backwards-compat alias — points to the highest 5.x tier.
+# Consumers should prefer TRANSFORMERS_510_VERSION / TRANSFORMERS_550_VERSION /
+# TRANSFORMERS_530_VERSION.
+TRANSFORMERS_5_VERSION = TRANSFORMERS_510_VERSION
 
 # Pre-installed directories — created by setup.sh / setup.ps1.
 from utils.paths.storage_roots import studio_root as _studio_root  # noqa: E402
 
 _VENV_T5_530_DIR = str(_studio_root() / ".venv_t5_530")
 _VENV_T5_550_DIR = str(_studio_root() / ".venv_t5_550")
+_VENV_T5_510_DIR = str(_studio_root() / ".venv_t5_510")
 # Backwards-compat alias
 _VENV_T5_DIR = _VENV_T5_550_DIR
 
@@ -108,15 +147,34 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
     resolved = _resolve_base_model(model_name)
     tier = get_transformers_tier(resolved)
 
-    if tier == "550":
+    if tier == "510":
+        if not _ensure_venv_t5_510_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers {TRANSFORMERS_510_VERSION}: "
+                f".venv_t5_510 missing at {_VENV_T5_510_DIR}"
+            )
+        if _VENV_T5_510_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_510_DIR)
+        logger.info(
+            "Activated transformers %s from %s",
+            TRANSFORMERS_510_VERSION,
+            _VENV_T5_510_DIR,
+        )
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_510_DIR + (os.pathsep + _pp if _pp else "")
+    elif tier == "550":
         if not _ensure_venv_t5_550_exists():
             raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: "
+                f"Cannot activate transformers {TRANSFORMERS_550_VERSION}: "
                 f".venv_t5_550 missing at {_VENV_T5_550_DIR}"
             )
         if _VENV_T5_550_DIR not in sys.path:
             sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
+        logger.info(
+            "Activated transformers %s from %s",
+            TRANSFORMERS_550_VERSION,
+            _VENV_T5_550_DIR,
+        )
         _pp = os.environ.get("PYTHONPATH", "")
         os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
     elif tier == "530":
@@ -260,6 +318,67 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
+def _load_config_json(model_name: str) -> dict | None:
+    """Return parsed ``config.json`` for *model_name*, checking local files first."""
+    if model_name in _config_json_cache:
+        return _config_json_cache[model_name]
+
+    local_cfg = Path(model_name) / "config.json"
+    if local_cfg.is_file():
+        try:
+            with open(local_cfg) as f:
+                cfg = json.load(f)
+            _config_json_cache[model_name] = cfg
+            return cfg
+        except Exception as exc:
+            logger.debug("Could not read %s: %s", local_cfg, exc)
+            _config_json_cache[model_name] = None
+            return None
+
+    if _env_offline():
+        _config_json_cache[model_name] = None
+        return None
+
+    import urllib.request
+
+    url = f"https://huggingface.co/{model_name}/raw/main/config.json"
+    try:
+        req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
+        with urllib.request.urlopen(req, timeout = 10) as resp:
+            cfg = json.loads(resp.read().decode())
+        _config_json_cache[model_name] = cfg
+        return cfg
+    except Exception as exc:
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
+        _config_json_cache[model_name] = None
+        return None
+
+
+def _config_matches_tier(cfg: dict, architectures: set[str], model_types: set[str]) -> bool:
+    archs = cfg.get("architectures", [])
+    if any(a in architectures for a in archs):
+        return True
+    if cfg.get("model_type") in model_types:
+        return True
+    return False
+
+
+def _config_needs_550(cfg: dict) -> bool:
+    return _config_matches_tier(
+        cfg,
+        _TRANSFORMERS_550_ARCHITECTURES,
+        _TRANSFORMERS_550_MODEL_TYPES,
+    )
+
+
+def _config_needs_510(cfg: dict) -> bool:
+    return _config_matches_tier(
+        cfg,
+        _TRANSFORMERS_510_ARCHITECTURES,
+        _TRANSFORMERS_510_MODEL_TYPES,
+    )
+
+
 def _check_config_needs_550(model_name: str) -> bool:
     """True if ``config.json`` has architectures/model_type needing transformers
     5.5.0 (e.g. Gemma 4).
@@ -270,81 +389,88 @@ def _check_config_needs_550(model_name: str) -> bool:
     if model_name in _config_needs_550_cache:
         return _config_needs_550_cache[model_name]
 
-    def _check_cfg(cfg: dict) -> bool:
-        archs = cfg.get("architectures", [])
-        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-            return True
-        if cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES:
-            return True
-        return False
-
-    # --- Check local config.json first ------------------------------------
-    local_path = Path(model_name)
-    local_cfg = local_path / "config.json"
-    if local_cfg.is_file():
-        try:
-            with open(local_cfg) as f:
-                cfg = json.load(f)
-            result = _check_cfg(cfg)
-            if result:
-                logger.info(
-                    "Local config.json check: %s needs transformers 5.5.0 "
-                    "(architectures=%s, model_type=%s)",
-                    model_name,
-                    cfg.get("architectures", []),
-                    cfg.get("model_type"),
-                )
-            _config_needs_550_cache[model_name] = result
-            return result
-        except Exception as exc:
-            logger.debug("Could not read %s: %s", local_cfg, exc)
-
-    # Offline: skip the 10s urllib fetch (fail-open to lower tier).
-    if _env_offline():
+    cfg = _load_config_json(model_name)
+    if cfg is None:
         _config_needs_550_cache[model_name] = False
         return False
 
-    # --- Fall back to fetching from HuggingFace ---------------------------
-    import urllib.request
+    result = _config_needs_550(cfg)
+    if result:
+        logger.info(
+            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
+            model_name,
+            TRANSFORMERS_550_VERSION,
+            cfg.get("architectures", []),
+            cfg.get("model_type"),
+        )
+    _config_needs_550_cache[model_name] = result
+    return result
 
-    url = f"https://huggingface.co/{model_name}/raw/main/config.json"
-    try:
-        req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
-        with urllib.request.urlopen(req, timeout = 10) as resp:
-            cfg = json.loads(resp.read().decode())
-        result = _check_cfg(cfg)
-        if result:
-            logger.info(
-                "Dynamic config.json check: %s needs transformers 5.5.0 "
-                "(architectures=%s, model_type=%s)",
-                model_name,
-                cfg.get("architectures", []),
-                cfg.get("model_type"),
-            )
-        _config_needs_550_cache[model_name] = result
-        return result
-    except Exception as exc:
-        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
-        _config_needs_550_cache[model_name] = False
+
+def _check_config_needs_510(model_name: str) -> bool:
+    """Check ``config.json`` for Gemma 4 Unified / 12B architectures."""
+    if model_name in _config_needs_510_cache:
+        return _config_needs_510_cache[model_name]
+
+    cfg = _load_config_json(model_name)
+    if cfg is None:
+        _config_needs_510_cache[model_name] = False
         return False
 
+    result = _config_needs_510(cfg)
+    if result:
+        logger.info(
+            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
+            model_name,
+            TRANSFORMERS_510_VERSION,
+            cfg.get("architectures", []),
+            cfg.get("model_type"),
+        )
+    _config_needs_510_cache[model_name] = result
+    return result
+
 
 def get_transformers_tier(model_name: str) -> str:
     """Return the transformers tier required for *model_name*.
 
-    ``"550"`` for transformers 5.5.0 (e.g. Gemma 4), ``"530"`` for 5.3.0
-    (e.g. Ministral-3, Qwen3 MoE), or ``"default"`` for everything else (4.57.x).
-    The 5.5.0 check runs first, then 5.3.0.
+    Returns ``"510"`` for models needing transformers 5.10.x (Gemma 4 Unified),
+    ``"550"`` for models needing transformers 5.5.0 (Gemma 4),
+    ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
+    or ``"default"`` for everything else (4.57.x).
+
+    Higher 5.x tiers run first.
     """
     lowered = model_name.lower()
 
+    # Local checkpoint names can contain architecture substrings in their
+    # directory names (for example a pytest temp dir). If config.json exists,
+    # trust it before using name heuristics.
+    local_cfg = Path(model_name) / "config.json"
+    if local_cfg.is_file():
+        cfg = _load_config_json(model_name)
+        if cfg is not None and _config_needs_510(cfg):
+            return "510"
+        if cfg is not None and _config_needs_550(cfg):
+            return "550"
+        if cfg is not None:
+            local_tc = Path(model_name) / "tokenizer_config.json"
+            if local_tc.is_file() and _check_tokenizer_config_needs_v5(model_name):
+                return "530"
+            return "default"
+
     # --- Fast substring checks (no I/O) ------------------------------------
+    if "assistant" in lowered and ("gemma-4" in lowered or "gemma4" in lowered):
+        return "510"
+    if any(sub in lowered for sub in TRANSFORMERS_510_MODEL_SUBSTRINGS):
+        return "510"
     if any(sub in lowered for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS):
         return "550"
     if any(sub in lowered for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS):
         return "530"
 
-    # --- Slow config fallbacks (local file first, then network) -----------
+    # --- Slow config fallbacks (network for HF IDs) ------------------------
+    if _check_config_needs_510(model_name):
+        return "510"
     if _check_config_needs_550(model_name):
         return "550"
     if _check_tokenizer_config_needs_v5(model_name):
@@ -419,6 +545,13 @@ _VENV_T5_530_PACKAGES = (
     "tiktoken",
 )
 
+_VENV_T5_510_PACKAGES = (
+    f"transformers=={TRANSFORMERS_510_VERSION}",
+    "huggingface_hub==1.8.0",
+    "hf_xet==1.4.2",
+    "tiktoken",
+)
+
 _VENV_T5_550_PACKAGES = (
     f"transformers=={TRANSFORMERS_550_VERSION}",
     "huggingface_hub==1.8.0",
@@ -475,7 +608,7 @@ def _venv_dir_is_valid(venv_dir: str, packages: tuple[str, ...]) -> bool:
 
 
 def _venv_t5_is_valid() -> bool:
-    """Backwards-compat: check the 5.5.0 venv."""
+    """Backwards-compat: check the Gemma 4 sidecar venv."""
     return _venv_dir_is_valid(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES)
 
 
@@ -553,11 +686,24 @@ def _ensure_venv_t5_530_exists() -> bool:
 
 def _ensure_venv_t5_550_exists() -> bool:
     """Ensure .venv_t5_550/ exists with transformers 5.5.0."""
-    return _ensure_venv_dir(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES, "transformers 5.5.0")
+    return _ensure_venv_dir(
+        _VENV_T5_550_DIR,
+        _VENV_T5_550_PACKAGES,
+        f"transformers {TRANSFORMERS_550_VERSION}",
+    )
+
+
+def _ensure_venv_t5_510_exists() -> bool:
+    """Ensure .venv_t5_510/ exists with transformers 5.10.x."""
+    return _ensure_venv_dir(
+        _VENV_T5_510_DIR,
+        _VENV_T5_510_PACKAGES,
+        f"transformers {TRANSFORMERS_510_VERSION}",
+    )
 
 
 def _ensure_venv_t5_exists() -> bool:
-    """Backwards-compat: ensure the 5.5.0 venv exists."""
+    """Backwards-compat: ensure the Gemma 4 5.5 sidecar venv exists."""
     return _ensure_venv_t5_550_exists()
 
 
@@ -577,7 +723,7 @@ def _activate_venv(venv_dir: str, label: str) -> None:
 
 def _deactivate_5x() -> None:
     """Remove all .venv_t5_*/ dirs from sys.path, purge stale modules, reimport."""
-    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR):
+    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR):
         while d in sys.path:
             sys.path.remove(d)
     logger.info("Removed venv_t5 dirs from sys.path")
@@ -593,7 +739,9 @@ def _deactivate_5x() -> None:
 def ensure_transformers_version(model_name: str) -> None:
     """Ensure the correct ``transformers`` version is active for *model_name*.
 
-    Uses sys.path with .venv_t5_530/ or .venv_t5_550/ (pre-installed by setup.sh):
+    Uses sys.path with .venv_t5_510/, .venv_t5_550/, or .venv_t5_530/
+    (pre-installed by setup.sh):
+      • Need 5.10.x → prepend .venv_t5_510/ to sys.path, purge modules.
       • Need 5.5.0 → prepend .venv_t5_550/ to sys.path, purge modules.
       • Need 5.3.0 → prepend .venv_t5_530/ to sys.path, purge modules.
       • Need 4.x  → remove all .venv_t5_*/ from sys.path, purge modules.
@@ -608,7 +756,11 @@ def ensure_transformers_version(model_name: str) -> None:
     resolved = _resolve_base_model(model_name)
     tier = get_transformers_tier(resolved)
 
-    if tier == "550":
+    if tier == "510":
+        target_version = TRANSFORMERS_510_VERSION
+        venv_dir = _VENV_T5_510_DIR
+        ensure_fn = _ensure_venv_t5_510_exists
+    elif tier == "550":
         target_version = TRANSFORMERS_550_VERSION
         venv_dir = _VENV_T5_550_DIR
         ensure_fn = _ensure_venv_t5_550_exists
@@ -643,7 +795,7 @@ def ensure_transformers_version(model_name: str) -> None:
                 model_name,
             )
             return
-        # Different 5.x → must switch (e.g. 5.3.0 loaded but need 5.5.0).
+        # Different 5.x -> need to switch (e.g. 5.3.0 loaded but need 5.10.x).
         in_memory_major = int(in_memory.split(".")[0])
         if in_memory_major == target_major and venv_dir is None:
             # Both are default (4.x) — close enough.
diff --git a/studio/frontend/.npmrc b/studio/frontend/.npmrc
index 8e21abe7a2..f2d15a4f15 100644
--- a/studio/frontend/.npmrc
+++ b/studio/frontend/.npmrc
@@ -6,11 +6,8 @@
 # upstream removal. npm interprets the bare integer as DAYS; do not
 # append `d`, npm 11.x will parse `7d` as a Date string and abort.
 min-release-age=7
-# Defensive alias: `minimum-release-age` takes minutes (10080 = 7 days).
-# Some npm versions / wrappers consult one key but not the other; setting
-# both means a single setting-name parse change upstream cannot silently
-# disable the cooldown. The two keys MUST agree; do not let them drift.
-minimum-release-age=10080
+# Do not re-add the old `minimum-release-age` alias: npm >=11.16 warns on
+# unknown project configs and npm 12 stops accepting them.
 # Belt-and-braces: refuse to write back loose `^x.y.z` ranges into
 # package.json when a maintainer runs `npm install <pkg>` locally. This
 # does NOT rewrite already-present ranges (those need an explicit
diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index 690da8c0c4..a521552d79 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -10264,9 +10264,9 @@
       }
     },
     "node_modules/hono": {
-      "version": "4.12.18",
-      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.18.tgz",
-      "integrity": "sha512-RWzP96k/yv0PQfyXnWjs6zot20TqfpfsNXhOnev8d1InAxubW93L11/oNUc3tQqn2G0bSdAOBpX+2uDFHV7kdQ==",
+      "version": "4.12.21",
+      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.21.tgz",
+      "integrity": "sha512-uV63apnb0kyPtAUwoWgaGh9HyIFcv8lgmzPZSiTBQAFOFGIzka5EZ1dZocmGnn0XdX0+XTqJ6Tqv7selMuGLRQ==",
       "license": "MIT",
       "engines": {
         "node": ">=16.9.0"
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index 1e527813f9..c49b62ab50 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -86,7 +86,7 @@
     "@tanstack/router-core": "1.169.2",
     "@tanstack/history": "1.161.6",
     "mermaid": "11.15.0",
-    "hono": "4.12.18",
+    "hono": "4.12.21",
     "qs": "6.15.2",
     "ip-address": "10.1.1",
     "brace-expansion@5.0.5": "5.0.6"
@@ -108,5 +108,10 @@
     "typescript": "~5.9.3",
     "typescript-eslint": "^8.55.0",
     "vite": "^8.0.1"
+  },
+  "allowScripts": {
+    "@biomejs/biome@1.9.4": true,
+    "msw@2.14.3": true,
+    "fsevents": true
   }
 }
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 25dbfdc780..91d99e238e 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -11,6 +11,7 @@ import {
 import { Toaster } from "@/components/ui/sonner";
 import { TooltipProvider } from "@/components/ui/tooltip";
 import { WebUpdateBanner } from "@/components/web/update-banner";
+import { LlamaUpdateBanner } from "@/components/llama-update-banner";
 import { DownloadManagerPanel } from "@/features/hub/download-manager";
 import { getTauriAuthFailure, tauriAutoAuth } from "@/features/auth";
 import { NativeIntentDrain } from "@/features/native-intents/native-intent-drain";
@@ -259,6 +260,7 @@ function TauriWrapper({ children }: { children: ReactNode }) {
         {children}
         <DownloadManagerPanel />
         <WebUpdateBanner enabled={!WEB_UPDATE_HIDDEN_ROUTES.has(pathname)} />
+        <LlamaUpdateBanner enabled={!WEB_UPDATE_HIDDEN_ROUTES.has(pathname)} />
       </>
     );
   }
@@ -294,7 +296,18 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     />
   );
 
-  if (!shouldUseCustomWindowTitlebar()) return content;
+  if (!shouldUseCustomWindowTitlebar()) {
+    // macOS desktop uses the native titlebar and returns here before the
+    // custom-titlebar branch, so mount the updater banner on this path too.
+    return (
+      <>
+        {content}
+        <LlamaUpdateBanner
+          enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
+        />
+      </>
+    );
+  }
 
   const showSidebarSurface =
     showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname);
@@ -305,6 +318,9 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       <div className="min-h-0 flex-1 overflow-hidden">
         {content}
       </div>
+      <LlamaUpdateBanner
+        enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
+      />
     </div>
   );
 }
diff --git a/studio/frontend/src/app/router.tsx b/studio/frontend/src/app/router.tsx
index dbb74ee1a1..5c18e637e2 100644
--- a/studio/frontend/src/app/router.tsx
+++ b/studio/frontend/src/app/router.tsx
@@ -3,6 +3,7 @@
 
 import { Link, createRouter, useRouterState } from "@tanstack/react-router";
 import { Button } from "@/components/ui/button";
+import { MascotImg } from "@/components/mascot-img";
 import { useT } from "@/i18n";
 import { Route as rootRoute } from "./routes/__root";
 import { Route as dataRecipesRoute } from "./routes/data-recipes";
@@ -41,11 +42,7 @@ function DefaultNotFound() {
 
   return (
     <div className="flex flex-1 flex-col items-center justify-center gap-4 p-8 text-center">
-      <img
-        src="/Sloth%20emojis/sloth%20shy%20large.png"
-        alt="Sloth mascot"
-        className="size-24"
-      />
+      <MascotImg src="Sloth emojis/sloth shy large.png" className="size-24" />
       <div className="flex flex-col items-center gap-1">
         <h1 className="font-heading font-semibold text-2xl tracking-tight">
           {t("shell.notFound.title")}
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index c8e47902b9..c05b0e6451 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -140,9 +140,14 @@ function RootLayout() {
           <SidebarInset className={isChatRoute ? "overflow-hidden" : "overflow-y-auto"}>
             <Navbar />
             <div
-              className={`flex min-h-0 min-w-0 flex-1 basis-0 flex-col ${isChatRoute ? "overflow-hidden" : "overflow-visible"} ${isChatRoute ? "" : "pt-14 md:pt-0"}`}
+              className={`relative flex min-h-0 min-w-0 flex-1 basis-0 flex-col ${isChatRoute ? "overflow-hidden" : "overflow-visible"} ${isChatRoute ? "" : "pt-14 md:pt-0"}`}
             >
-              <AnimatePresence initial={false} mode="wait">
+              {/* Use mode="popLayout" instead of "wait" to prevent UI freezes when
+                  switching from heavy pages (like Export with many checkpoints).
+                  "popLayout" allows the new route to mount immediately while the
+                  old one animates out, avoiding blocking on expensive exit renders.
+                  See issue #5850. */}
+              <AnimatePresence initial={false} mode="popLayout">
                 <motion.div
                   key={pathname}
                   initial={{ opacity: 0 }}
diff --git a/studio/frontend/src/asset-queries.d.ts b/studio/frontend/src/asset-queries.d.ts
new file mode 100644
index 0000000000..2c7a1cd936
--- /dev/null
+++ b/studio/frontend/src/asset-queries.d.ts
@@ -0,0 +1,9 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Vite `?inline` imports (data URIs); vite/client only types bare extensions.
+declare module "*?inline" {
+  const src: string;
+  // biome-ignore lint/style/noDefaultExport: Vite asset modules export default.
+  export default src;
+}
diff --git a/studio/frontend/src/assets/mascot-fallback.webp b/studio/frontend/src/assets/mascot-fallback.webp
new file mode 100644
index 0000000000..ad50361dbb
Binary files /dev/null and b/studio/frontend/src/assets/mascot-fallback.webp differ
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 9cb28a198b..f1a200b09b 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -51,7 +51,6 @@ import {
   Delete02Icon,
   Download01Icon,
   DownloadSquare01Icon,
-  Upload01Icon,
   Edit03Icon,
   FolderAddIcon,
   FolderExportIcon,
@@ -61,6 +60,7 @@ import {
   Logout05Icon,
   MoreVerticalIcon,
   Search01Icon,
+  PlusSignIcon,
   PowerIcon,
   PencilEdit02Icon,
   LayoutAlignLeftIcon,
@@ -72,11 +72,6 @@ import {
   exportConversationRawJsonl,
   exportConversationCsv,
   exportConversationShareGPT,
-  exportBulkConversationsMerged,
-  exportBulkConversationsSeparate,
-  importConversationsFromFile,
-  EXPORT_FORMATS_LIST,
-  type ConvExportFormat,
 } from "@/features/chat/prompt-storage/prompt-storage-dialog";
 import { listStoredChatThreads } from "@/features/chat/utils/chat-history-storage";
 import {
@@ -223,7 +218,7 @@ function NavItem({
           onClick={onClick}
           isActive={active}
           data-tour={dataTour}
-          className="sidebar-nav-btn h-[33px] rounded-[11px] gap-[8.5px] px-2.5 font-medium group-data-[collapsible=icon]:!w-[32px] group-data-[collapsible=icon]:!rounded-[10px] group-data-[collapsible=icon]:mx-auto"
+          className="sidebar-nav-btn h-[33px] rounded-full gap-[8.5px] pl-3 pr-2.5 font-medium group-data-[collapsible=icon]:px-2.5 group-data-[collapsible=icon]:!w-[32px] group-data-[collapsible=icon]:!rounded-full group-data-[collapsible=icon]:mx-auto"
         >
           <HugeiconsIcon icon={icon} strokeWidth={1.75} className="size-icon! shrink-0 group-hover/menu-button:animate-icon-pop" />
           <span className="text-[14.5px] leading-[19px] tracking-nav">{label}</span>
@@ -259,40 +254,6 @@ export function AppSidebar() {
   const isStudioRoute = pathname === "/studio" || pathname.startsWith("/studio/");
   const [chatOpen, setChatOpen] = useState(true);
 
-  const recentsImportInputRef = useRef<HTMLInputElement>(null);
-
-  async function handleImportToRecents(file: File) {
-    try {
-      const count = await importConversationsFromFile(file, null);
-      if (count === 0) {
-        toast.info("No conversations found in file.");
-      } else {
-        toast.success(`Imported ${count} conversation${count === 1 ? "" : "s"} to Recents.`);
-      }
-    } catch {
-      toast.error("Import failed.");
-    }
-  }
-
-  async function handleBulkExport(scope: "recents" | "all", fmt: ConvExportFormat, merged: boolean) {
-    try {
-      const threads = await listStoredChatThreads({
-        includeArchived: false,
-        ...(scope === "recents" ? { projectId: null } : {}),
-      });
-      const ids = [...new Set(threads.map((t) => t.id))];
-      if (ids.length === 0) { toast.info("No conversations to export."); return; }
-      const ts = new Date().toISOString().slice(0, 10);
-      const basename = scope === "all" ? `all-chats-${ts}` : `recents-${ts}`;
-      if (merged) {
-        await exportBulkConversationsMerged(ids, fmt, basename);
-      } else {
-        await exportBulkConversationsSeparate(ids, fmt, basename);
-      }
-    } catch {
-      toast.error("Export failed.");
-    }
-  }
   const [trainOpen, setTrainOpen] = useState(true);
   const [runsOpen, setRunsOpen] = useState(true);
 
@@ -597,8 +558,9 @@ export function AppSidebar() {
         ? "sidebar-row-action group-hover/project-chat-item:opacity-100 group-hover/project-chat-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
         : "sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto";
     const buttonClass = cn(
-      "sidebar-nav-btn h-[33px] cursor-pointer rounded-[11px] pr-4 text-[14.5px] leading-[19px] tracking-nav font-medium",
-      variant === "project" ? "pl-[37px]" : "pl-2.5",
+      "sidebar-nav-btn h-[33px] cursor-pointer rounded-full pr-4 text-[14.5px] leading-[19px] tracking-nav font-medium",
+      // pl-3.5 starts the title at the same x as the Recents label text.
+      variant === "project" ? "pl-[39px]" : "pl-3.5",
       variant === "project"
         ? "group-hover/project-chat-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/project-chat-item:pr-8"
         : "group-hover/recent-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/recent-item:pr-8",
@@ -647,7 +609,7 @@ export function AppSidebar() {
           <DropdownMenuContent
             side="bottom"
             align="start"
-            sideOffset={6}
+            sideOffset={0}
             className="unsloth-plus-menu menu-flat-destructive w-52"
           >
             <DropdownMenuItem onSelect={() => openRenameChat(item)}>
@@ -660,7 +622,7 @@ export function AppSidebar() {
                 <span>Move to project</span>
               </DropdownMenuSubTrigger>
               <DropdownMenuSubContent
-                sideOffset={8}
+                sideOffset={0}
                 alignOffset={-4}
                 className="unsloth-plus-menu w-52"
               >
@@ -719,6 +681,15 @@ export function AppSidebar() {
                     {label}
                   </DropdownMenuItem>
                 ))}
+                <DropdownMenuSeparator />
+                {/* Bulk export and import live in Settings -> Chat -> Data. */}
+                <DropdownMenuItem
+                  onSelect={() =>
+                    useSettingsDialogStore.getState().openDialog("chat")
+                  }
+                >
+                  Export all chats…
+                </DropdownMenuItem>
               </DropdownMenuSubContent>
             </DropdownMenuSub>
             <DropdownMenuItem
@@ -736,18 +707,6 @@ export function AppSidebar() {
 
   return (
     <>
-    {/* Hidden file inputs for chat import */}
-    <input
-      ref={recentsImportInputRef}
-      type="file"
-      accept=".jsonl,.ndjson,.csv"
-      className="hidden"
-      onChange={(e) => {
-        const file = e.target.files?.[0];
-        if (file) void handleImportToRecents(file);
-        e.target.value = "";
-      }}
-    />
     <Sidebar
       collapsible="icon"
       variant="sidebar"
@@ -827,7 +786,7 @@ export function AppSidebar() {
         )}
       </SidebarHeader>
 
-      <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-2 pt-[9px] pb-px shrink-0">
+      <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-1.5 pt-[9px] pb-px shrink-0">
         <SidebarGroupContent>
           <SidebarMenu>
             <NavItem
@@ -867,7 +826,7 @@ export function AppSidebar() {
           scrolled && "is-scrolled",
         )}
       >
-        <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-2 py-0 shrink-0">
+        <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-1.5 py-0 shrink-0">
           <SidebarGroupContent>
             <SidebarMenu>
               <NavItem
@@ -880,7 +839,28 @@ export function AppSidebar() {
                   navigate({ to: "/projects" });
                   closeMobileIfOpen();
                 }}
-              />
+                className="group/projects-item relative"
+              >
+                <button
+                  type="button"
+                  aria-label="New project"
+                  onClick={(e) => {
+                    e.stopPropagation();
+                    setProjectCreateMoveTarget(null);
+                    setProjectNameDraft("");
+                    setCreatingProject(true);
+                  }}
+                  className="sidebar-row-action group-hover/projects-item:opacity-100 group-hover/projects-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto group-data-[collapsible=icon]:hidden"
+                >
+                  <span className="sidebar-row-action-glyph">
+                    <HugeiconsIcon
+                      icon={PlusSignIcon}
+                      strokeWidth={1.75}
+                      className="size-4"
+                    />
+                  </span>
+                </button>
+              </NavItem>
               <NavItem
                 icon={DashboardCircleIcon}
                 label={t("shell.navigation.hub")}
@@ -918,7 +898,7 @@ export function AppSidebar() {
               </CollapsibleTrigger>
             </SidebarGroupLabel>
             <CollapsibleContent>
-              <SidebarGroupContent className="px-2">
+              <SidebarGroupContent className="px-1.5">
                 <SidebarMenu>
                   <NavItem
                     icon={TestTubeOutlineIcon}
@@ -966,66 +946,10 @@ export function AppSidebar() {
                     {t("shell.navigation.recents")}
                     <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg] [[data-state=closed]_&]:opacity-100" />
                   </CollapsibleTrigger>
-                  <DropdownMenu>
-                    <DropdownMenuTrigger asChild>
-                      <button
-                        type="button"
-                        className="ml-auto flex items-center justify-center rounded-sm p-0.5 text-muted-foreground hover:bg-accent hover:text-foreground focus:outline-none focus-visible:ring-0"
-                        title="Export recents"
-                      >
-                        <MoreHorizontalIcon className="size-3.5" />
-                      </button>
-                    </DropdownMenuTrigger>
-                    <DropdownMenuContent side="bottom" align="start" className="w-56">
-                      <DropdownMenuItem onSelect={() => recentsImportInputRef.current?.click()}>
-                        <HugeiconsIcon icon={Upload01Icon} strokeWidth={1.75} className="size-icon mr-1" />
-                        Import chats
-                      </DropdownMenuItem>
-                      <DropdownMenuSeparator />
-                      <DropdownMenuSub>
-                        <DropdownMenuSubTrigger>
-                          <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} className="size-icon mr-1" />
-                          Export Recents
-                        </DropdownMenuSubTrigger>
-                        <DropdownMenuSubContent avoidCollisions={false} className="unsloth-plus-menu w-52">
-                          {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                            <DropdownMenuItem key={`r-m-${fmt}`} onSelect={() => void handleBulkExport("recents", fmt, true)}>
-                              {label} — combined
-                            </DropdownMenuItem>
-                          ))}
-                          <DropdownMenuSeparator />
-                          {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                            <DropdownMenuItem key={`r-s-${fmt}`} onSelect={() => void handleBulkExport("recents", fmt, false)}>
-                              {label} — per chat
-                            </DropdownMenuItem>
-                          ))}
-                        </DropdownMenuSubContent>
-                      </DropdownMenuSub>
-                      <DropdownMenuSub>
-                        <DropdownMenuSubTrigger>
-                          <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} className="size-icon mr-1" />
-                          Export Recents + Projects
-                        </DropdownMenuSubTrigger>
-                        <DropdownMenuSubContent avoidCollisions={false} className="unsloth-plus-menu w-52">
-                          {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                            <DropdownMenuItem key={`a-m-${fmt}`} onSelect={() => void handleBulkExport("all", fmt, true)}>
-                              {label} — combined
-                            </DropdownMenuItem>
-                          ))}
-                          <DropdownMenuSeparator />
-                          {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                            <DropdownMenuItem key={`a-s-${fmt}`} onSelect={() => void handleBulkExport("all", fmt, false)}>
-                              {label} — per chat
-                            </DropdownMenuItem>
-                          ))}
-                        </DropdownMenuSubContent>
-                      </DropdownMenuSub>
-                    </DropdownMenuContent>
-                  </DropdownMenu>
                 </div>
               </SidebarGroupLabel>
               <CollapsibleContent>
-                <SidebarGroupContent className="px-2">
+                <SidebarGroupContent className="px-1.5">
                   <SidebarMenu>
                     {recentChatItems.map((item) =>
                       renderChatSidebarItem(item, "recent"),
@@ -1065,7 +989,7 @@ export function AppSidebar() {
                       >
                         <SidebarMenuButton
                           isActive={isActiveRun}
-                          className="sidebar-nav-btn h-auto flex-col items-start gap-0.5 py-[5px] rounded-[11px] pl-2.5 pr-7 text-[14.5px] tracking-nav font-medium"
+                          className="sidebar-nav-btn h-auto flex-col items-start gap-0.5 py-[5px] rounded-full pl-3 pr-7 text-[14.5px] tracking-nav font-medium"
                           onClick={() => {
                             setSelectedHistoryRunId(run.id);
                             closeMobileIfOpen();
@@ -1106,8 +1030,8 @@ export function AppSidebar() {
                           <DropdownMenuContent
                             side="bottom"
                             align="end"
-                            sideOffset={4}
-                            className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
+                            sideOffset={0}
+                            className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-full border-0"
                           >
                             <DropdownMenuItem onSelect={() => openRenameRun(run)}>
                               <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
@@ -1154,9 +1078,9 @@ export function AppSidebar() {
                 <SidebarMenuButton
                   size="lg"
                   aria-label={t("shell.accountMenu", { name: displayTitle })}
-                  className="sidebar-nav-btn !h-[44px] gap-[9px] px-2 py-[3px] rounded-[14px]"
+                  className="sidebar-nav-btn !h-[44px] -my-[3px] gap-[9px] px-2 py-[3px] rounded-[14px]"
                 >
-                  <div className="shrink-0">
+                  <div className="flex shrink-0 items-center">
                     <UserAvatar
                       name={displayTitle}
                       imageUrl={avatarDataUrl}
@@ -1164,7 +1088,7 @@ export function AppSidebar() {
                       className="!size-[32px]"
                     />
                   </div>
-                  <div className="flex flex-col gap-0.5 leading-tight group-data-[collapsible=icon]:hidden">
+                  <div className="flex flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
                     <span className="truncate font-heading text-[13.5px] tracking-[0.025em] dark:tracking-[0.04em] font-semibold text-nav-fg">{displayTitle}</span>
                     <span className="truncate text-[11.5px] tracking-nav text-muted-foreground">Unsloth</span>
                   </div>
@@ -1174,7 +1098,7 @@ export function AppSidebar() {
               <DropdownMenuContent
                 side="top"
                 align="center"
-                sideOffset={6}
+                sideOffset={8}
                 className="app-user-menu menu-soft-surface-up ring-0 w-[16rem] px-1.5 py-2.5 font-heading rounded-[20px] border-0"
               >
                 <DropdownMenuGroup>
@@ -1268,7 +1192,7 @@ export function AppSidebar() {
         }
       }}
     >
-      <DialogContent className="menu-flat-destructive corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+      <DialogContent className="menu-flat-destructive corner-squircle dialog-soft-surface sm:max-w-md">
         <DialogHeader>
           <DialogTitle>
             {confirmingDelete?.kind === "run"
@@ -1348,7 +1272,7 @@ export function AppSidebar() {
         if (!open) setRenamingTarget(null);
       }}
     >
-      <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+      <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
         <DialogHeader>
           <DialogTitle>
             {renamingTarget?.kind === "run"
@@ -1413,7 +1337,7 @@ export function AppSidebar() {
         }
       }}
     >
-      <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+      <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
         <DialogHeader>
           <DialogTitle>
             {projectCreateMoveTarget ? "Move to new project" : "New project"}
diff --git a/studio/frontend/src/components/assistant-ui/markdown-text.tsx b/studio/frontend/src/components/assistant-ui/markdown-text.tsx
index fc547da0b1..ef9b33fced 100644
--- a/studio/frontend/src/components/assistant-ui/markdown-text.tsx
+++ b/studio/frontend/src/components/assistant-ui/markdown-text.tsx
@@ -13,7 +13,7 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import { createMathPlugin } from "@streamdown/math";
 import { mermaid } from "@streamdown/mermaid";
 import { useEffect, useMemo, useRef, useState } from "react";
-import { Block, type BlockProps, Streamdown } from "streamdown";
+import { Block, type BlockProps, Streamdown, defaultUrlTransform, type UrlTransform } from "streamdown";
 import { createCodePlugin } from "./code-plugin";
 import "katex/dist/katex.min.css";
 import { AudioPlayer } from "./audio-player";
@@ -424,6 +424,22 @@ function useRafCoalescedText(text: string, isStreaming: boolean): string {
   return text;
 }
 
+const safeImageUrl: UrlTransform = (url, _key, node) => {
+  // Only images are restricted; links/other nodes use the default transform.
+  if (node.tagName !== "img") return defaultUrlTransform(url, _key, node);
+
+  // Strip ASCII controls first: browsers drop them mid-parse, so a value like
+  // "\t//attacker.com" would otherwise slip past the guards below.
+  // eslint-disable-next-line no-control-regex
+  const normalized = url.replace(/[\x00-\x1f\x7f]/g, "").trim();
+  const lower = normalized.toLowerCase();
+
+  if (lower.startsWith("data:") || lower.startsWith("blob:")) return normalized;
+  if (/^[/\\]{2}/.test(normalized)) return null; // protocol-relative: // \\ /\ \/
+  if (/^[a-zA-Z][a-zA-Z0-9+\-.]*:/.test(normalized)) return null; // scheme prefix (colon later in path is fine)
+  return normalized; // relative -> same-origin
+};
+
 const MarkdownTextImpl = () => {
   const { text, status } = useMessagePartText();
   const displayText = useRafCoalescedText(text, status.type === "running");
@@ -444,6 +460,7 @@ const MarkdownTextImpl = () => {
         isAnimating={status.type === "running"}
         plugins={{ code, math, mermaid }}
         components={STREAMDOWN_COMPONENTS}
+        urlTransform={safeImageUrl}
         controls={{
           code: false,
           mermaid: {
diff --git a/studio/frontend/src/components/assistant-ui/model-selector.tsx b/studio/frontend/src/components/assistant-ui/model-selector.tsx
index f8a3ce4e20..b8d1bc0f00 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector.tsx
@@ -148,12 +148,12 @@ function ModelSelectorTrigger({
         className={cn(
           "unsloth-model-selector-trigger flex min-w-0 items-center gap-2 transition-colors",
           variant === "outline" &&
-          "rounded-[10px] border border-border/60 hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
-          variant === "ghost" && "rounded-[10px] hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
-          variant === "muted" && "rounded-[10px] bg-muted hover:bg-muted/80",
-          size === "sm" && "h-8 px-3 text-xs",
-          size === "default" && "h-9 px-3.5 text-sm",
-          size === "lg" && "h-10 px-4 text-sm",
+          "rounded-full border border-border/60 hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
+          variant === "ghost" && "rounded-full hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
+          variant === "muted" && "rounded-full bg-muted hover:bg-muted/80",
+          size === "sm" && "h-8 px-2.5 text-xs",
+          size === "default" && "h-9 px-3 text-sm",
+          size === "lg" && "h-10 px-3.5 text-sm",
           className,
         )}
       >
@@ -244,6 +244,7 @@ function ModelSelectorContent({
   return (
     <PopoverContent
       align="start"
+      alignOffset={10}
       data-tour={dataTour}
       className={cn(
         "unsloth-model-selector-menu menu-soft-surface ring-0 w-[min(440px,calc(100vw-1rem))] max-w-[calc(100vw-1rem)] min-w-0 gap-0 p-3",
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx b/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
index 4a9b22e103..d6afa30e05 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
@@ -12,7 +12,8 @@ import {
   AlertDialogTitle,
 } from "@/components/ui/alert-dialog";
 import { cn } from "@/lib/utils";
-import { Trash2Icon } from "lucide-react";
+import { Delete02Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
 import { useCallback, useState, type ReactNode } from "react";
 import { toast } from "@/lib/toast";
 
@@ -77,7 +78,11 @@ export function ModelDeleteAction({
           buttonClassName,
         )}
       >
-        <Trash2Icon className={cn("size-3.5", iconClassName)} />
+        <HugeiconsIcon
+          icon={Delete02Icon}
+          strokeWidth={1.75}
+          className={cn("size-3.5", iconClassName)}
+        />
       </button>
 
       <AlertDialog
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 4fd86ad789..be66f23a20 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -146,7 +146,7 @@ function ModelRow({
       type="button"
       onClick={onClick}
       className={cn(
-        "flex w-full items-center gap-2 rounded-[8px] px-2.5 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d44]",
+        "flex w-full items-center gap-2 rounded-full px-3 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d44]",
         selected && "bg-[#ececec] dark:bg-[#3a3d44]",
       )}
     >
@@ -386,7 +386,7 @@ function GgufVariantExpander({
                 handleVariantClick(v.quant, v.downloaded, v.size_bytes)
               }
               className={cn(
-                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-[8px] px-2.5 py-1 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d44]",
+                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full px-3 py-1 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d44]",
               )}
             >
               <span className="min-w-0 flex-1 truncate font-mono text-xs">
@@ -887,8 +887,8 @@ export function HubModelPicker({
         <Input
           value={query}
           onChange={(event) => setQuery(event.target.value)}
-          placeholder="Search Hugging Face models"
-          className="h-9 pl-8 pr-8"
+          placeholder="Search models"
+          className="h-9 border-[#f2f2f2] dark:border-input pl-8 pr-8"
         />
         {isLoading && (
           <Spinner className="pointer-events-none absolute right-2.5 top-2.5 size-4 text-muted-foreground" />
@@ -896,7 +896,7 @@ export function HubModelPicker({
       </div>
 
       <div ref={scrollRef} className="max-h-64 overflow-y-auto">
-        <div className="p-1">
+        <div className="py-1">
           {!cachedReady && !showHfSection ? (
             <div className="flex items-center gap-2 px-5 py-3">
               <Spinner className="size-3 text-muted-foreground" />
@@ -1478,12 +1478,12 @@ export function LoraModelPicker({
           value={query}
           onChange={(event) => setQuery(event.target.value)}
           placeholder="Search trained models"
-          className="h-9 pl-8"
+          className="h-9 border-[#f2f2f2] dark:border-input pl-8"
         />
       </div>
 
       <div className="max-h-64 overflow-y-auto">
-        <div className="p-1">
+        <div className="py-1">
           {grouped.length === 0 ? (
             <div className="px-2.5 py-2 text-xs text-muted-foreground">
               No trained models found.
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index b6fc554423..84e1654079 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -36,6 +36,7 @@ import {
   useScrollThreadToBottom,
 } from "@/components/assistant-ui/use-intent-aware-autoscroll";
 import { Button } from "@/components/ui/button";
+import { MascotImg } from "@/components/mascot-img";
 import { Spinner } from "@/components/ui/spinner";
 import {
   DropdownMenu,
@@ -64,7 +65,7 @@ import { NewProjectDialog } from "@/features/chat/components/new-project-dialog"
 import { parseExternalModelId } from "@/features/chat/external-providers";
 import { McpComposerButton } from "@/features/chat/mcp-composer-button";
 import { getExternalReasoningCapabilities } from "@/features/chat/provider-capabilities";
-import { useRagToolAvailable } from "@/features/chat/hooks/use-rag-tool-available";
+import { useRagToolDisabled } from "@/features/chat/hooks/use-rag-tool-disabled";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import { useExternalProvidersStore } from "@/features/chat/stores/external-providers-store";
 import { deleteThreadMessage } from "@/features/chat/utils/delete-thread-message";
@@ -74,7 +75,6 @@ import { DocumentPreviewMount } from "@/features/rag/components/document-preview
 import { useUserProfileStore } from "@/features/profile/stores/user-profile-store";
 import { applyQwenThinkingParams } from "@/features/chat/utils/qwen-params";
 import { isTauri } from "@/lib/api-base";
-import { AUDIO_ACCEPT, MAX_AUDIO_SIZE, fileToBase64 } from "@/lib/audio-utils";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
@@ -740,8 +740,8 @@ function buildWelcome(hour: number, name: string): Welcome {
   // Use the name on ~a third of lines (only direct salutations where it reads
   // naturally); the rest stay name-free so greetings don't feel repetitive.
   const base: Welcome[] = [
-    g(name ? `Good to see you, ${name}.` : "Good to see you.", "large sloth wave.png"),
-    g("Ready when you are.", "large sloth thumbs.png"),
+    g(name ? `Good to see you, ${name}` : "Good to see you", "large sloth wave.png"),
+    g("Ready when you are", "large sloth thumbs.png"),
     DEFAULT_WELCOME,
     g("How can I help?", "sloth sir large.png"),
   ];
@@ -781,17 +781,16 @@ const ThreadWelcome: FC<{
     setWelcome(buildWelcome(new Date().getHours(), name));
   }, [displayName]);
 
-  const currentEmojiSrc = `/Sloth emojis/${welcome.sloth}`;
+  const currentEmojiSrc = `Sloth emojis/${welcome.sloth}`;
 
   return (
     <div className="aui-thread-welcome-root mx-auto my-auto flex w-full max-w-(--thread-max-width) grow flex-col">
-      <div className="aui-thread-welcome-center flex w-full grow flex-col items-center justify-start pt-[28vh]">
+      <div className="aui-thread-welcome-center flex w-full grow flex-col items-center justify-start pt-[27.5vh]">
         <div className="aui-thread-welcome-message flex w-full flex-col justify-center gap-9 px-4">
           {/* Center the greeting (sloth + title) over the composer. */}
           <div className="flex flex-row items-center justify-center gap-[15px]">
-            <img
+            <MascotImg
               src={currentEmojiSrc}
-              alt="Sloth mascot"
               className="size-[44px] -translate-y-[2px]"
             />
             <h1 className="aui-thread-welcome-message-inner unsloth-welcome-title fade-in slide-in-from-bottom-1 animate-in text-3xl tracking-[-0.02em] duration-200">
@@ -876,12 +875,11 @@ const Composer: FC<{
   const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
   const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
-  const ragToolAvailable = useRagToolAvailable();
   // More than 4 pills: collapse to icons only. Search and Code always show;
-  // Images, Canvas and MCP are conditional.
+  // Images, RAG, Canvas and MCP are conditional.
   const pillsCompact =
     2 +
-      (ragEnabled && ragToolAvailable ? 1 : 0) +
+      (ragEnabled ? 1 : 0) +
       (supportsBuiltinImageGeneration ? 1 : 0) +
       (artifactsEnabled ? 1 : 0) +
       (mcpEnabledForChat ? 1 : 0) >
@@ -1389,63 +1387,6 @@ function useImeComposerInputHandlers() {
   };
 }
 
-// Audio upload row, only for audio-input models.
-const ComposerAudioMenuItem: FC = () => {
-  const setPendingAudio = useChatRuntimeStore((s) => s.setPendingAudio);
-  const activeModel = useChatRuntimeStore((s) => {
-    const checkpoint = s.params.checkpoint;
-    return s.models.find((m) => m.id === checkpoint);
-  });
-
-  const handleAudioFile = useCallback(
-    async (file: File) => {
-      if (file.size > MAX_AUDIO_SIZE) {
-        return;
-      }
-      try {
-        const base64 = await fileToBase64(file);
-        setPendingAudio(base64, file.name);
-      } catch {
-        // skip
-      }
-    },
-    [setPendingAudio],
-  );
-
-  // Build the input on document.body, not in the menu: selecting the item
-  // closes the dropdown, unmounting a menu-rendered input before the OS picker
-  // returns and dropping the file.
-  const pickAudio = useCallback(() => {
-    const input = document.createElement("input");
-    input.type = "file";
-    input.accept = AUDIO_ACCEPT;
-    input.hidden = true;
-    document.body.appendChild(input);
-    input.onchange = () => {
-      const file = input.files?.[0];
-      if (file) handleAudioFile(file);
-      document.body.removeChild(input);
-    };
-    input.oncancel = () => {
-      if (!input.files || input.files.length === 0) {
-        document.body.removeChild(input);
-      }
-    };
-    input.click();
-  }, [handleAudioFile]);
-
-  if (!activeModel?.hasAudioInput) {
-    return null;
-  }
-
-  return (
-    <DropdownMenuItem onSelect={() => pickAudio()}>
-      <HeadphonesIcon />
-      Upload audio
-    </DropdownMenuItem>
-  );
-};
-
 // Phosphor microphone. Inlined to avoid a new icon dependency.
 const MicIcon: FC<{ className?: string }> = ({ className }) => (
   <svg
@@ -1826,6 +1767,7 @@ const WebSearchToggle: FC = () => {
         }
       }}
       className="composer-pill-btn"
+      data-pill-label="Search"
       data-active={toolsEnabled && !disabled ? "true" : "false"}
       aria-label={toolsEnabled ? "Disable web search" : "Enable web search"}
     >
@@ -1861,6 +1803,7 @@ const CodeToolsToggle: FC = () => {
       disabled={disabled}
       onClick={() => setCodeToolsEnabled(!codeToolsEnabled)}
       className="composer-pill-btn"
+      data-pill-label="Code"
       data-active={codeToolsEnabled && !disabled ? "true" : "false"}
       aria-label={
         codeToolsEnabled ? "Disable code execution" : "Enable code execution"
@@ -1902,6 +1845,7 @@ const ImagesToggle: FC = () => {
       disabled={disabled}
       onClick={() => setImageToolsEnabled(!imageToolsEnabled)}
       className="composer-pill-btn"
+      data-pill-label="Images"
       data-active={imageToolsEnabled && !disabled ? "true" : "false"}
       aria-label={
         imageToolsEnabled
@@ -1928,6 +1872,7 @@ const ArtifactsToggle: FC = () => {
       type="button"
       onClick={() => setArtifactsEnabled(false)}
       className="composer-pill-btn"
+      data-pill-label="Canvas"
       data-active="true"
       aria-label="Disable canvas"
     >
@@ -2001,6 +1946,21 @@ const ToolStatusDisplay: FC = () => {
 };
 // Plus menu: attachment and workflow actions. Opens downward in the welcome
 // composer; the docked composer passes side="top" to open upward.
+const AUDIO_ACCEPT_TOKEN_RE =
+  /^(audio\/|\.(?:wav|mp3|m4a|ogg|oga|flac)$)/i;
+
+function attachmentAcceptForPicker(accept: string, audioEnabled: boolean): string {
+  if (audioEnabled || accept === "*") {
+    return accept;
+  }
+  const filtered = accept
+    .split(",")
+    .map((token) => token.trim())
+    .filter((token) => token && !AUDIO_ACCEPT_TOKEN_RE.test(token))
+    .join(",");
+  return filtered || accept;
+}
+
 const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   side = "bottom",
 }) => {
@@ -2017,13 +1977,21 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   );
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
   const setRagEnabled = useChatRuntimeStore((s) => s.setRagEnabled);
-  // Shared gate so the menu row agrees with the RAG pill and Add Files bar.
-  const ragAvailable = useRagToolAvailable();
+  // Shared gate so the menu row agrees with the RAG pill.
+  const ragDisabled = useRagToolDisabled();
   // Capability gating mirrors the visible pills so menu and pills agree on
   // what a loaded model supports (a tool the backend drops must not look on).
   const modelLoaded = useChatRuntimeStore(
     (s) => !!s.params.checkpoint && !s.modelLoading,
   );
+  const audioAttachmentsEnabled = useChatRuntimeStore((s) => {
+    const activeCheckpoint = s.params.checkpoint;
+    if (!activeCheckpoint || s.modelLoading) {
+      return false;
+    }
+    const activeModel = s.models.find((m) => m.id === activeCheckpoint);
+    return Boolean(activeModel?.hasAudioInput);
+  });
   const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
   const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
   const supportsBuiltinWebSearch = useChatRuntimeStore(
@@ -2086,6 +2054,40 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   const [promptStorageOpen, setPromptStorageOpen] = useState(false);
   const activeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const aui = useAui();
+  const composerCanAddAttachments = useAuiState(
+    ({ composer }) => composer.isEditing,
+  );
+  const pickAttachment = useCallback(() => {
+    const input = document.createElement("input");
+    input.type = "file";
+    input.multiple = true;
+    input.hidden = true;
+
+    const attachmentAccept = attachmentAcceptForPicker(
+      aui.composer().getState().attachmentAccept,
+      audioAttachmentsEnabled,
+    );
+    if (attachmentAccept !== "*") {
+      input.accept = attachmentAccept;
+    }
+
+    document.body.appendChild(input);
+    input.onchange = (event) => {
+      const files = (event.target as HTMLInputElement).files;
+      if (files) {
+        for (const file of files) {
+          void aui.composer().addAttachment(file);
+        }
+      }
+      document.body.removeChild(input);
+    };
+    input.oncancel = () => {
+      if (!input.files || input.files.length === 0) {
+        document.body.removeChild(input);
+      }
+    };
+    input.click();
+  }, [aui, audioAttachmentsEnabled]);
   // Disable Export chat until the thread has content.
   const messageCount = useAuiState(({ thread }) => thread.messages.length);
   const { startQueue } = useContext(PromptQueueContext);
@@ -2131,19 +2133,19 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
       <DropdownMenuContent
         side={side}
         align="start"
-        sideOffset={2}
+        sideOffset={0}
         avoidCollisions={true}
         className="unsloth-plus-menu w-[212px]"
         // Don't refocus the + on close; restored focus showed a stray ring.
         onCloseAutoFocus={(event) => event.preventDefault()}
       >
-        <ComposerPrimitive.AddAttachment asChild={true}>
-          <DropdownMenuItem>
-            <HugeiconsIcon icon={AttachmentIcon} strokeWidth={2} />
-            Add photos &amp; files
-          </DropdownMenuItem>
-        </ComposerPrimitive.AddAttachment>
-        <ComposerAudioMenuItem />
+        <DropdownMenuItem
+          disabled={!composerCanAddAttachments}
+          onSelect={() => pickAttachment()}
+        >
+          <HugeiconsIcon icon={AttachmentIcon} strokeWidth={2} />
+          Add photos &amp; files
+        </DropdownMenuItem>
         <DropdownMenuItem
           disabled={searchDisabled}
           className={
@@ -2180,10 +2182,11 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           }
           onSelect={() => setCodeToolsEnabled(!codeToolsEnabled)}
         >
+          {/* Scale, not width: an oversized box pushed the label out of line. */}
           <HugeiconsIcon
             icon={CodeIcon}
             strokeWidth={2}
-            className="size-[1.175rem]!"
+            className="scale-[1.12]"
           />
           Code
           {codeToolsEnabled && !codeDisabled ? (
@@ -2217,15 +2220,15 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
         )}
         <DropdownMenuSeparator />
         <DropdownMenuItem
-          disabled={!ragAvailable}
+          disabled={ragDisabled}
           className={
-            ragEnabled && ragAvailable ? "text-primary font-medium" : undefined
+            ragEnabled && !ragDisabled ? "text-primary font-medium" : undefined
           }
           onSelect={() => setRagEnabled(!ragEnabled)}
         >
           <HugeiconsIcon icon={FileDatabaseIcon} strokeWidth={2} />
-          RAG
-          {ragEnabled && ragAvailable ? (
+          Chat with Files
+          {ragEnabled && !ragDisabled ? (
             <HugeiconsIcon
               icon={Tick02Icon}
               strokeWidth={2}
@@ -2252,80 +2255,82 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
             />
           ) : null}
         </DropdownMenuItem>
-        {/* Top-level so it stays one click away (not buried in More). */}
-        <DropdownMenuItem onSelect={() => startCompare()}>
-          <Columns2Icon />
-          Compare chat
-        </DropdownMenuItem>
-        <DropdownMenuSub>
-          <DropdownMenuSubTrigger>
-            <HugeiconsIcon icon={Bookmark02Icon} strokeWidth={2} />
-            Saved prompts
-          </DropdownMenuSubTrigger>
-          <DropdownMenuSubContent className="unsloth-plus-menu w-[176px]">
-            {recentPrompts.map((p) => (
-              <DropdownMenuItem
-                key={p.id}
-                onSelect={() => aui.composer().setText(p.text)}
-              >
-                <span className="truncate">{p.name}</span>
-              </DropdownMenuItem>
-            ))}
-            {recentPrompts.length > 0 ? <DropdownMenuSeparator /> : null}
-            <DropdownMenuItem onSelect={() => setPromptStorageOpen(true)}>
-              All saved prompts…
-            </DropdownMenuItem>
-          </DropdownMenuSubContent>
-        </DropdownMenuSub>
-        {/* Top-level: a third-level submenu collision-flips at narrow widths
-            and is awkward to reach. */}
-        <DropdownMenuSub>
-          <DropdownMenuSubTrigger disabled={!activeThreadId || messageCount === 0}>
-            <HugeiconsIcon icon={Download01Icon} strokeWidth={2} />
-            Export chat
-          </DropdownMenuSubTrigger>
-          <DropdownMenuSubContent
-            collisionPadding={16}
-            className="unsloth-plus-menu w-[176px]"
-          >
-            <DropdownMenuItem
-              onSelect={() => {
-                if (!activeThreadId) return;
-                exportConversationRawJsonl(activeThreadId).catch(() =>
-                  toast.error("Export failed."),
-                );
-              }}
-            >
-              Raw JSONL
-            </DropdownMenuItem>
-            <DropdownMenuItem
-              onSelect={() => {
-                if (!activeThreadId) return;
-                exportConversationCsv(activeThreadId).catch(() =>
-                  toast.error("Export failed."),
-                );
-              }}
-            >
-              CSV
-            </DropdownMenuItem>
-            <DropdownMenuItem
-              onSelect={() => {
-                if (!activeThreadId) return;
-                exportConversationShareGPT(activeThreadId).catch(() =>
-                  toast.error("Export failed."),
-                );
-              }}
-            >
-              ShareGPT JSONL
-            </DropdownMenuItem>
-          </DropdownMenuSubContent>
-        </DropdownMenuSub>
         <DropdownMenuSub>
           <DropdownMenuSubTrigger>
             <MoreHorizontalIcon className="size-4" />
             More
           </DropdownMenuSubTrigger>
-          <DropdownMenuSubContent className="w-[200px]">
+          <DropdownMenuSubContent className="unsloth-plus-menu w-[200px]">
+            <DropdownMenuSub>
+              <DropdownMenuSubTrigger>
+                <HugeiconsIcon icon={Bookmark02Icon} strokeWidth={2} />
+                Saved prompts
+              </DropdownMenuSubTrigger>
+              <DropdownMenuSubContent
+                collisionPadding={16}
+                className="unsloth-plus-menu w-[176px]"
+              >
+                {recentPrompts.map((p) => (
+                  <DropdownMenuItem
+                    key={p.id}
+                    onSelect={() => aui.composer().setText(p.text)}
+                  >
+                    <span className="truncate">{p.name}</span>
+                  </DropdownMenuItem>
+                ))}
+                {recentPrompts.length > 0 ? <DropdownMenuSeparator /> : null}
+                <DropdownMenuItem onSelect={() => setPromptStorageOpen(true)}>
+                  All saved prompts…
+                </DropdownMenuItem>
+              </DropdownMenuSubContent>
+            </DropdownMenuSub>
+            <DropdownMenuItem onSelect={() => startCompare()}>
+              <Columns2Icon />
+              Compare chat
+            </DropdownMenuItem>
+            <DropdownMenuSub>
+              <DropdownMenuSubTrigger
+                disabled={!activeThreadId || messageCount === 0}
+              >
+                <HugeiconsIcon icon={Download01Icon} strokeWidth={2} />
+                Export chat
+              </DropdownMenuSubTrigger>
+              <DropdownMenuSubContent
+                collisionPadding={16}
+                className="unsloth-plus-menu w-[176px]"
+              >
+                <DropdownMenuItem
+                  onSelect={() => {
+                    if (!activeThreadId) return;
+                    exportConversationRawJsonl(activeThreadId).catch(() =>
+                      toast.error("Export failed."),
+                    );
+                  }}
+                >
+                  Raw JSONL
+                </DropdownMenuItem>
+                <DropdownMenuItem
+                  onSelect={() => {
+                    if (!activeThreadId) return;
+                    exportConversationCsv(activeThreadId).catch(() =>
+                      toast.error("Export failed."),
+                    );
+                  }}
+                >
+                  CSV
+                </DropdownMenuItem>
+                <DropdownMenuItem
+                  onSelect={() => {
+                    if (!activeThreadId) return;
+                    exportConversationShareGPT(activeThreadId).catch(() =>
+                      toast.error("Export failed."),
+                    );
+                  }}
+                >
+                  ShareGPT JSONL
+                </DropdownMenuItem>
+              </DropdownMenuSubContent>
+            </DropdownMenuSub>
             <DropdownMenuItem
               className={
                 artifactsEnabled ? "text-primary font-medium" : undefined
@@ -2633,7 +2638,7 @@ const AssistantActionBar: FC = () => {
   return (
     <ActionBarPrimitive.Root
       hideWhenRunning={true}
-      className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-[10px] [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
+      className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
     >
       <CopyButton />
       <ActionBarPrimitive.Reload asChild={true}>
@@ -2655,7 +2660,7 @@ const AssistantActionBar: FC = () => {
           side="bottom"
           align="start"
           onCloseAutoFocus={(e) => e.preventDefault()}
-          className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-md [--radius:1.1rem] bg-popover p-1 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(27,27,31,0.16)] dark:shadow-none"
+          className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-md [--radius:1.1rem] bg-popover p-1 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
         >
           <ActionBarPrimitive.ExportMarkdown asChild={true}>
             <ActionBarMorePrimitive.Item className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-sm px-2 py-1.5 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground">
@@ -2713,7 +2718,7 @@ const UserActionBar: FC = () => {
   return (
     <ActionBarPrimitive.Root
       autohide="always"
-      className="aui-user-action-bar-root flex gap-1 text-chat-icon-fg [&_button]:size-8 [&_button]:!rounded-[10px] [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
+      className="aui-user-action-bar-root flex gap-1 text-chat-icon-fg [&_button]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
     >
       <CopyButton />
       <ActionBarPrimitive.Edit asChild={true}>
diff --git a/studio/frontend/src/components/llama-update-banner.tsx b/studio/frontend/src/components/llama-update-banner.tsx
new file mode 100644
index 0000000000..83320f8e15
--- /dev/null
+++ b/studio/frontend/src/components/llama-update-banner.tsx
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
+import { useShowLlamaUpdateBanner } from "@/hooks/use-llama-update-pref";
+import { toast } from "@/lib/toast";
+import { AnimatePresence, motion } from "motion/react";
+import type { ReactElement } from "react";
+
+const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
+
+interface LlamaUpdateBannerProps {
+  enabled?: boolean;
+}
+
+/**
+ * Non-invasive "Update llama.cpp" affordance. Appears bottom-right ~1s after a
+ * newer prebuilt is detected and stays up until the user explicitly acts on it
+ * (X, Update, or Remind me later). Clicking Update swaps the prebuilt in place
+ * via POST /api/llama/update. Can be turned off entirely in Settings ->
+ * General -> Notifications (on by default).
+ */
+export function LlamaUpdateBanner({
+  enabled = true,
+}: LlamaUpdateBannerProps): ReactElement | null {
+  const showBannerPref = useShowLlamaUpdateBanner();
+  const { status, visible, applying, apply, dismiss, snooze } =
+    useLlamaUpdateCheck({
+      enabled: enabled && showBannerPref,
+    });
+
+  async function handleUpdate() {
+    const result = await apply();
+    if (result?.ok) {
+      toast.success(
+        `llama.cpp updated to ${result.tag ?? "the latest build"}. Reload your model to use it.`,
+      );
+    } else if (result) {
+      toast.error(
+        `llama.cpp update failed: ${result.error ?? "unknown error"}`,
+      );
+    }
+  }
+
+  const show =
+    visible && status != null && (status.update_available || applying);
+  const updateProgress = status?.job.progress ?? null;
+
+  return (
+    <AnimatePresence>
+      {show ? (
+        <motion.div
+          initial={{ opacity: 0, y: 12, scale: 0.96 }}
+          animate={{ opacity: 1, y: 0, scale: 1 }}
+          exit={{ opacity: 0, y: 8, scale: 0.97 }}
+          transition={{ duration: 0.35, ease: EASE_OUT_QUART }}
+          className="fixed bottom-4 right-4 z-[9998] w-[calc(100vw-2rem)] max-w-[340px]"
+          data-testid="llama-update-banner"
+        >
+          <div className="relative overflow-hidden rounded-[24px] bg-white px-4 pb-[22px] pl-6 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
+            {applying ? null : (
+              <button
+                type="button"
+                onClick={dismiss}
+                className="absolute top-2.5 right-3 flex size-6 items-center justify-center rounded-full text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
+                aria-label="Dismiss llama.cpp update notification"
+              >
+                <svg
+                  aria-hidden="true"
+                  width="12"
+                  height="12"
+                  viewBox="0 0 14 14"
+                  fill="none"
+                  xmlns="http://www.w3.org/2000/svg"
+                >
+                  <path
+                    d="M11 3L3 11M3 3l8 8"
+                    stroke="currentColor"
+                    strokeWidth="1.5"
+                    strokeLinecap="round"
+                  />
+                </svg>
+              </button>
+            )}
+
+            <div className="min-w-0 pr-6">
+              <p className="font-heading text-base font-medium text-foreground">
+                {applying ? "Updating llama.cpp..." : "New llama.cpp version"}
+              </p>
+              <p className="mt-0.5 text-xs text-muted-foreground">
+                {status?.installed_tag ?? "unknown"} &rarr;{" "}
+                <span className="font-medium text-foreground">
+                  {status?.latest_tag ?? ""}
+                </span>
+              </p>
+            </div>
+
+            {applying ? (
+              <div
+                className="mb-1.5 mt-4 h-1 overflow-hidden rounded-full bg-muted"
+                role="progressbar"
+                aria-label="Updating llama.cpp"
+                aria-valuemin={0}
+                aria-valuemax={100}
+                aria-valuenow={
+                  updateProgress != null
+                    ? Math.round(updateProgress * 100)
+                    : undefined
+                }
+                data-testid="llama-update-progress"
+              >
+                {updateProgress != null && updateProgress > 0 ? (
+                  <div
+                    className="h-full rounded-full bg-primary transition-[width] duration-700 ease-out"
+                    style={{ width: `${Math.round(updateProgress * 100)}%` }}
+                  />
+                ) : (
+                  // No percent yet (resolving the release): sweep until the
+                  // first download progress arrives.
+                  <div className="loading-bar-slide h-full w-1/3 rounded-full bg-primary" />
+                )}
+              </div>
+            ) : (
+              <div className="mt-3 flex items-center gap-2">
+                <Button
+                  size="sm"
+                  className="h-auto rounded-full px-3.5 py-2 text-[13px]"
+                  onClick={handleUpdate}
+                  data-testid="llama-update-button"
+                >
+                  Update
+                </Button>
+                <Button
+                  size="sm"
+                  variant="ghost"
+                  className="h-auto rounded-full px-2.5 py-2 text-[13px] text-muted-foreground hover:text-foreground"
+                  onClick={snooze}
+                  data-testid="llama-update-snooze-button"
+                >
+                  Remind me later
+                </Button>
+              </div>
+            )}
+          </div>
+        </motion.div>
+      ) : null}
+    </AnimatePresence>
+  );
+}
diff --git a/studio/frontend/src/components/mascot-img.tsx b/studio/frontend/src/components/mascot-img.tsx
new file mode 100644
index 0000000000..cff4d54c4d
--- /dev/null
+++ b/studio/frontend/src/components/mascot-img.tsx
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { type ComponentProps, useState } from "react";
+
+// Bundled as a data URI so the fallback ships inside the JS bundle and can
+// never 404, unlike the public-folder originals.
+import fallbackMascot from "@/assets/mascot-fallback.webp?inline";
+
+const LEADING_SLASH = /^\//;
+
+// Resolve a public-folder asset against the deploy base (subpath mounts) and
+// encode the spaced sloth filenames.
+export function publicAssetUrl(path: string): string {
+  return encodeURI(import.meta.env.BASE_URL + path.replace(LEADING_SLASH, ""));
+}
+
+type MascotImgProps = { src: string } & Omit<
+  ComponentProps<"img">,
+  "src" | "onError"
+>;
+
+// Keying on src remounts the inner component, so the retry state resets
+// whenever the source changes (greeting sloths rotate).
+export function MascotImg(props: MascotImgProps) {
+  return <MascotImgInner key={props.src} {...props} />;
+}
+
+type Stage = "primary" | "retry" | "fallback";
+
+// Decorative mascot that degrades gracefully: a failed load retries once with
+// a cache-buster (transient blips, stale caches), then swaps to the bundled
+// fallback sloth. Empty alt by default so a broken image never paints text.
+function MascotImgInner({ src, alt = "", ...rest }: MascotImgProps) {
+  const [stage, setStage] = useState<Stage>("primary");
+
+  const url = publicAssetUrl(src);
+  const effectiveSrc =
+    stage === "primary"
+      ? url
+      : stage === "retry"
+        ? `${url}${url.includes("?") ? "&" : "?"}retry=1`
+        : fallbackMascot;
+
+  return (
+    <img
+      draggable={false}
+      {...rest}
+      src={effectiveSrc}
+      alt={alt}
+      onError={() => setStage((s) => (s === "primary" ? "retry" : "fallback"))}
+    />
+  );
+}
diff --git a/studio/frontend/src/components/tauri/startup-screen.tsx b/studio/frontend/src/components/tauri/startup-screen.tsx
index 1eb3f81d10..8e373c82a0 100644
--- a/studio/frontend/src/components/tauri/startup-screen.tsx
+++ b/studio/frontend/src/components/tauri/startup-screen.tsx
@@ -228,7 +228,7 @@ function RepairingContent({
       </div>
       <div className="mb-10 flex flex-col items-center gap-2">
         <TealSpinner />
-        <p className="text-sm font-bold text-foreground">Updating existing Studio install...</p>
+        <p className="text-sm font-bold text-foreground">Updating existing Unsloth install...</p>
         {latest && (
           <p className="max-w-xs text-center text-xs text-muted-foreground">{latest}</p>
         )}
diff --git a/studio/frontend/src/components/ui/button.tsx b/studio/frontend/src/components/ui/button.tsx
index 9e27446989..51fc31697e 100644
--- a/studio/frontend/src/components/ui/button.tsx
+++ b/studio/frontend/src/components/ui/button.tsx
@@ -10,7 +10,7 @@ import type * as React from "react";
 import { cn } from "@/lib/utils";
 
 export const buttonVariants = cva(
-  "focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 rounded-4xl border border-transparent text-sm font-medium focus-visible:ring-[3px] aria-invalid:ring-[3px] [&_svg:not([class*='size-'])]:size-4 inline-flex items-center justify-center whitespace-nowrap transition-all disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none shrink-0 [&_svg]:shrink-0 outline-none group/button select-none cursor-pointer",
+  "focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 rounded-full border border-transparent text-sm font-medium focus-visible:ring-[3px] aria-invalid:ring-[3px] [&_svg:not([class*='size-'])]:size-4 inline-flex items-center justify-center whitespace-nowrap transition-all disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none shrink-0 [&_svg]:shrink-0 outline-none group/button select-none cursor-pointer",
   {
     variants: {
       variant: {
diff --git a/studio/frontend/src/components/ui/combobox.tsx b/studio/frontend/src/components/ui/combobox.tsx
index 53d77415e5..7eb8b07d34 100644
--- a/studio/frontend/src/components/ui/combobox.tsx
+++ b/studio/frontend/src/components/ui/combobox.tsx
@@ -146,7 +146,7 @@ function ComboboxInput({
 function ComboboxContent({
   className,
   side = "bottom",
-  sideOffset = 6,
+  sideOffset = 0,
   align = "start",
   alignOffset = 0,
   anchor,
diff --git a/studio/frontend/src/components/ui/dropdown-menu.tsx b/studio/frontend/src/components/ui/dropdown-menu.tsx
index 0ce4fd87c8..a39651d939 100644
--- a/studio/frontend/src/components/ui/dropdown-menu.tsx
+++ b/studio/frontend/src/components/ui/dropdown-menu.tsx
@@ -36,7 +36,7 @@ function DropdownMenuTrigger({
 function DropdownMenuContent({
   className,
   align = "start",
-  sideOffset = 4,
+  sideOffset = 0,
   ...props
 }: React.ComponentProps<typeof DropdownMenuPrimitive.Content>) {
   return (
@@ -46,7 +46,10 @@ function DropdownMenuContent({
         sideOffset={sideOffset}
         align={align}
         className={cn(
-          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-48 rounded-lg p-1 duration-100 z-50 max-h-(--radix-dropdown-menu-content-available-height) w-(--radix-dropdown-menu-trigger-width) origin-(--radix-dropdown-menu-content-transform-origin) overflow-x-hidden overflow-y-auto data-[state=closed]:overflow-hidden",
+          // The 3px alignment nudge must be margin, not translate: a transform
+          // here makes this scroll container the containing block for nested
+          // position:fixed submenu wrappers, clipping every submenu.
+          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-48 rounded-lg p-1 duration-100 z-50 max-h-(--radix-dropdown-menu-content-available-height) w-[calc(var(--radix-dropdown-menu-trigger-width)_+_6px)] data-[align=start]:-ml-[3px] data-[align=end]:ml-[3px] origin-(--radix-dropdown-menu-content-transform-origin) overflow-x-hidden overflow-y-auto data-[state=closed]:overflow-hidden",
           className,
         )}
         {...props}
@@ -78,7 +81,7 @@ function DropdownMenuItem({
       data-inset={inset}
       data-variant={variant}
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground data-[variant=destructive]:text-destructive data-[variant=destructive]:focus:bg-destructive/10 dark:data-[variant=destructive]:focus:bg-destructive/20 data-[variant=destructive]:focus:text-destructive data-[variant=destructive]:*:[svg]:text-destructive not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 group/dropdown-menu-item relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground data-[variant=destructive]:text-destructive data-[variant=destructive]:focus:bg-destructive/10 dark:data-[variant=destructive]:focus:bg-destructive/20 data-[variant=destructive]:focus:text-destructive data-[variant=destructive]:*:[svg]:text-destructive not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-full px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 group/dropdown-menu-item relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
@@ -96,7 +99,7 @@ function DropdownMenuCheckboxItem({
     <DropdownMenuPrimitive.CheckboxItem
       data-slot="dropdown-menu-checkbox-item"
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-full py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       checked={checked}
@@ -135,7 +138,7 @@ function DropdownMenuRadioItem({
     <DropdownMenuPrimitive.RadioItem
       data-slot="dropdown-menu-radio-item"
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-full py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
@@ -221,7 +224,7 @@ function DropdownMenuSubTrigger({
       data-slot="dropdown-menu-sub-trigger"
       data-inset={inset}
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground data-open:bg-accent data-open:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground data-open:bg-accent data-open:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2 rounded-full px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
@@ -241,14 +244,19 @@ function DropdownMenuSubContent({
   ...props
 }: React.ComponentProps<typeof DropdownMenuPrimitive.SubContent>) {
   return (
-    <DropdownMenuPrimitive.SubContent
-      data-slot="dropdown-menu-sub-content"
-      className={cn(
-        "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-36 rounded-lg p-1 duration-100 z-50 origin-(--radix-dropdown-menu-content-transform-origin) overflow-hidden",
-        className,
-      )}
-      {...props}
-    />
+    // Portaled like DropdownMenuContent: rendered inline, the fixed popper
+    // wrapper is a descendant of the parent menu's scroll container, so any
+    // transform there turns on overflow clipping and hides the submenu.
+    <DropdownMenuPrimitive.Portal>
+      <DropdownMenuPrimitive.SubContent
+        data-slot="dropdown-menu-sub-content"
+        className={cn(
+          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-36 rounded-lg p-1 duration-100 z-50 origin-(--radix-dropdown-menu-content-transform-origin) overflow-hidden",
+          className,
+        )}
+        {...props}
+      />
+    </DropdownMenuPrimitive.Portal>
   );
 }
 
diff --git a/studio/frontend/src/components/ui/menubar.tsx b/studio/frontend/src/components/ui/menubar.tsx
index 14befd8b7c..35b2843102 100644
--- a/studio/frontend/src/components/ui/menubar.tsx
+++ b/studio/frontend/src/components/ui/menubar.tsx
@@ -72,7 +72,7 @@ function MenubarContent({
   className,
   align = "start",
   alignOffset = -4,
-  sideOffset = 8,
+  sideOffset = 0,
   ...props
 }: React.ComponentProps<typeof MenubarPrimitive.Content>) {
   return (
diff --git a/studio/frontend/src/components/ui/popover.tsx b/studio/frontend/src/components/ui/popover.tsx
index a3fdac2100..800bd9c7da 100644
--- a/studio/frontend/src/components/ui/popover.tsx
+++ b/studio/frontend/src/components/ui/popover.tsx
@@ -23,7 +23,7 @@ function PopoverTrigger({
 function PopoverContent({
   className,
   align = "center",
-  sideOffset = 4,
+  sideOffset = 0,
   ...props
 }: React.ComponentProps<typeof PopoverPrimitive.Content>) {
   return (
diff --git a/studio/frontend/src/components/ui/select.tsx b/studio/frontend/src/components/ui/select.tsx
index 03c584fdd1..0ec21850ac 100644
--- a/studio/frontend/src/components/ui/select.tsx
+++ b/studio/frontend/src/components/ui/select.tsx
@@ -63,7 +63,9 @@ function SelectTrigger({
   children,
   icon,
   iconClassName,
-  animateRadius = true,
+  /* Off by default: the radius morph matched the old item-aligned popup
+     that overlaid the trigger; popper menus drop below it instead. */
+  animateRadius = false,
   ...props
 }: React.ComponentProps<typeof SelectPrimitive.Trigger> & {
   size?: "sm" | "default";
@@ -111,7 +113,7 @@ function SelectTrigger({
 function SelectContent({
   className,
   children,
-  position = "item-aligned",
+  position = "popper",
   align = "center",
   container,
   ...props
@@ -125,9 +127,8 @@ function SelectContent({
         data-slot="select-content"
         data-align-trigger={position === "item-aligned"}
         className={cn(
-          "bg-popover text-popover-foreground data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 min-w-36 rounded-xl p-1 corner-squircle duration-100 relative z-50 max-h-(--radix-select-content-available-height) origin-(--radix-select-content-transform-origin) overflow-x-hidden overflow-y-auto ",
-          position === "popper" &&
-            "data-[side=bottom]:translate-y-1 data-[side=left]:-translate-x-1 data-[side=right]:translate-x-1 data-[side=top]:-translate-y-1",
+          "bg-popover text-popover-foreground font-heading data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 min-w-36 rounded-xl p-1 corner-squircle duration-100 relative z-50 max-h-(--radix-select-content-available-height) origin-(--radix-select-content-transform-origin) overflow-x-hidden overflow-y-auto ",
+          // No popper translate offset: the menu sits flush against the trigger.
           className,
         )}
         position={position}
@@ -172,7 +173,7 @@ function SelectItem({
     <SelectPrimitive.Item
       data-slot="select-item"
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-xl corner-squircle py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 *:[span]:last:flex *:[span]:last:items-center *:[span]:last:gap-2 relative flex w-full cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-full py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 *:[span]:last:flex *:[span]:last:items-center *:[span]:last:gap-2 relative flex w-full cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/sidebar.tsx b/studio/frontend/src/components/ui/sidebar.tsx
index b3bae96ce7..eb1e1e1eaf 100644
--- a/studio/frontend/src/components/ui/sidebar.tsx
+++ b/studio/frontend/src/components/ui/sidebar.tsx
@@ -305,9 +305,8 @@ function Sidebar({
           data-sidebar="sidebar"
           data-slot="sidebar-inner"
           className={cn(
-            "bg-sidebar flex size-full flex-col overflow-hidden border-r border-sidebar-border",
+            "bg-sidebar flex size-full flex-col overflow-hidden border-r border-sidebar-border dark:border-r-0",
             "group-data-[variant=floating]:ring-sidebar-border group-data-[variant=floating]:rounded-lg group-data-[variant=floating]:shadow-sm group-data-[variant=floating]:ring-1",
-            hasPinMode && "ring-1 ring-sidebar-border/60",
           )}
         >
           {children}
diff --git a/studio/frontend/src/components/ui/tooltip.tsx b/studio/frontend/src/components/ui/tooltip.tsx
index 954b1871f3..cbc1a09a8f 100644
--- a/studio/frontend/src/components/ui/tooltip.tsx
+++ b/studio/frontend/src/components/ui/tooltip.tsx
@@ -98,9 +98,27 @@ function TooltipContent({
 }: React.ComponentProps<typeof TooltipPrimitive.Content> & {
   variant?: TooltipVariant;
 }) {
+  // Single-line compact tooltips render as a full pill; wrapped ones keep
+  // the squarer corners so tall pills do not look like capsules. A ref
+  // callback measures on mount: Radix mounts the portal content without
+  // re-rendering this wrapper, so an effect here would never see the node.
+  const measureRef = useCallback(
+    (el: HTMLDivElement | null) => {
+      if (!el || variant !== "default") return;
+      const cs = getComputedStyle(el);
+      const lineHeight = Number.parseFloat(cs.lineHeight) || 16;
+      const innerHeight =
+        el.clientHeight -
+        Number.parseFloat(cs.paddingTop) -
+        Number.parseFloat(cs.paddingBottom);
+      el.classList.toggle("rounded-full!", innerHeight < lineHeight * 1.5);
+    },
+    [variant],
+  );
   return (
     <TooltipPrimitive.Portal>
       <TooltipPrimitive.Content
+        ref={measureRef}
         data-slot="tooltip-content"
         sideOffset={sideOffset}
         className={cn(
diff --git a/studio/frontend/src/components/web/update-banner.tsx b/studio/frontend/src/components/web/update-banner.tsx
index 685a8aa3dc..164d922baa 100644
--- a/studio/frontend/src/components/web/update-banner.tsx
+++ b/studio/frontend/src/components/web/update-banner.tsx
@@ -2,13 +2,16 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { usePlatformStore } from "@/config/env";
 import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { AnimatePresence, motion } from "motion/react";
 import { type ReactElement, useEffect, useRef, useState } from "react";
 
-const STUDIO_UPDATE_CMD = "unsloth studio update";
+const STUDIO_INSTALL_UNIX_CMD =
+  "curl -fsSL https://unsloth.ai/install.sh | sh";
+const STUDIO_INSTALL_WINDOWS_CMD = "irm https://unsloth.ai/install.ps1 | iex";
 const RELEASE_NOTES_URL = "https://unsloth.ai/docs/new/changelog";
 const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
 
@@ -20,6 +23,11 @@ export function WebUpdateBanner({
   enabled = true,
 }: WebUpdateBannerProps): ReactElement | null {
   const { status, dismiss } = useWebUpdateCheck({ enabled });
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const installCmd =
+    deviceType === "windows"
+      ? STUDIO_INSTALL_WINDOWS_CMD
+      : STUDIO_INSTALL_UNIX_CMD;
   const [copiedVersion, setCopiedVersion] = useState<string | null>(null);
   const dismissTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
@@ -36,7 +44,7 @@ export function WebUpdateBanner({
   }
 
   async function handleCopyCommand() {
-    if (!(await copyToClipboard(STUDIO_UPDATE_CMD))) {
+    if (!(await copyToClipboard(installCmd))) {
       return;
     }
     setCopiedVersion(status?.latestVersion ?? null);
@@ -89,8 +97,8 @@ export function WebUpdateBanner({
                   Package update available: {status.latestVersion}
                 </p>
                 <p className="mt-1 text-xs leading-relaxed text-muted-foreground">
-                  Installed package: {status.currentVersion}. To update Studio,
-                  run this in your terminal, then restart Studio.
+                  Installed package: {status.currentVersion}. To update Unsloth,
+                  run this in your terminal, then restart Unsloth.
                 </p>
               </div>
             </div>
diff --git a/studio/frontend/src/config/env.ts b/studio/frontend/src/config/env.ts
index 92dad2ae27..61ab111d45 100644
--- a/studio/frontend/src/config/env.ts
+++ b/studio/frontend/src/config/env.ts
@@ -46,12 +46,28 @@ export async function fetchDeviceType(): Promise<DeviceType> {
   if (fetched) return usePlatformStore.getState().deviceType;
 
   try {
-    const res = await fetch(apiUrl("/api/health"));
+    // /api/health only reports the server's device_type to authed callers.
+    // Read the token from storage directly: importing features/auth here
+    // would be an import cycle (auth/session imports this store).
+    const token =
+      typeof window === "undefined"
+        ? null
+        : localStorage.getItem("unsloth_auth_token");
+    const res = await fetch(apiUrl("/api/health"), {
+      headers: token ? { Authorization: `Bearer ${token}` } : undefined,
+    });
     if (res.ok) {
       const data = (await res.json()) as { device_type?: string; chat_only?: boolean };
       const deviceType = data.device_type ?? detectLocalPlatform();
       const chatOnly = data.chat_only ?? false;
-      usePlatformStore.setState({ deviceType, chatOnly, fetched: true });
+      // Cache only a server-reported platform. Unauthenticated responses fall
+      // back to the browser platform, which can differ from the host (WSL,
+      // SSH); keeping fetched=false retries once a token exists.
+      usePlatformStore.setState({
+        deviceType,
+        chatOnly,
+        fetched: data.device_type !== undefined,
+      });
       return deviceType;
     }
   } catch {
diff --git a/studio/frontend/src/features/auth/api.ts b/studio/frontend/src/features/auth/api.ts
index 6356f89378..144e85d40a 100644
--- a/studio/frontend/src/features/auth/api.ts
+++ b/studio/frontend/src/features/auth/api.ts
@@ -175,7 +175,7 @@ export async function authFetch(
           "You appear to be offline. Check your network connection and try again.",
         );
       }
-      throw new Error("Studio isn't running -- please relaunch it.");
+      throw new Error("Unsloth isn't running -- please relaunch it.");
     }
     throw err;
   }
diff --git a/studio/frontend/src/features/auth/change-password-page.tsx b/studio/frontend/src/features/auth/change-password-page.tsx
index 96ee8dd1a8..30f840faca 100644
--- a/studio/frontend/src/features/auth/change-password-page.tsx
+++ b/studio/frontend/src/features/auth/change-password-page.tsx
@@ -10,11 +10,11 @@ export function ChangePasswordPage() {
     <div className="relative flex min-h-[calc(100dvh-var(--studio-titlebar-height,0px))] items-center justify-center overflow-hidden bg-background px-4 py-8 sm:px-6 sm:py-10 md:px-10">
       <LightRays
         count={6}
-        color="rgba(34, 197, 94, 0.25)"
+        color="rgba(34, 197, 94, 0.1)"
         blur={34}
         speed={15}
         length="70vh"
-        style={{ opacity: 0.4 }}
+        className="opacity-35 dark:opacity-15"
       />
       <Card className="relative z-10 w-full max-w-sm px-5 py-6 shadow-border ring-1 ring-border sm:px-6 sm:py-8">
         <AuthForm mode="change-password" />
diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx
index f2017ff0b6..d109b5f196 100644
--- a/studio/frontend/src/features/auth/components/auth-form.tsx
+++ b/studio/frontend/src/features/auth/components/auth-form.tsx
@@ -3,6 +3,7 @@
 
 import { apiUrl } from "@/lib/api-base";
 import { Button } from "@/components/ui/button";
+import { MascotImg } from "@/components/mascot-img";
 import { Input } from "@/components/ui/input";
 import { Label } from "@/components/ui/label";
 import { Link, useNavigate } from "@tanstack/react-router";
@@ -311,9 +312,8 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
   return (
     <div className="w-full max-w-sm space-y-6">
       <div className="space-y-1.5 text-center">
-        <img
-          src="/Sloth emojis/large sloth wave.png"
-          alt="Unsloth waving mascot"
+        <MascotImg
+          src="Sloth emojis/large sloth wave.png"
           className="mx-auto mb-2 h-20 w-20 object-contain"
         />
         <h2 className="text-2xl font-semibold text-foreground">{title}</h2>
diff --git a/studio/frontend/src/features/auth/login-page.tsx b/studio/frontend/src/features/auth/login-page.tsx
index 6f4d03f1a1..638df7d362 100644
--- a/studio/frontend/src/features/auth/login-page.tsx
+++ b/studio/frontend/src/features/auth/login-page.tsx
@@ -10,13 +10,13 @@ export function LoginPage() {
     <div className="relative flex min-h-[calc(100dvh-var(--studio-titlebar-height,0px))] items-center justify-center overflow-hidden bg-background px-4 py-8 sm:px-6 sm:py-10 md:px-10">
       <LightRays
         count={6}
-        color="rgba(34, 197, 94, 0.25)"
+        color="rgba(34, 197, 94, 0.1)"
         blur={34}
         speed={15}
         length="70vh"
-        style={{ opacity: 0.4 }}
+        className="opacity-35 dark:opacity-15"
       />
-      <Card className="relative z-10 w-full max-w-sm px-5 py-6 shadow-border ring-1 ring-border sm:px-6 sm:py-8">
+      <Card className="relative z-10 w-full max-w-sm px-5 py-6 shadow-border ring-0 sm:px-6 sm:py-8">
         <AuthForm mode="login" />
       </Card>
     </div>
diff --git a/studio/frontend/src/features/auth/tauri-auto-auth.ts b/studio/frontend/src/features/auth/tauri-auto-auth.ts
index 760af26395..e825d5de31 100644
--- a/studio/frontend/src/features/auth/tauri-auto-auth.ts
+++ b/studio/frontend/src/features/auth/tauri-auto-auth.ts
@@ -26,7 +26,7 @@ let pending: { promise: Promise<boolean>; force: boolean } | null = null;
 let lastTauriAuthFailure: string | null = null;
 
 const TAURI_AUTH_FAILURE_FALLBACK =
-  "Desktop authentication failed. Update or repair the managed Studio install, then restart Studio.";
+  "Desktop authentication failed. Update or repair the managed Unsloth install, then restart Unsloth.";
 const BACKEND_NOT_READY_MESSAGE = "Backend is not ready";
 
 function authFailureMessage(error: unknown): string {
diff --git a/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
index a679e8ed83..146a12bd40 100644
--- a/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
+++ b/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
@@ -18,7 +18,7 @@ const describeMediaError = (error: unknown): string => {
     return "Dictation could not access the microphone.";
   }
   if (error.name === "NotAllowedError") {
-    return "Microphone access is blocked. Allow microphone access for this Studio page, then try again.";
+    return "Microphone access is blocked. Allow microphone access for this Unsloth page, then try again.";
   }
   if (error.name === "NotFoundError") {
     return "No microphone was found for dictation.";
@@ -31,7 +31,7 @@ const describeMediaError = (error: unknown): string => {
 
 const describeSpeechError = (error: string, message?: string): string => {
   if (error === "not-allowed") {
-    return "Speech recognition was blocked by the browser. Check microphone permissions for this Studio page.";
+    return "Speech recognition was blocked by the browser. Check microphone permissions for this Unsloth page.";
   }
   if (error === "service-not-allowed") {
     return "Speech recognition is blocked by the browser speech service.";
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index f0d229b412..d944c7ab74 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -986,24 +986,51 @@ function findLatestUserImageBase64(messages: RunMessages): string | undefined {
   return undefined;
 }
 
-function findLatestUserAudioBase64(messages: RunMessages): string | undefined {
-  // Message content parts (compare view CompareMessagePart type: "audio").
+function extractAudioPartBase64(
+  part: { type: string } | null | undefined,
+): string | undefined {
+  if (!part || part.type !== "audio" || !("audio" in part)) return undefined;
+  const audioPart = (
+    part as unknown as {
+      type: "audio";
+      audio: string | { data: string; format: string };
+    }
+  ).audio;
+  const raw = typeof audioPart === "string" ? audioPart : audioPart?.data;
+  if (!raw) return undefined;
+  return raw.startsWith("data:") ? raw.split(",")[1] : raw;
+}
+
+// Exported for tests.
+export function findLatestUserAudioBase64(
+  messages: RunMessages,
+): string | undefined {
   for (let i = messages.length - 1; i >= 0; i -= 1) {
     const message = messages[i];
     if (!message || message.role !== "user") continue;
 
+    // Message content parts (from compare view's CompareMessagePart with type: "audio")
     for (const part of message.content ?? []) {
-      if (part.type === "audio" && "audio" in part) {
-        const audioPart = (
-          part as unknown as {
-            type: "audio";
-            audio: string | { data: string; format: string };
-          }
-        ).audio;
-        const raw = typeof audioPart === "string" ? audioPart : audioPart?.data;
-        if (raw) return raw.startsWith("data:") ? raw.split(",")[1] : raw;
+      const base64 = extractAudioPartBase64(part);
+      if (base64) return base64;
+    }
+
+    // Attachment content parts (from AudioAttachmentAdapter)
+    if ("attachments" in message) {
+      for (const attachment of message.attachments ?? []) {
+        for (const part of attachment.content ?? []) {
+          const base64 = extractAudioPartBase64(part);
+          if (base64) return base64;
+        }
       }
     }
+
+    // Only the newest user message counts. audio_base64 switches the
+    // backend onto the audio generation path, so replaying audio from an
+    // older turn would hijack text follow-ups (Whisper would retranscribe
+    // the stale clip). Matches the consumed-on-send semantics of the
+    // legacy pendingAudio path.
+    break;
   }
 
   // Runtime store (main composer's audio upload).
diff --git a/studio/frontend/src/features/chat/api/mcp-servers-api.ts b/studio/frontend/src/features/chat/api/mcp-servers-api.ts
index be88d12664..0dc0719da4 100644
--- a/studio/frontend/src/features/chat/api/mcp-servers-api.ts
+++ b/studio/frontend/src/features/chat/api/mcp-servers-api.ts
@@ -21,6 +21,12 @@ export interface McpServerProbeResult {
   error: string | null;
 }
 
+export interface McpServerImportResult {
+  created: McpServerConfig[];
+  skipped: string[];
+  errors: string[];
+}
+
 function parseErrorText(status: number, body: unknown): string {
   if (body && typeof body === "object") {
     const { detail, message } = body as { detail?: unknown; message?: unknown };
@@ -114,3 +120,12 @@ export function testMcpServer(payload: {
     },
   });
 }
+
+// Bulk-import servers from a standard mcpServers JSON config (Claude Desktop,
+// Cursor, Cline, VS Code). The backend skips duplicates and reports per-entry
+// errors instead of failing the whole batch.
+export function importMcpServers(
+  config: unknown,
+): Promise<McpServerImportResult> {
+  return mcpRequest("/import", { method: "POST", body: { config } });
+}
diff --git a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
index a685515253..22066b6fc3 100644
--- a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
+++ b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
@@ -9,6 +9,7 @@ import {
   unslothDarkTheme,
   unslothLightTheme,
 } from "@/components/assistant-ui/code-themes";
+import { MascotImg } from "@/components/mascot-img";
 import { Button } from "@/components/ui/button";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { cn } from "@/lib/utils";
@@ -78,9 +79,8 @@ function ArtifactGeneratingPanel() {
   return (
     <div className="flex h-full min-h-0 flex-col items-center justify-center bg-muted/10 px-6 text-center">
       <div className="max-w-[30ch] space-y-1.5">
-        <img
-          src="/Sloth%20emojis/sloth%20w%20pc%20transparent.png"
-          alt=""
+        <MascotImg
+          src="Sloth emojis/sloth w pc transparent.png"
           aria-hidden={true}
           className="mx-auto mb-3 size-20 object-contain"
         />
diff --git a/studio/frontend/src/features/chat/audio-attachment-adapter.ts b/studio/frontend/src/features/chat/audio-attachment-adapter.ts
new file mode 100644
index 0000000000..148f025e57
--- /dev/null
+++ b/studio/frontend/src/features/chat/audio-attachment-adapter.ts
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  AUDIO_ACCEPT,
+  MAX_AUDIO_SIZE,
+  fileToBase64,
+} from "@/lib/audio-utils";
+import type {
+  Attachment,
+  AttachmentAdapter,
+  CompleteAttachment,
+  PendingAttachment,
+} from "@assistant-ui/react";
+import { toast } from "sonner";
+import { useChatRuntimeStore } from "./stores/chat-runtime-store";
+
+// crypto.randomUUID is undefined in non-secure contexts (HTTP over a LAN IP).
+function newAttachmentId(): string {
+  if (typeof globalThis.crypto?.randomUUID === "function") {
+    return globalThis.crypto.randomUUID();
+  }
+  return `${Date.now()}-${Math.random().toString(36).slice(2, 10)}`;
+}
+
+// Audio shares the "Add photos & files" picker. Like VisionImageAdapter,
+// unsupported models are rejected at add() time with a toast.
+export class AudioAttachmentAdapter implements AttachmentAdapter {
+  // MIME is unreliable for some containers (m4a), so also match by
+  // extension. No .webm extension: it would claim video/webm files; real
+  // audio webm (MediaRecorder) always reports the audio/webm MIME.
+  accept = `${AUDIO_ACCEPT},audio/x-m4a,.wav,.mp3,.m4a,.ogg,.oga,.flac`;
+  private readonly attachmentIds = new Set<string>();
+
+  async add({ file }: { file: File }): Promise<PendingAttachment> {
+    const state = useChatRuntimeStore.getState();
+    const checkpoint = state.params.checkpoint;
+    const activeModel = state.models.find((m) => m.id === checkpoint);
+    const modelLoaded = !!checkpoint && !state.modelLoading;
+    let unavailableReason: string | null = null;
+    if (!modelLoaded) {
+      unavailableReason = "Load a model before adding audio files.";
+    } else if (!activeModel?.hasAudioInput) {
+      const label = activeModel?.name || checkpoint || "Current model";
+      unavailableReason = `${label} cannot accept audio. Load an audio-input model before attaching audio files.`;
+    }
+    if (unavailableReason) {
+      toast.error(unavailableReason);
+      throw new Error(unavailableReason);
+    }
+    if (file.size > MAX_AUDIO_SIZE) {
+      const sizeReason = "Audio size exceeds 50MB limit";
+      toast.error(sizeReason);
+      throw new Error(sizeReason);
+    }
+    if (this.attachmentIds.size > 0 || state.pendingAudioBase64) {
+      const duplicateReason = "Only one audio file can be attached per message.";
+      toast.error(duplicateReason);
+      throw new Error(duplicateReason);
+    }
+
+    const id = newAttachmentId();
+    this.attachmentIds.add(id);
+    return {
+      id,
+      type: "file",
+      name: file.name,
+      contentType: file.type,
+      file,
+      status: { type: "requires-action", reason: "composer-send" },
+    };
+  }
+
+  async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
+    try {
+      const base64 = await fileToBase64(attachment.file);
+      // Backend takes raw base64; format only satisfies the part type.
+      const format = attachment.contentType === "audio/mpeg" ? "mp3" : "wav";
+      return {
+        id: attachment.id,
+        type: "file",
+        name: attachment.name,
+        contentType: attachment.contentType,
+        content: [{ type: "audio", audio: { data: base64, format } }],
+        status: { type: "complete" },
+      };
+    } finally {
+      this.attachmentIds.delete(attachment.id);
+    }
+  }
+
+  remove(attachment: Attachment): Promise<void> {
+    this.attachmentIds.delete(attachment.id);
+    return Promise.resolve();
+  }
+}
diff --git a/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx b/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
index f78784b436..17d6d1e9ec 100644
--- a/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx
@@ -1,11 +1,11 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { useCallback, useEffect, useState } from "react";
+import { type ChangeEvent, useCallback, useEffect, useRef, useState } from "react";
 import { toast } from "sonner";
 import { Delete02Icon, Edit03Icon, PlusSignIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { RefreshCwIcon } from "lucide-react";
+import { RefreshCwIcon, UploadIcon } from "lucide-react";
 
 import { Button } from "@/components/ui/button";
 import {
@@ -23,6 +23,7 @@ import {
   type McpServerConfig,
   createMcpServer,
   deleteMcpServer,
+  importMcpServers,
   listMcpServers,
   refreshMcpServerTools,
   testMcpServer,
@@ -194,7 +195,9 @@ export function ChatMcpServersDialog({
   const [form, setForm] = useState<FormState>(EMPTY_FORM);
   const [saving, setSaving] = useState(false);
   const [testing, setTesting] = useState(false);
+  const [importing, setImporting] = useState(false);
   const [refreshingId, setRefreshingId] = useState<string | null>(null);
+  const fileInputRef = useRef<HTMLInputElement>(null);
 
   const refresh = useCallback(async () => {
     setLoading(true);
@@ -315,6 +318,49 @@ export function ChatMcpServersDialog({
     }
   }
 
+  async function onImportFile(e: ChangeEvent<HTMLInputElement>) {
+    const file = e.target.files?.[0];
+    e.target.value = ""; // let the user re-pick the same file later
+    if (!file) return;
+    let config: unknown;
+    try {
+      config = JSON.parse(await file.text());
+    } catch {
+      toast.error("Invalid JSON file");
+      return;
+    }
+    setImporting(true);
+    try {
+      const result = await importMcpServers(config);
+      const parts = [`${result.created.length} added`];
+      if (result.skipped.length) parts.push(`${result.skipped.length} skipped`);
+      if (result.errors.length) {
+        parts.push(
+          `${result.errors.length} error${result.errors.length === 1 ? "" : "s"}`,
+        );
+      }
+      const summary = parts.join(", ");
+      if (result.errors.length) {
+        toast.warning(summary, {
+          description: (
+            <div className="whitespace-pre-line">
+              {result.errors.slice(0, 5).join("\n")}
+            </div>
+          ),
+        });
+      } else {
+        toast.success(summary);
+      }
+      await refresh();
+    } catch (err) {
+      toast.error("Import failed", {
+        description: err instanceof Error ? err.message : String(err),
+      });
+    } finally {
+      setImporting(false);
+    }
+  }
+
   async function removeServer(server: McpServerConfig) {
     const ok = window.confirm(`Delete MCP server "${server.display_name}"?`);
     if (!ok) return;
@@ -385,9 +431,35 @@ export function ChatMcpServersDialog({
             Register remote (HTTP) or local (stdio command) MCP servers.
           </DialogDescription>
         </DialogHeader>
+        <input
+          ref={fileInputRef}
+          type="file"
+          accept="application/json,.json"
+          className="hidden"
+          onChange={onImportFile}
+        />
 
         {showForm ? (
           <div className="flex flex-col gap-4">
+            {view.kind === "create" && (
+              <div className="flex items-center justify-between gap-3 rounded-md border border-dashed px-3 py-2">
+                <span className="text-xs text-muted-foreground">
+                  Import servers from a config file.
+                </span>
+                <Button
+                  type="button"
+                  size="sm"
+                  variant="outline"
+                  className="shrink-0"
+                  onClick={() => fileInputRef.current?.click()}
+                  disabled={importing}
+                  title="Import servers from a mcpServers JSON config (Claude Desktop, Cursor, VS Code…)"
+                >
+                  {importing ? <Spinner /> : <UploadIcon size={14} />}
+                  Import config
+                </Button>
+              </div>
+            )}
             <div className="grid gap-2">
               <Label htmlFor="mcp-display-name">Display name</Label>
               <Input
@@ -467,7 +539,17 @@ export function ChatMcpServersDialog({
           </div>
         ) : (
           <div className="flex min-w-0 flex-col gap-3">
-            <div className="flex justify-end">
+            <div className="flex justify-end gap-2">
+              <Button
+                size="sm"
+                variant="outline"
+                onClick={() => fileInputRef.current?.click()}
+                disabled={importing}
+                title="Import servers from a mcpServers JSON config (Claude Desktop, Cursor, VS Code…)"
+              >
+                {importing ? <Spinner /> : <UploadIcon size={14} />}
+                Import config
+              </Button>
               <Button size="sm" onClick={startCreate}>
                 <HugeiconsIcon icon={PlusSignIcon} size={14} />
                 Add server
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 4472fffe7f..41451d06b1 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -1009,7 +1009,7 @@ function ProjectLanding({
                             : { compare: item.id, project: projectId },
                       });
                     }}
-                    className="group flex min-h-[58px] w-full items-center gap-4 rounded-[10px] px-4 py-2 text-left transition-colors hover:bg-nav-surface-hover"
+                    className="group flex min-h-[58px] w-full items-center gap-4 rounded-full px-4 py-2 text-left transition-colors hover:bg-nav-surface-hover"
                   >
                     <div className="min-w-0 flex-1">
                       <div className="truncate text-[15px] font-semibold leading-5 text-foreground">
diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
index e35a8e35e0..0594b55083 100644
--- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
@@ -1511,7 +1511,7 @@ export function ChatProvidersSettings({
         <div className="flex min-w-0 flex-col gap-1">
           <h1 className="font-heading text-lg font-semibold">Connections</h1>
           <p className="text-xs leading-relaxed text-muted-foreground">
-            Manage model connections for chat through the Studio proxy.
+            Manage model connections for chat.
           </p>
         </div>
       </header>
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index fc5d5f8b5c..5751a1a426 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -50,10 +50,12 @@ import {
   TooltipTrigger,
 } from "@/components/ui/tooltip";
 import { useIsMobile } from "@/hooks/use-mobile";
+import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
 import { cn } from "@/lib/utils";
 import {
   ArrowDown01Icon,
   ArrowTurnBackwardIcon,
+  Edit03Icon,
   InformationCircleIcon,
   LayoutAlignRightIcon,
 } from "@hugeicons/core-free-icons";
@@ -61,7 +63,7 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import { ChevronDown, ExternalLink } from "lucide-react";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import { Fragment, type ReactNode } from "react";
-import { useEffect, useMemo, useRef, useState } from "react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import { toast } from "@/lib/toast";
 import { OpenAICodeExecSection } from "./components/openai-code-exec-section";
 import {
@@ -187,12 +189,16 @@ function NumericValueInput({
     }
   };
 
+  const displayed = focused ? draft : (displayValue ?? String(value));
+
   return (
     <input
       type="text"
       inputMode="decimal"
       size={sizeAttr}
-      value={focused ? draft : (displayValue ?? String(value))}
+      /* Fixed 4ch pill; grows only when a longer value would clip. */
+      style={{ width: `calc(${Math.max(displayed.length, 4)}ch + 18px)` }}
+      value={displayed}
       aria-label={ariaLabel}
       onFocus={(e) => {
         cancelBlurCommitRef.current = false;
@@ -263,7 +269,7 @@ function ParamSlider({
           onChange={onChange}
           displayValue={displayValue}
           ariaLabel={label}
-          size={valueSize ?? 6}
+          size={valueSize ?? 4}
         />
       </div>
       <Slider
@@ -318,6 +324,8 @@ function saveCollapsibleOpen(label: string, open: boolean) {
 function CollapsibleSection({
   label,
   labelHref,
+  headerAction,
+  onLabelClick,
   children,
   defaultOpen = false,
   first = false,
@@ -329,6 +337,14 @@ function CollapsibleSection({
    * are siblings rather than an <a> nested in a <button> (invalid HTML).
    */
   labelHref?: string;
+  /**
+   * Optional control rendered before the chevron (e.g. an edit icon). The
+   * label and chevron become sibling toggles so the action is not a button
+   * nested in a button.
+   */
+  headerAction?: ReactNode;
+  /** When set, clicking the label runs this instead of toggling collapse. */
+  onLabelClick?: () => void;
   children?: ReactNode;
   defaultOpen?: boolean;
   first?: boolean;
@@ -378,6 +394,29 @@ function CollapsibleSection({
             />
           </button>
         </div>
+      ) : headerAction ? (
+        <div className={headerClasses}>
+          <button
+            type="button"
+            onClick={onLabelClick ?? toggle}
+            className="flex min-w-0 flex-1 cursor-pointer items-center text-left leading-none transition-colors hover:text-nav-fg"
+          >
+            <span className="leading-none">{label}</span>
+          </button>
+          <span className="flex shrink-0 items-center gap-1">
+            {headerAction}
+            <button
+              type="button"
+              onClick={toggle}
+              aria-label={open ? `Collapse ${label}` : `Expand ${label}`}
+              className="flex shrink-0 cursor-pointer items-center leading-none transition-colors hover:text-nav-fg"
+            >
+              <ChevronDown
+                className={cn("size-3.5", open ? "rotate-0" : "-rotate-90")}
+              />
+            </button>
+          </span>
+        </div>
       ) : (
         <button
           type="button"
@@ -452,6 +491,27 @@ export function ChatSettingsPanel({
   const loadedSpeculativeType = useChatRuntimeStore(
     (s) => s.loadedSpeculativeType,
   );
+  const specFallbackReason = useChatRuntimeStore((s) => s.specFallbackReason);
+  // "binary_no_mtp" / "binary_outdated" mean a newer prebuilt would re-enable
+  // MTP; "runtime_error" means the current build cannot run it (no update push).
+  const mtpUpdatable =
+    specFallbackReason === "binary_no_mtp" ||
+    specFallbackReason === "binary_outdated";
+  const {
+    status: llamaUpdateStatus,
+    applying: llamaUpdating,
+    apply: applyLlamaUpdate,
+  } = useLlamaUpdateCheck({ enabled: mtpUpdatable });
+  const handleMtpUpdate = useCallback(async () => {
+    const result = await applyLlamaUpdate();
+    if (result.ok) {
+      toast.success(
+        `llama.cpp updated to ${result.tag ?? "the latest build"}. Reload your model to enable MTP.`,
+      );
+    } else {
+      toast.error(`llama.cpp update failed: ${result.error ?? "unknown error"}`);
+    }
+  }, [applyLlamaUpdate]);
   const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax);
   const setSpecDraftNMax = useChatRuntimeStore((s) => s.setSpecDraftNMax);
   const loadedSpecDraftNMax = useChatRuntimeStore(
@@ -492,19 +552,18 @@ export function ChatSettingsPanel({
   const specDirty = speculativeType !== loadedSpeculativeType;
   const specDraftDirty = specDraftNMax !== loadedSpecDraftNMax;
   const modelSettingsDirty = kvDirty || ctxDirty || specDirty || specDraftDirty;
-  const chatTemplateOverride = useChatRuntimeStore(
-    (s) => s.chatTemplateOverride,
-  );
   const loadedChatTemplateOverride = useChatRuntimeStore(
     (s) => s.loadedChatTemplateOverride,
   );
   const setChatTemplateOverride = useChatRuntimeStore(
     (s) => s.setChatTemplateOverride,
   );
-  const templateDirty = chatTemplateOverride !== loadedChatTemplateOverride;
   const [presetNameInput, setPresetNameInput] = useState(activePreset);
   const [systemPromptEditorOpen, setSystemPromptEditorOpen] = useState(false);
   const [systemPromptDraft, setSystemPromptDraft] = useState("");
+  // When the prompt overflows the inline box, clicking opens the popup editor.
+  const systemPromptBoxRef = useRef<HTMLTextAreaElement>(null);
+  const [systemPromptOverflows, setSystemPromptOverflows] = useState(false);
   const [activePresetBaseline, setActivePresetBaseline] = useState(params);
   const presets = useMemo(() => {
     return getOrderedPresets(customPresets);
@@ -709,6 +768,16 @@ export function ChatSettingsPanel({
     }
   }, [open]);
 
+  useEffect(() => {
+    const el = systemPromptBoxRef.current;
+    setSystemPromptOverflows(
+      params.systemPrompt.length > 0 &&
+        el != null &&
+        el.clientHeight > 0 &&
+        el.scrollHeight > el.clientHeight + 1,
+    );
+  }, [params.systemPrompt, open]);
+
   const settingsScrollRef = useRef<HTMLDivElement>(null);
 
   const settingsContent = (
@@ -837,7 +906,7 @@ export function ChatSettingsPanel({
                         animateRadius={false}
                         icon={ArrowDown01Icon}
                         iconClassName="size-3.5"
-                        className="grid h-7 w-[60px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-[10px] border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.07] px-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
+                        className="grid h-7 w-[64px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
                       >
                         <SelectValue />
                       </SelectTrigger>
@@ -877,7 +946,7 @@ export function ChatSettingsPanel({
                         animateRadius={false}
                         icon={ArrowDown01Icon}
                         iconClassName="size-3.5"
-                        className="grid h-7 w-[120px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-[10px] border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.07] px-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
+                        className="grid h-7 w-[124px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
                         data-test-id="speculative-type-select"
                       >
                         <SelectValue />
@@ -892,6 +961,32 @@ export function ChatSettingsPanel({
                     </Select>
                   </div>
                 </div>
+                {specFallbackReason &&
+                  (speculativeType === "auto" ||
+                    speculativeType === "mtp" ||
+                    speculativeType === "mtp+ngram") && (
+                    <div className="rounded-lg bg-amber-500/[0.08] px-3 py-2 text-[12px] leading-[1.4] text-nav-fg/80">
+                      <p>
+                        {specFallbackReason === "runtime_error"
+                          ? "MTP could not start for this model on the installed llama.cpp build, so it is running without speculative decoding."
+                          : "MTP is not available in the installed llama.cpp build, so this model is running without it." +
+                            (llamaUpdateStatus?.update_available
+                              ? " Update llama.cpp to enable it."
+                              : "")}
+                      </p>
+                      {mtpUpdatable && llamaUpdateStatus?.update_available && (
+                        <Button
+                          size="sm"
+                          className="corner-squircle mt-2 h-7 text-[12px]"
+                          onClick={handleMtpUpdate}
+                          disabled={llamaUpdating}
+                          data-test-id="mtp-update-button"
+                        >
+                          {llamaUpdating ? "Updating..." : "Update llama.cpp"}
+                        </Button>
+                      )}
+                    </div>
+                  )}
                 {(speculativeType === "mtp" ||
                   speculativeType === "mtp+ngram") && (
                   <div className="flex items-center justify-between gap-3">
@@ -928,7 +1023,7 @@ export function ChatSettingsPanel({
                       }}
                       data-test-id="spec-draft-n-max-input"
                       aria-label="Speculative decoding draft tokens"
-                      className="h-7 w-[72px] rounded-[10px] border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.07] px-2 py-0 text-[13px] font-medium text-nav-fg outline-none focus-visible:ring-0"
+                      className="h-7 w-[76px] rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px] font-medium text-nav-fg outline-none focus-visible:ring-0"
                     />
                   </div>
                 )}
@@ -967,7 +1062,7 @@ export function ChatSettingsPanel({
               </>
             )}
             <ChatTemplateFields />
-            {(modelSettingsDirty || templateDirty) && (
+            {modelSettingsDirty && (
               <div className="flex flex-wrap gap-1.5 pt-1">
                 <Button
                   type="button"
@@ -1040,7 +1135,7 @@ export function ChatSettingsPanel({
                     />
                     <InputGroupAddon
                       align="inline-end"
-                      className="min-h-0 shrink-0 gap-0 self-stretch border-0 py-0 pl-0 !pr-1 has-[>button]:mr-0"
+                      className="min-h-0 shrink-0 gap-0 self-stretch border-0 py-0 pl-0 !pr-1 has-[>button]:mr-0 !cursor-pointer"
                     >
                       <span
                         className="!h-7 min-h-7 !w-7 min-w-7 shrink-0 self-center inline-flex items-center justify-center rounded-full border-0 px-0 text-[#a0a097] dark:text-nav-fg pointer-events-none"
@@ -1058,7 +1153,7 @@ export function ChatSettingsPanel({
               </DropdownMenuTrigger>
               <DropdownMenuContent
                 align="start"
-                sideOffset={6}
+                sideOffset={0}
                 className="menu-soft-surface ring-0 border-0 rounded-lg p-1.5"
               >
                 {presets.map((p, index) => (
@@ -1092,7 +1187,7 @@ export function ChatSettingsPanel({
                 variant={presetSaveState.isSaveReady ? "default" : "outline"}
                 size="sm"
                 className={cn(
-                  "h-9 w-full rounded-[10px] text-[13px] font-medium tracking-nav",
+                  "h-9 w-full rounded-full text-[13px] font-medium tracking-nav",
                   presetSaveState.isSaveReady &&
                     "bg-primary text-primary-foreground hover:bg-primary/90",
                 )}
@@ -1107,7 +1202,7 @@ export function ChatSettingsPanel({
                 disabled={!(settingsHydrated && activeCustomPreset)}
                 variant="outline"
                 size="sm"
-                className="h-9 w-full rounded-[10px] text-[13px] font-medium tracking-nav text-muted-foreground"
+                className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav text-muted-foreground"
                 title={
                   activeCustomPreset
                     ? activeBuiltinPreset
@@ -1218,23 +1313,66 @@ export function ChatSettingsPanel({
           </CollapsibleSection>
         ) : null}
 
-        <CollapsibleSection label="System Prompt" defaultOpen={true}>
-          <button
-            type="button"
-            onClick={openSystemPromptEditor}
-            aria-label="Edit system prompt"
+        <CollapsibleSection
+          label="System Prompt"
+          defaultOpen={true}
+          onLabelClick={openSystemPromptEditor}
+          headerAction={
+            <Tooltip>
+              <TooltipPrimitive.Trigger asChild>
+                <button
+                  type="button"
+                  onClick={openSystemPromptEditor}
+                  className="nav-icon-btn text-nav-icon-idle hover:bg-panel-surface-hover hover:text-black dark:hover:text-white"
+                  aria-label="Edit system prompt"
+                >
+                  <HugeiconsIcon
+                    icon={Edit03Icon}
+                    strokeWidth={1.75}
+                    className="size-3"
+                  />
+                </button>
+              </TooltipPrimitive.Trigger>
+              <TooltipContent
+                side="top"
+                sideOffset={6}
+                className="tooltip-compact"
+              >
+                Edit prompt
+              </TooltipContent>
+            </Tooltip>
+          }
+        >
+          {/* Rounded wrapper clips overflowing text and the scrollbar. */}
+          <div
             className={cn(
-              "panel-text-surface mt-1 flex w-full h-20 overflow-hidden cursor-pointer items-start px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed corner-squircle focus-visible:outline-none focus-visible:border-ring focus-visible:ring-[1px] focus-visible:ring-ring/40",
-              params.systemPrompt
-                ? "text-nav-fg"
-                : "text-muted-foreground",
+              "panel-text-surface -mt-1 h-20 w-full overflow-hidden corner-squircle",
+              systemPromptOverflows && "cursor-pointer",
             )}
           >
-            <span className="block line-clamp-3 whitespace-pre-wrap break-words">
-              {params.systemPrompt ||
-                "Example: You are a helpful assistant..."}
-            </span>
-          </button>
+            <textarea
+              ref={systemPromptBoxRef}
+              value={params.systemPrompt}
+              onChange={(e) => set("systemPrompt")(e.target.value)}
+              onMouseDown={(e) => {
+                // Overflowing prompt: click opens the popup editor instead.
+                // While focused, clicks still move the caret normally.
+                if (
+                  systemPromptOverflows &&
+                  document.activeElement !== e.currentTarget
+                ) {
+                  e.preventDefault();
+                  openSystemPromptEditor();
+                }
+              }}
+              placeholder="Example: You are a helpful assistant..."
+              aria-label="System prompt"
+              className={cn(
+                "block size-full resize-none bg-transparent px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed text-nav-fg outline-none placeholder:text-muted-foreground",
+                systemPromptOverflows && "cursor-pointer",
+              )}
+            />
+          </div>
         </CollapsibleSection>
 
         <CollapsibleSection label="Sampling" defaultOpen={true}>
@@ -1379,7 +1517,7 @@ export function ChatSettingsPanel({
         }}
       >
         <DialogContent
-          className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-3xl"
+          className="corner-squircle dialog-soft-surface sm:max-w-3xl"
           overlayClassName="bg-background/35 supports-backdrop-filter:backdrop-blur-[1px]"
         >
           <DialogHeader>
@@ -1402,7 +1540,7 @@ export function ChatSettingsPanel({
               onChange={(event) => setSystemPromptDraft(event.target.value)}
               placeholder="You are a helpful assistant..."
               fieldSizing="fixed"
-              className="min-h-[24rem] max-h-[50vh] overflow-y-auto text-sm leading-6 corner-squircle focus-visible:border-input focus-visible:ring-0"
+              className="min-h-[24rem] max-h-[50vh] overflow-y-auto border-0 text-sm leading-6 corner-squircle focus-visible:ring-0"
               rows={14}
             />
           </div>
@@ -1410,20 +1548,31 @@ export function ChatSettingsPanel({
             <Button
               type="button"
               variant="ghost"
-              onClick={() => {
-                setSystemPromptDraft(params.systemPrompt);
-                setSystemPromptEditorOpen(false);
-              }}
+              onClick={() => setSystemPromptDraft("")}
+              disabled={systemPromptDraft.length === 0}
+              className="text-muted-foreground"
             >
-              Cancel
-            </Button>
-            <Button
-              type="button"
-              onClick={saveSystemPromptEditor}
-              disabled={!systemPromptEditorDirty}
-            >
-              Save
+              Reset
             </Button>
+            <div className="flex flex-wrap gap-2">
+              <Button
+                type="button"
+                variant="ghost"
+                onClick={() => {
+                  setSystemPromptDraft(params.systemPrompt);
+                  setSystemPromptEditorOpen(false);
+                }}
+              >
+                Cancel
+              </Button>
+              <Button
+                type="button"
+                onClick={saveSystemPromptEditor}
+                disabled={!systemPromptEditorDirty}
+              >
+                Save
+              </Button>
+            </div>
           </DialogFooter>
         </DialogContent>
       </Dialog>
@@ -1551,19 +1700,27 @@ function ChatTemplateFields() {
     setEditorOpen(true);
   };
   const saveEditor = () => {
-    setOverride(
-      draft.trim().length === 0 || draft === defaultTemplate ? null : draft,
-    );
+    const cleared = draft.trim().length === 0 || draft === defaultTemplate;
+    setOverride(cleared ? null : draft);
     setEditorOpen(false);
+    toast.success(
+      cleared
+        ? "Chat template reset to default. It applies on the next model reload."
+        : "Chat template saved. It applies on the next model reload.",
+    );
   };
 
   return (
     <>
-      <div className="flex flex-col gap-2">
-        <div className="flex items-center justify-between gap-2">
-          <span className="text-[13px] font-medium tracking-nav text-nav-fg">
-            Chat Template
-          </span>
+      <div className="-mb-1.5 flex items-center justify-between gap-2">
+        <button
+          type="button"
+          onClick={openEditor}
+          className="cursor-pointer text-left text-[13px] font-medium tracking-nav text-nav-fg"
+        >
+          Chat Template
+        </button>
+        <div className="flex items-center gap-1">
           {isModified && (
             <Tooltip>
               <TooltipPrimitive.Trigger asChild>
@@ -1589,21 +1746,34 @@ function ChatTemplateFields() {
               </TooltipContent>
             </Tooltip>
           )}
+          <Tooltip>
+            <TooltipPrimitive.Trigger asChild>
+              <button
+                type="button"
+                onClick={openEditor}
+                className="nav-icon-btn text-nav-icon-idle hover:bg-panel-surface-hover hover:text-black dark:hover:text-white"
+                aria-label="Edit chat template"
+              >
+                <HugeiconsIcon
+                  icon={Edit03Icon}
+                  strokeWidth={1.75}
+                  className="size-3"
+                />
+              </button>
+            </TooltipPrimitive.Trigger>
+            <TooltipContent
+              side="top"
+              sideOffset={6}
+              className="tooltip-compact"
+            >
+              Edit template
+            </TooltipContent>
+          </Tooltip>
         </div>
-        <button
-          type="button"
-          onClick={openEditor}
-          aria-label="Edit chat template"
-          className="panel-text-surface mt-1 flex w-full h-20 overflow-hidden cursor-pointer items-start px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed text-nav-fg corner-squircle focus-visible:outline-none focus-visible:border-ring focus-visible:ring-[1px] focus-visible:ring-ring/40"
-        >
-          <span className="block line-clamp-3 whitespace-pre-wrap break-words">
-            {displayValue}
-          </span>
-        </button>
       </div>
       <Dialog open={editorOpen} onOpenChange={setEditorOpen}>
         <DialogContent
-          className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-3xl"
+          className="corner-squircle dialog-soft-surface sm:max-w-3xl"
           overlayClassName="bg-background/35 supports-backdrop-filter:backdrop-blur-[1px]"
         >
           <DialogHeader>
@@ -1624,7 +1794,7 @@ function ChatTemplateFields() {
               value={draft}
               onChange={(event) => setDraft(event.target.value)}
               fieldSizing="fixed"
-              className="min-h-[24rem] max-h-[50vh] overflow-y-auto font-mono text-xs leading-5 corner-squircle focus-visible:border-input focus-visible:ring-0"
+              className="min-h-[24rem] max-h-[50vh] overflow-y-auto border-0 font-mono text-xs leading-5 corner-squircle focus-visible:ring-0"
               rows={14}
               spellCheck={false}
             />
@@ -1633,13 +1803,24 @@ function ChatTemplateFields() {
             <Button
               type="button"
               variant="ghost"
-              onClick={() => setEditorOpen(false)}
+              onClick={() => setDraft(defaultTemplate)}
+              disabled={draft === defaultTemplate}
+              className="text-muted-foreground"
             >
-              Cancel
-            </Button>
-            <Button type="button" onClick={saveEditor} disabled={!draftDirty}>
-              Save
+              Reset
             </Button>
+            <div className="flex flex-wrap gap-2">
+              <Button
+                type="button"
+                variant="ghost"
+                onClick={() => setEditorOpen(false)}
+              >
+                Cancel
+              </Button>
+              <Button type="button" onClick={saveEditor} disabled={!draftDirty}>
+                Save
+              </Button>
+            </div>
           </DialogFooter>
         </DialogContent>
       </Dialog>
diff --git a/studio/frontend/src/features/chat/components/chat-search-dialog.tsx b/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
index 3d867b07f7..b0fd3c5f27 100644
--- a/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
+++ b/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
@@ -64,10 +64,10 @@ export function ChatSearchDialog() {
     <CommandDialog
       open={isOpen}
       onOpenChange={setOpen}
-      className="chat-search-surface corner-squircle top-[25%] w-[635px] max-w-[calc(100%-2rem)] gap-0 p-0 sm:max-w-[635px]"
+      className="chat-search-surface rounded-3xl! top-1/2 -translate-y-1/2 w-[635px] max-w-[calc(100%-2rem)] gap-0 p-0 ring-0 sm:max-w-[635px]"
       overlayClassName="bg-transparent"
     >
-      <Command className="rounded-4xl p-0" filter={chatSearchFilter}>
+      <Command className="rounded-3xl p-0" filter={chatSearchFilter}>
         <div className="flex items-center gap-3 border-b border-border/40 px-4 py-3">
           <HugeiconsIcon
             icon={SearchIcon}
@@ -117,7 +117,7 @@ export function ChatSearchDialog() {
                   });
                   close();
                 }}
-                className="relative flex cursor-pointer select-none items-center gap-3 rounded-lg px-3 py-2.5 text-sm outline-hidden data-selected:bg-muted data-selected:text-foreground"
+                className="relative flex cursor-pointer select-none items-center gap-3 rounded-full px-3 py-2.5 text-sm outline-hidden data-selected:bg-muted data-selected:text-foreground"
               >
                 <HugeiconsIcon
                   icon={Message01Icon}
diff --git a/studio/frontend/src/features/chat/components/new-project-dialog.tsx b/studio/frontend/src/features/chat/components/new-project-dialog.tsx
index 38b0983e10..880129ac6c 100644
--- a/studio/frontend/src/features/chat/components/new-project-dialog.tsx
+++ b/studio/frontend/src/features/chat/components/new-project-dialog.tsx
@@ -56,7 +56,7 @@ export function NewProjectDialog({
         onOpenChange(next);
       }}
     >
-      <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+      <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
         <DialogHeader>
           <DialogTitle>New project</DialogTitle>
         </DialogHeader>
diff --git a/studio/frontend/src/features/chat/components/project-switcher.tsx b/studio/frontend/src/features/chat/components/project-switcher.tsx
index 3deaf4701d..08e252a2c2 100644
--- a/studio/frontend/src/features/chat/components/project-switcher.tsx
+++ b/studio/frontend/src/features/chat/components/project-switcher.tsx
@@ -71,7 +71,7 @@ export function ProjectSwitcher({
       <DropdownMenuContent
         side="bottom"
         align="start"
-        sideOffset={6}
+        sideOffset={0}
         className="app-user-menu menu-soft-surface ring-0 min-w-56 max-w-72 max-h-72 py-2 font-heading rounded-[14px] border-0"
       >
         {showLoadingRow ? (
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index a152cbb358..4bd36e00d4 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -142,6 +142,50 @@ function toChatModelSummary(model: {
   };
 }
 
+// Merge capability flags from a load/status response into the matching
+// models[] entry. /api/models/list omits audio capability for default and
+// active-GGUF entries, so the attach gates (`activeModel?.hasAudioInput`)
+// would otherwise stay false. Mirrors the compare composer's sync.
+// Exported for tests.
+export function syncModelCapabilities(
+  modelId: string,
+  resp: {
+    display_name?: string | null;
+    is_vision?: boolean;
+    is_lora?: boolean;
+    is_gguf?: boolean;
+    is_audio?: boolean;
+    audio_type?: string | null;
+    has_audio_input?: boolean;
+  },
+): void {
+  const store = useChatRuntimeStore.getState();
+  const models = store.models;
+  const synced = {
+    isVision: Boolean(resp.is_vision),
+    isGguf: Boolean(resp.is_gguf),
+    isAudio: Boolean(resp.is_audio),
+    audioType: resp.audio_type ?? null,
+    hasAudioInput: Boolean(resp.has_audio_input),
+  };
+  const idx = models.findIndex((m) => m.id === modelId);
+  if (idx === -1) {
+    store.setModels([
+      ...models,
+      {
+        id: modelId,
+        name: resp.display_name || modelId,
+        isLora: Boolean(resp.is_lora),
+        ...synced,
+      },
+    ]);
+  } else {
+    const next = [...models];
+    next[idx] = { ...next[idx], ...synced };
+    store.setModels(next);
+  }
+}
+
 function toLoraSummary(lora: {
   display_name: string;
   adapter_path: string;
@@ -368,6 +412,7 @@ export function useChatModelRuntime() {
             statusRes.requires_trust_remote_code ?? false,
           defaultChatTemplate: nextDefaultChatTemplate,
           loadedIsMultimodal: isMultimodalResponse(statusRes),
+          specFallbackReason: statusRes.spec_fallback_reason ?? null,
           ...(prevState.loadedSpeculativeType === null && {
             speculativeType: currentSpecType,
             loadedSpeculativeType: currentSpecType,
@@ -390,6 +435,9 @@ export function useChatModelRuntime() {
               loadedChatTemplateOverride: statusRes.chat_template_override,
             }),
         });
+        // setModels(listRes...) above used catalog data, which omits audio
+        // capability. Re-apply live status so attach gates survive a refresh.
+        syncModelCapabilities(statusRes.active_model, statusRes);
 
         // Set reasoning default for Qwen3.5/3.6 small models
         if (
@@ -736,6 +784,8 @@ export function useChatModelRuntime() {
               loadedIsMultimodal: isMultimodalResponse(loadResponse),
               activeNativePathToken: nativePathToken ?? null,
             });
+            // Unlock attach menus for capabilities the catalog entry lacked.
+            syncModelCapabilities(modelId, loadResponse);
             // Qwen3/3.5/3.6: apply thinking-mode-specific params after load
             if (
               modelId.toLowerCase().includes("qwen3") &&
diff --git a/studio/frontend/src/features/chat/hooks/use-rag-tool-available.ts b/studio/frontend/src/features/chat/hooks/use-rag-tool-disabled.ts
similarity index 60%
rename from studio/frontend/src/features/chat/hooks/use-rag-tool-available.ts
rename to studio/frontend/src/features/chat/hooks/use-rag-tool-disabled.ts
index 4463aa3cd4..4f0912e435 100644
--- a/studio/frontend/src/features/chat/hooks/use-rag-tool-available.ts
+++ b/studio/frontend/src/features/chat/hooks/use-rag-tool-disabled.ts
@@ -4,15 +4,16 @@
 import { parseExternalModelId } from "../external-providers";
 import { useChatRuntimeStore } from "../stores/chat-runtime-store";
 
-// Single source of truth for the RAG pill's disabled gate and the Add Files bar's
-// visibility so the bar never shows while the pill is inert.
-export function useRagToolAvailable(): boolean {
+// Pre-select gate for the RAG toggle, mirroring Web search/Code/MCP: armable
+// with no model; disabled only when a loaded model can't run
+// search_knowledge_base. The send path checks supportsTools independently.
+export function useRagToolDisabled(): boolean {
   const modelLoaded = useChatRuntimeStore(
     (s) => !!s.params.checkpoint && !s.modelLoading,
   );
   const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
   const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
   return (
-    modelLoaded && parseExternalModelId(checkpoint) === null && supportsTools
+    modelLoaded && (parseExternalModelId(checkpoint) !== null || !supportsTools)
   );
 }
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index ac05c8c089..dde3a75475 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -19,6 +19,7 @@ export {
 export { useChatRuntimeStore } from "./stores/chat-runtime-store";
 export { useChatSearchStore } from "./stores/chat-search-store";
 export { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
+export { isExternalModelId } from "./external-providers";
 export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
 export type { ProjectRecord } from "./types";
@@ -29,6 +30,11 @@ export {
   useSelectedChatArtifact,
 } from "./artifacts/store";
 export { downloadChatExport } from "./utils/export-chat-history";
+export {
+  EXPORT_FORMATS_LIST,
+  bulkExportConversationsByScope,
+  importConversationsFromFile,
+} from "./prompt-storage/prompt-storage-dialog";
 export {
   deleteChatItem,
   renameChatItem,
diff --git a/studio/frontend/src/features/chat/mcp-composer-button.tsx b/studio/frontend/src/features/chat/mcp-composer-button.tsx
index 348c744b2c..aecfe9b430 100644
--- a/studio/frontend/src/features/chat/mcp-composer-button.tsx
+++ b/studio/frontend/src/features/chat/mcp-composer-button.tsx
@@ -3,6 +3,7 @@
 
 import { McpServerIcon, Tick02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
+import { XIcon } from "lucide-react";
 import { type FC, useCallback, useEffect, useState } from "react";
 import { toast } from "sonner";
 
@@ -245,14 +246,36 @@ export function McpComposerButton({
             <button
               type="button"
               className="composer-pill-btn"
+              data-pill-label="MCP"
               data-active={active ? "true" : "false"}
               aria-label="MCP servers"
             >
-              <HugeiconsIcon
-                icon={McpServerIcon}
-                className="size-[15px]"
-                strokeWidth={2}
-              />
+              {/* Icon doubles as an off switch: hover swaps to an X; clicking
+                  it turns MCP off without opening the menu. In compact
+                  icon-only mode the glyph is the whole button, so clicks fall
+                  through to the trigger and open the menu instead. */}
+              <span
+                role="button"
+                aria-label="Turn off MCP"
+                tabIndex={-1}
+                onPointerDown={(e) => {
+                  if (e.currentTarget.closest('[data-pill-compact="true"]')) return;
+                  e.stopPropagation();
+                }}
+                onClick={(e) => {
+                  if (e.currentTarget.closest('[data-pill-compact="true"]')) return;
+                  e.stopPropagation();
+                  setMcpEnabledForChat(false);
+                }}
+                className="composer-pill-glyph cursor-pointer"
+              >
+                <HugeiconsIcon
+                  icon={McpServerIcon}
+                  className="size-[15px]"
+                  strokeWidth={2}
+                />
+                <XIcon className="composer-pill-x" />
+              </span>
               <span>MCP</span>
               <ArrowDownStandardIcon className="composer-pill-caret size-[15px]" />
             </button>
@@ -260,7 +283,7 @@ export function McpComposerButton({
           <DropdownMenuContent
             side={side}
             align="start"
-            sideOffset={2}
+            sideOffset={0}
             avoidCollisions={true}
             className="unsloth-plus-menu mcp-menu w-[232px]"
           >
diff --git a/studio/frontend/src/features/chat/projects-page.tsx b/studio/frontend/src/features/chat/projects-page.tsx
index c4cd911fdb..724cd0a755 100644
--- a/studio/frontend/src/features/chat/projects-page.tsx
+++ b/studio/frontend/src/features/chat/projects-page.tsx
@@ -258,7 +258,7 @@ export function ProjectsPage() {
               value={sortMode}
               onValueChange={(v) => setSortMode(v as SortMode)}
             >
-              <SelectTrigger className="h-9 w-[130px]">
+              <SelectTrigger className="h-9 w-[130px] rounded-full border-none bg-muted shadow-none dark:bg-card">
                 <SelectValue />
               </SelectTrigger>
               <SelectContent>
@@ -269,7 +269,12 @@ export function ProjectsPage() {
           </div>
           <DropdownMenu>
             <DropdownMenuTrigger asChild>
-              <Button variant="outline" size="icon" title="Import / Export projects">
+              <Button
+                variant="outline"
+                size="icon"
+                title="Import / Export projects"
+                className="rounded-full border-none bg-muted shadow-none dark:bg-card"
+              >
                 <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} className="size-icon" />
               </Button>
             </DropdownMenuTrigger>
@@ -324,15 +329,15 @@ export function ProjectsPage() {
         </div>
       </div>
 
-      <div className="relative mt-6">
-        <span className="pointer-events-none absolute left-3 top-1/2 -translate-y-1/2 text-muted-foreground">
+      <div className="relative mx-auto mt-10 w-full max-w-[720px]">
+        <span className="pointer-events-none absolute left-5 top-1/2 -translate-y-1/2 text-muted-foreground">
           <HugeiconsIcon icon={Search01Icon} strokeWidth={1.75} className="size-icon" />
         </span>
         <Input
           value={query}
           onChange={(e) => setQuery(e.target.value)}
           placeholder="Search projects..."
-          className="h-11 pl-10"
+          className="h-14 rounded-full border-none bg-background pl-13 pr-5 shadow-[0_0_20px_0_rgba(0,0,0,0.07)] dark:bg-card dark:shadow-none"
           aria-label="Search projects"
         />
       </div>
@@ -342,7 +347,7 @@ export function ProjectsPage() {
           {Array.from({ length: 6 }).map((_, index) => (
             <div
               key={index}
-              className="min-h-[160px] rounded-[14px] border border-border/70 bg-card p-5"
+              className="min-h-[160px] rounded-[14px] bg-card p-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
             >
               <Skeleton className="h-5 w-2/3 rounded-[6px]" />
               <Skeleton className="mt-3 h-4 w-full rounded-[6px]" />
@@ -361,7 +366,7 @@ export function ProjectsPage() {
           {projects.length === 0 && (
             <Button
               variant="outline"
-              className="mt-2"
+              className="mt-2 border-none bg-background shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-none"
               onClick={() => {
                 setNameDraft("");
                 setCreating(true);
@@ -402,7 +407,7 @@ export function ProjectsPage() {
                   openProject(project.id);
                 }
               }}
-              className="group/project-card relative flex min-h-[160px] cursor-pointer flex-col rounded-[14px] border border-border/70 bg-card p-5 text-left transition-colors hover:border-border hover:bg-accent/40 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              className="group/project-card relative flex min-h-[160px] cursor-pointer flex-col rounded-[14px] bg-card p-5 text-left shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-accent/40 dark:shadow-none focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
             >
               <div className="flex items-start justify-between gap-2">
                 <h2 className="truncate pr-2 text-[16px] font-semibold text-foreground">
@@ -422,7 +427,7 @@ export function ProjectsPage() {
                   <DropdownMenuContent
                     side="bottom"
                     align="end"
-                    sideOffset={4}
+                    sideOffset={0}
                     onClick={(e) => e.stopPropagation()}
                     onKeyDown={(e) => e.stopPropagation()}
                     className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
@@ -496,7 +501,7 @@ export function ProjectsPage() {
           if (!open) setCreating(false);
         }}
       >
-        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
           <DialogHeader>
             <DialogTitle>New project</DialogTitle>
           </DialogHeader>
@@ -533,7 +538,7 @@ export function ProjectsPage() {
           if (!open) setRenaming(null);
         }}
       >
-        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
           <DialogHeader>
             <DialogTitle>Rename project</DialogTitle>
           </DialogHeader>
@@ -569,7 +574,7 @@ export function ProjectsPage() {
 
       {/* Import destination picker */}
       <Dialog open={importFile !== null} onOpenChange={(open) => { if (!open) setImportFile(null); }}>
-        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
           <DialogHeader>
             <DialogTitle>Import chats</DialogTitle>
           </DialogHeader>
@@ -604,7 +609,7 @@ export function ProjectsPage() {
           if (!open) setDeleting(null);
         }}
       >
-        <DialogContent className="menu-flat-destructive corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+        <DialogContent className="menu-flat-destructive corner-squircle dialog-soft-surface sm:max-w-md">
           <DialogHeader>
             <DialogTitle>Delete project</DialogTitle>
           </DialogHeader>
diff --git a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
index 80c2de54a2..e62f96720d 100644
--- a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
+++ b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
@@ -48,6 +48,7 @@ import {
 } from "../api/prompts-api";
 import {
   listStoredChatMessages,
+  listStoredChatThreads,
   saveStoredChatThread,
   syncStoredChatMessages,
 } from "../utils/chat-history-storage";
@@ -331,7 +332,7 @@ function messageToOpenAI(msg: { role: unknown; content: unknown; attachments?: u
   return contentParts.length > 0 ? [{ role: role as "user" | "system", content: contentParts }] : [];
 }
 
-// ShareGPT training JSONL (human/gpt turns).
+// ShareGPT training JSONL (human/system/gpt turns).
 export async function exportConversationShareGPT(threadId: string): Promise<void> {
   const messages = await loadConversationMessages(threadId);
   if (!messages) return;
@@ -339,7 +340,7 @@ export async function exportConversationShareGPT(threadId: string): Promise<void
   const conversations: Array<{ from: string; value: string }> = [];
   for (const msg of messages) {
     const role = msg.role as string;
-    const from = role === "user" ? "human" : "gpt";
+    const from = role === "user" ? "human" : role === "system" ? "system" : "gpt";
     const value = messageToText(msg);
     if (value.trim()) conversations.push({ from, value });
   }
@@ -413,7 +414,7 @@ async function buildThreadContent(
     for (const msg of messages) {
       const role = msg.role as string;
       const value = messageToText(msg);
-      if (value.trim()) conversations.push({ from: role === "user" ? "human" : "gpt", value });
+      if (value.trim()) conversations.push({ from: role === "user" ? "human" : role === "system" ? "system" : "gpt", value });
     }
     if (conversations.length === 0) return null;
     return JSON.stringify({ conversations });
@@ -493,6 +494,36 @@ export async function exportBulkConversationsSeparate(
   );
 }
 
+// Scope-level bulk export shared by the sidebar Recents menu and
+// Settings -> Chat -> Data. "recents" = chats outside projects; "all" adds
+// project chats.
+export async function bulkExportConversationsByScope(
+  scope: "recents" | "all",
+  format: ConvExportFormat,
+  merged: boolean,
+): Promise<void> {
+  try {
+    const threads = await listStoredChatThreads({
+      includeArchived: false,
+      ...(scope === "recents" ? { projectId: null } : {}),
+    });
+    const ids = [...new Set(threads.map((t) => t.id))];
+    if (ids.length === 0) {
+      toast.info("No conversations to export.");
+      return;
+    }
+    const ts = new Date().toISOString().slice(0, 10);
+    const basename = scope === "all" ? `all-chats-${ts}` : `recents-${ts}`;
+    if (merged) {
+      await exportBulkConversationsMerged(ids, format, basename);
+    } else {
+      await exportBulkConversationsSeparate(ids, format, basename);
+    }
+  } catch {
+    toast.error("Export failed.");
+  }
+}
+
 export async function exportProjectConversations(
   threadIds: string[],
   format: ConvExportFormat,
@@ -626,30 +657,16 @@ function sharegptToRecords(
 }
 
 function csvToRecords(csvText: string, threadId: string, baseTs: number): MessageRecord[] {
-  const lines = csvText.split(/\r?\n/);
+  // parseCsv handles quoted newlines, so multi-line message content
+  // round-trips; a naive per-line split would break those records.
+  const rows = parseCsv(csvText).slice(1);
   const records: MessageRecord[] = [];
   let prevId: string | null = null;
   let idx = 0;
-  for (let i = 1; i < lines.length; i++) {
-    const line = lines[i].trim();
-    if (!line) continue;
-    let role = "";
-    let content = "";
-    if (line.startsWith('"')) {
-      const m = line.match(/^"((?:[^"]|"")*)"\s*,\s*([\s\S]*)$/);
-      if (!m) continue;
-      role = m[1].replace(/""/g, '"');
-      content = m[2].startsWith('"')
-        ? m[2].slice(1, m[2].endsWith('"') ? -1 : undefined).replace(/""/g, '"')
-        : m[2];
-    } else {
-      const comma = line.indexOf(",");
-      if (comma === -1) continue;
-      role = line.slice(0, comma);
-      content = line.slice(comma + 1).startsWith('"')
-        ? line.slice(comma + 2, line.endsWith('"') ? -1 : undefined).replace(/""/g, '"')
-        : line.slice(comma + 1);
-    }
+  for (const row of rows) {
+    if (row.length < 2) continue;
+    const role = row[0].trim();
+    const content = row.slice(1).join(",");
     if (!content.trim()) continue;
     const validRole = role === "user" || role === "assistant" || role === "system" ? role : "user";
     const id = crypto.randomUUID();
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 8d9e5a10e3..359301e182 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -51,6 +51,7 @@ import {
   readActiveOpenDocumentAttachmentContent,
   readOpenDocumentAttachmentContent,
 } from "./open-document";
+import { AudioAttachmentAdapter } from "./audio-attachment-adapter";
 import { useChatRuntimeStore } from "./stores/chat-runtime-store";
 import type { MessageRecord, ModelType, ThreadRecord } from "./types";
 import {
@@ -983,6 +984,7 @@ function useStudioRuntimeAdapters(
     () =>
       new CompositeAttachmentAdapter([
         new VisionImageAdapter(),
+        new AudioAttachmentAdapter(),
         new TextAttachmentAdapter(),
         new HtmlAttachmentAdapter(),
         new PDFAttachmentAdapter(),
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index 6919d5e7a5..74ab4a1b9e 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -656,8 +656,10 @@ export function SharedComposer({
   // Fetch pill: Anthropic-only (web_fetch_20250910 / web_fetch_20260209).
   const webFetchDisabled = !modelLoaded || !supportsBuiltinWebFetch;
   const showWebFetchPill = supportsBuiltinWebFetch;
-  // Docs (RAG) pill is local-only: search_knowledge_base needs the local runtime.
-  const ragDisabled = !modelLoaded || isExternalModel || !supportsTools;
+  // Docs (RAG) is local-only: search_knowledge_base needs the local runtime.
+  // Disable only when a loaded model can't run it; with no model the toggle
+  // can still be pre-selected, matching Web search/Code/MCP.
+  const ragDisabled = modelLoaded && (isExternalModel || !supportsTools);
   const showRagPill = !isExternalModel;
   // Above 4 pills, collapse to icons only to cut clutter. Compare, Search and
   // Code always show; the rest are conditional.
@@ -1270,7 +1272,7 @@ export function SharedComposer({
             <DropdownMenuContent
               side="top"
               align="start"
-              sideOffset={2}
+              sideOffset={0}
               avoidCollisions={true}
               className="unsloth-plus-menu w-[212px]"
               onCloseAutoFocus={(event) => event.preventDefault()}
@@ -1323,10 +1325,12 @@ export function SharedComposer({
                 }
                 onSelect={() => setCodeToolsEnabled(!codeToolsEnabled)}
               >
+                {/* Scale, not width: an oversized box pushed the label out of
+                    line. */}
                 <HugeiconsIcon
                   icon={CodeIcon}
                   strokeWidth={2}
-                  className="size-[1.175rem]!"
+                  className="scale-[1.12]"
                 />
                 Code
                 {codeToolsEnabled && !codeDisabled ? (
@@ -1369,7 +1373,7 @@ export function SharedComposer({
                 onSelect={() => setRagEnabled(!ragEnabled)}
               >
                 <HugeiconsIcon icon={FileDatabaseIcon} strokeWidth={2} />
-                RAG
+                Chat with Files
                 {ragEnabled && !ragDisabled ? (
                   <HugeiconsIcon
                     icon={Tick02Icon}
@@ -1396,88 +1400,95 @@ export function SharedComposer({
                 ) : null}
               </DropdownMenuItem>
               {/* RAG hidden temporarily */}
-              {/* Always active: this menu only renders in compare mode. Ticked
-                  like Web search/Code; click toggles it off. */}
-              <DropdownMenuItem
-                className="text-primary font-medium"
-                onSelect={handleExitCompare}
-              >
-                <Columns2Icon />
-                Compare chat
-                <HugeiconsIcon
-                  icon={Tick02Icon}
-                  strokeWidth={2}
-                  className="ml-auto"
-                />
-              </DropdownMenuItem>
-              <DropdownMenuSub>
-                <DropdownMenuSubTrigger>
-                  <HugeiconsIcon icon={Bookmark02Icon} strokeWidth={2} />
-                  Saved prompts
-                </DropdownMenuSubTrigger>
-                <DropdownMenuSubContent className="unsloth-plus-menu w-[176px]">
-                  {recentPrompts.map((p) => (
-                    <DropdownMenuItem
-                      key={p.id}
-                      onSelect={() => {
-                        setText(p.text);
-                        requestAnimationFrame(() =>
-                          textareaRef.current?.focus(),
-                        );
-                      }}
-                    >
-                      <span className="truncate">{p.name}</span>
-                    </DropdownMenuItem>
-                  ))}
-                  {recentPrompts.length > 0 ? <DropdownMenuSeparator /> : null}
-                  <DropdownMenuItem onSelect={() => setPromptStorageOpen(true)}>
-                    All saved prompts…
-                  </DropdownMenuItem>
-                </DropdownMenuSubContent>
-              </DropdownMenuSub>
-              {/* Top-level: a third-level submenu collision-flips at narrow
-                  widths and is awkward to reach. */}
-              <DropdownMenuSub>
-                <DropdownMenuSubTrigger disabled={exportThreadIds.length === 0}>
-                  <HugeiconsIcon icon={Download01Icon} strokeWidth={2} />
-                  Export chat
-                </DropdownMenuSubTrigger>
-                <DropdownMenuSubContent
-                  collisionPadding={16}
-                  className="unsloth-plus-menu w-[176px]"
-                >
-                  {[
-                    { label: "Raw JSONL", fn: exportConversationRawJsonl },
-                    { label: "CSV", fn: exportConversationCsv },
-                    {
-                      label: "ShareGPT JSONL",
-                      fn: exportConversationShareGPT,
-                    },
-                  ].map(({ label, fn }) => (
-                    <DropdownMenuItem
-                      key={label}
-                      disabled={exportThreadIds.length === 0}
-                      onSelect={() => {
-                        if (!exportThreadIds.length) {
-                          toast.error("No conversation to export yet.");
-                          return;
-                        }
-                        Promise.all(exportThreadIds.map((id) => fn(id))).catch(
-                          () => toast.error("Export failed."),
-                        );
-                      }}
-                    >
-                      {label}
-                    </DropdownMenuItem>
-                  ))}
-                </DropdownMenuSubContent>
-              </DropdownMenuSub>
               <DropdownMenuSub>
                 <DropdownMenuSubTrigger>
                   <MoreHorizontalIcon className="size-4" />
                   More
                 </DropdownMenuSubTrigger>
                 <DropdownMenuSubContent className="unsloth-plus-menu w-[200px]">
+                  {/* Always active: this menu only renders in compare mode. Ticked
+                      like Web search/Code; click toggles it off. */}
+                  <DropdownMenuItem
+                    className="text-primary font-medium"
+                    onSelect={handleExitCompare}
+                  >
+                    <Columns2Icon />
+                    Compare chat
+                    <HugeiconsIcon
+                      icon={Tick02Icon}
+                      strokeWidth={2}
+                      className="ml-auto"
+                    />
+                  </DropdownMenuItem>
+                  <DropdownMenuSub>
+                    <DropdownMenuSubTrigger>
+                      <HugeiconsIcon icon={Bookmark02Icon} strokeWidth={2} />
+                      Saved prompts
+                    </DropdownMenuSubTrigger>
+                    <DropdownMenuSubContent
+                      collisionPadding={16}
+                      className="unsloth-plus-menu w-[176px]"
+                    >
+                      {recentPrompts.map((p) => (
+                        <DropdownMenuItem
+                          key={p.id}
+                          onSelect={() => {
+                            setText(p.text);
+                            requestAnimationFrame(() =>
+                              textareaRef.current?.focus(),
+                            );
+                          }}
+                        >
+                          <span className="truncate">{p.name}</span>
+                        </DropdownMenuItem>
+                      ))}
+                      {recentPrompts.length > 0 ? (
+                        <DropdownMenuSeparator />
+                      ) : null}
+                      <DropdownMenuItem
+                        onSelect={() => setPromptStorageOpen(true)}
+                      >
+                        All saved prompts…
+                      </DropdownMenuItem>
+                    </DropdownMenuSubContent>
+                  </DropdownMenuSub>
+                  <DropdownMenuSub>
+                    <DropdownMenuSubTrigger
+                      disabled={exportThreadIds.length === 0}
+                    >
+                      <HugeiconsIcon icon={Download01Icon} strokeWidth={2} />
+                      Export chat
+                    </DropdownMenuSubTrigger>
+                    <DropdownMenuSubContent
+                      collisionPadding={16}
+                      className="unsloth-plus-menu w-[176px]"
+                    >
+                      {[
+                        { label: "Raw JSONL", fn: exportConversationRawJsonl },
+                        { label: "CSV", fn: exportConversationCsv },
+                        {
+                          label: "ShareGPT JSONL",
+                          fn: exportConversationShareGPT,
+                        },
+                      ].map(({ label, fn }) => (
+                        <DropdownMenuItem
+                          key={label}
+                          disabled={exportThreadIds.length === 0}
+                          onSelect={() => {
+                            if (!exportThreadIds.length) {
+                              toast.error("No conversation to export yet.");
+                              return;
+                            }
+                            Promise.all(
+                              exportThreadIds.map((id) => fn(id)),
+                            ).catch(() => toast.error("Export failed."));
+                          }}
+                        >
+                          {label}
+                        </DropdownMenuItem>
+                      ))}
+                    </DropdownMenuSubContent>
+                  </DropdownMenuSub>
                   <DropdownMenuItem
                     className={
                       artifactsEnabled ? "text-primary font-medium" : undefined
@@ -1556,6 +1567,7 @@ export function SharedComposer({
               }
             }}
             className="composer-pill-btn"
+            data-pill-label="Search"
             data-active={toolsEnabled && !searchDisabled ? "true" : "false"}
             aria-label={
               toolsEnabled ? "Disable web search" : "Enable web search"
@@ -1571,6 +1583,7 @@ export function SharedComposer({
             disabled={codeDisabled}
             onClick={() => setCodeToolsEnabled(!codeToolsEnabled)}
             className="composer-pill-btn"
+            data-pill-label="Code"
             data-active={codeToolsEnabled && !codeDisabled ? "true" : "false"}
             aria-label={
               codeToolsEnabled
@@ -1593,6 +1606,7 @@ export function SharedComposer({
               disabled={imageDisabled}
               onClick={() => setImageToolsEnabled(!imageToolsEnabled)}
               className="composer-pill-btn"
+              data-pill-label="Images"
               data-active={
                 imageToolsEnabled && !imageDisabled ? "true" : "false"
               }
@@ -1619,6 +1633,7 @@ export function SharedComposer({
               disabled={webFetchDisabled}
               onClick={() => setWebFetchToolsEnabled(!webFetchToolsEnabled)}
               className="composer-pill-btn"
+              data-pill-label="Fetch"
               data-active={
                 webFetchToolsEnabled && !webFetchDisabled ? "true" : "false"
               }
@@ -1637,6 +1652,7 @@ export function SharedComposer({
               type="button"
               onClick={() => setArtifactsEnabled(false)}
               className="composer-pill-btn"
+              data-pill-label="Canvas"
               data-active="true"
               aria-label="Disable canvas"
             >
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index f9fb2abe2b..3ef63d2945 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -36,7 +36,6 @@ export const CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY =
 export const CHAT_MCP_ENABLED_KEY = "unsloth_chat_mcp_enabled";
 export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
-export const CHAT_RAG_ENABLED_KEY = "unsloth_chat_rag_enabled";
 export const CHAT_RAG_SOURCE_KEY = "unsloth_chat_rag_source";
 export const CHAT_RAG_MODE_KEY = "unsloth_chat_rag_mode";
 export const CHAT_RAG_TOP_K_KEY = "unsloth_chat_rag_top_k";
@@ -417,6 +416,11 @@ type ChatRuntimeStore = {
   loadedKvCacheDtype: string | null;
   speculativeType: string | null;
   loadedSpeculativeType: string | null;
+  /**
+   * Why MTP was disabled on the loaded model despite being requested, or null.
+   * Mirrors InferenceStatusResponse.spec_fallback_reason.
+   */
+  specFallbackReason: string | null;
   /** User --spec-draft-n-max override (null = platform default). */
   specDraftNMax: number | null;
   loadedSpecDraftNMax: number | null;
@@ -480,7 +484,7 @@ type ChatRuntimeStore = {
   setAllowArtifactNetworkAccess: (enabled: boolean) => void;
   setMcpEnabledForChat: (enabled: boolean) => void;
   setWebFetchToolsEnabled: (enabled: boolean) => void;
-  setRagEnabled: (enabled: boolean, options?: { persist?: boolean }) => void;
+  setRagEnabled: (enabled: boolean) => void;
   setRagSource: (source: RagSource) => void;
   setRagMode: (mode: RagMode) => void;
   setRagTopK: (topK: number) => void;
@@ -746,7 +750,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   ),
   mcpEnabledForChat: loadBool(CHAT_MCP_ENABLED_KEY, false),
   webFetchToolsEnabled: loadBool(CHAT_WEB_FETCH_TOOLS_ENABLED_KEY, false),
-  ragEnabled: loadBool(CHAT_RAG_ENABLED_KEY, false),
+  // RAG is opt-in per session: always starts off, never restored from storage.
+  ragEnabled: false,
   ragSource: loadRagSource(),
   ragMode: loadRagMode(),
   ragTopK: loadRagTopK(),
@@ -765,6 +770,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   loadedKvCacheDtype: null,
   speculativeType: "auto",
   loadedSpeculativeType: null,
+  specFallbackReason: null,
   specDraftNMax: null,
   loadedSpecDraftNMax: null,
   loadedIsMultimodal: false,
@@ -977,6 +983,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       loadedKvCacheDtype: null,
       speculativeType: "auto",
       loadedSpeculativeType: null,
+      specFallbackReason: null,
       specDraftNMax: null,
       loadedSpecDraftNMax: null,
       loadedIsMultimodal: false,
@@ -1072,13 +1079,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       saveBool(CHAT_WEB_FETCH_TOOLS_ENABLED_KEY, webFetchToolsEnabled);
       return { webFetchToolsEnabled };
     }),
-  setRagEnabled: (ragEnabled, options) =>
-    set(() => {
-      if (options?.persist !== false) {
-        saveBool(CHAT_RAG_ENABLED_KEY, ragEnabled);
-      }
-      return { ragEnabled };
-    }),
+  setRagEnabled: (ragEnabled) => set(() => ({ ragEnabled })),
   setRagSource: (ragSource) =>
     set(() => {
       saveRagSource(ragSource);
diff --git a/studio/frontend/src/features/chat/thread-sidebar.tsx b/studio/frontend/src/features/chat/thread-sidebar.tsx
index 2b255c6c14..0d35facb2a 100644
--- a/studio/frontend/src/features/chat/thread-sidebar.tsx
+++ b/studio/frontend/src/features/chat/thread-sidebar.tsx
@@ -331,7 +331,7 @@ export function ThreadSidebar({
 
       {/* Rename dialog */}
       <Dialog open={renamingItem !== null} onOpenChange={(open) => { if (!open) setRenamingItem(null); }}>
-        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-sm">
+        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-sm">
           <DialogHeader>
             <DialogTitle>Rename chat</DialogTitle>
           </DialogHeader>
diff --git a/studio/frontend/src/features/chat/tour/steps.tsx b/studio/frontend/src/features/chat/tour/steps.tsx
index de27e93c88..cbbb7bae4f 100644
--- a/studio/frontend/src/features/chat/tour/steps.tsx
+++ b/studio/frontend/src/features/chat/tour/steps.tsx
@@ -28,7 +28,7 @@ export function buildChatTourSteps({
       body: (
         <>
           This selects what’s loaded for inference. Hub = base models. Fine-tuned
-          = trained Studio outputs, including LoRA adapters and full finetunes.
+          = trained Unsloth outputs, including LoRA adapters and full finetunes.
         </>
       ),
     },
@@ -38,7 +38,7 @@ export function buildChatTourSteps({
       title: "Two tabs",
       body: (
         <>
-          Hub: search Hugging Face models. Fine-tuned: local Studio outputs you’ve
+          Hub: search Hugging Face models. Fine-tuned: local Unsloth outputs you’ve
           trained or exported. If results look off, compare base vs fine-tuned
           outputs to see what changed.
         </>
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 49524c17b4..a0e5d5355e 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -174,6 +174,12 @@ export interface InferenceStatusResponse {
   /** Canonical UI-facing mode currently active. See LoadModelRequest. */
   speculative_type?: string | null;
   spec_draft_n_max?: number | null;
+  /**
+   * Why MTP was disabled on the loaded model despite being requested.
+   * "binary_no_mtp" / "binary_outdated" -> updating llama.cpp would re-enable
+   * it; "runtime_error" -> the current build could not run it. Null otherwise.
+   */
+  spec_fallback_reason?: string | null;
 }
 
 export interface AudioGenerationResponse {
diff --git a/studio/frontend/src/features/chat/utils/chat-history-storage.ts b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
index 0e9b30a5cc..2023ae30b1 100644
--- a/studio/frontend/src/features/chat/utils/chat-history-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
@@ -134,7 +134,7 @@ export function isExpectedBackgroundChatStorageError(error: unknown): boolean {
     (error.message === "Invalid or expired token" ||
       error.message === "Not authenticated" ||
       error.message === "Request failed (401)" ||
-      error.message === "Studio isn't running -- please relaunch it.")
+      error.message === "Unsloth isn't running -- please relaunch it.")
   );
 }
 
diff --git a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
index 79566f21bf..5fee63d12d 100644
--- a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
+++ b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
@@ -75,7 +75,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Easy",
     learningBadges: ["Seed Dataset", "LLM Text", "Prompting"],
     surfaceClassName:
-      "from-emerald-500/15 via-green-500/5 to-transparent dark:from-emerald-400/30 dark:via-green-400/14 dark:to-emerald-950/16",
+      "from-emerald-500/15 via-green-500/5 to-transparent dark:from-emerald-400/20 dark:via-green-400/10 dark:to-emerald-950/16",
     shineColor: [
       "rgb(16 185 129 / 0.45)",
       "rgb(34 197 94 / 0.4)",
@@ -91,7 +91,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Easy",
     learningBadges: ["Unstructured", "LLM Text"],
     surfaceClassName:
-      "from-violet-500/15 via-fuchsia-500/5 to-transparent dark:from-violet-400/30 dark:via-fuchsia-400/14 dark:to-violet-950/16",
+      "from-violet-500/15 via-fuchsia-500/5 to-transparent dark:from-violet-400/20 dark:via-fuchsia-400/10 dark:to-violet-950/16",
     shineColor: [
       "rgb(139 92 246 / 0.45)",
       "rgb(217 70 239 / 0.4)",
@@ -107,7 +107,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Starter",
     learningBadges: ["Vision", "LLM Text", "Image Context"],
     surfaceClassName:
-      "from-lime-500/15 via-emerald-500/5 to-transparent dark:from-lime-400/30 dark:via-emerald-400/14 dark:to-lime-950/16",
+      "from-lime-500/15 via-emerald-500/5 to-transparent dark:from-lime-400/20 dark:via-emerald-400/10 dark:to-lime-950/16",
     shineColor: [
       "rgb(132 204 22 / 0.45)",
       "rgb(16 185 129 / 0.4)",
@@ -123,7 +123,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Intermediate",
     learningBadges: ["LLM Judge", "LLM Code", "Subcategory", "Category"],
     surfaceClassName:
-      "from-amber-500/15 via-orange-500/5 to-transparent dark:from-amber-400/30 dark:via-orange-400/14 dark:to-amber-950/16",
+      "from-amber-500/15 via-orange-500/5 to-transparent dark:from-amber-400/20 dark:via-orange-400/10 dark:to-amber-950/16",
     shineColor: [
       "rgb(245 158 11 / 0.45)",
       "rgb(249 115 22 / 0.4)",
@@ -139,7 +139,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Intermediate",
     learningBadges: ["LLM Code", "Prompting", "Drop Columns"],
     surfaceClassName:
-      "from-blue-500/15 via-indigo-500/5 to-transparent dark:from-blue-400/30 dark:via-indigo-400/14 dark:to-blue-950/16",
+      "from-blue-500/15 via-indigo-500/5 to-transparent dark:from-blue-400/20 dark:via-indigo-400/10 dark:to-blue-950/16",
     shineColor: [
       "rgb(59 130 246 / 0.45)",
       "rgb(99 102 241 / 0.4)",
@@ -155,7 +155,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Advanced",
     learningBadges: ["Structured LLM", "Expression", "Jinja"],
     surfaceClassName:
-      "from-cyan-500/15 via-sky-500/5 to-transparent dark:from-cyan-400/30 dark:via-sky-400/14 dark:to-cyan-950/16",
+      "from-cyan-500/15 via-sky-500/5 to-transparent dark:from-cyan-400/20 dark:via-sky-400/10 dark:to-cyan-950/16",
     shineColor: [
       "rgb(6 182 212 / 0.45)",
       "rgb(56 189 248 / 0.4)",
@@ -171,7 +171,7 @@ const TEMPLATE_CARDS: TemplateCard[] = [
     difficulty: "Intermediate",
     learningBadges: ["GitHub", "LLM Text", "Structured LLM"],
     surfaceClassName:
-      "from-slate-500/15 via-zinc-500/5 to-transparent dark:from-slate-400/30 dark:via-zinc-400/14 dark:to-slate-950/16",
+      "from-slate-500/15 via-zinc-500/5 to-transparent dark:from-slate-400/20 dark:via-zinc-400/10 dark:to-slate-950/16",
     shineColor: [
       "rgb(71 85 105 / 0.45)",
       "rgb(100 116 139 / 0.4)",
@@ -241,7 +241,7 @@ function LearningRecipeCards({
             type="button"
             disabled={isDisabled}
             onClick={() => onSelect(template)}
-            className={`group shadow-border relative overflow-hidden rounded-2xl bg-gradient-to-br text-left transition-transform ${template.surfaceClassName} enabled:cursor-pointer enabled:hover:-translate-y-0.5 enabled:hover:shadow-md disabled:cursor-not-allowed disabled:opacity-70`}
+            className={`group shadow-border relative overflow-hidden rounded-2xl bg-gradient-to-br dark:bg-white/[0.05] text-left transition-transform ${template.surfaceClassName} enabled:cursor-pointer enabled:hover:-translate-y-0.5 disabled:cursor-not-allowed disabled:opacity-70`}
           >
             <ShineBorder
               borderWidth={1.2}
@@ -267,7 +267,7 @@ function LearningRecipeCards({
                 <p className="line-clamp-2 text-sm font-semibold leading-tight text-foreground">
                   {template.title}
                 </p>
-                <p className="line-clamp-2 text-xs text-muted-foreground">
+                <p className="line-clamp-2 text-xs text-muted-foreground dark:text-zinc-300">
                   {template.description}
                 </p>
               </div>
@@ -280,7 +280,7 @@ function LearningRecipeCards({
                       <Badge
                         key={`${template.title}-${badge}`}
                         variant="outline"
-                        className="h-5 shrink-0 px-1.5 text-[10px]"
+                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
                       >
                         {badge}
                       </Badge>
@@ -288,7 +288,7 @@ function LearningRecipeCards({
                     {extraLearningBadgeCount > 0 ? (
                       <Badge
                         variant="outline"
-                        className="h-5 shrink-0 px-1.5 text-[10px]"
+                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
                       >
                         +{extraLearningBadgeCount}
                       </Badge>
@@ -296,7 +296,7 @@ function LearningRecipeCards({
                     {isReady ? null : (
                       <Badge
                         variant="secondary"
-                        className="h-5 shrink-0 px-1.5 text-[10px]"
+                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
                       >
                         Soon
                       </Badge>
@@ -449,7 +449,7 @@ export function DataRecipesPage(): ReactElement {
             </p>
           </div>
         ) : recipes.length === 0 ? (
-          <Empty className="mt-8 border border-dashed border-border/70">
+          <Empty className="mt-8 border border-dashed border-border/70 dark:border-none">
             <EmptyHeader>
               <EmptyMedia variant="icon">
                 <HugeiconsIcon icon={CookBookIcon} className="size-5" />
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index b1a379b582..aa18be37f7 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -539,7 +539,7 @@ export function ExportPage() {
           description="Select source, method, and quantization"
           accent="emerald"
           featured={true}
-          className="shadow-border ring-1 ring-border"
+          className="ring-0 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
         >
           {/* Loading / error states */}
           {loadingCheckpoints && (
diff --git a/studio/frontend/src/features/hub/catalog/catalog-states.tsx b/studio/frontend/src/features/hub/catalog/catalog-states.tsx
index e4c8554773..1df85a6f81 100644
--- a/studio/frontend/src/features/hub/catalog/catalog-states.tsx
+++ b/studio/frontend/src/features/hub/catalog/catalog-states.tsx
@@ -58,7 +58,7 @@ export function NetworkErrorState({
         <button
           type="button"
           onClick={onRetry}
-          className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
+          className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.1]"
         >
           <HugeiconsIcon
             icon={RefreshIcon}
@@ -113,7 +113,7 @@ export function DiscoverFetchMoreState({
           type="button"
           onClick={onFetchMore}
           disabled={isLoadingMore}
-          className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-white/[0.05]"
+          className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-white/[0.1]"
         >
           <HugeiconsIcon
             icon={RefreshIcon}
@@ -191,7 +191,7 @@ export function InventoryErrorState({
       <button
         type="button"
         onClick={onRetry}
-        className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
+        className="inline-flex h-8 items-center gap-1.5 rounded-[10px] bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.1]"
       >
         <HugeiconsIcon icon={RefreshIcon} strokeWidth={1.75} className="size-3.5" />
         Try again
diff --git a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
index c9be7b549b..e0178c3472 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
@@ -20,7 +20,7 @@ import {
 import { type GgufVariantDetail, deleteCachedModel } from "../inventory";
 import { formatBytes } from "../lib/format";
 import { type GgufFitClass, classifyGgufFit } from "../lib/gguf-fit";
-import { HUB_POST_DOWNLOAD_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
+import { HUB_GGUF_RUN_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
 import {
   ggufVariantsMatch,
   normalizeGgufVariantIdentity,
@@ -666,7 +666,7 @@ export function GgufDownloadCard({
                 e.preventDefault();
                 setOpen((o) => !o);
               }}
-              className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2.5 rounded-[12px] px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] dark:hover:bg-white/[0.04] dark:data-[state=open]:bg-white/[0.06]"
+              className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2.5 rounded-[12px] px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] dark:hover:bg-white/[0.1] dark:data-[state=open]:bg-white/[0.06]"
             >
               {selected ? (
                 <QuantBadge
@@ -723,7 +723,7 @@ export function GgufDownloadCard({
           <PopoverContent
             align="start"
             side="bottom"
-            sideOffset={8}
+            sideOffset={0}
             avoidCollisions={false}
             className="hub-menu-instant menu-soft-surface w-[var(--radix-popover-trigger-width)] min-w-[200px] gap-0 overflow-hidden p-0 py-2 ring-0"
           >
@@ -794,8 +794,7 @@ export function GgufDownloadCard({
             downloadingThisVariant &&
               !cancelling &&
               "hover:bg-rose-500/10 hover:text-rose-600 dark:hover:text-rose-400",
-            // Hide post-download CTAs (Run / New Chat) for this PR.
-            !HUB_POST_DOWNLOAD_ACTIONS_VISIBLE &&
+            !HUB_GGUF_RUN_ACTIONS_VISIBLE &&
               !downloadingThisVariant &&
               !cancelling &&
               !isLoadingThisModel &&
diff --git a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
index 76539d43de..a6596a3d54 100644
--- a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
@@ -181,7 +181,7 @@ export function HubOptionMenu<T extends string>({
       <PopoverContent
         align={align}
         side="bottom"
-        sideOffset={8}
+        sideOffset={0}
         collisionPadding={12}
         onCloseAutoFocus={(event) => event.preventDefault()}
         className={cn(
diff --git a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
index c5e754d982..0dd50f0a41 100644
--- a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
@@ -2,7 +2,10 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { TrainIcon } from "../components/train-icon";
-import { HUB_POST_DOWNLOAD_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
+import {
+  HUB_GGUF_RUN_ACTIONS_VISIBLE,
+  HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
+} from "../lib/hub-feature-flags";
 import {
   Popover,
   PopoverContent,
@@ -295,6 +298,9 @@ export function LocalOnDeviceCard({
   const formatTone =
     modelFormat === "adapter" ? "adapter" : isGguf ? "gguf" : "checkpoint";
   const showOldCacheHint = source === "hf_cache" && !!unsupportedReason;
+  const runActionsVisible = isGguf
+    ? HUB_GGUF_RUN_ACTIONS_VISIBLE
+    : HUB_POST_DOWNLOAD_ACTIONS_VISIBLE;
 
   return (
     <div className="flex w-full flex-col gap-2">
@@ -352,7 +358,7 @@ export function LocalOnDeviceCard({
                   <PopoverContent
                     align="start"
                     side="bottom"
-                    sideOffset={8}
+                    sideOffset={0}
                     avoidCollisions={false}
                     className="hub-menu-instant menu-soft-surface w-[var(--radix-popover-trigger-width)] min-w-[220px] gap-0 overflow-hidden p-0 py-2 ring-0"
                   >
@@ -440,11 +446,10 @@ export function LocalOnDeviceCard({
               className="ml-1 mr-0 h-5 w-px shrink-0 bg-foreground/[0.06] opacity-100 transition-opacity duration-150 group-hover/dl:opacity-0 dark:bg-white/[0.04]"
             />
           )}
-          {/* Run + Train pair hidden until Hub->chat/train pickers ship. */}
           <div
             className={cn(
               "group/pair flex h-9 shrink-0 items-stretch gap-1.5",
-              !HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && "hidden",
+              !runActionsVisible && "hidden",
             )}
           >
             {onTrain && (
diff --git a/studio/frontend/src/features/hub/catalog/model-inspector.tsx b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
index 54165aa5e6..a3a73f2714 100644
--- a/studio/frontend/src/features/hub/catalog/model-inspector.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
@@ -725,8 +725,6 @@ export const ModelInspector = memo(function ModelInspector({
           />
         )}
       </div>
-
-      <div aria-hidden="true" className="h-3 shrink-0" />
     </div>
   );
 });
diff --git a/studio/frontend/src/features/hub/download-manager/api.ts b/studio/frontend/src/features/hub/download-manager/api.ts
index ca987f1643..2c55b90edd 100644
--- a/studio/frontend/src/features/hub/download-manager/api.ts
+++ b/studio/frontend/src/features/hub/download-manager/api.ts
@@ -12,14 +12,14 @@ function parseErrorText(status: number, body: unknown): string {
     const detail = (body as { detail?: unknown }).detail;
     const formatted = formatFastApiDetail(detail);
     if (status === 405) {
-      return `${formatted || "Method Not Allowed"} - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.`;
+      return `${formatted || "Method Not Allowed"} - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.`;
     }
     if (formatted) return formatted;
     const message = (body as { message?: unknown }).message;
     if (typeof message === "string" && message) return message;
   }
   if (status === 405) {
-    return "Method Not Allowed - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.";
+    return "Method Not Allowed - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.";
   }
   return `Request failed (${status})`;
 }
@@ -137,7 +137,7 @@ const DOWNLOAD_TRANSPORT_CAPABILITIES_FALLBACK: DownloadTransportCapabilities =
   http: { available: true, reason: null },
   xet: {
     available: null,
-    reason: "Couldn't verify Xet support with the Studio backend.",
+    reason: "Couldn't verify Xet support with the Unsloth backend.",
   },
 };
 let downloadTransportCapabilitiesCache: DownloadTransportCapabilities | null =
diff --git a/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts b/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
index 6942dfbbb9..a6518ec656 100644
--- a/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
+++ b/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
@@ -184,7 +184,7 @@ export async function effectiveTransportMode(
     return preferred;
   }
   const reason =
-    capabilities.xet.reason ?? "Studio will use HTTP downloads instead.";
+    capabilities.xet.reason ?? "Unsloth will use HTTP downloads instead.";
   if (lastXetUnavailableWarningReason !== reason) {
     lastXetUnavailableWarningReason = reason;
     toast.warning("Xet download transport unavailable", {
diff --git a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
index e01cd6b5cc..9833513c97 100644
--- a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
+++ b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
@@ -139,7 +139,7 @@ function DownloadRow({ jobKey }: { jobKey: string }) {
               }
               className={cn(
                 "inline-flex size-6 shrink-0 cursor-pointer items-center justify-center rounded-[7px] text-muted-foreground transition-colors",
-                "hover:bg-foreground/[0.06] hover:text-foreground disabled:cursor-default disabled:opacity-50 dark:hover:bg-white/[0.06]",
+                "hover:bg-foreground/[0.06] hover:text-foreground disabled:cursor-default disabled:opacity-50 dark:hover:bg-white/[0.1]",
               )}
             >
               <HugeiconsIcon
@@ -234,7 +234,7 @@ export function DownloadManagerPanel() {
               type="button"
               aria-label="Collapse downloads"
               onClick={() => setCollapsed(true)}
-              className="inline-flex size-6 shrink-0 cursor-pointer items-center justify-center rounded-[7px] text-muted-foreground transition-colors hover:bg-foreground/[0.06] hover:text-foreground dark:hover:bg-white/[0.06]"
+              className="inline-flex size-6 shrink-0 cursor-pointer items-center justify-center rounded-[7px] text-muted-foreground transition-colors hover:bg-foreground/[0.06] hover:text-foreground dark:hover:bg-white/[0.1]"
             >
               <HugeiconsIcon
                 icon={ArrowDown01Icon}
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index f606a9941a..7324bed262 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Hub page: browse + download surface only. Train CTA hidden until Hub->train picker ships.
 import { useHubInventory } from "@/features/hub/inventory";
 import { useDebouncedValue } from "@/hooks/use-debounced-value";
 import { useGpuInfo } from "@/hooks/use-gpu-info";
@@ -12,9 +11,15 @@ import {
 import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
 import { useIsHubDesktop } from "@/features/hub/hooks/use-is-hub-desktop";
 import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
-import { modelIdsMatch } from "@/features/hub/lib/model-identity";
+import { ggufVariantsMatch, modelIdsMatch } from "@/features/hub/lib/model-identity";
 import { cn } from "@/lib/utils";
 import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import {
+  getInferenceStatus,
+  isExternalModelId,
+  useChatModelRuntime,
+  useChatRuntimeStore,
+} from "@/features/chat";
 import { ArrowLeft01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate, useSearch } from "@tanstack/react-router";
@@ -179,14 +184,35 @@ export function ModelsPage() {
   const gpu = useGpuInfo();
   const online = useOnlineStatus();
 
-  // "Use in chat" just navigates to /chat; checkpoint/variant handoff lands with the Hub-aware picker.
-  const activeCheckpoint: string | null = null;
-  const activeGgufVariant: string | null = null;
-  // Stubs so the inspector + load-status memo still typecheck; cast keeps the shape from narrowing to `never`.
-  const loadingModel = null as { id: string } | null;
-  const loadProgress = null as {
-    phase?: "downloading" | "starting";
-  } | null;
+  const { selectModel, loadingModel, loadProgress, ejectModel } =
+    useChatModelRuntime();
+  const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
+  const activeCheckpoint =
+    checkpoint && !isExternalModelId(checkpoint) ? checkpoint : null;
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+
+  useEffect(() => {
+    let cancelled = false;
+    void getInferenceStatus()
+      .then((status) => {
+        if (cancelled || !status.active_model) return;
+        const store = useChatRuntimeStore.getState();
+        if (
+          !isExternalModelId(store.params.checkpoint) &&
+          (!modelIdsMatch(store.params.checkpoint, status.active_model) ||
+            !ggufVariantsMatch(
+              store.activeGgufVariant,
+              status.gguf_variant ?? null,
+            ))
+        ) {
+          store.setCheckpoint(status.active_model, status.gguf_variant ?? null);
+        }
+      })
+      .catch(() => undefined);
+    return () => {
+      cancelled = true;
+    };
+  }, []);
 
   const { tab, setTab: setModelsTab } = useModelsTabState();
   const [query, setQuery] = useState("");
@@ -522,8 +548,8 @@ export function ModelsPage() {
   );
 
   const isLoadingThisModel = useMemo(() => {
-    if (!loadingModel) return false;
-    return selectedRepoMatchesRuntime(selectedModel, loadingModel.id, null);
+    if (!loadingModel || !selectedModel) return false;
+    return modelIdsMatch(loadingModel.id, selectedModel.resource.runId);
   }, [loadingModel, selectedModel]);
 
   const { vramInfo, minMemory } = useHubModelVram(selectedModel, gpu);
@@ -536,22 +562,41 @@ export function ModelsPage() {
       ? `${Math.floor(gpu.systemRamAvailableGb)} GB`
       : "Unavailable";
 
-  // Load / use-in-chat / train handlers are stubs; they wire in once the
-  // Hub-aware chat and train pickers ship.
+  const openNewChat = useCallback(() => {
+    void navigate({ to: "/chat", search: { new: crypto.randomUUID() } });
+  }, [navigate]);
+  const runSelectedModel = useCallback(
+    (opts: ModelLoadOptions, isDownloaded: boolean) => {
+      if (!selectedModel) return;
+      const runId = selectedModel.resource.runId;
+      void selectModel({
+        id: runId,
+        ggufVariant: opts.ggufVariant,
+        isDownloaded,
+        expectedBytes: opts.expectedBytes,
+        throwOnError: true,
+      })
+        .then(() => {
+          const store = useChatRuntimeStore.getState();
+          if (!modelIdsMatch(store.params.checkpoint, runId)) {
+            store.setCheckpoint(runId, opts.ggufVariant ?? null);
+          }
+        })
+        .catch(() => undefined);
+      openNewChat();
+    },
+    [openNewChat, selectModel, selectedModel],
+  );
   const handleLoad = useCallback(
-    (_opts: ModelLoadOptions) => undefined,
-    [],
+    (opts: ModelLoadOptions) =>
+      runSelectedModel(opts, selectedModel?.isDownloaded ?? true),
+    [runSelectedModel, selectedModel],
   );
   const handleLoadLocal = useCallback(
-    (_opts: ModelLoadOptions = {}) => undefined,
-    [],
+    (opts: ModelLoadOptions = {}) => runSelectedModel(opts, true),
+    [runSelectedModel],
   );
-  const handleUseInChat = useCallback(() => {
-    void navigate({ to: "/chat" });
-  }, [navigate]);
-  const handleTrain = useCallback(() => {
-    // Train CTA hidden until Hub->train picker ships.
-  }, []);
+  const handleTrain = useCallback(() => undefined, []);
 
   const inspectorRuntime = useMemo(
     () => ({
@@ -582,14 +627,14 @@ export function ModelsPage() {
     () => ({
       onLoad: handleLoad,
       onLoadLocal: handleLoadLocal,
-      onUseInChat: handleUseInChat,
+      onUseInChat: openNewChat,
       onTrain: handleTrain,
       onInventoryChange: refreshInventory,
     }),
     [
       handleLoad,
       handleLoadLocal,
-      handleUseInChat,
+      openNewChat,
       handleTrain,
       refreshInventory,
     ],
@@ -675,7 +720,7 @@ export function ModelsPage() {
   );
 
   return (
-    <div className="hub-page flex h-full min-h-0 flex-col bg-background">
+    <div className="hub-page flex h-full min-h-0 flex-col">
       <div className="mx-auto flex w-full max-w-[1180px] flex-1 min-h-0 flex-col gap-6 px-5 pt-8 pb-16 sm:px-9 sm:pt-10 sm:pb-24">
         <ModelsHeader
           cachedCount={effectiveCachedRows.length}
@@ -685,7 +730,7 @@ export function ModelsPage() {
           ramLabel={ramLabel}
           activeCheckpoint={activeCheckpoint}
           activeGgufVariant={activeGgufVariant}
-          onEject={() => undefined}
+          onEject={() => void ejectModel()}
         />
 
         <section className="elevated-card flex min-h-0 flex-1 flex-col overflow-hidden bg-card">
diff --git a/studio/frontend/src/features/hub/hub.css b/studio/frontend/src/features/hub/hub.css
index 66e732c7b6..7dab214e04 100644
--- a/studio/frontend/src/features/hub/hub.css
+++ b/studio/frontend/src/features/hub/hub.css
@@ -97,9 +97,8 @@
 	.hub-page .elevated-card {
 		border: 1px solid color-mix(in srgb, var(--foreground) 5%, transparent);
 		border-radius: 20px;
-		box-shadow:
-			0 12px 48px -8px color-mix(in srgb, var(--foreground) 8%, transparent),
-			0 0 100px 20px color-mix(in srgb, var(--foreground) 3%, transparent);
+		/* No ambient glow: it tinted the white page gray around the card. */
+		box-shadow: 0 12px 48px -8px color-mix(in srgb, var(--foreground) 8%, transparent);
 	}
 
 	@media (min-width: 640px) {
@@ -131,7 +130,7 @@
 	}
 
 	html:not(.dark) .hub-page {
-		background-color: var(--background);
+		background-color: #fff;
 	}
 
 	html:not(.dark) .hub-page > div > section .border-b,
diff --git a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts b/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
index cf28403725..d0d527b720 100644
--- a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
+++ b/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
@@ -1,8 +1,5 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Hub feature flags for staged rollout. Each gates a CTA whose integration ships
-// in a later PR; JSX/wiring stay in place so flipping here is the only re-enable edit.
-
-// Post-download Run / chat / Train CTAs, hidden until the Hub-aware chat and train pickers ship.
 export const HUB_POST_DOWNLOAD_ACTIONS_VISIBLE = false;
+export const HUB_GGUF_RUN_ACTIONS_VISIBLE = true;
diff --git a/studio/frontend/src/features/onboarding/components/splash-screen.tsx b/studio/frontend/src/features/onboarding/components/splash-screen.tsx
index 70438a04d8..47d9c25efe 100644
--- a/studio/frontend/src/features/onboarding/components/splash-screen.tsx
+++ b/studio/frontend/src/features/onboarding/components/splash-screen.tsx
@@ -3,6 +3,7 @@
 
 import { Button } from "@/components/ui/button";
 import { Card } from "@/components/ui/card";
+import { MascotImg } from "@/components/mascot-img";
 import { motion } from "motion/react";
 
 interface SplashScreenProps {
@@ -19,10 +20,7 @@ export function SplashScreen({
       <Card className="w-full max-w-md px-8 py-8 shadow-border ring-1 ring-border">
         {/* Mascot */}
         <div className="flex justify-center">
-          <motion.img
-            src={`${import.meta.env.BASE_URL}Sloth emojis/Sloth loca pc.png`}
-            alt="Sloth mascot"
-            className="size-30"
+          <motion.div
             initial={{ opacity: 0, y: 40, scale: 0.95 }}
             animate={{ opacity: 1, y: 0, scale: 1 }}
             transition={{
@@ -31,7 +29,9 @@ export function SplashScreen({
               bounce: 0.3,
               delay: 0.1,
             }}
-          />
+          >
+            <MascotImg src="Sloth emojis/Sloth loca pc.png" className="size-30" />
+          </motion.div>
         </div>
 
         {/* Brand text */}
diff --git a/studio/frontend/src/features/onboarding/components/wizard-content.tsx b/studio/frontend/src/features/onboarding/components/wizard-content.tsx
index 6906b87f8f..40d290a73e 100644
--- a/studio/frontend/src/features/onboarding/components/wizard-content.tsx
+++ b/studio/frontend/src/features/onboarding/components/wizard-content.tsx
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { MascotImg } from "@/components/mascot-img";
 import { STEPS } from "@/config/training";
 import { useTrainingConfigStore } from "@/features/training";
 import type { StepNumber } from "@/types/training";
@@ -19,11 +20,11 @@ const STEP_COMPONENTS = {
 } as const;
 
 const STEP_MASCOTS: Record<StepNumber, string> = {
-  1: `${import.meta.env.BASE_URL}Sloth emojis/large sloth wave.png`,
-  2: `${import.meta.env.BASE_URL}Sloth emojis/sloth magnify final.png`,
-  3: `${import.meta.env.BASE_URL}Sloth emojis/sloth huglove large.png`,
-  4: `${import.meta.env.BASE_URL}Sloth emojis/large sloth glasses.png`,
-  5: `${import.meta.env.BASE_URL}Sloth emojis/large sloth yay.png`,
+  1: "Sloth emojis/large sloth wave.png",
+  2: "Sloth emojis/sloth magnify final.png",
+  3: "Sloth emojis/sloth huglove large.png",
+  4: "Sloth emojis/large sloth glasses.png",
+  5: "Sloth emojis/large sloth yay.png",
 };
 
 export function WizardContent() {
@@ -35,7 +36,7 @@ export function WizardContent() {
   return (
     <main className="flex-1 flex flex-col overflow-y-auto">
       <header className="flex flex-wrap items-start gap-3 p-4 pb-3 sm:p-6 sm:pb-4">
-        <img src={mascotSrc} alt="Unsloth mascot" className="size-12 sm:size-14" />
+        <MascotImg src={mascotSrc} className="size-12 sm:size-14" />
         <div className="flex flex-col min-w-0">
           <h1 className="text-lg font-semibold sm:text-xl">{stepConfig.title}</h1>
           <p className="text-sm text-muted-foreground">
diff --git a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
index 3b9b60e5dd..050595f636 100644
--- a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
+++ b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
@@ -118,7 +118,7 @@ export function ProfilePersonalizationPanel() {
         <button
           type="button"
           onClick={() => fileInputRef.current?.click()}
-          className="absolute right-0 bottom-0 -translate-x-[15.625%] -translate-y-[15.625%] flex size-8 items-center justify-center rounded-full border border-border bg-background text-foreground shadow-[0_2px_8px_-2px_rgba(27,27,31,0.16)] transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
+          className="absolute right-0 bottom-0 -translate-x-[15.625%] -translate-y-[15.625%] flex size-8 items-center justify-center rounded-full border border-border bg-background text-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
           aria-label={t("settings.profile.changePicture")}
         >
           <HugeiconsIcon icon={Camera01Icon} className="size-3.5" strokeWidth={2} />
diff --git a/studio/frontend/src/features/rag/components/knowledge-base-composer-button.tsx b/studio/frontend/src/features/rag/components/knowledge-base-composer-button.tsx
index b703fca347..e3c9c9b41f 100644
--- a/studio/frontend/src/features/rag/components/knowledge-base-composer-button.tsx
+++ b/studio/frontend/src/features/rag/components/knowledge-base-composer-button.tsx
@@ -14,7 +14,7 @@ import {
   DropdownMenuSeparator,
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
-import { useRagToolAvailable } from "@/features/chat/hooks/use-rag-tool-available";
+import { useRagToolDisabled } from "@/features/chat/hooks/use-rag-tool-disabled";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 
 import { listKnowledgeBases } from "../api/rag-api";
@@ -38,8 +38,8 @@ const ArrowDownStandardIcon: FC<{ className?: string }> = ({ className }) => (
   </svg>
 );
 
-// Picks the retrieval source. Only rendered when retrieval is on and the loaded
-// model can run search_knowledge_base.
+// Picks the retrieval source. Shown whenever retrieval is on; dims but stays
+// interactive (so it can be turned off) while the loaded model can't run it.
 export function KnowledgeBaseComposerButton({
   side = "bottom",
 }: {
@@ -47,7 +47,7 @@ export function KnowledgeBaseComposerButton({
 } = {}) {
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
   const setRagEnabled = useChatRuntimeStore((s) => s.setRagEnabled);
-  const ragAvailable = useRagToolAvailable();
+  const ragDisabled = useRagToolDisabled();
   const ragSource = useChatRuntimeStore((s) => s.ragSource);
   const setRagSource = useChatRuntimeStore((s) => s.setRagSource);
 
@@ -85,7 +85,7 @@ export function KnowledgeBaseComposerButton({
     }
   }, [kbs, kbsLoaded, ragSource, setRagSource]);
 
-  if (!ragEnabled || !ragAvailable) return null;
+  if (!ragEnabled) return null;
 
   return (
     <>
@@ -100,17 +100,24 @@ export function KnowledgeBaseComposerButton({
           <button
             type="button"
             className="composer-pill-btn"
-            data-active="true"
+            data-pill-label="RAG"
+            data-active={ragDisabled ? "false" : "true"}
             aria-label="Retrieval source"
           >
             {/* Icon doubles as an off switch: hover swaps to an X; clicking it
-                turns RAG off without opening the menu. */}
+                turns RAG off without opening the menu. In compact icon-only
+                mode the glyph is the whole button, so clicks fall through to
+                the trigger and open the menu instead. */}
             <span
               role="button"
               aria-label="Turn off retrieval"
               tabIndex={-1}
-              onPointerDown={(e) => e.stopPropagation()}
+              onPointerDown={(e) => {
+                if (e.currentTarget.closest('[data-pill-compact="true"]')) return;
+                e.stopPropagation();
+              }}
               onClick={(e) => {
+                if (e.currentTarget.closest('[data-pill-compact="true"]')) return;
                 e.stopPropagation();
                 setRagEnabled(false);
               }}
diff --git a/studio/frontend/src/features/rag/components/thread-documents-bar.tsx b/studio/frontend/src/features/rag/components/thread-documents-bar.tsx
index 6ba0440699..38677c27c4 100644
--- a/studio/frontend/src/features/rag/components/thread-documents-bar.tsx
+++ b/studio/frontend/src/features/rag/components/thread-documents-bar.tsx
@@ -7,7 +7,6 @@ import { AttachmentIcon, FileDatabaseIcon } from "@hugeicons/core-free-icons";
 import { useAui } from "@assistant-ui/react";
 import { cn } from "@/lib/utils";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
-import { useRagToolAvailable } from "@/features/chat/hooks/use-rag-tool-available";
 import { toast } from "@/lib/toast";
 import { listKnowledgeBases, listThreadDocuments } from "../api/rag-api";
 import { RAG_UPLOAD_ACCEPT } from "../types/rag";
@@ -55,7 +54,6 @@ export function ThreadDocumentsBar({
   onIndexingChange?: (active: boolean) => void;
 }) {
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
-  const ragAvailable = useRagToolAvailable();
   const ragSource = useChatRuntimeStore((s) => s.ragSource);
   const aui = useAui();
   const fileInputRef = useRef<HTMLInputElement>(null);
@@ -137,8 +135,9 @@ export function ThreadDocumentsBar({
     fileInputRef.current?.click();
   }, []);
 
-  // Only when the RAG pill is on: enabled AND a tool-capable model loaded.
-  if (!ragEnabled || !ragAvailable) return null;
+  // Shown whenever the RAG pill is on: ingestion only needs the embedder, so
+  // files can index before a chat model loads.
+  if (!ragEnabled) return null;
   // A KB source uploads via the KB dialog, not here; show which KB is active.
   if (ragSource.type === "kb") {
     return <KnowledgeBaseSourceChip kbId={ragSource.kbId} />;
diff --git a/studio/frontend/src/features/recipe-studio/components/executions/execution-data-tab.tsx b/studio/frontend/src/features/recipe-studio/components/executions/execution-data-tab.tsx
index a3e5dbe962..1616f6f77d 100644
--- a/studio/frontend/src/features/recipe-studio/components/executions/execution-data-tab.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/executions/execution-data-tab.tsx
@@ -45,7 +45,7 @@ function formatGitHubSourceMessage(execution: RecipeExecutionRecord): string {
     return "Collecting repository threads before rows are available.";
   }
   if (source.status === "rate_limited") {
-    return source.message ?? "Waiting for GitHub rate limit. Studio will resume automatically.";
+    return source.message ?? "Waiting for GitHub rate limit. Unsloth will resume automatically.";
   }
   return source.message ?? "Collecting repository threads before rows are available.";
 }
diff --git a/studio/frontend/src/features/recipe-studio/components/executions/execution-overview-tab.tsx b/studio/frontend/src/features/recipe-studio/components/executions/execution-overview-tab.tsx
index 72624738fc..5ff9d644e7 100644
--- a/studio/frontend/src/features/recipe-studio/components/executions/execution-overview-tab.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/executions/execution-overview-tab.tsx
@@ -41,7 +41,7 @@ function formatSourceMessage(execution: RecipeExecutionRecord): string {
       typeof source.retry_after_sec === "number" && source.retry_after_sec > 0
         ? ` Waiting ~${formatMetricValue(source.retry_after_sec)}s.`
         : "";
-    return `Waiting for GitHub rate limit. Studio will resume automatically.${wait}`;
+    return `Waiting for GitHub rate limit. Unsloth will resume automatically.${wait}`;
   }
   return source.message ?? "Crawling GitHub source.";
 }
diff --git a/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx b/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
index 563280c55f..1a194ba857 100644
--- a/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
@@ -68,7 +68,7 @@ function formatGitHubSourceSummary(
       typeof source.retry_after_sec === "number" && source.retry_after_sec > 0
         ? ` ~${formatMetricValue(source.retry_after_sec)}s`
         : "";
-    return `Waiting for GitHub rate limit${wait}. Studio will resume automatically.`;
+    return `Waiting for GitHub rate limit${wait}. Unsloth will resume automatically.`;
   }
   if (source.status === "retrying") {
     return source.message ?? "GitHub request failed; retrying automatically.";
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
index 9b91bfe334..c333fd84f8 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
@@ -591,7 +591,7 @@ export function LocalRecipeModelSelector({
       </PopoverTrigger>
       <PopoverContent
         align="start"
-        sideOffset={6}
+        sideOffset={0}
         className="menu-soft-surface nodrag nowheel gap-0 overflow-hidden p-0"
         style={{
           width:
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
index 878f9eba7c..0ed7eeed75 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
@@ -334,8 +334,8 @@ export function GithubRepoSeedForm({
         />
         <p id={tokenHelpId} className="text-xs text-muted-foreground">
           {usingEnvToken
-            ? "Studio detected a server env token, so saved/shared recipes can leave this blank."
-            : "Blank is safest for saved/shared recipes because Studio will read the server environment at run time."}
+            ? "Unsloth detected a server env token, so saved/shared recipes can leave this blank."
+            : "Blank is safest for saved/shared recipes because Unsloth will read the server environment at run time."}
         </p>
         {hasToken && (
           <p className="rounded-md bg-amber-500/10 px-2 py-1.5 text-xs text-amber-700 dark:text-amber-300">
@@ -450,7 +450,7 @@ export function GithubRepoSeedForm({
       </fieldset>
 
       <p className="text-xs text-muted-foreground">
-        Backed by Studio's built-in <code>github_repo</code> seed reader. Large
+        Backed by Unsloth's built-in <code>github_repo</code> seed reader. Large
         repos can take minutes, so start with small limits for previews.
       </p>
     </div>
diff --git a/studio/frontend/src/features/settings/components/update-studio-instructions.tsx b/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
index 90f66483b7..ef6bbf3c18 100644
--- a/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
+++ b/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
@@ -4,21 +4,29 @@
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { useT } from "@/i18n";
 import { cn } from "@/lib/utils";
-import { Copy01Icon, Tick02Icon } from "@hugeicons/core-free-icons";
+import {
+  ArrowUpRight01Icon,
+  Copy01Icon,
+  Tick02Icon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { AnimatePresence, motion, useReducedMotion } from "motion/react";
 import type { ReactElement } from "react";
 import { useEffect, useRef, useState } from "react";
 
-const STUDIO_UPDATE_CMD = "unsloth studio update";
-const STUDIO_UPDATE_FALLBACK_UNIX_CMD =
+const STUDIO_INSTALL_UNIX_CMD =
   "curl -fsSL https://unsloth.ai/install.sh | sh";
-const STUDIO_UPDATE_FALLBACK_WINDOWS_CMD =
-  "irm https://unsloth.ai/install.ps1 | iex";
+const STUDIO_INSTALL_WINDOWS_CMD = "irm https://unsloth.ai/install.ps1 | iex";
 const STUDIO_LOCAL_PULL_CMD = "git pull --ff-only";
-const STUDIO_LOCAL_UPDATE_CMD = "unsloth studio update --local";
-const STUDIO_LOCAL_FALLBACK_UNIX_CMD = "./install.sh --local";
-const STUDIO_LOCAL_FALLBACK_WINDOWS_CMD = ".\\install.ps1 --local";
+const STUDIO_LOCAL_INSTALL_UNIX_CMD = "./install.sh --local";
+const STUDIO_LOCAL_INSTALL_WINDOWS_CMD = ".\\install.ps1 --local";
+
+const DOCS_INSTALL_URL = "https://unsloth.ai/docs/get-started/install";
+const DOCS_UPDATING_URL =
+  "https://unsloth.ai/docs/get-started/install/updating";
+const DOCS_MAC_URL = "https://unsloth.ai/docs/get-started/install/mac";
+const DOCS_WINDOWS_URL =
+  "https://unsloth.ai/docs/get-started/install/windows-installation";
 
 export type UpdateShell = "windows" | "unix";
 export type UpdateInstallSource =
@@ -30,15 +38,6 @@ export type UpdateInstallSource =
   | "unknown";
 type UpdateInstallSourceState = UpdateInstallSource | "loading";
 
-function getStudioUpdateInstructionLine(
-  shell: UpdateShell,
-  t: ReturnType<typeof useT>,
-): string {
-  return shell === "windows"
-    ? t("settings.about.update.openPowerShell")
-    : t("settings.about.update.openTerminal");
-}
-
 function isLocalInstallSource(
   installSource?: UpdateInstallSourceState | null,
 ): boolean {
@@ -128,6 +127,77 @@ function CopyableCommand({
   );
 }
 
+function DocsLink({
+  href,
+  label,
+}: {
+  href: string;
+  label: string;
+}): ReactElement {
+  return (
+    <a
+      href={href}
+      target="_blank"
+      rel="noopener noreferrer"
+      className="inline-flex items-center gap-0.5 font-medium text-muted-foreground hover:text-foreground"
+    >
+      {label}
+      <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+    </a>
+  );
+}
+
+function UpdateDocsLinks(): ReactElement {
+  const t = useT();
+  return (
+    <p className="flex flex-wrap items-center gap-x-2 gap-y-1 text-xs text-muted-foreground leading-relaxed">
+      {t("settings.about.update.docs")}
+      <DocsLink
+        href={DOCS_INSTALL_URL}
+        label={t("settings.about.update.docsInstall")}
+      />
+      <DocsLink
+        href={DOCS_UPDATING_URL}
+        label={t("settings.about.update.docsUpdating")}
+      />
+      <DocsLink
+        href={DOCS_MAC_URL}
+        label={t("settings.about.update.docsMac")}
+      />
+      <DocsLink
+        href={DOCS_WINDOWS_URL}
+        label={t("settings.about.update.docsWindows")}
+      />
+    </p>
+  );
+}
+
+function ShellToggleButton({
+  active,
+  label,
+  onClick,
+}: {
+  active: boolean;
+  label: string;
+  onClick: () => void;
+}): ReactElement {
+  return (
+    <button
+      type="button"
+      onClick={onClick}
+      aria-pressed={active}
+      className={cn(
+        "rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors",
+        active
+          ? "bg-foreground/[0.08] text-foreground dark:bg-white/[0.12]"
+          : "text-muted-foreground hover:text-foreground",
+      )}
+    >
+      {label}
+    </button>
+  );
+}
+
 // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: keep source-specific update guidance in one component so the command matrix stays visible.
 export function UpdateStudioInstructions({
   className,
@@ -145,6 +215,9 @@ export function UpdateStudioInstructions({
   const shell = shellOverride ?? defaultShell;
   const prefersReducedMotion = useReducedMotion();
   const windows = shell === "windows";
+  // null means the desktop app: its bundled backend updates through the
+  // built-in updater, so terminal commands would target the wrong install.
+  const desktopManaged = installSource === null;
   const localInstallSource = isLocalInstallSource(installSource);
   const checkoutInstallSource =
     installSource === "editable" || installSource === "local_repo";
@@ -163,6 +236,22 @@ export function UpdateStudioInstructions({
     ? { opacity: 1 }
     : { opacity: 0, y: -2 };
 
+  if (desktopManaged) {
+    return (
+      <div className={cn("flex flex-col gap-3", className)}>
+        {showTitle ? (
+          <p className="shrink-0 whitespace-nowrap text-sm font-semibold font-heading">
+            {t("settings.about.update.title")}
+          </p>
+        ) : null}
+        <p className="text-xs text-muted-foreground leading-relaxed">
+          {t("settings.about.update.desktopManaged")}
+        </p>
+        <UpdateDocsLinks />
+      </div>
+    );
+  }
+
   return (
     <div className={cn("flex flex-col gap-3", className)}>
       <div
@@ -176,42 +265,55 @@ export function UpdateStudioInstructions({
             {t("settings.about.update.title")}
           </p>
         ) : null}
-        <div className="flex shrink-0 items-center gap-0.5 text-[11px]">
-          <button
-            type="button"
-            onClick={() => setShellOverride("windows")}
-            className={cn(
-              "px-0.5 py-0.5 font-medium transition-colors",
-              windows
-                ? "text-foreground"
-                : "text-muted-foreground hover:text-emerald-600",
-            )}
-            aria-pressed={windows}
-          >
-            Windows
-          </button>
-          <span className="text-border">/</span>
-          <button
-            type="button"
+        <div className="flex shrink-0 items-center gap-1.5">
+          <ShellToggleButton
+            active={!windows}
+            label="MacOS / Linux"
             onClick={() => setShellOverride("unix")}
-            className={cn(
-              "px-0.5 py-0.5 font-medium transition-colors",
-              windows
-                ? "text-muted-foreground hover:text-emerald-600"
-                : "text-foreground",
-            )}
-            aria-pressed={!windows}
-          >
-            macOS/Linux
-          </button>
+          />
+          <ShellToggleButton
+            active={windows}
+            label="Windows"
+            onClick={() => setShellOverride("windows")}
+          />
         </div>
       </div>
       {loadingInstallSource ? (
         <p className="text-xs text-muted-foreground leading-relaxed">
           {t("settings.about.update.checkingInstall")}
         </p>
-      ) : localInstallSource ? (
+      ) : (
         <>
+          <p className="text-xs text-muted-foreground leading-relaxed">
+            {t("settings.about.update.installIntro")}
+          </p>
+          <AnimatePresence mode="wait" initial={false}>
+            <motion.div
+              key={`install-${shell}`}
+              initial={fadeInitial}
+              animate={fadeAnimate}
+              exit={fadeExit}
+              transition={fadeTransition}
+            >
+              <CopyableCommand
+                command={
+                  windows ? STUDIO_INSTALL_WINDOWS_CMD : STUDIO_INSTALL_UNIX_CMD
+                }
+                copyLabel={
+                  windows
+                    ? t("settings.about.update.installCommandWindows")
+                    : t("settings.about.update.installCommandUnix")
+                }
+              />
+            </motion.div>
+          </AnimatePresence>
+        </>
+      )}
+      {loadingInstallSource ? null : localInstallSource ? (
+        <>
+          <p className="text-xs font-semibold text-foreground">
+            {t("settings.about.update.localUpdateHeading")}
+          </p>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.localInstallDetected")}
           </p>
@@ -224,16 +326,9 @@ export function UpdateStudioInstructions({
                 command={STUDIO_LOCAL_PULL_CMD}
                 copyLabel={t("settings.about.update.gitPullCommand")}
               />
-              <CopyableCommand
-                command={STUDIO_LOCAL_UPDATE_CMD}
-                copyLabel={t("settings.about.update.localUpdateCommand")}
-              />
-              <p className="text-xs text-muted-foreground leading-relaxed">
-                {t("settings.about.update.localInstallerFallback")}
-              </p>
               <AnimatePresence mode="wait" initial={false}>
                 <motion.div
-                  key={`local-fallback-${shell}`}
+                  key={`local-installer-${shell}`}
                   initial={fadeInitial}
                   animate={fadeAnimate}
                   exit={fadeExit}
@@ -242,8 +337,8 @@ export function UpdateStudioInstructions({
                   <CopyableCommand
                     command={
                       windows
-                        ? STUDIO_LOCAL_FALLBACK_WINDOWS_CMD
-                        : STUDIO_LOCAL_FALLBACK_UNIX_CMD
+                        ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                        : STUDIO_LOCAL_INSTALL_UNIX_CMD
                     }
                     copyLabel={t("settings.about.update.localInstallerCommand")}
                   />
@@ -261,7 +356,7 @@ export function UpdateStudioInstructions({
               </p>
               <AnimatePresence mode="wait" initial={false}>
                 <motion.div
-                  key={`source-fallback-${shell}`}
+                  key={`source-installer-${shell}`}
                   initial={fadeInitial}
                   animate={fadeAnimate}
                   exit={fadeExit}
@@ -270,8 +365,8 @@ export function UpdateStudioInstructions({
                   <CopyableCommand
                     command={
                       windows
-                        ? STUDIO_LOCAL_FALLBACK_WINDOWS_CMD
-                        : STUDIO_LOCAL_FALLBACK_UNIX_CMD
+                        ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                        : STUDIO_LOCAL_INSTALL_UNIX_CMD
                     }
                     copyLabel={t("settings.about.update.localInstallerCommand")}
                   />
@@ -282,54 +377,22 @@ export function UpdateStudioInstructions({
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.restartAfterUpdate")}
           </p>
+          <UpdateDocsLinks />
         </>
       ) : unknownInstallSource ? (
         <>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.unknownInstall")}
           </p>
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.curlOrPypi")}
+          <p className="text-xs font-semibold text-foreground">
+            {t("settings.about.update.localUpdateHeading")}
           </p>
-          <CopyableCommand
-            command={STUDIO_UPDATE_CMD}
-            copyLabel={t("settings.about.update.updateCommand")}
-          />
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.localCheckout")}
           </p>
-          <CopyableCommand
-            command={STUDIO_LOCAL_UPDATE_CMD}
-            copyLabel={t("settings.about.update.localUpdateCommand")}
-          />
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.restartAfterUpdate")}
-          </p>
-        </>
-      ) : (
-        <>
-          <AnimatePresence mode="wait" initial={false}>
-            <motion.p
-              key={`instruction-${shell}`}
-              initial={fadeInitial}
-              animate={fadeAnimate}
-              exit={fadeExit}
-              transition={fadeTransition}
-              className="text-xs text-muted-foreground leading-relaxed"
-            >
-              {getStudioUpdateInstructionLine(shell, t)}
-            </motion.p>
-          </AnimatePresence>
-          <CopyableCommand
-            command={STUDIO_UPDATE_CMD}
-            copyLabel={t("settings.about.update.updateCommand")}
-          />
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.fallbackInstruction")}
-          </p>
           <AnimatePresence mode="wait" initial={false}>
             <motion.div
-              key={`fallback-${shell}`}
+              key={`local-installer-${shell}`}
               initial={fadeInitial}
               animate={fadeAnimate}
               exit={fadeExit}
@@ -338,16 +401,24 @@ export function UpdateStudioInstructions({
               <CopyableCommand
                 command={
                   windows
-                    ? STUDIO_UPDATE_FALLBACK_WINDOWS_CMD
-                    : STUDIO_UPDATE_FALLBACK_UNIX_CMD
+                    ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                    : STUDIO_LOCAL_INSTALL_UNIX_CMD
                 }
-                copyLabel={t("settings.about.update.fallbackCommand")}
+                copyLabel={t("settings.about.update.localInstallerCommand")}
               />
             </motion.div>
           </AnimatePresence>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.restartAfterUpdate")}
           </p>
+          <UpdateDocsLinks />
+        </>
+      ) : (
+        <>
+          <p className="text-xs text-muted-foreground leading-relaxed">
+            {t("settings.about.update.restartAfterUpdate")}
+          </p>
+          <UpdateDocsLinks />
         </>
       )}
     </div>
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 26efcde7ea..0d10ff722a 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -155,18 +155,18 @@ export function SettingsDialog() {
                     type="button"
                     onClick={() => setActiveTab(tab.id)}
                     className={cn(
-                      "relative flex h-[32px] items-center gap-2.5 rounded-[11px] pl-3 pr-2.5 text-[14.5px] leading-[19px] tracking-nav font-medium transition-colors",
+                      "relative flex h-[32px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-[14.5px] leading-[19px] tracking-nav font-medium transition-colors",
                       "max-sm:shrink-0",
                       "focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background",
                       active
                         ? "text-black dark:text-white"
-                        : "text-[#383835] dark:text-[#c7c7c4] hover:bg-[#ececec] dark:hover:bg-[#2d2f33] hover:text-black dark:hover:text-white",
+                        : "text-[#383835] dark:text-[#c7c7c4] hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white",
                     )}
                   >
                     {active && (
                       <motion.span
                         layoutId="settings-active-pill"
-                        className="absolute inset-0 rounded-[11px] bg-[#ececec] dark:bg-[#2d2f33]"
+                        className="absolute inset-0 rounded-full bg-[#ececec] dark:bg-[#3a3d43]"
                         transition={
                           reduced
                             ? { duration: 0 }
@@ -202,7 +202,7 @@ export function SettingsDialog() {
             <button
               type="button"
               onClick={closeDialog}
-              className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-[#ececec] dark:hover:bg-[#2d2f33] hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
               aria-label={t("settings.dialog.closeAriaLabel")}
             >
               <HugeiconsIcon icon={Cancel01Icon} className="size-4" />
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index e73ad802de..69c4aa6ef1 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -142,7 +142,7 @@ export function AboutTab() {
         </p>
       </header>
 
-      <SettingsSection title="Studio">
+      <SettingsSection title="Unsloth">
         <SettingsRow label={t("settings.about.studioVersion")}>
           <code className="font-mono text-xs text-muted-foreground">
             {studioVersion}
@@ -207,6 +207,37 @@ export function AboutTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection title={t("settings.about.license.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.about.license.studioLabel")}
+          description={t("settings.about.license.studioDescription")}
+        >
+          <a
+            href="https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="inline-flex items-center gap-1 font-mono text-xs font-medium text-muted-foreground hover:text-foreground"
+          >
+            {t("settings.about.license.studioLicense")}
+            <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+          </a>
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.about.license.libraryLabel")}
+          description={t("settings.about.license.libraryDescription")}
+        >
+          <a
+            href="https://github.com/unslothai/unsloth/blob/main/LICENSE"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="inline-flex items-center gap-1 font-mono text-xs font-medium text-muted-foreground hover:text-foreground"
+          >
+            {t("settings.about.license.libraryLicense")}
+            <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+          </a>
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.about.dangerZone")}>
         <SettingsRow
           destructive={true}
diff --git a/studio/frontend/src/features/settings/tabs/chat-tab.tsx b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
index e37d56e2f7..9a81bdab0d 100644
--- a/studio/frontend/src/features/settings/tabs/chat-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
@@ -13,15 +13,32 @@ import {
   DialogTitle,
 } from "@/components/ui/dialog";
 import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuSub,
+  DropdownMenuSubContent,
+  DropdownMenuSubTrigger,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import {
+  EXPORT_FORMATS_LIST,
+  bulkExportConversationsByScope,
   clearAllChats,
   countAllChats,
   downloadChatExport,
+  importConversationsFromFile,
   useChatRuntimeStore,
 } from "@/features/chat";
 import { useT } from "@/i18n";
-import { Delete02Icon, Download01Icon } from "@hugeicons/core-free-icons";
+import {
+  Delete02Icon,
+  Download01Icon,
+  Upload01Icon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useState } from "react";
+import { useEffect, useRef, useState } from "react";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 
@@ -61,6 +78,25 @@ export function ChatTab() {
     }
   };
 
+  const importInputRef = useRef<HTMLInputElement>(null);
+  const handleImport = async (file: File) => {
+    try {
+      const imported = await importConversationsFromFile(file, null);
+      if (imported === 0) {
+        toast.info(t("settings.chat.importNoConversations"));
+      } else {
+        toast.success(
+          imported === 1
+            ? t("settings.chat.importedOneChat")
+            : t("settings.chat.importedChatCount", { count: imported }),
+        );
+        setCount(await countAllChats().catch(() => count));
+      }
+    } catch {
+      toast.error(t("settings.chat.importFailed"));
+    }
+  };
+
   const handleClear = async () => {
     setClearing(true);
     try {
@@ -172,6 +208,89 @@ export function ChatTab() {
           </Button>
         </SettingsRow>
 
+        <SettingsRow
+          label={t("settings.chat.exportConversations")}
+          description={t("settings.chat.exportConversationsDescription")}
+        >
+          <DropdownMenu>
+            <DropdownMenuTrigger asChild={true}>
+              <Button variant="outline" size="sm" disabled={count === 0}>
+                <HugeiconsIcon
+                  icon={Download01Icon}
+                  className="size-3.5 mr-1.5"
+                />
+                {t("settings.chat.exportConversationsAction")}
+              </Button>
+            </DropdownMenuTrigger>
+            <DropdownMenuContent align="end" className="w-56">
+              {(
+                [
+                  { scope: "recents", label: "exportScopeRecents" },
+                  { scope: "all", label: "exportScopeAll" },
+                ] as const
+              ).map(({ scope, label }) => (
+                <DropdownMenuSub key={scope}>
+                  <DropdownMenuSubTrigger>
+                    <HugeiconsIcon
+                      icon={Download01Icon}
+                      className="size-3.5 mr-1"
+                    />
+                    {t(`settings.chat.${label}`)}
+                  </DropdownMenuSubTrigger>
+                  <DropdownMenuSubContent className="w-56">
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
+                      <DropdownMenuItem
+                        key={`${scope}-m-${fmt}`}
+                        onSelect={() =>
+                          void bulkExportConversationsByScope(scope, fmt, true)
+                        }
+                      >
+                        {fmtLabel} {t("settings.chat.exportCombinedSuffix")}
+                      </DropdownMenuItem>
+                    ))}
+                    <DropdownMenuSeparator />
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
+                      <DropdownMenuItem
+                        key={`${scope}-s-${fmt}`}
+                        onSelect={() =>
+                          void bulkExportConversationsByScope(scope, fmt, false)
+                        }
+                      >
+                        {fmtLabel} {t("settings.chat.exportPerChatSuffix")}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuSubContent>
+                </DropdownMenuSub>
+              ))}
+            </DropdownMenuContent>
+          </DropdownMenu>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.chat.importChats")}
+          description={t("settings.chat.importChatsDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => importInputRef.current?.click()}
+          >
+            <HugeiconsIcon icon={Upload01Icon} className="size-3.5 mr-1.5" />
+            {t("settings.chat.importChatsAction")}
+          </Button>
+          <input
+            ref={importInputRef}
+            type="file"
+            accept=".jsonl,.ndjson,.csv"
+            className="hidden"
+            onChange={(e) => {
+              const file = e.target.files?.[0];
+              e.target.value = "";
+              if (file) void handleImport(file);
+            }}
+          />
+        </SettingsRow>
+
         <SettingsRow
           destructive
           label={t("settings.chat.clearAllChats")}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 8a15eb4985..135f21caf5 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -15,6 +15,10 @@ import { Switch } from "@/components/ui/switch";
 import { usePlatformStore } from "@/config/env";
 import { resetOnboardingDone } from "@/features/auth";
 import { useChatRuntimeStore } from "@/features/chat";
+import {
+  setShowLlamaUpdateBanner,
+  useShowLlamaUpdateBanner,
+} from "@/hooks/use-llama-update-pref";
 import {
   loadHelperPrecacheSettings,
   updateHelperPrecacheSettings,
@@ -68,6 +72,8 @@ const PREFS_KEYS: string[] = [
   "unsloth_user_profile",
   // Guided tour flags
   "tour:studio:v1",
+  // Update notifications
+  "unsloth_show_llama_update_banner",
 ];
 
 // Set by resetAllPrefs so the unmount-commit effect skips writing back the
@@ -106,6 +112,7 @@ export function GeneralTab() {
   const autoTitle = useChatRuntimeStore((s) => s.autoTitle);
   const setAutoTitle = useChatRuntimeStore((s) => s.setAutoTitle);
   const chatOnly = usePlatformStore((s) => s.chatOnly);
+  const showLlamaUpdates = useShowLlamaUpdateBanner();
   const redirectTo = `${pathname}${search}`;
 
   const [draftToken, setDraftToken] = useState(hfToken ?? "");
@@ -315,6 +322,20 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection title={t("settings.general.notifications.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.notifications.showLlamaUpdates")}
+          description={t(
+            "settings.general.notifications.showLlamaUpdatesDescription",
+          )}
+        >
+          <Switch
+            checked={showLlamaUpdates}
+            onCheckedChange={setShowLlamaUpdateBanner}
+          />
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.general.uploads.sectionTitle")}>
         <SettingsRow
           label={t("settings.general.uploads.maxUploadSize")}
diff --git a/studio/frontend/src/features/studio/tour/steps/nav.tsx b/studio/frontend/src/features/studio/tour/steps/nav.tsx
index 41c107888c..8a8da8cab0 100644
--- a/studio/frontend/src/features/studio/tour/steps/nav.tsx
+++ b/studio/frontend/src/features/studio/tour/steps/nav.tsx
@@ -9,7 +9,7 @@ export const studioNavStep: TourStep = {
   title: "Quick orientation",
   body: (
     <>
-      Studio: pick base model, dataset, hyperparams, then start training. After
+      Unsloth: pick base model, dataset, hyperparams, then start training. After
       you start, you’ll see a Training view with live loss/metrics. Chat is for
       testing base vs LoRA adapters. Export packages checkpoints for deployment.{" "}
       <ReadMore href="https://unsloth.ai/docs/get-started/fine-tuning-for-beginners" />
diff --git a/studio/frontend/src/features/studio/training-start-overlay.tsx b/studio/frontend/src/features/studio/training-start-overlay.tsx
index ba80432896..3b0cce658b 100644
--- a/studio/frontend/src/features/studio/training-start-overlay.tsx
+++ b/studio/frontend/src/features/studio/training-start-overlay.tsx
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { MascotImg } from "@/components/mascot-img";
 import {
   AlertDialog,
   AlertDialogAction,
@@ -311,12 +312,8 @@ export function TrainingStartOverlay({
 
   return (
     <div className="pointer-events-none absolute inset-0 z-30 flex items-center justify-center rounded-2xl bg-background/45 backdrop-blur-[1px]">
-      <div className="pointer-events-auto relative flex w-[860px] max-w-[calc(100%-2rem)] flex-col items-center gap-4">
-        <img
-          src="/unsloth-gem.png"
-          alt="Unsloth Studio"
-          className="size-24 object-contain"
-        />
+      <div className="pointer-events-auto relative flex w-[860px] max-w-[calc(100%-2rem)] flex-col items-center">
+        <MascotImg src="unsloth-gem.png" className="size-24 object-contain" />
         <div className="relative w-full">
           <AlertDialog open={cancelDialogOpen} onOpenChange={setCancelDialogOpen}>
             <Button
@@ -358,7 +355,7 @@ export function TrainingStartOverlay({
             </AlertDialogContent>
           </AlertDialog>
           <Terminal
-            className="w-full min-h-[390px] rounded-2xl px-7 py-6 text-left"
+            className="w-full min-h-[390px] rounded-2xl border-0 px-7 py-6 text-left"
             startOnView={false}
             instant={alreadyAnimated}
           >
diff --git a/studio/frontend/src/features/tour/components/guided-tour.tsx b/studio/frontend/src/features/tour/components/guided-tour.tsx
index c13e2fefcc..a7b29fb925 100644
--- a/studio/frontend/src/features/tour/components/guided-tour.tsx
+++ b/studio/frontend/src/features/tour/components/guided-tour.tsx
@@ -332,7 +332,7 @@ export function GuidedTour({
                       <Button
                         variant="ghost"
                         size="icon-sm"
-                        className="text-foreground/60 hover:text-foreground hover:bg-black/[0.05] dark:text-zinc-300/70 dark:hover:text-zinc-100 dark:hover:bg-white/[0.08]"
+                        className="text-foreground/60 hover:text-foreground hover:bg-black/[0.05] dark:text-zinc-300/70 dark:hover:text-zinc-100 dark:hover:bg-white/[0.1]"
                         onClick={() => requestClose("skip")}
                         aria-label="Skip tour"
                       >
@@ -343,7 +343,7 @@ export function GuidedTour({
                     <div className="mt-5 flex items-center justify-between gap-3">
                       <Button
                         variant="ghost"
-                        className="text-foreground/60 hover:text-foreground hover:bg-black/[0.05] dark:text-zinc-300/70 dark:hover:text-zinc-100 dark:hover:bg-white/[0.08]"
+                        className="text-foreground/60 hover:text-foreground hover:bg-black/[0.05] dark:text-zinc-300/70 dark:hover:text-zinc-100 dark:hover:bg-white/[0.1]"
                         onClick={() => requestClose("skip")}
                       >
                         Skip
diff --git a/studio/frontend/src/features/training/stores/training-runtime-store.ts b/studio/frontend/src/features/training/stores/training-runtime-store.ts
index f1966f03fb..97fbd32d57 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -274,7 +274,10 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
         jobId: payload.job_id || state.jobId,
         currentStep: step,
         totalSteps: Math.max(payload.total_steps, state.totalSteps),
-        currentLoss: currentLoss ?? state.currentLoss,
+        // A null loss at a new step means the backend reported a non-finite
+        // loss; clear the display instead of keeping the stale value.
+        currentLoss:
+          currentLoss ?? (step > state.currentStep ? null : state.currentLoss),
         currentLearningRate: currentLearningRate ?? state.currentLearningRate,
         progressPercent: payload.progress_percent,
         currentEpoch: payload.epoch ?? state.currentEpoch,
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index d30c3f75c2..5fd8286d22 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -90,7 +90,8 @@ export interface TrainingRuntimeState {
   currentStep: number;
   totalSteps: number;
   currentEpoch: number;
-  currentLoss: number;
+  // null while the latest reported loss is non-finite
+  currentLoss: number | null;
   currentLearningRate: number;
   progressPercent: number;
   elapsedSeconds: number | null;
diff --git a/studio/frontend/src/hooks/use-llama-update-check.ts b/studio/frontend/src/hooks/use-llama-update-check.ts
new file mode 100644
index 0000000000..5c609558d6
--- /dev/null
+++ b/studio/frontend/src/hooks/use-llama-update-check.ts
@@ -0,0 +1,246 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch, getAuthToken } from "@/features/auth";
+import { useCallback, useEffect, useRef, useState } from "react";
+
+// First check shortly after load, then re-surface as an hourly reminder. The
+// banner stays up until the user explicitly acts on it (X, Update, or
+// Remind me later).
+const FIRST_CHECK_DELAY_MS = 1000;
+const REMINDER_INTERVAL_MS = 60 * 60 * 1000; // ~1 hour
+// "Remind me later" re-surfaces sooner than the hourly reminder.
+const SNOOZE_DELAY_MS = 15 * 60 * 1000; // ~15 minutes
+// While an update is applying, poll the job state at this cadence.
+const JOB_POLL_INTERVAL_MS = 1500;
+
+export interface LlamaUpdateJob {
+  state: "idle" | "running" | "success" | "error";
+  message: string;
+  from_tag: string | null;
+  to_tag: string | null;
+  error: string | null;
+  // Download fraction (0..1) while running, 1 on success, null when unknown.
+  progress: number | null;
+}
+
+export interface LlamaUpdateStatus {
+  supported: boolean;
+  update_available: boolean;
+  installed_tag: string | null;
+  latest_tag: string | null;
+  job: LlamaUpdateJob;
+}
+
+function parseStatus(value: unknown): LlamaUpdateStatus | null {
+  if (!value || typeof value !== "object") return null;
+  const s = value as Record<string, unknown>;
+  const job = (s.job ?? {}) as Record<string, unknown>;
+  return {
+    supported: s.supported === true,
+    update_available: s.update_available === true,
+    installed_tag: typeof s.installed_tag === "string" ? s.installed_tag : null,
+    latest_tag: typeof s.latest_tag === "string" ? s.latest_tag : null,
+    job: {
+      state: (job.state as LlamaUpdateJob["state"]) ?? "idle",
+      message: typeof job.message === "string" ? job.message : "",
+      from_tag: typeof job.from_tag === "string" ? job.from_tag : null,
+      to_tag: typeof job.to_tag === "string" ? job.to_tag : null,
+      error: typeof job.error === "string" ? job.error : null,
+      progress: typeof job.progress === "number" ? job.progress : null,
+    },
+  };
+}
+
+async function fetchStatus(
+  forceRefresh = false,
+): Promise<LlamaUpdateStatus | null> {
+  if (!getAuthToken()) return null;
+  try {
+    const res = await authFetch(
+      `/api/llama/update-status${forceRefresh ? "?force_refresh=true" : ""}`,
+    );
+    if (!res.ok) return null;
+    return parseStatus(await res.json());
+  } catch {
+    return null;
+  }
+}
+
+interface UseLlamaUpdateCheckOptions {
+  enabled?: boolean;
+}
+
+export interface LlamaApplyResult {
+  ok: boolean;
+  tag?: string | null;
+  error?: string | null;
+}
+
+/**
+ * Polls the backend for a newer llama.cpp prebuilt. When one exists, `visible`
+ * becomes true ~1s after load and stays up until the user dismisses it (X),
+ * snoozes it ("Remind me later", ~15 min), or updates; it re-surfaces every
+ * ~hour as a reminder. `apply()` triggers the in-place swap and tracks the
+ * job.
+ */
+export function useLlamaUpdateCheck({
+  enabled = true,
+}: UseLlamaUpdateCheckOptions = {}) {
+  const [status, setStatus] = useState<LlamaUpdateStatus | null>(null);
+  const [visible, setVisible] = useState(false);
+  const [applying, setApplying] = useState(false);
+  const pollTimer = useRef<ReturnType<typeof setInterval> | null>(null);
+  const snoozeTimer = useRef<ReturnType<typeof setTimeout> | null>(null);
+
+  const clearPollTimer = useCallback(() => {
+    if (pollTimer.current) {
+      clearInterval(pollTimer.current);
+      pollTimer.current = null;
+    }
+  }, []);
+
+  // Poll the job to completion. Shared by apply() and surfaceIfAvailable() so a
+  // job is tracked once whoever noticed it; onDone resolves with the result.
+  const startJobPoll = useCallback(
+    (onDone?: (result: LlamaApplyResult) => void) => {
+      clearPollTimer();
+      pollTimer.current = setInterval(async () => {
+        const s = await fetchStatus();
+        if (!s) return;
+        setStatus(s);
+        if (s.job.state === "running") return;
+        clearPollTimer();
+        setApplying(false);
+        if (s.job.state === "success") {
+          setVisible(false);
+          onDone?.({ ok: true, tag: s.job.to_tag });
+        } else if (s.job.state === "error") {
+          // Leave the banner up so the user can retry; clearing applying drops
+          // the "Updating..." state.
+          onDone?.({ ok: false, error: s.job.error });
+        } else {
+          // idle without a terminal result (job reset): stop tracking.
+          onDone?.({ ok: false, error: "update did not complete" });
+        }
+      }, JOB_POLL_INTERVAL_MS);
+    },
+    [clearPollTimer],
+  );
+
+  // Surface the banner when an update is available; it stays up until dismissed.
+  const surfaceIfAvailable = useCallback(
+    (next: LlamaUpdateStatus | null) => {
+      if (!next) return;
+      setStatus(next);
+      if (next.job.state === "running") {
+        // Swap in progress (e.g. another tab): keep the banner up and track the
+        // job so "Updating..." clears when it finishes instead of sticking.
+        setApplying(true);
+        setVisible(true);
+        if (!pollTimer.current) startJobPoll();
+        return;
+      }
+      if (next.update_available) {
+        setVisible(true);
+      }
+    },
+    [startJobPoll],
+  );
+
+  useEffect(() => {
+    if (!enabled) return;
+    let canceled = false;
+
+    const firstTimer = setTimeout(() => {
+      fetchStatus().then((s) => {
+        if (!canceled) surfaceIfAvailable(s);
+      });
+    }, FIRST_CHECK_DELAY_MS);
+
+    const reminder = setInterval(() => {
+      fetchStatus().then((s) => {
+        if (!canceled) surfaceIfAvailable(s);
+      });
+    }, REMINDER_INTERVAL_MS);
+
+    return () => {
+      canceled = true;
+      clearTimeout(firstTimer);
+      clearInterval(reminder);
+      clearPollTimer();
+      if (snoozeTimer.current) {
+        clearTimeout(snoozeTimer.current);
+        snoozeTimer.current = null;
+      }
+    };
+  }, [enabled, surfaceIfAvailable, clearPollTimer]);
+
+  const dismiss = useCallback(() => {
+    setVisible(false);
+  }, []);
+
+  // Hide now, re-check and re-surface after SNOOZE_DELAY_MS.
+  const snooze = useCallback(() => {
+    setVisible(false);
+    if (snoozeTimer.current) clearTimeout(snoozeTimer.current);
+    snoozeTimer.current = setTimeout(() => {
+      snoozeTimer.current = null;
+      fetchStatus().then(surfaceIfAvailable);
+    }, SNOOZE_DELAY_MS);
+  }, [surfaceIfAvailable]);
+
+  const apply = useCallback(async (): Promise<LlamaApplyResult> => {
+    if (applying) return { ok: false, error: "already running" };
+    setApplying(true);
+    setVisible(true);
+    let action: {
+      started?: boolean;
+      reason?: string | null;
+      message?: string | null;
+    } | null = null;
+    try {
+      const res = await authFetch("/api/llama/update", { method: "POST" });
+      if (!res.ok) {
+        setApplying(false);
+        return { ok: false, error: `HTTP ${res.status}` };
+      }
+      try {
+        action = await res.json();
+      } catch {
+        action = null;
+      }
+    } catch (e) {
+      setApplying(false);
+      return { ok: false, error: String(e) };
+    }
+
+    // 200 without a started job (no marker / installer missing) leaves it idle,
+    // so surface the reason instead of polling forever. already_running is the
+    // exception: a job is in flight, so track it to completion below.
+    if (
+      action &&
+      action.started === false &&
+      action.reason !== "already_running"
+    ) {
+      setApplying(false);
+      return {
+        ok: false,
+        error: action.message ?? action.reason ?? "update was not started",
+      };
+    }
+
+    return await new Promise<LlamaApplyResult>((resolve) =>
+      startJobPoll(resolve),
+    );
+  }, [applying, startJobPoll]);
+
+  return {
+    status: enabled ? status : null,
+    visible: enabled && visible,
+    applying,
+    apply,
+    dismiss,
+    snooze,
+  };
+}
diff --git a/studio/frontend/src/hooks/use-llama-update-pref.ts b/studio/frontend/src/hooks/use-llama-update-pref.ts
new file mode 100644
index 0000000000..33a21965c4
--- /dev/null
+++ b/studio/frontend/src/hooks/use-llama-update-pref.ts
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useSyncExternalStore } from "react";
+
+// Whether the llama.cpp update banner may appear. On by default; only an
+// explicit "false" (Settings -> General -> Notifications) disables it.
+const STORAGE_KEY = "unsloth_show_llama_update_banner";
+
+const listeners = new Set<() => void>();
+
+export function getShowLlamaUpdateBanner(): boolean {
+  try {
+    return localStorage.getItem(STORAGE_KEY) !== "false";
+  } catch {
+    return true;
+  }
+}
+
+export function setShowLlamaUpdateBanner(show: boolean): void {
+  try {
+    if (show) {
+      // Remove rather than store "true" so the default stays on.
+      localStorage.removeItem(STORAGE_KEY);
+    } else {
+      localStorage.setItem(STORAGE_KEY, "false");
+    }
+  } catch {
+    // storage unavailable
+  }
+  for (const listener of listeners) listener();
+}
+
+function subscribe(listener: () => void): () => void {
+  listeners.add(listener);
+  // Sync toggles made in another tab.
+  const onStorage = (event: StorageEvent) => {
+    if (event.key === STORAGE_KEY) listener();
+  };
+  window.addEventListener("storage", onStorage);
+  return () => {
+    listeners.delete(listener);
+    window.removeEventListener("storage", onStorage);
+  };
+}
+
+export function useShowLlamaUpdateBanner(): boolean {
+  return useSyncExternalStore(subscribe, getShowLlamaUpdateBanner);
+}
diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts
index b376cb3a37..086cfc8090 100644
--- a/studio/frontend/src/hooks/use-tauri-backend.ts
+++ b/studio/frontend/src/hooks/use-tauri-backend.ts
@@ -56,23 +56,23 @@ function wait(ms: number) {
 function externalConflictMessage(preflight: DesktopPreflightResult) {
   if (preflight.reason === "desktop_owned_backend_active") {
     return preflight.port
-      ? `A desktop-owned Studio server for this install is already running on port ${preflight.port}. Quit the other desktop app instance, then try again.`
-      : "A desktop-owned Studio server for this install is already running. Quit the other desktop app instance, then try again.";
+      ? `A desktop-owned Unsloth server for this install is already running on port ${preflight.port}. Quit the other desktop app instance, then try again.`
+      : "A desktop-owned Unsloth server for this install is already running. Quit the other desktop app instance, then try again.";
   }
 
   if (preflight.reason === "desktop_owned_backend_starting") {
-    return "The desktop-owned Studio backend is still starting. Wait a moment, then try again.";
+    return "The desktop-owned Unsloth backend is still starting. Wait a moment, then try again.";
   }
 
   if (preflight.reason?.startsWith("desktop_owned_backend_unmanageable:")) {
     return preflight.port
-      ? `A desktop-owned Studio backend on port ${preflight.port} cannot be safely controlled by this desktop app. Stop that backend, then reopen Studio.`
-      : "A desktop-owned Studio backend cannot be safely controlled by this desktop app. Stop that backend, then reopen Studio.";
+      ? `A desktop-owned Unsloth backend on port ${preflight.port} cannot be safely controlled by this desktop app. Stop that backend, then reopen Unsloth.`
+      : "A desktop-owned Unsloth backend cannot be safely controlled by this desktop app. Stop that backend, then reopen Unsloth.";
   }
 
   return preflight.port
-    ? `A Studio server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
-    : "A Studio server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
+    ? `A Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
+    : "A Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
 }
 
 async function waitForManagedServerReady(
@@ -248,8 +248,8 @@ export function useTauriBackend() {
           } else {
             setBackendError(
               preflight.disposition === "owned_stale"
-                ? "Desktop-owned Studio backend is too old for this desktop app. Run `unsloth studio update`, then restart Studio."
-                : "Managed Studio install is too old. Run `unsloth studio update`.",
+                ? "Desktop-owned Unsloth backend is too old for this desktop app. Run `unsloth studio update`, then restart Unsloth."
+                : "Managed Unsloth install is too old. Run `unsloth studio update`.",
             );
           }
           return;
@@ -304,7 +304,7 @@ export function useTauriBackend() {
       if (msg.includes("already running")) {
         startingRef.current = false;
         setBackendError(
-          "Managed server is already running but did not report a port. Restart Studio and try again.",
+          "Managed server is already running but did not report a port. Restart Unsloth and try again.",
         );
         return;
       }
@@ -596,7 +596,7 @@ export function useTauriBackend() {
       const detail =
         event instanceof CustomEvent && typeof event.detail === "string"
           ? event.detail
-          : "Desktop authentication failed. Update or repair the managed Studio install, then restart Studio.";
+          : "Desktop authentication failed. Update or repair the managed Unsloth install, then restart Unsloth.";
       setAuthFailure(detail);
     };
     window.addEventListener("tauri-auth-failed", onAuthFailed);
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 1492c7bb2b..19863d42bb 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -83,7 +83,7 @@ export const en = {
     title: "Settings",
     dialog: {
       title: "Settings",
-      description: "Manage your Unsloth Studio preferences.",
+      description: "Manage your Unsloth preferences.",
       closeAriaLabel: "Close settings",
     },
     tabs: {
@@ -93,11 +93,11 @@ export const en = {
       chat: "Chat",
       connections: "Connections",
       apiKeys: "API",
-      about: "Help",
+      about: "About",
     },
     general: {
       title: "General",
-      description: "Global preferences for Unsloth Studio.",
+      description: "Global preferences for Unsloth.",
       account: "Account",
       huggingFaceToken: "Hugging Face token",
       huggingFaceTokenDescription:
@@ -112,38 +112,44 @@ export const en = {
         sectionTitle: "Helper LLM",
         preloadOnStartup: "Pre-cache Helper LLM on startup",
         preloadOnStartupDescription:
-          "Download and cache the AI Assist helper model in the background when Studio starts. Off by default; AI Assist can still download it on demand when clicked.",
+          "Download the AI Assist helper model in the background on startup. Off by default; AI Assist can still fetch it on demand.",
         disabledByEnv:
           "Disabled by UNSLOTH_HELPER_MODEL_DISABLE in the backend environment.",
         loadError: "Failed to load Helper LLM settings.",
         saveError: "Failed to save Helper LLM settings.",
       },
+      notifications: {
+        sectionTitle: "Notifications",
+        showLlamaUpdates: "llama.cpp update notifications",
+        showLlamaUpdatesDescription:
+          "Notify when a newer llama.cpp build is available. Turn off if you only train.",
+      },
       gettingStarted: "Getting started",
       startOnboarding: "Start onboarding",
       startOnboardingDescription:
-        "Open the setup wizard again without changing your account.",
+        "Reopen the setup wizard without changing your account.",
       startOnboardingAction: "Start onboarding",
       uploads: {
         sectionTitle: "Uploads",
         maxUploadSize: "Training dataset upload cap",
         maxUploadSizeDescription:
-          "Applies to training dataset uploads. Default is {defaultSize} MB.",
+          "Default is {defaultSize} MB.",
       },
       resetPreferences: {
         sectionTitle: "Danger zone",
         label: "Reset all local preferences",
         description:
-          "Clears local-only preferences. Chats, API access, and DB-backed settings are not affected.",
+          "Clears local-only preferences. Chats, API access, and DB-backed settings are kept.",
         action: "Reset preferences",
         confirmTitle: "Reset all local preferences?",
         confirmDescription:
-          "This clears local-only preferences, then reloads Studio. Chats, API access, and DB-backed settings are not affected.",
+          "Clears local-only preferences and reloads Unsloth. Chats, API access, and DB-backed settings are kept.",
         confirmAction: "Reset and reload",
       },
     },
     profile: {
       title: "Profile",
-      description: "Update how your profile appears in Studio.",
+      description: "How your profile appears in Unsloth.",
       changePicture: "Change profile picture",
       displayName: "Display name",
       nameSaved: "Profile name saved",
@@ -163,7 +169,7 @@ export const en = {
       theme: {
         title: "Theme",
         label: "Color scheme",
-        description: "Choose light, dark, or follow your system.",
+        description: "Light, dark, or follow your system.",
         system: "System",
         light: "Light",
         dark: "Dark",
@@ -171,7 +177,7 @@ export const en = {
       language: {
         title: "Language",
         label: "Display language",
-        description: "Choose the language used by Studio.",
+        description: "The language used by Unsloth.",
       },
       layout: {
         title: "Layout",
@@ -182,24 +188,40 @@ export const en = {
     },
     chat: {
       title: "Chat",
-      description: "Manage your chat history stored on this device.",
+      description: "Manage chat history stored on this device.",
       artifacts: {
         title: "Artifacts",
         collapseHtmlBlocks: "Collapse HTML blocks",
         collapseHtmlBlocksDescription:
-          "Artifacts mode collapses full HTML fallback automatically. Turn this on to also collapse full fenced HTML documents when Artifacts is off.",
+          "Artifacts mode collapses full HTML automatically. Turn on to also collapse fenced HTML documents when Artifacts is off.",
         allowNetworkAccess: "Allow artifact network access",
         allowNetworkAccessDescription:
-          "Let artifact previews load scripts, styles, fonts, media, fetch, and WebSocket resources from HTTP(S) CDNs. Keep off for fully offline previews.",
+          "Let artifact previews load scripts, styles, fonts, media, and network resources from CDNs. Keep off for fully offline previews.",
       },
       data: "Data",
       exportHistory: "Export chat history",
       exportHistoryDescription:
-        "Download all chats and messages as a JSON file.",
+        "Download all chats and messages as JSON.",
       exportAction: "Export",
       exportingAction: "Exporting...",
+      exportConversations: "Export Recents and Projects",
+      exportConversationsDescription:
+        "Download Recents or Recents plus project chats as Raw JSONL, CSV, or ShareGPT JSONL, combined or per chat.",
+      exportConversationsAction: "Export",
+      exportScopeRecents: "Recents",
+      exportScopeAll: "Recents + Projects",
+      exportCombinedSuffix: "(combined)",
+      exportPerChatSuffix: "(per chat)",
+      importChats: "Import chats",
+      importChatsDescription:
+        "Import a JSONL, NDJSON, or CSV export into Recents.",
+      importChatsAction: "Import",
+      importNoConversations: "No conversations found in file.",
+      importedOneChat: "Imported 1 conversation to Recents.",
+      importedChatCount: "Imported {count} conversations to Recents.",
+      importFailed: "Import failed.",
       clearHistory: "Clear chat history",
-      clearHistoryDescription: "Delete local chat history from this device.",
+      clearHistoryDescription: "Delete chat history from this device.",
       clearAction: "Clear",
       clearAllChats: "Clear all chats",
       clearAllChatsDescription: "Permanently delete every chat on this device.",
@@ -212,7 +234,7 @@ export const en = {
       clearOneChatTitle: "Clear 1 chat?",
       clearChatsTitle: "Clear {count} chats?",
       clearChatsConfirmDescription:
-        "This permanently deletes every chat and message stored on this device. This cannot be undone.",
+        "Permanently deletes every chat on this device. This cannot be undone.",
       clearingAction: "Clearing...",
       clearOneChatAction: "Clear 1 chat",
       clearChatCountAction: "Clear {count} chats",
@@ -235,12 +257,12 @@ export const en = {
     },
     connections: {
       title: "Connections",
-      description: "Manage providers and external service connections.",
+      description: "Manage providers and external connections.",
     },
     apiKeys: {
       title: "API",
       description:
-        "Access Unsloth programmatically via the OpenAI-compatible API.",
+        "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
       newBadge: "New",
@@ -280,62 +302,71 @@ export const en = {
       revokeToken: "Revoke token",
       revokeTitle: 'Revoke access token "{name}"?',
       revokeDescription:
-        "Applications using this token will immediately lose access. This cannot be undone.",
+        "Apps using this token immediately lose access. This cannot be undone.",
       revokeAction: 'Revoke "{name}"',
       revoking: "Revoking...",
     },
     about: {
       title: "About",
       description:
-        "Documentation, release notes, feedback, and Studio build info.",
-      studioVersion: "Studio Version",
+        "Docs, release notes, feedback, and build info.",
+      studioVersion: "Unsloth Version",
       packageVersion: "Package Version",
-      updates: "Updates",
+      updates: "Update",
       help: "Help",
       documentation: "Documentation",
       releaseNotes: "Release notes",
       whatsNew: "What's new",
       feedback: "Feedback",
       reportIssue: "Report an issue",
+      license: {
+        sectionTitle: "License",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription:
+          "Open source under the GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Licensed under Apache 2.0.",
+      },
       dangerZone: "Danger zone",
       shutDownStudio: "Shut down Unsloth Studio",
       shutDownStudioDescription:
-        "Stops the Studio server process and ends your session.",
+        "Stops the Unsloth server and ends your session.",
       shutDown: "Shut down",
       update: {
         title: "Update Unsloth Studio",
-        openPowerShell: "Open PowerShell and run:",
-        openTerminal: "Open Terminal and run:",
         commandText: "{label} text",
         copied: "Copied",
         copyCommand: "Copy command",
         commandCopied: "{label} copied",
         copyNamedCommand: "Copy {label}",
-        checkingInstall: "Checking how Studio was installed...",
+        checkingInstall: "Checking how Unsloth was installed...",
+        installIntro: "To install or update Unsloth:",
+        localUpdateHeading: "Local update",
+        installCommandUnix: "macOS/Linux install command",
+        installCommandWindows: "Windows install command",
         localInstallDetected:
-          "Source or local install detected. To avoid replacing it with PyPI, update from the checkout or source you originally installed from.",
-        pullThenUpdate:
-          "Pull latest changes from your Unsloth repo checkout, then update Studio locally:",
+          "Local install detected. Update from your original checkout to avoid replacing it with PyPI.",
+        pullThenUpdate: "Pull the latest changes, then run the local installer:",
         gitPullCommand: "git pull command",
-        localUpdateCommand: "local update command",
-        localInstallerFallback:
-          "If the Studio update command is unavailable, run the local installer from that checkout:",
         localInstallerCommand: "local installer command",
         sourceInstallDetected:
-          "This looks like a source or VCS package install. Reinstall from the original local path or Git URL you used.",
+          "Source or VCS package install detected. Reinstall from the original local path or Git URL.",
         repoCheckoutFallback:
-          "If you still have the Unsloth repo checkout, run the local installer from that checkout:",
-        restartAfterUpdate:
-          "Restart Studio after updating for changes to take effect.",
+          "If you still have the repo checkout, run the local installer from it:",
+        restartAfterUpdate: "Restart Unsloth after updating.",
+        desktopManaged:
+          "The desktop app keeps its bundled backend updated and will prompt when a new version is available.",
         unknownInstall:
-          "Studio could not detect how it was installed. Check how you installed Studio first, then choose the matching update path.",
-        curlOrPypi: "For curl or PyPI installs, run:",
-        updateCommand: "update command",
+          "Could not detect how Unsloth was installed. For installer or PyPI installs, use the commands above.",
         localCheckout:
-          "For local checkout installs, update from that checkout instead and use the local update command:",
-        fallbackInstruction:
-          "If that fails or unsloth studio update is unavailable, run:",
-        fallbackCommand: "fallback command",
+          "For local checkout installs, run the local installer from that checkout:",
+        docs: "Install docs:",
+        docsInstall: "Installation",
+        docsUpdating: "Updating",
+        docsMac: "Mac",
+        docsWindows: "Windows",
       },
     },
   },
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 3cfea1ae7a..5173f9fa74 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -93,7 +93,7 @@ export const zhCN = {
       chat: "聊天",
       connections: "连接",
       apiKeys: "API",
-      about: "帮助",
+      about: "关于",
     },
     general: {
       title: "通用",
@@ -106,6 +106,12 @@ export const zhCN = {
       chatDefaults: "聊天默认设置",
       autoTitleNewChats: "自动为新聊天命名",
       autoTitleNewChatsDescription: "根据第一条消息生成简短标题。",
+      notifications: {
+        sectionTitle: "通知",
+        showLlamaUpdates: "llama.cpp 更新通知",
+        showLlamaUpdatesDescription:
+          "有新的 llama.cpp 构建时提醒。如果只用于训练可以关闭。",
+      },
       gettingStarted: "入门",
       startOnboarding: "开始引导",
       startOnboardingDescription: "重新打开设置向导，不会更改你的账号。",
@@ -124,13 +130,13 @@ export const zhCN = {
         action: "重置偏好设置",
         confirmTitle: "重置所有本地偏好设置？",
         confirmDescription:
-          "这会清除仅保存在本地的偏好设置，然后重新加载 Studio。聊天、API 访问权限和数据库中的设置不会受到影响。",
+          "这会清除仅保存在本地的偏好设置，然后重新加载 Unsloth。聊天、API 访问权限和数据库中的设置不会受到影响。",
         confirmAction: "重置并重新加载",
       },
     },
     profile: {
       title: "个人资料",
-      description: "更新你在 Studio 中显示的个人资料。",
+      description: "更新你在 Unsloth 中显示的个人资料。",
       changePicture: "更换头像",
       displayName: "显示名称",
       nameSaved: "个人资料名称已保存",
@@ -150,7 +156,7 @@ export const zhCN = {
       language: {
         title: "语言",
         label: "显示语言",
-        description: "选择 Studio 使用的语言。",
+        description: "选择 Unsloth 使用的语言。",
       },
       theme: {
         title: "主题",
@@ -174,6 +180,22 @@ export const zhCN = {
       exportHistoryDescription: "将所有聊天和消息下载为 JSON 文件。",
       exportAction: "导出",
       exportingAction: "导出中...",
+      exportConversations: "导出最近对话与项目",
+      exportConversationsDescription:
+        "将最近对话（或加上项目内对话）导出为 Raw JSONL、CSV 或 ShareGPT JSONL，可合并为单个文件或每个对话一个文件。",
+      exportConversationsAction: "导出",
+      exportScopeRecents: "最近对话",
+      exportScopeAll: "最近对话 + 项目",
+      exportCombinedSuffix: "（合并）",
+      exportPerChatSuffix: "（每个对话）",
+      importChats: "导入对话",
+      importChatsDescription:
+        "从 JSONL、NDJSON 或 CSV 导出文件将对话导入到最近对话。",
+      importChatsAction: "导入",
+      importNoConversations: "文件中未找到对话。",
+      importedOneChat: "已导入 1 个对话到最近对话。",
+      importedChatCount: "已导入 {count} 个对话到最近对话。",
+      importFailed: "导入失败。",
       clearHistory: "清除聊天记录",
       clearHistoryDescription: "从此设备删除本地聊天记录。",
       clearAction: "清除",
@@ -258,9 +280,9 @@ export const zhCN = {
       revoking: "撤销中...",
     },
     about: {
-      title: "帮助",
-      description: "文档、发布说明、反馈和 Studio 构建信息。",
-      studioVersion: "Studio 版本",
+      title: "关于",
+      description: "文档、发布说明、反馈和 Unsloth 构建信息。",
+      studioVersion: "Unsloth 版本",
       packageVersion: "包版本",
       updates: "更新",
       help: "帮助",
@@ -269,43 +291,52 @@ export const zhCN = {
       whatsNew: "最新内容",
       feedback: "反馈",
       reportIssue: "报告问题",
+      license: {
+        sectionTitle: "许可证",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "基于 GNU AGPL v3.0 开源。",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "基于 Apache License 2.0 许可。",
+      },
       dangerZone: "危险区域",
       shutDownStudio: "关闭 Unsloth Studio",
-      shutDownStudioDescription: "停止 Studio 服务进程并结束你的会话。",
+      shutDownStudioDescription: "停止 Unsloth 服务进程并结束你的会话。",
       shutDown: "关闭",
       update: {
         title: "更新 Unsloth Studio",
-        openPowerShell: "打开 PowerShell 并运行：",
-        openTerminal: "打开终端并运行：",
         commandText: "{label} 文本",
         copied: "已复制",
         copyCommand: "复制命令",
         commandCopied: "{label} 已复制",
         copyNamedCommand: "复制 {label}",
-        checkingInstall: "正在检查 Studio 的安装方式...",
+        checkingInstall: "正在检查 Unsloth 的安装方式...",
+        installIntro: "安装或更新 Unsloth：",
+        localUpdateHeading: "本地更新",
+        installCommandUnix: "macOS/Linux 安装命令",
+        installCommandWindows: "Windows 安装命令",
         localInstallDetected:
-          "检测到源码或本地安装。为避免替换为 PyPI 版本，请从最初安装时使用的 checkout 或源码位置更新。",
-        pullThenUpdate:
-          "从你的 Unsloth 仓库 checkout 拉取最新变更，然后本地更新 Studio：",
+          "检测到本地安装。请从最初的 checkout 更新，以免被 PyPI 版本替换。",
+        pullThenUpdate: "拉取最新变更，然后运行本地安装器：",
         gitPullCommand: "git pull 命令",
-        localUpdateCommand: "本地更新命令",
-        localInstallerFallback:
-          "如果 Studio 更新命令不可用，请从该 checkout 运行本地安装器：",
         localInstallerCommand: "本地安装器命令",
         sourceInstallDetected:
           "这看起来是源码或 VCS 包安装。请从最初使用的本地路径或 Git URL 重新安装。",
         repoCheckoutFallback:
           "如果你仍保留 Unsloth 仓库 checkout，请从该 checkout 运行本地安装器：",
-        restartAfterUpdate: "更新后重启 Studio，使变更生效。",
+        restartAfterUpdate: "更新后请重启 Unsloth。",
+        desktopManaged:
+          "桌面应用会自动更新其内置后端，有新版本时会提示。",
         unknownInstall:
-          "Studio 无法检测安装方式。请先确认你如何安装 Studio，然后选择匹配的更新方式。",
-        curlOrPypi: "对于 curl 或 PyPI 安装，请运行：",
-        updateCommand: "更新命令",
+          "Unsloth 无法检测安装方式。如果你使用一键安装器或 PyPI 安装，请使用上面的命令。",
         localCheckout:
-          "对于本地 checkout 安装，请改为从该 checkout 更新并使用本地更新命令：",
-        fallbackInstruction:
-          "如果失败，或 unsloth studio update 不可用，请运行：",
-        fallbackCommand: "备用命令",
+          "对于本地 checkout 安装，请改为从该 checkout 运行本地安装器：",
+        docs: "安装文档：",
+        docsInstall: "安装",
+        docsUpdating: "更新",
+        docsMac: "Mac",
+        docsWindows: "Windows",
       },
     },
   },
diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index 07bef147d7..9e394a31e7 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -113,7 +113,8 @@
 	--ease-out-quart: cubic-bezier(0.165, 0.84, 0.44, 1);
 	--ease-out-cubic: cubic-bezier(0.215, 0.61, 0.355, 1);
 
-	--background: oklch(1 0 0);
+	/* Slightly warm off white; every page (and the chat surface) shares it. */
+	--background: #fefefd;
 	--foreground: oklch(0.2686 0 0);
 	--card: oklch(1 0 0);
 	--card-foreground: oklch(0.1281 0.0179 169.2764);
@@ -137,15 +138,15 @@
 	--chart-4: oklch(0.6926 0.1112 346.5775);
 	--chart-5: oklch(0.7497 0.1003 85.0057);
 	--radius: 1.1rem;
-	/* Match the page background so the sidebar reads as one surface with the
-	   content; the faint --sidebar-border on its right edge is the separator. */
-	--sidebar: oklch(1 0 0);
+	/* White sidebar against the warm off-white page; the tone difference is
+	   the separator now that the right-edge divider is gone. */
+	--sidebar: #ffffff;
 	--sidebar-foreground: oklch(0.1281 0.0179 169.2764);
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
 	--sidebar-accent: oklch(0.96 0.0279 166.55);
 	--sidebar-accent-foreground: oklch(0.2868 0.0649 159.9823);
-	--sidebar-border: oklch(0.92 0 0);
+	--sidebar-border: #f2f2f2;
 	--sidebar-ring: #17b88b;
 	--destructive-foreground: oklch(1 0 0);
 	--font-sans: "Inter Variable", ui-sans-serif, sans-serif, system-ui;
@@ -233,7 +234,8 @@
 	/* Exact palette from apps/studio/index_chat.html mockup. Using hex so the
 	   rendered surface matches the mockup pixel-for-pixel — OKLCH conversion
 	   drifted ~3% darker and shifted the neutral hue. */
-	--background: #1f2023;
+	/* Slightly darker than --sidebar so the sidebar reads as its own surface. */
+	--background: #1a1b1e;
 	--foreground: #ececee;
 	--card: #2d2e32;
 	--card-foreground: #ececee;
@@ -245,7 +247,8 @@
 	--secondary-foreground: #ececee;
 	--muted: #2e3035;
 	--muted-foreground: #999999;
-	--accent: #2e3035;
+	/* One step lighter than --popover so menu item hover reads clearly. */
+	--accent: #3a3d43;
 	--accent-foreground: #ececee;
 	--destructive: oklch(0.6368 0.2078 25.3313);
 	/* --border / --input one step lighter than --muted so outlines and form
@@ -259,14 +262,14 @@
 	--chart-3: oklch(0.7554 0.1285 197.339);
 	--chart-4: oklch(0.7503 0.1199 346.7805);
 	--chart-5: oklch(0.799 0.1196 84.6633);
-	/* Match the page background (dark); --sidebar-border is the right-edge separator. */
+	/* A step lighter than --background; the tone difference is the separator. */
 	--sidebar: #1f2023;
 	--sidebar-foreground: #ececee;
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
-	--sidebar-accent: #2f2f31;
+	--sidebar-accent: #2e3035;
 	--sidebar-accent-foreground: #ececee;
-	--sidebar-border: #2d2d2f;
+	--sidebar-border: #2d2e32;
 	--sidebar-ring: #17b88b;
 	--destructive-foreground: oklch(1 0 0);
 	/* Match light's radius so every rounded-* element is the same in both themes. */
@@ -293,7 +296,7 @@
 
 	--nav-fg: #c7c7c4;
 	--nav-fg-muted: #96979b;
-	--nav-surface-hover: #2e2e30;
+	--nav-surface-hover: #2e3035;
 	--nav-icon-idle: #5c5c5c;
 	--nav-beta-border: #3a3c3f;
 	--panel-surface-hover: #3a3c42;
@@ -305,7 +308,7 @@
 	--panel-surface: var(--background);
 	--panel-surface-fg: var(--foreground);
 	--panel-input-surface: #2a2b2e;
-	--panel-input-surface-hover: #2e3033;
+	--panel-input-surface-hover: #34373c;
 	/* Soft neutral gray for muted text and sliders. Pure-ish #ababab
 	   reads as quiet on the dark panel without going colored. */
 	--panel-surface-fg-muted: #ababab;
@@ -521,7 +524,7 @@
 	}
 
 	.nav-icon-btn {
-		@apply inline-flex h-7 w-7 items-center justify-center rounded-[10px] transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring;
+		@apply inline-flex h-7 w-7 items-center justify-center rounded-full transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring;
 	}
 
 	/* Standard icon size — drives every nav/menu/action-bar icon
@@ -541,31 +544,28 @@
 	.sidebar-nav-btn {
 		color: var(--nav-fg);
 	}
+	/* Gemini-style: hovering or opening the row action (3 dots) shows only
+	   the action's circle; the row highlights only when hovering the row
+	   itself. */
 	.sidebar-nav-btn:hover,
 	.sidebar-nav-btn[data-active="true"],
 	.sidebar-nav-btn[data-state="open"],
-	.group\/project-item:hover .sidebar-nav-btn,
-	.group\/project-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.group\/project-chat-item:hover .sidebar-nav-btn,
-	.group\/project-chat-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.group\/recent-item:hover .sidebar-nav-btn,
-	.group\/recent-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.group\/run-item:hover .sidebar-nav-btn,
-	.group\/run-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn {
+	.group\/project-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.group\/project-chat-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.group\/projects-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.group\/recent-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.group\/run-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn {
 		background-color: var(--nav-surface-hover) !important;
 		color: #000 !important;
 	}
 	.dark .sidebar-nav-btn:hover,
 	.dark .sidebar-nav-btn[data-active="true"],
 	.dark .sidebar-nav-btn[data-state="open"],
-	.dark .group\/project-item:hover .sidebar-nav-btn,
-	.dark .group\/project-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.dark .group\/project-chat-item:hover .sidebar-nav-btn,
-	.dark .group\/project-chat-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.dark .group\/recent-item:hover .sidebar-nav-btn,
-	.dark .group\/recent-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
-	.dark .group\/run-item:hover .sidebar-nav-btn,
-	.dark .group\/run-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn {
+	.dark .group\/project-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.dark .group\/project-chat-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.dark .group\/projects-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.dark .group\/recent-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.dark .group\/run-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn {
 		color: #fff !important;
 	}
 
@@ -576,19 +576,21 @@
 		@apply opacity-100 pointer-events-auto;
 	}
 	.sidebar-row-action-glyph {
-		@apply inline-flex size-6 items-center justify-center rounded-[10px] text-sidebar-foreground/55;
+		@apply inline-flex size-6 items-center justify-center rounded-full text-sidebar-foreground/55;
 	}
 
 	/* Branch picker chevron buttons sit beside action bar icon buttons
-	   (size-8, rounded-[10px]). Height + radius match for visual
+	   (size-8, rounded-full). Height + radius match for visual
 	   alignment, but width is tighter so the small chevron glyph reads
 	   as a compact control rather than a full-size icon button. */
 	.aui-branch-chevron-btn {
-		@apply inline-flex h-8 w-6 cursor-pointer items-center justify-center rounded-[10px] p-0 text-chat-icon-fg transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring disabled:cursor-not-allowed disabled:opacity-30 disabled:hover:bg-transparent;
+		@apply inline-flex h-8 w-6 cursor-pointer items-center justify-center rounded-full p-0 text-chat-icon-fg transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring disabled:cursor-not-allowed disabled:opacity-30 disabled:hover:bg-transparent;
 	}
 	.sidebar-row-action:hover .sidebar-row-action-glyph,
 	.sidebar-row-action[data-state="open"] .sidebar-row-action-glyph {
-		@apply bg-nav-surface-hover text-nav-fg;
+		@apply text-nav-fg;
+		/* A step darker than the row hover so the circle reads on top of it. */
+		background-color: color-mix(in srgb, var(--foreground) 11%, transparent);
 	}
 	.dark .sidebar-row-action:hover .sidebar-row-action-glyph,
 	.dark .sidebar-row-action[data-state="open"] .sidebar-row-action-glyph {
@@ -596,7 +598,7 @@
 	}
 
 	.sidebar-sticky-label {
-		@apply rounded-none bg-sidebar pt-0 pb-[8px] pl-[18px] pr-4 text-[14px]! leading-[17px] font-medium normal-case focus-visible:ring-0! focus-visible:outline-none transition-shadow duration-150;
+		@apply rounded-none bg-sidebar pt-0 pb-[8px] pl-[20px] pr-4 text-[14px]! leading-[17px] font-medium normal-case focus-visible:ring-0! focus-visible:outline-none transition-shadow duration-150;
 		/* Muted section-header gray, matching Gemini's "Notebooks"/"Recents".
 		   Lightened from #5f6368 so the label reads as a header, clearly
 		   lighter than the near-black nav items. */
@@ -618,7 +620,7 @@
 	   prompt button, and the chat-template textarea so all three read
 	   as one quiet gray family, matching the focused-edit-number tint. */
 	.panel-input-group {
-		@apply !h-9 min-h-9 min-w-0 items-stretch gap-0 rounded-[10px] pr-0 transition-colors focus-within:ring-0 focus-within:shadow-none;
+		@apply !h-9 min-h-9 min-w-0 items-stretch gap-0 rounded-full pr-0 transition-colors focus-within:ring-0 focus-within:shadow-none;
 		border: 0 !important;
 		background-color: var(--panel-input-surface);
 	}
@@ -714,13 +716,12 @@
 	/* Inline numeric input — used for slider values and Context Length.
 	   Designed to read as *editable text* rather than a pill button so
 	   it doesn't mirror the Select/dropdown components on the panel.
-	   Default: transparent box, no border, no ring, sized by the
-	   `size` HTML attribute (each consumer picks 6 / 8 / etc. chars).
-	   Hover/focus: very light bg fade-in to signal editability — just
-	   enough to read as interactive, not enough to compete with the
-	   slider row's quiet aesthetic. */
+	   Default: transparent box, no border, no ring; the component sets
+	   an inline width that hugs the value. Hover/focus: very light bg
+	   fade-in, just enough to read as interactive without competing
+	   with the slider row's quiet aesthetic. */
 	.panel-number-input {
-		@apply h-7 rounded-md border-0 bg-transparent px-1.5 text-right text-[13px]! font-medium tabular-nums text-nav-fg shadow-none transition-colors hover:bg-black/[0.04] focus:bg-black/[0.05] focus-visible:ring-0! focus-visible:outline-none md:text-[13px]!;
+		@apply h-7 min-w-8 shrink-0 rounded-full border-0 bg-transparent px-2 text-right text-[13px]! font-medium tabular-nums text-nav-fg shadow-none transition-colors hover:bg-black/[0.04] focus:bg-black/[0.05] focus-visible:ring-0! focus-visible:outline-none md:text-[13px]!;
 	}
 	.dark .panel-number-input {
 		@apply hover:bg-white/[0.04] focus:bg-white/[0.06];
@@ -742,15 +743,14 @@
 		background-color: #ffffff !important;
 	}
 
-	/* Compact tooltip — small black pill with white text. Used for short
-	   hover labels on chat-area icon buttons (Copy, Edit, Delete,
-	   Refresh, More, code-block actions) and the panel's info-icon
-	   hints. Corner radius is fixed at 8px so it tracks with the
-	   underlying icon button corners (also 8px) — keeps the tooltip
-	   visually anchored to its trigger rather than reading as a much
-	   larger floating pill. */
 	.tooltip-compact {
-		@apply rounded-[10px] border-transparent bg-black px-2 py-1.5 text-[11px] font-medium leading-snug text-white shadow-md;
+		@apply rounded-[9px] border-transparent bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug text-white shadow-md;
+	}
+
+	/* Dialog popups: borderless; chatbox shadow in light, flat card
+	   surface in dark. */
+	.dialog-soft-surface {
+		@apply border-0 ring-0 bg-background/98 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-none;
 	}
 
 	/* Rich tooltip — used for the context-usage and token-counter
@@ -779,7 +779,7 @@
 		height: 36px;
 		padding: 0 0.75rem !important;
 		gap: 9.5px !important;
-		border-radius: 12px;
+		border-radius: 9999px;
 		font-weight: 500;
 		font-size: 15px;
 		line-height: 20px;
@@ -805,6 +805,9 @@
 	.dark .app-user-menu [data-slot="dropdown-menu-item"]:focus,
 	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"]:focus,
 	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"][data-state="open"] {
+		/* --nav-surface-hover is tuned for the page bg; on the lighter menu
+		   surface it is invisible, so use the menu hover token instead. */
+		background-color: var(--accent);
 		color: #fff;
 	}
 	.app-user-menu [data-slot="dropdown-menu-item"]:focus *,
@@ -837,14 +840,15 @@
 
 	/* Elevated surface shadow (use ring-* for borders) */
 	.shadow-border {
-		--tw-shadow: 0 4px 16px rgba(0, 0, 0, 0.1);
-		--tw-shadow-colored: 0 4px 16px var(--tw-shadow-color);
+		/* Chatbox shadow; dark mode drops it. */
+		--tw-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
+		--tw-shadow-colored: 0 2px 8px -2px var(--tw-shadow-color);
 		box-shadow: var(--tw-ring-offset-shadow, 0 0 #0000),
 			var(--tw-ring-shadow, 0 0 #0000), var(--tw-shadow);
 	}
 
 	.dark .shadow-border {
-		--tw-shadow: 0 4px 16px rgba(0, 0, 0, 0.3);
+		box-shadow: none;
 	}
 
 	/* Settings dialog: no shadow in dark mode. */
@@ -852,6 +856,12 @@
 		box-shadow: none;
 	}
 
+	/* Lift the dialog off the page bg; halfway to --card keeps the inner
+	   bg-muted / bg-background fills readable. */
+	.dark .settings-surface {
+		background-color: color-mix(in srgb, var(--background) 50%, var(--card));
+	}
+
 	/* App-wide: every clickable control shows the hand cursor. Disabled ones
 	   (native :disabled, aria-disabled, Radix data-disabled) keep the default. */
 	button:not(:disabled):not([aria-disabled="true"]),
@@ -889,38 +899,43 @@
 	/* Chat search box: borderless, soft Gemini-style elevation. */
 	.chat-search-surface {
 		border: none;
-		/* Pin to the dark --radius so the corners are the same (less round) in
-		   both themes; rounded-4xl here and on the inner Command follow this. */
+		/* Pin to the dark --radius so rounded-3xl corners stay consistent. */
 		--radius: 0.625rem;
-		/* Match the chat box: soft elevation in light, none in dark. */
-		box-shadow: 0 2px 8px -2px rgba(27, 27, 31, 0.16);
+		/* Match the chat box: soft elevation. */
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
 	}
 
+	/* Dark: page-bg shadow so the dialog does not merge into same-color
+	   content behind it (same convention as the dropdowns). */
 	.dark .chat-search-surface {
-		box-shadow: none;
+		box-shadow: 0 4px 14px var(--background);
 	}
 
 	.menu-soft-surface,
 	.menu-soft-surface-up {
+		/* Light: chatbox shadow geometry. */
 		--menu-soft-edge: rgba(0, 0, 0, 0.14);
-		--menu-soft-shadow: rgba(0, 0, 0, 0.18);
-		--menu-soft-offset-y: 8px;
-		--menu-soft-blur: 28px;
-		--menu-soft-spread: -6px;
+		--menu-soft-shadow: rgba(0, 0, 0, 0.16);
+		--menu-soft-offset-y: 2px;
+		--menu-soft-blur: 8px;
+		--menu-soft-spread: -2px;
 		@apply bg-popover text-popover-foreground;
 		box-shadow:
 			inset 0 0 0 1px var(--menu-soft-edge),
 			0 var(--menu-soft-offset-y) var(--menu-soft-blur)
 				var(--menu-soft-spread) var(--menu-soft-shadow);
 	}
-	.menu-soft-surface-up {
-		--menu-soft-offset-y: -6px;
-		--menu-soft-spread: -8px;
-	}
 	.dark .menu-soft-surface,
 	.dark .menu-soft-surface-up {
 		--menu-soft-edge: rgba(255, 255, 255, 0.07);
 		--menu-soft-shadow: rgba(0, 0, 0, 0.28);
+		--menu-soft-offset-y: 8px;
+		--menu-soft-blur: 28px;
+		--menu-soft-spread: -6px;
+	}
+	.dark .menu-soft-surface-up {
+		--menu-soft-offset-y: -6px;
+		--menu-soft-spread: -8px;
 	}
 
 	/* Model selector: drop the inset edge ring, keep the soft drop shadow.
@@ -935,6 +950,20 @@
 		box-shadow: none;
 	}
 
+	/* List menus: one radius in both themes, slightly tighter than --radius-lg
+	   and matching the sidebar buttons. !important beats per-menu rounded-*. */
+	[data-slot="dropdown-menu-content"],
+	[data-slot="dropdown-menu-sub-content"],
+	[data-slot="select-content"],
+	[data-slot="combobox-content"] {
+		border-radius: 14px !important;
+	}
+
+	/* Account menu: a touch rounder than standard list menus. */
+	[data-slot="dropdown-menu-content"].app-user-menu.menu-soft-surface-up {
+		border-radius: 18px !important;
+	}
+
 	/* Every dropdown/menu/select/popover: borderless, chatbox shadow in light,
 	   none in dark. !important so it also overrides the bespoke menu shadows. */
 	[data-slot="dropdown-menu-content"],
@@ -942,7 +971,7 @@
 	[data-slot="select-content"],
 	[data-slot="combobox-content"],
 	[data-slot="popover-content"] {
-		box-shadow: 0 2px 8px -2px rgba(27, 27, 31, 0.16) !important;
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16) !important;
 	}
 	/* Dark: keep the chat-box dropdown shadow so menus don't merge into a
 	   same-color surface (e.g. a menu overlapping the composer). */
@@ -954,6 +983,15 @@
 		box-shadow: 0 4px 14px var(--background) !important;
 	}
 
+	/* Model selector: shadow in the page-bg color so it lifts off the
+	   composer it overlaps (chatbox geometry). Account menu stays flat. */
+	.dark .unsloth-model-selector-menu {
+		box-shadow: 0 2px 8px -2px var(--background) !important;
+	}
+	.dark .app-user-menu.menu-soft-surface-up {
+		box-shadow: none !important;
+	}
+
 	/* Context menus (chat, run, project): drop the inset edge ring, keep the soft shadow. */
 	.app-user-menu.menu-soft-surface {
 		box-shadow: 0 var(--menu-soft-offset-y) var(--menu-soft-blur)
@@ -962,7 +1000,7 @@
 
 	/* Account menu: drop the inset edge ring and use the composer's soft shadow. */
 	.app-user-menu.menu-soft-surface-up {
-		box-shadow: 0 0 20px 0 rgba(0, 0, 0, 0.09);
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
 	}
 	.dark .app-user-menu.menu-soft-surface-up {
 		box-shadow: 0 8px 28px -6px rgba(0, 0, 0, 0.28);
@@ -972,16 +1010,25 @@
 		@apply relative flex w-full flex-col rounded-[32px] bg-background dark:bg-card px-1 pt-2 outline-none transition-shadow;
 		font-family: var(--font-sans);
 		background-clip: padding-box;
-		box-shadow: 0 2px 8px -2px rgba(27, 27, 31, 0.16);
+		background-color: #ffffff;
+		/* Gemini-style composer shadow. */
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
+		transition: box-shadow 0.1s;
 	}
 
 	.dark .chat-composer-surface {
-		background-color: #2a2a2c;
+		background-color: var(--card);
 		box-shadow: none;
 	}
 
 	.composer-pill-btn {
-		@apply flex cursor-pointer items-center gap-1.5 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.08] disabled:cursor-not-allowed disabled:opacity-40;
+		@apply flex cursor-pointer items-center gap-1.5 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
+	}
+
+	/* Caret pills (RAG, MCP): the chevron carries its own whitespace, so the
+	   trailing padding is tightened to keep the hover pill balanced. */
+	.composer-pill-btn:has(.composer-pill-caret) {
+		@apply pr-1.5;
 	}
 
 	/* Icon-only (compact) pills drop their label, so make the button a square
@@ -1029,6 +1076,35 @@
 		@apply size-[15px] bg-transparent p-0 dark:bg-transparent;
 	}
 
+	/* Compact pills hide their labels, so surface the name as a hover
+	   tooltip styled like .tooltip-compact. */
+	[data-pill-compact="true"]
+		.composer-pill-btn:not([data-keep-label])[data-pill-label] {
+		position: relative;
+	}
+	[data-pill-compact="true"]
+		.composer-pill-btn:not([data-keep-label])[data-pill-label]:hover::after {
+		content: attr(data-pill-label);
+		/* Always one nowrap line, so always a full pill. */
+		@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-full bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug whitespace-nowrap text-white shadow-md;
+	}
+
+	/* Compact caret pills (RAG, MCP) open their menu on click instead of
+	   toggling off, so keep the icon and skip the X swap. No data-active
+	   requirement: the off-switch glyph rules below hide the icon on hover
+	   even for inactive pills, which left a blank slot in compact mode. */
+	[data-pill-compact="true"]
+		.composer-pill-btn:not([data-keep-label]):has(.composer-pill-caret):hover
+		.composer-pill-glyph
+		> :not(.composer-pill-x) {
+		opacity: 1;
+	}
+	[data-pill-compact="true"]
+		.composer-pill-btn:not([data-keep-label]):has(.composer-pill-caret)
+		.composer-pill-x {
+		display: none;
+	}
+
 	/* Hovering an active pill swaps the icon for an X (click to turn off). */
 	.composer-pill-btn[data-active="true"]:hover .composer-pill-glyph > :not(.composer-pill-x),
 	.unsloth-thinking-pill[data-active="true"]:hover .composer-pill-glyph > :not(.composer-pill-x) {
@@ -1040,6 +1116,20 @@
 		@apply opacity-100;
 	}
 
+	/* Pills whose icon is an off switch (RAG, MCP) only render while the
+	   feature is on, so reveal the X on hover even when the pill is not in
+	   its active look (MCP with no servers enabled, RAG without a model). */
+	.composer-pill-btn:has(> .composer-pill-glyph[role="button"]):hover
+		.composer-pill-glyph
+		> :not(.composer-pill-x) {
+		@apply opacity-0;
+	}
+
+	.composer-pill-btn:has(> .composer-pill-glyph[role="button"]):hover
+		.composer-pill-x {
+		@apply opacity-100;
+	}
+
 	.composer-input {
 		@apply mt-2 mb-1 mx-3 min-h-12 w-[calc(100%-1.5rem)] resize-none overflow-y-auto bg-transparent pl-2 pr-4 pt-2 pb-3 text-sm font-[450] outline-none placeholder:text-muted-foreground focus-visible:ring-0;
 	}
@@ -1058,16 +1148,18 @@
 		@apply relative flex w-full flex-col rounded-[32px] bg-background dark:bg-card px-3 py-3 outline-none transition-shadow;
 		font-family: var(--font-sans);
 		background-clip: padding-box;
-		/* Gemini search pill: soft, centered shadow, no downward offset. */
-		box-shadow: 0 0 20px 0 rgba(0, 0, 0, 0.09);
+		background-color: #ffffff;
+		/* Gemini-style composer shadow. */
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
+		transition: box-shadow 0.1s;
 	}
 
 	.unsloth-composer-surface:focus-within {
-		box-shadow: 0 0 20px 0 rgba(0, 0, 0, 0.09);
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
 	}
 
 	.dark .unsloth-composer-surface {
-		background-color: #2a2a2c;
+		background-color: var(--card);
 		box-shadow: none;
 	}
 
@@ -1177,7 +1269,7 @@
 	/* Right-side Thinking pill (toggle or dropdown). pl-2 matches the left
 	   pills so the hover X is not pushed in too far. */
 	.unsloth-thinking-pill {
-		@apply inline-flex shrink-0 cursor-pointer items-center gap-1 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.08] disabled:cursor-not-allowed disabled:opacity-40;
+		@apply inline-flex shrink-0 cursor-pointer items-center gap-1 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
 		/* Reserve a text line so the icon-only (inactive) pill matches the
 		   text pills' height instead of collapsing to the icon. */
 		min-height: calc(1lh + 0.75rem);
@@ -1202,23 +1294,22 @@
 
 	/* Soft elevation; [data-slot] outranks the component ring-1, dropping the border. */
 	.unsloth-plus-menu[data-slot] {
-		/* Pin radius so dark matches light (rounded-lg resolves smaller in dark). */
-		border-radius: 18px;
+		/* Pin radius so dark matches light (rounded-lg resolves smaller in dark).
+		   !important beats the global 14px dropdown radius. */
+		border-radius: 18px !important;
 		padding-top: 0.5rem;
 		padding-bottom: 0.5rem;
-		box-shadow:
-			0 1px 2px oklch(0 0 0 / 0.04),
-			0 6px 18px oklch(0 0 0 / 0.08);
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
 	}
 
 	/* MCP menu is wider (232px), so nudge the radius up so it reads as round as
 	   the narrower + menu (18 * 232/212 ~= 20px). */
 	.unsloth-plus-menu.mcp-menu[data-slot] {
-		border-radius: 20px;
+		border-radius: 20px !important;
 	}
 
 	.dark .unsloth-plus-menu[data-slot] {
-		background-color: #2a2a2c;
+		background-color: var(--card);
 		/* Shadow tinted to the page bg so it blends, not a dark halo. */
 		box-shadow: 0 4px 14px var(--background);
 	}
@@ -1257,7 +1348,7 @@
 			[data-slot="dropdown-menu-item"],
 			[data-slot="dropdown-menu-sub-trigger"]
 		):is(:hover, :focus, :focus-visible, [data-highlighted], [data-state="open"]) {
-		background-color: rgba(255, 255, 255, 0.08) !important;
+		background-color: rgba(255, 255, 255, 0.1) !important;
 	}
 
 	/* Hover tints only the background; the item and every descendant (text, icons,
@@ -1302,7 +1393,7 @@
 		white-space: nowrap;
 	}
 	.artifact-panel-shell {
-		box-shadow: 0 2px 8px -2px rgba(27, 27, 31, 0.16);
+		box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
 	}
 
 	.dark .artifact-panel-shell {
@@ -1431,6 +1522,20 @@
 		}
 	}
 
+	/* Indeterminate loading bar: a 1/3-width segment sweeping the track. */
+	.loading-bar-slide {
+		animation: loading-bar-slide 1.3s ease-in-out infinite;
+	}
+
+	@keyframes loading-bar-slide {
+		from {
+			transform: translateX(-110%);
+		}
+		to {
+			transform: translateX(420%);
+		}
+	}
+
 	@keyframes artifact-loading-line {
 		0% {
 			transform: translate3d(-125%, 0, 0) scaleX(0.78);
@@ -1634,7 +1739,7 @@
    !important because Sonner injects its base rules at runtime. */
 [data-sonner-toast][data-styled='true'] {
 	padding: 12px 18px !important;
-	box-shadow: 0 2px 6px rgba(0, 0, 0, 0.08) !important;
+	box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16) !important;
 }
 
 [data-sonner-toast][data-styled='true']:has([data-close-button]):not(:has([data-cancel])) {
@@ -1669,10 +1774,11 @@
 	margin: 0 !important;
 }
 
-/* Boost shadow on dark surfaces; mirrors .shadow-border / .menu-soft-surface pattern. */
+/* Dark toasts share the chatbox surface color (.chat-composer-surface
+   uses var(--card) in dark); the composer shadow lifts them off the page. */
 .dark [data-sonner-toast][data-styled='true'] {
-	background-color: #2a2a2c !important;
-	box-shadow: none !important;
+	background-color: var(--card) !important;
+	box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16) !important;
 }
 
 /* Selectable toast text; non-selectable toast buttons. */
@@ -1923,7 +2029,7 @@
 
 /* Keep the (borderless) close button blended with the dark toast surface. */
 .dark [data-sonner-toast][data-styled="true"] [data-close-button] {
-	background: #2a2a2c !important;
+	background: var(--card) !important;
 }
 
 [data-sonner-toast][data-styled="true"] [data-close-button] svg {
@@ -2017,6 +2123,11 @@
 		animation-iteration-count: infinite !important;
 	}
 
+	.loading-bar-slide {
+		animation-duration: 1.3s !important;
+		animation-iteration-count: infinite !important;
+	}
+
 	/* Keep the plus/x morph animating under reduced motion (small rotation,
 	   like the spinners above). */
 	.unsloth-composer-plus svg {
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 5d3f0ed339..bed7a27a63 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -47,6 +47,38 @@ EXIT_FALLBACK = 2
 EXIT_ERROR = 1
 EXIT_BUSY = 3
 
+# DiskPart-prompt suppression. RunAsInvoker does NOT stop amd-smi's runtime
+# elevation (its manifest is asInvoker), so this is just harmless belt-and-
+# suspenders for manifest-elevating tools. The real guard is _amd_smi_allowed():
+# we don't spawn amd-smi on Windows w/o a HIP SDK (or opt-in).
+if platform.system() == "Windows":
+    os.environ.setdefault("__COMPAT_LAYER", "RunAsInvoker")
+
+
+def _amd_smi_allowed() -> bool:
+    """Whether it is safe to spawn amd-smi here.
+
+    On Windows w/o a working HIP runtime, amd-smi elevates a child and pops a
+    UAC/DiskPart prompt RunAsInvoker can't suppress. Only call it on Windows
+    when a HIP SDK is detectable (hipinfo present) or UNSLOTH_ENABLE_AMD_SMI=1;
+    Linux/macOS always allowed. When skipped, the gfx arch still arrives via the
+    forwarded --rocm-gfx, so prebuilt selection is unaffected.
+    """
+    if platform.system() != "Windows":
+        return True
+    flag = os.environ.get("UNSLOTH_ENABLE_AMD_SMI", "").strip().lower()
+    if flag in ("1", "true", "yes", "on"):
+        return True
+    if flag in ("0", "false", "no", "off"):
+        return False
+    if shutil.which("hipinfo"):
+        return True
+    for _var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
+        _root = os.environ.get(_var)
+        if _root and os.path.isfile(os.path.join(_root, "bin", "hipinfo.exe")):
+            return True
+    return False
+
 
 def windows_hidden_subprocess_kwargs() -> dict[str, object]:
     """Return Windows-only subprocess kwargs that suppress console windows."""
@@ -89,9 +121,9 @@ def env_int(
     return value
 
 
-# Prefer "latest" over "master": "master" bypasses the prebuilt resolver
+# Prefer "latest" over "master" -- "master" bypasses the prebuilt resolver
 # (no matching GitHub release), forces a source build, and causes HTTP 422
-# errors. Use "master" only temporarily when the latest release lacks
+# errors. Only use "master" temporarily when the latest release is missing
 # support for a new model architecture.
 DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "latest")
 # Default published repo for prebuilt release resolution. Linux uses
@@ -113,13 +145,21 @@ LEMONADE_ROCM_RELEASES_API = f"https://api.github.com/repos/{LEMONADE_ROCM_REPO}
 
 
 def _lemonade_release_api_for(llama_tag: str) -> str:
-    """GitHub API URL for the lemonade release matching a llama.cpp tag.
+    """Return the GitHub API URL for the lemonade release that matches a
+    requested llama.cpp tag.
 
-    "latest"/unset -> /releases/latest; a pinned tag -> the same lemonade tag.
-    Lemonade tracks ggml-org build tags but may lag, so a tag it skipped 404s
-    and the caller falls back to the upstream tarball (keeps pinned installs
-    reproducible). Do NOT pass a fork tag -- the fork namespace always 404s.
-    Tag is URL-encoded (safe="") to prevent URL injection.
+    When llama_tag is unset or "latest", point at /releases/latest. When the
+    caller has pinned a specific tag (e.g. "b1260"), point at the same tag in
+    lemonade. Lemonade tracks `ggml-org/llama.cpp` build tags (e.g. "b1260")
+    but is NOT guaranteed to publish every upstream build -- lemonade may be
+    several builds behind ggml-org. Pinning to a specific tag that lemonade
+    skipped will produce a 404 and the caller falls through to the upstream
+    tarball; that is intentional so pinned installs stay reproducible.
+    Do NOT pass a `unslothai/llama.cpp` fork tag -- the fork uses its own
+    namespace and will always 404 against lemonade.
+
+    The tag is URL-encoded with `safe=""` so an unexpected slash / hash / query
+    character cannot reshape the URL.
     """
     normalized = (llama_tag or "").strip()
     if not normalized or normalized.lower() == "latest":
@@ -137,6 +177,7 @@ VALIDATION_MODEL_CACHE_FILENAME = "stories260K.gguf"
 INSTALL_LOCK_TIMEOUT_SECONDS = 300
 INSTALL_STAGING_ROOT_NAME = ".staging"
 GITHUB_AUTH_HOSTS = {"api.github.com", "github.com"}
+HF_AUTH_HOSTS = {"huggingface.co", "www.huggingface.co"}
 RETRYABLE_HTTP_STATUS = {408, 429, 500, 502, 503, 504}
 HTTP_FETCH_ATTEMPTS = 4
 HTTP_FETCH_BASE_DELAY_SECONDS = 0.75
@@ -154,90 +195,35 @@ DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS = env_int(
     2,
     minimum = 1,
 )
-# Deeper macOS-only walk-back: upstream can ship a run of prebuilts built for
-# a newer macOS than the host, caught only at validate time, so an older host
-# must skip the whole run. Free on new hosts (first plan validates).
+# Deeper macOS-only walk-back: upstream can ship a run of prebuilts built for a
+# newer macOS than the host, only caught at validate time, so an older host must
+# skip the whole run. Free on new hosts (first plan validates, extras unused).
 DEFAULT_MAX_MACOS_RELEASE_FALLBACKS = env_int(
     "UNSLOTH_LLAMA_MAX_MACOS_RELEASE_FALLBACKS",
     16,
     minimum = 1,
 )
 # Deterministic macOS pin. At b9428 ggml-org's macOS runner moved to macOS 26
-# (Tahoe), so b9428+ prebuilts load only on macOS 26+. b9415 is the last build
-# stamped below 26 (arm64 minos 14, x64 minos 13.3); loads on 13.3/14/15/26.
+# (Tahoe), so b9428+ prebuilts only load on macOS 26+. b9415 is the last build
+# stamped below 26 (arm64 minos 14, x64 minos 13.3); loads on macOS 13.3/14/15/26.
 _PINNED_MACOS_FALLBACK_TAG = "b9415"
 _PINNED_MACOS_LATEST_FLOOR = (26, 0)
 FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "master")
 
-# sm_103 (B300 / GB300 Blackwell Ultra) is not built natively but runs on the
-# bundled base compute_100 PTX, which the driver JIT-compiles forward to sm_103.
-# Listed in every bundle that ships the sm_100 build (the "newer" and
-# "portable" classes) so those hosts get a prebuilt, not a source compile.
-DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
-    "cuda12-older": {
-        "runtime_line": "cuda12",
-        "coverage_class": "older",
-        "supported_sms": ["70", "75", "80", "86", "89"],
-        "min_sm": 70,
-        "max_sm": 89,
-        "rank": 10,
-    },
-    "cuda12-newer": {
-        "runtime_line": "cuda12",
-        "coverage_class": "newer",
-        "supported_sms": ["86", "89", "90", "100", "103", "120"],
-        "min_sm": 86,
-        "max_sm": 120,
-        "rank": 20,
-    },
-    "cuda12-portable": {
-        "runtime_line": "cuda12",
-        "coverage_class": "portable",
-        "supported_sms": ["70", "75", "80", "86", "89", "90", "100", "103", "120"],
-        "min_sm": 70,
-        "max_sm": 120,
-        "rank": 30,
-    },
-    "cuda13-older": {
-        "runtime_line": "cuda13",
-        "coverage_class": "older",
-        "supported_sms": ["75", "80", "86", "89"],
-        "min_sm": 75,
-        "max_sm": 89,
-        "rank": 40,
-    },
-    "cuda13-newer": {
-        "runtime_line": "cuda13",
-        "coverage_class": "newer",
-        "supported_sms": ["86", "89", "90", "100", "103", "120"],
-        "min_sm": 86,
-        "max_sm": 120,
-        "rank": 50,
-    },
-    "cuda13-portable": {
-        "runtime_line": "cuda13",
-        "coverage_class": "portable",
-        "supported_sms": ["75", "80", "86", "89", "90", "100", "103", "120"],
-        "min_sm": 75,
-        "max_sm": 120,
-        "rank": 60,
-    },
-}
-
 # Lowest CUDA major we ship prebuilts for, and the highest major we probe for
 # installed runtime libraries. Detection and runtime-line derivation are
-# generated per major so a new toolkit (cuda14, ...) needs no code change as
-# long as llama.cpp keeps the cudart64_<major>.dll / libcudart.so.<major> naming.
+# generated per major so a new toolkit (cuda14, ...) needs no code change while
+# llama.cpp keeps the cudart64_<major>.dll / libcudart.so.<major> naming.
 _MIN_CUDA_MAJOR = 12
 _MAX_PROBE_CUDA_MAJOR = 19
 
 # Last ggml-org release whose Windows win-cuda-13 build is still sub-13.3
 # (cuda-13.1, b9360, 2026-05-27). Upstream bumped win-cuda-13 to 13.3 at b9365
 # and now ships only cuda-12.4 + cuda-13.3. cuda-12.4 predates Blackwell (ggml
-# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a
-# 13.0/13.1/13.2 driver is gated off 13.3 and would drop to CPU-only 12.4.
-# b9360 is immutable, so we pin its cuda-13.1 build (plus paired cudart) as a
-# GPU fallback for exactly those hosts. See unslothai/unsloth#5887.
+# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a 13.0/13.1/13.2
+# driver is gated off 13.3 and would drop to a CPU-only 12.4 build. b9360 is
+# immutable, so we pin its cuda-13.1 build (plus paired cudart) as a GPU
+# fallback for exactly those hosts. See unslothai/unsloth#5887.
 _PINNED_BLACKWELL_FALLBACK_TAG = "b9360"
 _PINNED_BLACKWELL_FALLBACK_RUNTIME = "13.1"
 # Floor at 13.0: b9360 ships native sm_120a SASS (no PTX/JIT) and a bundled
@@ -247,46 +233,19 @@ _PINNED_BLACKWELL_DRIVER_FLOOR = (13, 0)
 _BLACKWELL_MIN_SM = 120
 # ggml compiles Blackwell sm_120 only at toolkit >= 12.8, so an in-release
 # windows-cuda build at or above this already covers Blackwell and makes the
-# pinned 13.1 fallback unnecessary (cuda-12.4 is below it).
+# older pinned 13.1 fallback unnecessary (cuda-12.4 is below it).
 _BLACKWELL_MIN_TOOLKIT = (12, 8)
 _PINNED_BLACKWELL_LLAMA_SHA256 = "31ddb8b42d7ab4a47cab8c48c397519f580ca502df7e73f3ab396eacc16c8e8d"
 _PINNED_BLACKWELL_CUDART_SHA256 = "f96935e7e385e3b2d0189239077c10fe8fd7e95690fea4afec455b1b6c7e3f18"
 
 
 def _cuda_runtime_lines_for_major(major: int) -> list[str]:
-    """Runtime lines a driver of this CUDA major can use, newest first down to
-    the minimum we ship. A driver runs its own major and any older one
+    """Runtime lines a driver of this CUDA major can use, newest major first
+    down to the minimum we ship. A driver runs its own major and any older one
     (backward compatibility)."""
     return [f"cuda{m}" for m in range(major, _MIN_CUDA_MAJOR - 1, -1)]
 
 
-def _resolve_linux_bundle_profile(bundle_profile: str) -> "dict[str, Any] | None":
-    """Profile (runtime line + sm coverage) for a linux-x64-cuda<major>-<class>
-    bundle. Known majors use their published coverage; an unknown future major
-    reuses the newest known major's coverage for the same class as a forward
-    default, with the post-build GPU smoke test as backstop."""
-    known = DIRECT_LINUX_BUNDLE_PROFILES.get(bundle_profile)
-    if known is not None:
-        return known
-    m = re.fullmatch(r"cuda(?P<major>\d+)-(?P<klass>older|newer|portable)", bundle_profile)
-    if not m:
-        return None
-    base_key = max(
-        (
-            k
-            for k, v in DIRECT_LINUX_BUNDLE_PROFILES.items()
-            if v["coverage_class"] == m.group("klass")
-        ),
-        key = lambda k: int(re.match(r"cuda(\d+)-", k).group(1)),
-        default = None,
-    )
-    if base_key is None:
-        return None
-    profile = dict(DIRECT_LINUX_BUNDLE_PROFILES[base_key])
-    profile["runtime_line"] = f"cuda{m.group('major')}"
-    return profile
-
-
 @dataclass
 class HostInfo:
     system: str
@@ -317,8 +276,8 @@ class AssetChoice:
     url: str
     source_label: str
     # Paired runtime archive (Windows CUDA cudart bundle). When set,
-    # install_from_archives also downloads it and overlays its DLLs on top
-    # of the main install. See unslothai/unsloth#5106.
+    # install_from_archives also downloads it and overlays its DLLs on
+    # top of the main install. See unslothai/unsloth#5106.
     runtime_name: str | None = None
     runtime_url: str | None = None
     runtime_sha256: str | None = None
@@ -345,6 +304,10 @@ class PublishedLlamaArtifact:
     max_sm: int | None
     bundle_profile: str | None
     rank: int
+    # ROCm bundles only: the umbrella gfx target (e.g. "gfx110X") and the
+    # concrete gfx archs it covers (e.g. ["gfx1100", "gfx1101", ...]).
+    gfx_target: str | None = None
+    mapped_targets: list[str] = field(default_factory = list)
 
 
 @dataclass
@@ -497,8 +460,14 @@ def is_busy_lock_error(exc: BaseException) -> bool:
     return False
 
 
+# Status logs default to stderr so resolver modes keep stdout machine-readable
+# (setup.sh json.load()s the whole stdout). main() flips this for the install
+# path, where PowerShell otherwise renders stderr as NativeCommandError noise.
+_LOG_TO_STDOUT = False
+
+
 def log(message: str) -> None:
-    print(f"[llama-prebuilt] {message}", file = sys.stderr)
+    print(f"[llama-prebuilt] {message}", file = sys.stdout if _LOG_TO_STDOUT else sys.stderr)
 
 
 def log_lines(lines: Iterable[str]) -> None:
@@ -522,6 +491,10 @@ def should_send_github_auth(url: str | None) -> bool:
     return parsed_hostname(url) in GITHUB_AUTH_HOSTS
 
 
+def should_send_hf_auth(url: str | None) -> bool:
+    return parsed_hostname(url) in HF_AUTH_HOSTS
+
+
 def auth_headers(url: str | None = None) -> dict[str, str]:
     headers = {
         "User-Agent": "unsloth-studio-llama-prebuilt",
@@ -529,9 +502,35 @@ def auth_headers(url: str | None = None) -> dict[str, str]:
     token = os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")
     if token and should_send_github_auth(url):
         headers["Authorization"] = f"Bearer {token}"
+        return headers
+    # Anonymous huggingface.co fetches share a per-IP rate limit that CI
+    # fleets exhaust (HTTP 429), sinking the prebuilt path into a source
+    # build. Authenticate when a token is available.
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
+    if hf_token and should_send_hf_auth(url):
+        headers["Authorization"] = f"Bearer {hf_token}"
     return headers
 
 
+class _CrossHostAuthStrippingRedirectHandler(urllib.request.HTTPRedirectHandler):
+    """Drop Authorization when a redirect leaves the original host.
+
+    huggingface.co redirects file downloads to CDN hosts whose signed URLs
+    can reject foreign Authorization headers; urllib forwards headers to
+    redirect targets by default (requests/huggingface_hub strip them).
+    """
+
+    def redirect_request(self, req, fp, code, msg, headers, newurl):
+        new_request = super().redirect_request(req, fp, code, msg, headers, newurl)
+        if new_request is not None and parsed_hostname(newurl) != parsed_hostname(req.full_url):
+            new_request.headers.pop("Authorization", None)
+            new_request.unredirected_hdrs.pop("Authorization", None)
+        return new_request
+
+
+_URL_OPENER = urllib.request.build_opener(_CrossHostAuthStrippingRedirectHandler())
+
+
 def github_api_headers(url: str | None = None) -> dict[str, str]:
     return {
         "Accept": "application/vnd.github+json",
@@ -545,13 +544,14 @@ def is_github_api_url(url: str | None) -> bool:
 
 def is_retryable_url_error(exc: Exception) -> bool:
     if isinstance(exc, urllib.error.HTTPError):
-        # GitHub returns 403 (not 429) when the API rate limit is hit.
-        # Anonymous calls share a 60-req/hour bucket per runner IP, which
-        # CI fleets exhaust trivially. Treat 403 against api.github.com as
-        # retryable so we get a backoff cycle or two before the source-build
-        # fallback fires; sleep_backoff honours Retry-After /
-        # X-RateLimit-Reset for accurate waits. Real 403s on other hosts
-        # (private artefact downloads, auth failures) stay non-retryable.
+        # GitHub returns 403 (not the standard 429) when the API rate
+        # limit is hit. Anonymous calls share a 60-req/hour bucket per
+        # runner IP, which CI fleets can exhaust trivially. Treat 403
+        # against api.github.com as retryable so we get one or two
+        # backoff cycles before the source-build fallback fires; honour
+        # Retry-After / X-RateLimit-Reset in sleep_backoff for accurate
+        # waits. Real 403s on other hosts (private artefact downloads,
+        # auth failures) stay non-retryable.
         if exc.code == 403:
             return is_github_api_url(getattr(exc, "url", None))
         return exc.code in RETRYABLE_HTTP_STATUS
@@ -570,9 +570,9 @@ _RATE_LIMIT_WAIT_CAP_SECONDS = 60.0
 def _http_error_retry_delay(exc: Exception) -> float | None:
     """Extract a recommended wait from rate-limit headers on a 403/429.
 
-    Returns None when no header is present or the wait exceeds
-    _RATE_LIMIT_WAIT_CAP_SECONDS (the caller should not block then -- the
-    source-build fallback is faster).
+    Returns None when no header is present or the indicated wait is
+    longer than _RATE_LIMIT_WAIT_CAP_SECONDS (in which case the caller
+    should not block on it -- the source-build fallback is faster).
     """
     if not isinstance(exc, urllib.error.HTTPError):
         return None
@@ -781,12 +781,12 @@ def refs_match(candidate_ref: str | None, requested_ref: str | None) -> bool:
 
 
 def checkout_friendly_ref(ref_kind: str | None, ref: str | None) -> str | None:
-    """Normalize a source ref to a form ``git clone --branch`` accepts.
+    """Normalize a source ref to a form that ``git clone --branch`` accepts.
 
-    Fully qualified branch refs (``refs/heads/main``) are stripped to
-    ``main``; tag refs (``refs/tags/b8508``) to ``b8508``. Pull refs
-    (``refs/pull/123/head``) are left as-is since they are fetched
-    explicitly rather than cloned with ``--branch``.
+    Fully qualified branch refs like ``refs/heads/main`` are stripped to
+    ``main``; tag refs like ``refs/tags/b8508`` are stripped to ``b8508``.
+    Pull refs like ``refs/pull/123/head`` are left as-is since they are
+    always fetched explicitly rather than cloned with ``--branch``.
     """
     if not isinstance(ref, str) or not ref:
         return ref
@@ -836,8 +836,8 @@ def _published_windows_cuda_runtime(
     """Highest cuda-<major>.<minor> published upstream that `driver` can run by
     default CUDA compatibility, i.e. (major, minor) <= driver. None if nothing
     qualifies. Gating on the driver (not just the major) keeps a 13.3 build off
-    a 13.1-only driver, where it would rely on the unguaranteed
-    minor-version-compatibility path."""
+    a driver that only advertises 13.1, where it would otherwise rely on the
+    unguaranteed minor-version-compatibility path."""
     if driver is None:
         return None
     best: int | None = None
@@ -862,6 +862,16 @@ def format_byte_count(num_bytes: float) -> str:
     return f"{num_bytes:.1f} B"
 
 
+def _progress_percent_step() -> int:
+    """Non-tty milestone granularity. The in-app updater sets
+    UNSLOTH_PROGRESS_PERCENT_STEP=5 to stream finer progress lines."""
+    try:
+        step = int(os.environ.get("UNSLOTH_PROGRESS_PERCENT_STEP", "25"))
+    except ValueError:
+        return 25
+    return min(max(step, 1), 50)
+
+
 class DownloadProgress:
     def __init__(self, label: str, total_bytes: int | None) -> None:
         self.label = label
@@ -874,6 +884,7 @@ class DownloadProgress:
         )
         self.is_tty = term_ok and self.stream.isatty()
         self.completed = False
+        self.milestone_step = _progress_percent_step()
         self.last_milestone_percent = -1
         self.last_milestone_bytes = 0
         self.has_rendered_tty_progress = False
@@ -924,7 +935,8 @@ class DownloadProgress:
         should_emit = False
         if self.total_bytes is not None:
             percent = int((downloaded_bytes * 100) / max(self.total_bytes, 1))
-            milestone_percent = min((percent // 25) * 25, 100)
+            step = self.milestone_step
+            milestone_percent = min((percent // step) * step, 100)
             if milestone_percent > self.last_milestone_percent and milestone_percent < 100:
                 self.last_milestone_percent = milestone_percent
                 should_emit = True
@@ -973,7 +985,7 @@ def download_bytes(
     for attempt in range(1, attempts + 1):
         try:
             request = urllib.request.Request(url, headers = headers or auth_headers(url))
-            with urllib.request.urlopen(request, timeout = timeout) as response:
+            with _URL_OPENER.open(request, timeout = timeout) as response:
                 total_bytes: int | None = None
                 content_length = response.headers.get("Content-Length")
                 if content_length and content_length.isdigit():
@@ -1052,7 +1064,7 @@ def download_file(url: str, destination: Path) -> None:
                 delete = False,
             ) as handle:
                 tmp_path = Path(handle.name)
-                with urllib.request.urlopen(request, timeout = 120) as response:
+                with _URL_OPENER.open(request, timeout = 120) as response:
                     total_bytes: int | None = None
                     content_length = response.headers.get("Content-Length")
                     if content_length and content_length.isdigit():
@@ -1132,6 +1144,20 @@ def commit_source_archive_urls(repo: str, source_commit: str) -> list[str]:
     ]
 
 
+def release_asset_download_url(
+    repo: str | None, release_tag: str | None, asset_name: str | None
+) -> str | None:
+    """Direct download URL for a release asset, or None if any part is missing.
+    A mix build's merged commit is never pushed, so its source tree is only
+    reachable as this asset (codeload would 404 on the merge commit)."""
+    if not repo or not release_tag or not asset_name:
+        return None
+    return (
+        f"https://github.com/{repo}/releases/download/"
+        f"{urllib.parse.quote(release_tag, safe = '')}/{urllib.parse.quote(asset_name, safe = '')}"
+    )
+
+
 def github_release_assets(repo: str, tag: str) -> dict[str, str]:
     payload = fetch_json(
         f"https://api.github.com/repos/{repo}/releases/tags/{urllib.parse.quote(tag, safe = '')}"
@@ -1378,10 +1404,16 @@ def direct_linux_release_plan(
         )
         if lemonade_choice is not None:
             attempts.append(lemonade_choice)
-    else:
+    elif not host.has_usable_nvidia:
         cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
         if cpu_choice is not None:
             attempts.append(cpu_choice)
+    # NVIDIA hosts whose CUDA selection produced nothing fall through to the
+    # raise below (mirroring the ROCm policy above): the caller then walks
+    # back to an older release that still ships a usable CUDA line instead of
+    # silently installing a CPU binary on a GPU host. Today's walk-back only
+    # works because partial releases ship no CPU bundle; this keeps it working
+    # if a future partial release does.
     if not attempts:
         raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
     approved_checksums = synthetic_checksums_for_release(
@@ -1444,8 +1476,9 @@ def direct_upstream_release_plan(
                     torch_preference.selection_log,
                 )
             )
+            attempts[:] = _drop_blackwell_incapable_windows_cuda(host, attempts)
             # Blackwell on a 13.1/13.2 driver: prefer the pinned cuda-13.1 GPU
-            # build over the CPU-only cuda-12.4 left by in-release gating.
+            # build over the CPU-only cuda-12.4 the in-release gating leaves.
             pinned = _pinned_windows_cuda_fallback(host, attempts)
             if pinned is not None:
                 attempts.insert(0, pinned)
@@ -1483,9 +1516,9 @@ def direct_upstream_release_plan(
             )
     elif host.is_windows and host.is_arm64:
         # Upstream ggml-org/llama.cpp ships llama-bNNNN-bin-win-cpu-arm64.zip
-        # (in the b9334 release manifest). Without this branch the selector
-        # returned 0 attempts and fell back to a source build on every
-        # Windows ARM64 host.
+        # (visible in the b9334 release manifest). Without this branch the
+        # selector returned 0 attempts and the installer fell back to a
+        # source build on every Windows ARM64 host.
         cpu_asset = f"llama-{release_tag}-bin-win-cpu-arm64.zip"
         cpu_url = assets.get(cpu_asset)
         if cpu_url:
@@ -1528,6 +1561,14 @@ def direct_upstream_release_plan(
                 )
             )
     elif host.is_linux and host.is_x86_64 and not host.has_usable_nvidia:
+        if host.has_rocm:
+            # Lemonade first, mirroring the Windows ROCm branch above, so a
+            # ROCm host routed to ggml-org does not silently get the CPU build.
+            lemonade_choice = resolve_lemonade_rocm_choice(
+                host, "ubuntu", "linux-rocm", llama_tag = requested_tag
+            )
+            if lemonade_choice is not None:
+                attempts.append(lemonade_choice)
         asset_name = f"llama-{release_tag}-bin-ubuntu-x64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1543,10 +1584,10 @@ def direct_upstream_release_plan(
             )
     elif host.is_linux and host.is_arm64 and not host.has_usable_nvidia:
         # Upstream ggml-org/llama.cpp ships llama-bNNNN-bin-ubuntu-arm64.tar.gz
-        # (in the b9334 release manifest). Without this branch the selector
-        # returned 0 attempts and fell back to a source build on every Linux
-        # ARM64 host (DGX Spark, Ampere Altra, GitHub ubuntu-24.04-arm
-        # runners, etc.).
+        # (visible in the b9334 release manifest). Without this branch the
+        # selector returned 0 attempts and the installer fell back to a
+        # source build on every Linux ARM64 host (DGX Spark, Ampere
+        # Altra, GitHub-hosted ubuntu-24.04-arm runners, etc.).
         asset_name = f"llama-{release_tag}-bin-ubuntu-arm64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1576,9 +1617,9 @@ def direct_upstream_release_plan(
 
 
 def pinned_macos_release_tag(host: HostInfo, repo: str) -> str | None:
-    """Pin b9415 (last upstream macOS build that loads below macOS 26) for a
-    known pre-26 host on ggml-org upstream; None keeps latest selection. The
-    unslothai/llama.cpp fork ships its own prebuilts (arm64 minos 14, x64
+    """Pin b9415 (the last upstream macOS build that loads below macOS 26) for a
+    known pre-26 host on ggml-org upstream; return None to keep latest selection.
+    The unslothai/llama.cpp fork ships its own prebuilts (arm64 minos 14, x64
     minos 13.3) and needs no pin, so this is a no-op there and for macOS 26+,
     unknown version, non-macOS."""
     if repo != UPSTREAM_REPO:
@@ -1602,20 +1643,22 @@ def resolve_simple_install_release_plans(
     max_release_fallbacks: int = DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS,
 ) -> tuple[str, list[InstallReleasePlan]]:
     repo = published_repo or DEFAULT_PUBLISHED_REPO
-    requested_tag = normalized_requested_llama_tag(llama_tag)
-    # The unslothai/llama.cpp fork ships only linux-x64 bundles. An arm64 Linux
-    # host with a GPU (GH200/GB200/DGX Spark) routes here; it must not install
-    # an x64 binary, so fall back to a GPU-targeting source build rather than
-    # the wrong arch (or silently dropping to a CPU arm64 build).
-    if host.is_linux and not host.is_x86_64 and repo == DEFAULT_PUBLISHED_REPO:
-        raise PrebuiltFallback(
-            f"{repo} ships only linux-x64 prebuilts; "
-            f"{host.machine or 'non-x64'} Linux falls back to source build"
+    # The fork (unslothai) ships a manifest describing every bundle's GPU/arch
+    # coverage, so all fork hosts select from it. Upstream (ggml-org) ships no
+    # manifest and is selected by asset filename in the loop below.
+    if repo == DEFAULT_PUBLISHED_REPO:
+        return _fork_manifest_release_plans(
+            llama_tag,
+            host,
+            published_repo,
+            published_release_tag,
+            max_release_fallbacks = max_release_fallbacks,
         )
+    requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
     # macOS: pin the last upstream build that loads on a pre-26 host instead of
     # fetching the latest (macOS 26 only) build and walking back release by
-    # release. No-op on macOS 26+, unknown version, non-macOS, the fork.
+    # release. No-op on macOS 26+, unknown version, non-macOS, and the fork.
     if allow_older_release_fallback:
         pinned_macos = pinned_macos_release_tag(host, repo)
         if pinned_macos is not None:
@@ -1629,10 +1672,7 @@ def resolve_simple_install_release_plans(
         releases = iter_release_payloads_by_time(repo, published_release_tag, requested_tag)
         for release in releases:
             try:
-                if host.is_linux and repo == "unslothai/llama.cpp":
-                    plan = direct_linux_release_plan(release, host, repo, requested_tag)
-                else:
-                    plan = direct_upstream_release_plan(release, host, repo, requested_tag)
+                plan = direct_upstream_release_plan(release, host, repo, requested_tag)
                 if plan is None:
                     continue
             except PrebuiltFallback as exc:
@@ -1901,6 +1941,18 @@ def parse_published_artifact(raw: Any) -> PublishedLlamaArtifact | None:
         rank = int(rank_raw)
     except (TypeError, ValueError):
         raise ValueError(f"artifact {asset_name} rank was not an integer")
+    gfx_target_raw = raw.get("gfx_target")
+    gfx_target = (
+        gfx_target_raw.strip()
+        if isinstance(gfx_target_raw, str) and gfx_target_raw.strip()
+        else None
+    )
+    mapped_raw = raw.get("mapped_targets", [])
+    mapped_targets = (
+        [value.strip() for value in mapped_raw if isinstance(value, str) and value.strip()]
+        if isinstance(mapped_raw, (list, tuple))
+        else []
+    )
     return PublishedLlamaArtifact(
         asset_name = asset_name,
         install_kind = install_kind,
@@ -1915,6 +1967,8 @@ def parse_published_artifact(raw: Any) -> PublishedLlamaArtifact | None:
         if isinstance(bundle_profile, str) and bundle_profile
         else None,
         rank = rank,
+        gfx_target = gfx_target,
+        mapped_targets = mapped_targets,
     )
 
 
@@ -1930,8 +1984,8 @@ def parse_published_release_bundle(
     if not manifest_url:
         return None
 
-    # Mixed repos are filtered by an explicit release-side manifest, not by
-    # release tag or asset filename conventions.
+    # Mixed repos are filtered by an explicit release-side manifest rather than
+    # by release tag or asset filename conventions.
     manifest_bytes = download_bytes(
         manifest_url,
         timeout = 30,
@@ -2151,6 +2205,25 @@ def iter_published_release_bundles(
         yield bundle
 
 
+def _artifact_covers_sms(artifact: PublishedLlamaArtifact, host_sms: Iterable[str]) -> bool:
+    """True when every host SM is listed in the artifact's supported_sms and
+    falls within its [min_sm, max_sm] range."""
+    if not artifact.supported_sms or artifact.min_sm is None or artifact.max_sm is None:
+        return False
+    supported = {str(value) for value in artifact.supported_sms}
+    return all(sm in supported and artifact.min_sm <= int(sm) <= artifact.max_sm for sm in host_sms)
+
+
+def _sm_range(artifact: PublishedLlamaArtifact) -> int:
+    """SM-coverage span used as a sort key, where a tighter (smaller) range wins.
+    A bundle with no SM metadata (legacy/upstream-named) gets a max range so it
+    sorts last and can't outrank a real targeted bundle whose tight range would
+    otherwise sort first."""
+    if artifact.min_sm is not None and artifact.max_sm is not None:
+        return artifact.max_sm - artifact.min_sm
+    return 9999
+
+
 def linux_cuda_choice_from_release(
     host: HostInfo,
     release: PublishedReleaseBundle,
@@ -2190,8 +2263,12 @@ def linux_cuda_choice_from_release(
                 else "none"
             )
         )
+    # arm64 CUDA hosts (DGX Spark / Grace Hopper) consume linux-arm64-cuda
+    # bundles; x64 hosts consume linux-cuda. The SM / runtime-line matching
+    # below is arch-agnostic and applies to both.
+    cuda_install_kind = "linux-arm64-cuda" if host.is_arm64 else "linux-cuda"
     published_artifacts = [
-        artifact for artifact in release.artifacts if artifact.install_kind == "linux-cuda"
+        artifact for artifact in release.artifacts if artifact.install_kind == cuda_install_kind
     ]
     published_asset_names = sorted(artifact.asset_name for artifact in published_artifacts)
     selection_log.append(
@@ -2242,7 +2319,7 @@ def linux_cuda_choice_from_release(
                 url = asset_url,
                 source_label = "published",
                 is_ready_bundle = True,
-                install_kind = "linux-cuda",
+                install_kind = cuda_install_kind,
                 bundle_profile = artifact.bundle_profile,
                 runtime_line = artifact.runtime_line,
                 coverage_class = artifact.coverage_class,
@@ -2324,7 +2401,7 @@ def linux_cuda_choice_from_release(
             artifact, url = sorted(
                 coverage_candidates,
                 key = lambda item: (
-                    (item[0].max_sm or 0) - (item[0].min_sm or 0),
+                    _sm_range(item[0]),
                     item[0].rank,
                     item[0].max_sm or 0,
                 ),
@@ -2383,11 +2460,20 @@ def validated_checksums_for_bundle(
             raise PrebuiltFallback(
                 "published manifest checksum did not match the approved checksum asset"
             )
-    # Accept bundles carrying only an exact-commit source archive
-    # (llama.cpp-source-commit-<sha>.tar.gz) without requiring the legacy
-    # llama.cpp-source-<upstream_tag>.tar.gz entry.
+    # Accept bundles that carry only an exact-commit source archive
+    # (e.g. llama.cpp-source-commit-<sha>.tar.gz) without requiring the
+    # legacy llama.cpp-source-<upstream_tag>.tar.gz entry.
     if exact_source_archive_hash(checksums) is None:
         require_approved_source_hash(checksums, bundle.upstream_tag)
+    elif source_clone_url_for_release(checksums, bundle) is None:
+        # No source repo in either the checksum payload or the manifest bundle:
+        # preferred_source_archive would silently fall back to ggml-org source at
+        # the upstream tag, pairing the prebuilt with a possibly mismatched tree.
+        # Fail closed so the resolver skips this release.
+        raise PrebuiltFallback(
+            f"approved checksum asset for {repo}@{bundle.release_tag} declared an "
+            "exact source archive without a source repo to clone it from"
+        )
     return checksums
 
 
@@ -2521,22 +2607,23 @@ def resolve_requested_llama_tag(
 
     Resolution order:
       1. Concrete tag (e.g. "b8508") -- returned as-is.
-      2. "latest" with published_repo -- the latest usable Unsloth published
-         bundle's upstream_tag (matches the published prebuilt metadata).
-      3. "latest" without published_repo, or if (2) fails -- query upstream
-         ggml-org/llama.cpp. May return a newer, untested tag.
+      2. "latest" with published_repo -- resolve the latest usable Unsloth
+         published release bundle and return its upstream_tag. This is the
+         preferred version that matches the published prebuilt metadata.
+      3. "latest" without published_repo or if (2) fails -- query the upstream
+         ggml-org/llama.cpp repo. This may return a newer, untested tag.
 
-    The Unsloth repo is preferred because its releases are pinned to upstream
-    tags validated with Unsloth Studio; the upstream bleeding-edge tag risks
-    API/ABI incompatibilities.
+    The Unsloth repo is preferred because its releases are pinned to specific
+    upstream tags that have been validated with Unsloth Studio. Using the
+    upstream bleeding-edge tag risks API/ABI incompatibilities.
     """
     normalized_requested = normalized_requested_llama_tag(requested_tag)
     if normalized_requested != "latest":
         return normalized_requested
     # Prefer the Unsloth release repo tag (tested/approved) over bleeding-edge
-    # upstream. E.g. unslothai/llama.cpp may publish b8508 while ggml-org
-    # latest is b8514. The source-build fallback should compile the same
-    # version the prebuilt path would have installed.
+    # upstream. For example, unslothai/llama.cpp may publish b8508 while
+    # ggml-org/llama.cpp latest is b8514. The source-build fallback should
+    # compile the same version the prebuilt path would have installed.
     if published_repo:
         try:
             return resolve_published_release(
@@ -2546,7 +2633,7 @@ def resolve_requested_llama_tag(
             ).bundle.upstream_tag
         except Exception:
             pass
-    # Fall back to the upstream ggml-org latest release tag
+    # Fall back to upstream ggml-org latest release tag
     return latest_upstream_release_tag()
 
 
@@ -2568,8 +2655,16 @@ def exact_source_archive_hash(checksums: ApprovedReleaseChecksums) -> ApprovedAr
     return checksums.artifacts.get(exact_source_archive_logical_name(checksums.source_commit))
 
 
-def source_clone_url_from_checksums(checksums: ApprovedReleaseChecksums) -> str | None:
-    return source_repo_clone_url(checksums.source_repo, checksums.source_repo_url)
+def source_clone_url_for_release(
+    checksums: ApprovedReleaseChecksums, bundle: PublishedReleaseBundle
+) -> str | None:
+    # Single source of truth for "where do we clone source from", shared by the
+    # validation gate and source_build_plan_for_release: take the repo from the
+    # checksum payload, falling back field by field to the manifest bundle.
+    return source_repo_clone_url(
+        checksums.source_repo or bundle.source_repo,
+        checksums.source_repo_url or bundle.source_repo_url,
+    )
 
 
 def source_build_plan_for_release(release: ResolvedPublishedRelease) -> SourceBuildPlan:
@@ -2581,7 +2676,7 @@ def source_build_plan_for_release(release: ResolvedPublishedRelease) -> SourceBu
     resolved_source_ref = checksums.resolved_source_ref or release.bundle.resolved_source_ref
     source_commit = checksums.source_commit or release.bundle.source_commit
     source_ref_kind = checksums.source_ref_kind or release.bundle.source_ref_kind
-    source_url = source_repo_clone_url(source_repo, source_repo_url)
+    source_url = source_clone_url_for_release(checksums, release.bundle)
     if exact_source is not None and source_url and source_commit:
         return SourceBuildPlan(
             source_url = source_url,
@@ -2672,6 +2767,15 @@ def run_capture(
     check: bool = False,
     env: dict[str, str] | None = None,
 ) -> subprocess.CompletedProcess[str]:
+    # amd-smi on Windows auto-elevates and pops a UAC/DiskPart prompt mid-install;
+    # RunAsInvoker forces it un-elevated. Callers already fall back to WMI/name
+    # detection. Mirrors install.ps1's Invoke-AmdSmiNoElevate; Windows-only.
+    if (
+        command
+        and platform.system() == "Windows"
+        and os.path.basename(command[0]).lower().startswith("amd-smi")
+    ):
+        env = {**(os.environ if env is None else env), "__COMPAT_LAYER": "RunAsInvoker"}
     result = subprocess.run(
         command,
         capture_output = True,
@@ -2693,38 +2797,38 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
     A bare first-match picked the wrong device on mixed APU + dGPU hosts
     (e.g. Strix Halo gfx1151 + discrete RX 7900 gfx1100). Respect
     HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES so the
-    asset matches what HIP runs on; falls back to the first GPU with no env
-    var set.
+    asset matches what HIP actually runs on. Falls back to the first GPU when
+    no env var is set.
 
     rocminfo / hipinfo print the same gfx token multiple times per GPU (Name,
-    ISA, marketing-name). We first split the output on per-GPU section headers
-    (rocminfo: "Agent N" blocks, hipinfo: "device#N" entries) and take exactly
-    one gfx token per section -- this gives the correct per-GPU list even on
-    same-arch multi-GPU hosts (e.g. two RX 7900 XTX) where global dict.fromkeys
-    dedup would collapse both to one entry and make HIP_VISIBLE_DEVICES=1 point
-    out of range.
+    ISA, marketing-name). We first try to split the output on per-GPU section
+    headers (rocminfo: "Agent N" blocks, hipinfo: "device#N" entries) and take
+    exactly one gfx token per section. This gives the correct per-GPU list even
+    on same-arch multi-GPU hosts (e.g. two RX 7900 XTX cards) where global
+    dict.fromkeys dedup would collapse both cards to a single entry and make
+    HIP_VISIBLE_DEVICES=1 point out of range.
 
     Falls back to insertion-order dedup when the output has no recognisable
     section markers (flat gfx-string inputs, unit-test stubs, etc.).
 
     Empty / "-1" env values mean no AMD GPU is visible to HIP: return None.
     """
-    # Build a per-GPU token list by splitting on section boundaries. rocminfo
-    # sections start with "Agent N" lines (optionally between rows of
-    # asterisks); hipinfo sections with "device#N".
+    # Try to build a per-GPU token list by splitting on section boundaries.
+    # rocminfo sections are introduced by "Agent N" lines (optionally between
+    # rows of asterisks). hipinfo sections start with "device#N".
     _sections = re.split(
         r"(?mi)^\s*\*+\s*$\s*agent\s+\d+\s*$|\bdevice\s*#\s*\d+\b",
         out,
     )
     if len(_sections) > 1:
-        # One gfx token per GPU section preserves physical order.
+        # Section-based: one gfx token per GPU section preserves physical order.
         _tokens: list[str] = []
         for _sec in _sections[1:]:
             _m = re.search(r"gfx[1-9][0-9a-z]{2,3}", _sec.lower())
             if _m:
                 _tokens.append(_m.group(0))
     else:
-        # Fallback: insertion-order dedup (flat strings / unknown formats).
+        # Fallback: insertion-order dedup (handles flat strings / unknown formats).
         _raw = re.findall(r"gfx[1-9][0-9a-z]{2,3}", out.lower())
         _tokens = list(dict.fromkeys(_raw))
 
@@ -2740,7 +2844,7 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
             break
     if _vis_raw is not None:
         _vis = _vis_raw.strip()
-        # Empty or "-1" means "no AMD GPU visible" (matches the rest of Studio)
+        # Empty or "-1" means "no AMD GPU visible" (matches the rest of Studio).
         if _vis == "" or _vis == "-1":
             return None
         _first = _vis.split(",")[0].strip()
@@ -2772,10 +2876,11 @@ def detect_host() -> HostInfo:
     has_physical_nvidia = False
     has_usable_nvidia = False
     if nvidia_smi:
-        # Require `nvidia-smi -L` to list a GPU before treating the host as
-        # NVIDIA. The "NVIDIA-SMI ..." banner prints even when the command
-        # can't reach the driver (e.g. stale container leftovers), which
-        # would misclassify an AMD ROCm host as NVIDIA and skip the ROCm path.
+        # Require `nvidia-smi -L` to actually list a GPU before treating the
+        # host as NVIDIA. The banner text "NVIDIA-SMI ..." is printed even
+        # when the command fails to communicate with the driver (e.g. stale
+        # container leftovers), which would otherwise misclassify an AMD
+        # ROCm host as NVIDIA and short-circuit the ROCm path.
         try:
             listing = run_capture([nvidia_smi, "-L"], timeout = 20)
             gpu_lines = [line for line in listing.stdout.splitlines() if line.startswith("GPU ")]
@@ -2788,9 +2893,9 @@ def detect_host() -> HostInfo:
         try:
             result = run_capture([nvidia_smi], timeout = 20)
             merged = "\n".join(part for part in (result.stdout, result.stderr) if part)
-            # Newer NVIDIA drivers (e.g. 610.x on Windows) print "CUDA UMD
-            # Version: X.Y" instead of the legacy "CUDA Version: X.Y"; accept
-            # both spellings.
+            # Newer NVIDIA drivers (e.g. 610.x on Windows) print
+            # "CUDA UMD Version: X.Y" instead of the legacy
+            # "CUDA Version: X.Y"; accept both spellings.
             cuda_match = re.search(
                 r"CUDA(?: UMD)? Version:\s*(\d+)\.(\d+)",
                 merged,
@@ -2833,10 +2938,10 @@ def detect_host() -> HostInfo:
 
             if visible_gpu_rows:
                 has_usable_nvidia = True
-                # Older nvidia-smi (pre -L support) hits the except in the
-                # first try block but still succeeds here, leaving
-                # has_physical_nvidia unset. Mirror the -L path so downstream
-                # diagnostics on line ~4390 still run.
+                # Older nvidia-smi versions (pre -L support) hit the
+                # except in the first try block but still succeed here,
+                # leaving has_physical_nvidia unset. Mirror the -L path
+                # so downstream diagnostics on line ~4390 still run.
                 if not has_physical_nvidia:
                     has_physical_nvidia = True
             elif visible_device_tokens == []:
@@ -2848,7 +2953,27 @@ def detect_host() -> HostInfo:
         except Exception:
             pass
 
+    # Linux /proc/driver/nvidia/gpus fallback: the NVIDIA driver exposes one
+    # subdir per GPU here regardless of nvidia-smi state, so a host whose
+    # nvidia-smi is absent from PATH, wedged, or failing is still recognised as
+    # NVIDIA. Mirrors the fallback added to install.sh / install_python_stack.py
+    # in PR 6174 so the prebuilt installer does not misroute such hosts to ROCm
+    # or CPU. driver_cuda_version / compute_caps stay unset here; downstream
+    # CUDA asset selection treats unknown SMs as "prefer portable" and an
+    # unknown driver runtime line as "no published CUDA match" (returns None,
+    # no crash), so planning falls back to a source build with GGML_CUDA=ON.
+    if is_linux and not has_physical_nvidia:
+        try:
+            proc_gpu_dir = "/proc/driver/nvidia/gpus"
+            if os.path.isdir(proc_gpu_dir) and os.listdir(proc_gpu_dir):
+                has_physical_nvidia = True
+                has_usable_nvidia = visible_device_tokens != []
+        except OSError:
+            pass
+
     # Detect AMD ROCm (HIP) -- require actual GPU, not just tools installed
+    # NVIDIA takes precedence: when an NVIDIA GPU is usable, skip ROCm probing
+    # entirely so co-installed ROCm tools cannot misroute the host (PR 6174).
 
     def _amd_smi_has_gpu(stdout: str) -> bool:
         """Check for 'GPU: <number>' data rows, not just a table header."""
@@ -2856,12 +2981,19 @@ def detect_host() -> HostInfo:
 
     has_rocm = False
     rocm_gfx_target: str | None = None
-    if is_linux:
+    if is_linux and not has_usable_nvidia:
+        # WSL2 ROCDXG: the system rocminfo enumerates the GPU over /dev/dxg
+        # only when HSA_ENABLE_DXG_DETECTION=1 (a no-op on bare metal), and
+        # rocminfo can live only under /opt/rocm/bin (the profile.d PATH
+        # drop-in reaches login shells only). Probe accordingly or a ROCDXG
+        # WSL host is misdetected as CPU-only.
+        _dxg_probe_env = {**os.environ}
+        _dxg_probe_env.setdefault("HSA_ENABLE_DXG_DETECTION", "1")
         for _cmd, _check in (
-            # rocminfo: a real gfx GPU id (3-4 chars, nonzero first digit).
-            # gfx000 is the CPU agent; ROCm 6.1+ also emits generic ISA lines
-            # ("gfx11-generic", "gfx9-4-generic") with only 1-2 digits before
-            # the dash, which must not be treated as a real GPU.
+            # rocminfo: look for a real gfx GPU id (3-4 chars, nonzero first digit).
+            # gfx000 is the CPU agent; ROCm 6.1+ also emits generic ISA lines like
+            # "gfx11-generic" or "gfx9-4-generic" which only have 1-2 digits before
+            # the dash and must not be treated as a real GPU.
             (
                 ["rocminfo"],
                 lambda out: bool(re.search(r"gfx[1-9][0-9a-z]{2,3}", out.lower())),
@@ -2869,10 +3001,18 @@ def detect_host() -> HostInfo:
             (["amd-smi", "list"], _amd_smi_has_gpu),
         ):
             _exe = shutil.which(_cmd[0])
+            if not _exe and _cmd[0] == "rocminfo":
+                _opt_rocminfo = "/opt/rocm/bin/rocminfo"
+                if os.access(_opt_rocminfo, os.X_OK):
+                    _exe = _opt_rocminfo
             if not _exe:
                 continue
             try:
-                _result = run_capture([_exe, *_cmd[1:]], timeout = 10)
+                _result = run_capture(
+                    [_exe, *_cmd[1:]],
+                    timeout = 10,
+                    env = _dxg_probe_env if _cmd[0] == "rocminfo" else None,
+                )
             except Exception:
                 continue
             if _result.returncode == 0 and _result.stdout.strip():
@@ -2880,14 +3020,15 @@ def detect_host() -> HostInfo:
                     has_rocm = True
                     rocm_gfx_target = _pick_rocm_gfx_target(_result.stdout)
                     break
-    elif is_windows:
-        # Windows: prefer active probes that validate GPU presence. hipinfo /
-        # amd-smi are often NOT on PATH -- the HIP SDK installer sets HIP_PATH
-        # / ROCM_PATH but doesn't always add the bin dir to PATH. Mirror
-        # setup.ps1's fallback: check the env-var bin dirs before giving up so
-        # `has_rocm` isn't silently False when PATH isn't updated yet.
+    elif is_windows and not has_usable_nvidia:
+        # Windows: prefer active probes that validate GPU presence.
+        # hipinfo / amd-smi are often NOT on PATH -- the HIP SDK installer
+        # sets HIP_PATH / ROCM_PATH but does not always add the bin dir to
+        # the system PATH.  Mirror setup.ps1's fallback: check the env-var
+        # bin dirs before giving up so that `has_rocm` is not silently False
+        # on machines where the PATH is not yet updated.
         def _resolve_exe(name: str) -> str | None:
-            """Full path to `name`, checking PATH then HIP_PATH/ROCM_PATH bin."""
+            """Return full path to `name`, checking PATH then HIP_PATH/ROCM_PATH bin."""
             found = shutil.which(name)
             if found:
                 return found
@@ -2897,12 +3038,20 @@ def detect_host() -> HostInfo:
                     _candidate = os.path.join(_root, "bin", f"{name}.exe")
                     if os.path.isfile(_candidate):
                         return _candidate
+            # AMD torch wheels ship hipInfo.exe into the venv Scripts dir
+            # (next to python.exe) -- resolvable on driver-only hosts where no
+            # SDK dir exists, so a standalone rerun can still detect the GPU.
+            _venv_candidate = os.path.join(os.path.dirname(sys.executable), f"{name}.exe")
+            if os.path.isfile(_venv_candidate):
+                return _venv_candidate
             return None
 
-        for _cmd, _check in (
-            (["hipinfo"], lambda out: "gcnarchname" in out.lower()),
-            (["amd-smi", "list"], _amd_smi_has_gpu),
-        ):
+        _win_probes = [(["hipinfo"], lambda out: "gcnarchname" in out.lower())]
+        if _amd_smi_allowed():
+            # Skipped on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt);
+            # gfx arch still arrives via --rocm-gfx, so has_rocm is set by override.
+            _win_probes.append((["amd-smi", "list"], _amd_smi_has_gpu))
+        for _cmd, _check in _win_probes:
             _exe = _resolve_exe(_cmd[0])
             if not _exe:
                 continue
@@ -2916,8 +3065,8 @@ def detect_host() -> HostInfo:
                     # hipinfo reports "gcnArchName: gfx1100" -- extract if present
                     rocm_gfx_target = _pick_rocm_gfx_target(_result.stdout)
                     break
-        # Note: amdhip64.dll presence alone is NOT GPU evidence -- the HIP SDK
-        # can be installed without an AMD GPU.
+        # Note: amdhip64.dll presence alone is NOT treated as GPU evidence
+        # since the HIP SDK can be installed without an AMD GPU.
 
     return HostInfo(
         system = system,
@@ -2942,7 +3091,7 @@ def detect_host() -> HostInfo:
 def _normalize_forwarded_gfx(value: str | None) -> str | None:
     """Extract a single gfx token from a forwarded --rocm-gfx / env value.
     setup.sh/setup.ps1 already picked the active GPU, so take the token as-is
-    without re-applying visible-device selection. Ignore malformed input."""
+    without re-applying visible-device selection. Ignore anything malformed."""
     if not value:
         return None
     m = re.search(r"gfx[1-9][0-9a-z]{2,3}", value.lower())
@@ -2958,8 +3107,8 @@ def _apply_host_overrides(
 ) -> HostInfo:
     """Fold setup.sh/setup.ps1's forwarded detection into the host profile.
     A forwarded gfx (--rocm-gfx or UNSLOTH_ROCM_GFX_ARCH) is authoritative and
-    implies ROCm: the installer's own hipinfo/amd-smi probe can miss the arch
-    on amd-smi-only hosts or when setup inferred it from the GPU name, leaving
+    implies ROCm: the installer's own hipinfo/amd-smi probe can miss the arch on
+    amd-smi-only hosts or when setup inferred it from the GPU name, leaving
     rocm_gfx_target None and no lemonade prebuilt selected. force_cpu is the
     opposite explicit signal (arm64 Linux GPU host whose source build failed):
     drop GPU attributes so the CPU prebuilt for this OS/arch is selected."""
@@ -2979,6 +3128,21 @@ def _apply_host_overrides(
     return host
 
 
+def published_repo_for_host(host: HostInfo, *, linux_amd_tooling_present: bool = False) -> str:
+    """The release repo setup.sh / setup.ps1 pick for this host: macOS always the
+    fork (ggml-org macOS bundles need too-new macOS); else CPU-only Linux/Windows
+    -> ggml-org upstream (the fork ships no CPU bundle) and any usable GPU (NVIDIA
+    or ROCm) -> the fork. linux_amd_tooling_present mirrors setup.sh routing Linux
+    hosts that expose AMD tooling (rocminfo/amd-smi/hipconfig/hipinfo) to the fork
+    even when the probe cannot confirm an active GPU. Mirrors the shell routing."""
+    if host.is_macos:
+        return DEFAULT_PUBLISHED_REPO
+    has_gpu = (
+        host.has_usable_nvidia or host.has_rocm or (host.is_linux and linux_amd_tooling_present)
+    )
+    return DEFAULT_PUBLISHED_REPO if has_gpu else UPSTREAM_REPO
+
+
 def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     if not host.driver_cuda_version:
         return None
@@ -3001,7 +3165,7 @@ def compatible_linux_runtime_lines(host: HostInfo) -> list[str]:
 
 def windows_runtime_line_info() -> dict[str, tuple[str, ...]]:
     # Generated per CUDA major (newest first) so a new toolkit is detected
-    # without code changes while the cudart64_<major>.dll naming holds.
+    # without a code change while the cudart64_<major>.dll naming holds.
     return {
         f"cuda{m}": (
             f"cudart64_{m}*.dll",
@@ -3027,9 +3191,10 @@ def detected_windows_runtime_lines() -> tuple[list[str], dict[str, list[str]]]:
 def compatible_windows_runtime_lines(host: HostInfo) -> list[str]:
     if not host.driver_cuda_version:
         return []
-    major, minor = host.driver_cuda_version
-    # cuda12 prebuilts need a 12.4+ driver; cuda13+ any minor.
-    if major < _MIN_CUDA_MAJOR or (major == _MIN_CUDA_MAJOR and minor < 4):
+    major, _minor = host.driver_cuda_version
+    # cuda12 app bundles are toolkit-12.8 builds with bundled runtime libs; CUDA
+    # minor-version compatibility runs them on any 12.x driver, same as Linux.
+    if major < _MIN_CUDA_MAJOR:
         return []
     return _cuda_runtime_lines_for_major(major)
 
@@ -3168,8 +3333,8 @@ def windows_cuda_attempts(
     runtime_order.extend(
         runtime_line for runtime_line in normal_runtime_lines if runtime_line not in runtime_order
     )
-    # Keep every driver-compatible line reachable as a fallback, so a line
-    # gated out by driver version still drops to an older major (cuda13->cuda12).
+    # Keep every driver-compatible line reachable as a fallback, so a line gated
+    # out by the driver version still drops to an older major (cuda13 -> cuda12).
     runtime_order.extend(
         runtime_line
         for runtime_line in compatible_runtime_lines
@@ -3188,7 +3353,7 @@ def windows_cuda_attempts(
     for runtime_line in runtime_order:
         major = int(runtime_line.removeprefix("cuda"))
         # Track whatever minor llama.cpp actually ships for this major
-        # (cuda13 -> 13.1, 13.3, ...). Skip the line when the release lacks a
+        # (cuda13 -> 13.1, 13.3, ...). Skip the line when the release has no
         # matching asset instead of guessing a now-missing name.
         runtime = _published_windows_cuda_runtime(upstream_assets, major, host.driver_cuda_version)
         if runtime is None:
@@ -3209,10 +3374,10 @@ def windows_cuda_attempts(
                 + ",".join(windows_cuda_upstream_asset_names(llama_tag, runtime))
             )
             continue
-        # Pair the cudart bundle when upstream ships it; otherwise the binary
-        # needs a system CUDA toolkit on PATH at runtime (#5106). Only pair
-        # when the selected main archive is the binary archive, not the cudart
-        # archive itself.
+        # Pair the cudart bundle when upstream ships it. Without this
+        # the binary needs a system CUDA toolkit on PATH at runtime
+        # (#5106). Only pair when the selected main archive is the
+        # binary archive, not the cudart archive itself.
         runtime_archive_name: str | None = None
         runtime_archive_url: str | None = None
         if selected_name.startswith("llama-"):
@@ -3251,28 +3416,59 @@ def windows_cuda_attempts(
 
 
 def _windows_cuda_attempt_covers_blackwell(attempt: AssetChoice) -> bool:
-    """True if an in-release windows-cuda attempt's toolkit covers Blackwell
-    sm_120 (>= 12.8), read from its asset name's CUDA minor."""
+    """True if an in-release windows-cuda attempt yields a Blackwell sm_120
+    capable build. The fork's app-named bundles declare their SM coverage
+    directly; legacy upstream-named bundles instead encode their CUDA toolkit
+    minor in the filename (covers Blackwell at toolkit >= 12.8)."""
     if attempt.install_kind != "windows-cuda":
         return False
+    # Legacy upstream-named bundles encode their toolkit minor; it is the binding
+    # constraint (a 12.4 toolkit cannot offload sm_120 whatever its metadata says).
     m = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", attempt.name)
-    return m is not None and (int(m.group(1)), int(m.group(2))) >= _BLACKWELL_MIN_TOOLKIT
+    if m is not None:
+        return (int(m.group(1)), int(m.group(2))) >= _BLACKWELL_MIN_TOOLKIT
+    # App-named bundles carry no minor and declare their SM coverage directly.
+    return attempt.max_sm is not None and attempt.max_sm >= _BLACKWELL_MIN_SM
+
+
+def _host_is_blackwell(host: HostInfo) -> bool:
+    caps = normalize_compute_caps(host.compute_caps)
+    return bool(caps) and int(caps[-1]) >= _BLACKWELL_MIN_SM
+
+
+def _drop_blackwell_incapable_windows_cuda(
+    host: HostInfo, attempts: list[AssetChoice]
+) -> list[AssetChoice]:
+    """On a Blackwell host, drop windows-cuda attempts that cannot offload
+    sm_120 (e.g. upstream cuda-12.4, toolkit 12.4). Such a build loads and
+    passes the functional validator but runs the model on a slow non-native
+    path (an RTX 5090 measured 7.1 tok/s vs 551.2 on cuda-13.3), so it must
+    not sit in the fallback chain behind the pin or an in-release cuda13.
+    Non-cuda attempts (windows-cpu, windows-hip, ...) pass through so the
+    host still degrades to an honest CPU install when no CUDA 13 exists."""
+    if not _host_is_blackwell(host):
+        return attempts
+    return [
+        attempt
+        for attempt in attempts
+        if attempt.install_kind != "windows-cuda" or _windows_cuda_attempt_covers_blackwell(attempt)
+    ]
 
 
 def _pinned_windows_cuda_fallback(
     host: HostInfo, existing_cuda_attempts: list[AssetChoice]
 ) -> AssetChoice | None:
     """Pinned GPU fallback for a Blackwell host the in-release build gates off.
-    Upstream stopped publishing a sub-13.3 Windows cuda13 build after b9360,
-    and cuda-12.4 cannot offload sm_120, so a 13.1/13.2 driver would land on
-    CPU. b9360's cuda-13.1 build is immutable and runs on those drivers.
-    Returns None (dormant) whenever the in-release selection already offers a
-    Blackwell-capable build (toolkit >= 12.8, e.g. a runnable cuda13/cuda14),
-    so it self-disables once upstream ships a driver-runnable build again.
+    Upstream stopped publishing a sub-13.3 Windows cuda13 build after b9360, and
+    cuda-12.4 cannot offload sm_120, so a 13.1/13.2 driver would land on CPU.
+    b9360's cuda-13.1 build is immutable and runs on those drivers. Returns None
+    (dormant) whenever the in-release selection already offers a Blackwell-capable
+    build (toolkit >= 12.8, e.g. a runnable cuda13/cuda14), so it self-disables
+    once upstream ships a driver-runnable build again.
 
-    The b9360 binary reuses the current release's source tree and convert
-    scripts and is recorded via binary_release_tag, the same binary/source
-    split used for the lemonade prebuilt."""
+    The b9360 binary reuses the current release's source tree and convert scripts
+    and is recorded via binary_release_tag, the same binary/source split used for
+    the lemonade prebuilt."""
     if not (host.is_windows and host.is_x86_64 and host.has_usable_nvidia):
         return None
     driver = host.driver_cuda_version
@@ -3314,8 +3510,8 @@ def _pinned_windows_cuda_fallback(
 def _augment_checksums_with_pin(
     checksums: ApprovedReleaseChecksums, pin: AssetChoice
 ) -> ApprovedReleaseChecksums:
-    """Add the pin's verified hashes to a copy of the approved checksums so
-    apply_approved_hashes keeps it on the published path (b9360 isn't in the
+    """Add the pin's own verified hashes to a copy of the approved checksums so
+    apply_approved_hashes keeps it on the published path (b9360 is not in the
     release manifest)."""
     artifacts = dict(checksums.artifacts)
     if pin.expected_sha256:
@@ -3339,8 +3535,9 @@ def _with_pinned_windows_cuda_fallback(
     host: HostInfo, attempts: list[AssetChoice], checksums: ApprovedReleaseChecksums
 ) -> tuple[list[AssetChoice], ApprovedReleaseChecksums]:
     """Insert the Blackwell pin ahead of the Windows CUDA attempts and keep it
-    through apply_approved_hashes, or return inputs unchanged when dormant.
+    through apply_approved_hashes, or return the inputs unchanged when dormant.
     Gives the published install path the same GPU fallback as the simple path."""
+    attempts = _drop_blackwell_incapable_windows_cuda(host, attempts)
     pin = _pinned_windows_cuda_fallback(host, attempts)
     if pin is None:
         return attempts, checksums
@@ -3354,30 +3551,6 @@ def published_windows_cuda_attempts(
     selection_preamble: Iterable[str] = (),
 ) -> list[AssetChoice]:
     selection_log = list(release.selection_log) + list(selection_preamble)
-    # Seed the runtime-line ordering from the real published windows-cuda minors
-    # (their names encode the minor), so a future CUDA major published here is
-    # ordered too rather than a hardcoded cuda12/cuda13 pair. Keys mirror the
-    # upstream naming so windows_cuda_attempts can match them; fall back to the
-    # long-standing default when the release lists no windows-cuda asset.
-    published_minors: list[str] = []
-    for artifact in release.artifacts:
-        if artifact.install_kind != "windows-cuda":
-            continue
-        m = re.search(r"-bin-win-cuda-(\d+\.\d+)-x64\.zip$", artifact.asset_name)
-        if m:
-            published_minors.append(m.group(1))
-    if not published_minors:
-        published_minors = ["12.4", "13.1"]
-    runtime_order = windows_cuda_attempts(
-        host,
-        release.upstream_tag,
-        {
-            f"llama-{release.upstream_tag}-bin-win-cuda-{minor}-x64.zip": "published"
-            for minor in published_minors
-        },
-        preferred_runtime_line,
-        selection_log,
-    )
     published_artifacts = [
         artifact for artifact in release.artifacts if artifact.install_kind == "windows-cuda"
     ]
@@ -3387,61 +3560,147 @@ def published_windows_cuda_attempts(
             continue
         artifacts_by_runtime.setdefault(artifact.runtime_line, []).append(artifact)
 
+    # Order the runtime lines to try. Legacy upstream-named bundles encode a CUDA
+    # minor in the filename and are driver-gated per minor. The fork's app-named
+    # bundles carry no minor: their runtime line (cuda12/cuda13) is gated at the
+    # CUDA *major* level (a 13.0 driver runs any cuda13 build) and by what torch
+    # actually provides on the host, preferring the torch line. Routing them
+    # through the synthetic-minor path wrongly dropped cuda13 on a 13.0 driver.
+    legacy_minors: list[str] = []
+    for artifact in published_artifacts:
+        m = re.search(r"-bin-win-cuda-(\d+\.\d+)-x64\.zip$", artifact.asset_name)
+        if m:
+            legacy_minors.append(m.group(1))
+    if legacy_minors:
+        ordered_lines = [
+            attempt.runtime_line
+            for attempt in windows_cuda_attempts(
+                host,
+                release.upstream_tag,
+                {
+                    f"llama-{release.upstream_tag}-bin-win-cuda-{minor}-x64.zip": "published"
+                    for minor in legacy_minors
+                },
+                preferred_runtime_line,
+                selection_log,
+            )
+            if attempt.runtime_line
+        ]
+    else:
+        detected, _ = detected_windows_runtime_lines()
+        compatible = compatible_windows_runtime_lines(host)
+        # Prefer lines whose runtime DLLs are on disk, but fall back to the
+        # driver-derived order when none are detected (Windows torch bundles
+        # cudart in torch/lib, which probing misses) or when detected DLLs are
+        # incompatible with the driver. The app bundle ships its own runtime, so
+        # the driver major is the real constraint. Mirrors the legacy
+        # windows_cuda_attempts fallback; without it a torch-only host gets no
+        # fork attempt and silently drops to the upstream build.
+        ordered_lines = [line for line in compatible if line in detected] or list(compatible)
+        if preferred_runtime_line and preferred_runtime_line in ordered_lines:
+            ordered_lines = [preferred_runtime_line] + [
+                line for line in ordered_lines if line != preferred_runtime_line
+            ]
+        selection_log.append(
+            "windows_cuda_selection: app-bundle runtime lines (major-gated)="
+            + (",".join(ordered_lines) if ordered_lines else "none")
+        )
+
+    host_sms = normalize_compute_caps(host.compute_caps)
     attempts: list[AssetChoice] = []
-    for ordered_attempt in runtime_order:
-        runtime_line = ordered_attempt.runtime_line
+    for runtime_line in ordered_lines:
         if not runtime_line:
             continue
-        candidates = sorted(
-            artifacts_by_runtime.get(runtime_line, []),
-            key = lambda artifact: (artifact.rank, artifact.asset_name),
-        )
-        for artifact in candidates:
+        # Pick the artifact whose SM coverage fits the host, preferring the
+        # tightest targeted bundle and falling back to portable -- the same
+        # policy as linux_cuda_choice_from_release. Without this, app-named
+        # bundles (no minor in the filename) skip the SM filter and the
+        # lowest-rank "older" bundle is chosen for every host, breaking newer
+        # GPUs (e.g. Blackwell sm120 on a cuda12-older bundle capped at sm89).
+        targeted: list[tuple[PublishedLlamaArtifact, str, re.Match[str] | None]] = []
+        portable: tuple[PublishedLlamaArtifact, str, re.Match[str] | None] | None = None
+        for artifact in artifacts_by_runtime.get(runtime_line, []):
             asset_url = release.assets.get(artifact.asset_name)
             if not asset_url:
                 continue
             am = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", artifact.asset_name)
-            # Gate the published minor against the driver so it can never
-            # bypass the driver-version gate.
+            # Legacy upstream-named bundles encode the minor; gate it against the
+            # driver. app-named bundles carry no minor and are driver-gated at the
+            # runtime-line level by windows_cuda_attempts above.
             if (
                 am is not None
                 and host.driver_cuda_version is not None
                 and (int(am.group(1)), int(am.group(2))) > host.driver_cuda_version
             ):
                 continue
-            # See windows_cuda_attempts: pair the cudart bundle for the real minor.
-            runtime_archive_name: str | None = None
-            runtime_archive_url: str | None = None
-            if am is not None and artifact.asset_name.startswith("llama-"):
-                runtime = f"{am.group(1)}.{am.group(2)}"
-                cudart_name = f"cudart-llama-bin-win-cuda-{runtime}-x64.zip"
-                cudart_url = release.assets.get(cudart_name)
-                if cudart_url and cudart_url != asset_url:
-                    runtime_archive_name = cudart_name
-                    runtime_archive_url = cudart_url
-            attempt_log = list(ordered_attempt.selection_log or []) + [
-                "windows_cuda_selection: selected published asset "
-                f"{artifact.asset_name} for runtime_line={runtime_line}"
-            ]
-            if runtime_archive_name:
-                attempt_log.append(
-                    f"windows_cuda_selection: paired published runtime archive {runtime_archive_name}"
-                )
-            attempts.append(
-                AssetChoice(
-                    repo = release.repo,
-                    tag = release.release_tag,
-                    name = artifact.asset_name,
-                    url = asset_url,
-                    source_label = "published",
-                    install_kind = "windows-cuda",
-                    runtime_line = runtime_line,
-                    runtime_name = runtime_archive_name,
-                    runtime_url = runtime_archive_url,
-                    selection_log = attempt_log,
-                )
+            # Only SM-filter artifacts that declare full SM metadata (the app
+            # bundles). Legacy/upstream-named artifacts without it keep the old
+            # rank-based selection rather than being dropped.
+            has_sm_info = (
+                bool(artifact.supported_sms)
+                and artifact.min_sm is not None
+                and artifact.max_sm is not None
             )
-            break
+            if host_sms and has_sm_info and not _artifact_covers_sms(artifact, host_sms):
+                continue
+            if not host_sms and has_sm_info and artifact.coverage_class != "portable":
+                continue
+            if artifact.coverage_class == "portable":
+                portable = (artifact, asset_url, am)
+            else:
+                targeted.append((artifact, asset_url, am))
+        chosen: tuple[PublishedLlamaArtifact, str, re.Match[str] | None] | None = None
+        if targeted:
+            chosen = sorted(
+                targeted,
+                key = lambda item: (
+                    _sm_range(item[0]),
+                    item[0].rank,
+                    item[0].max_sm or 0,
+                ),
+            )[0]
+        elif portable is not None:
+            chosen = portable
+        if chosen is None:
+            continue
+        artifact, asset_url, am = chosen
+        # See windows_cuda_attempts: pair the cudart bundle for the real minor.
+        runtime_archive_name: str | None = None
+        runtime_archive_url: str | None = None
+        if am is not None and artifact.asset_name.startswith("llama-"):
+            runtime = f"{am.group(1)}.{am.group(2)}"
+            cudart_name = f"cudart-llama-bin-win-cuda-{runtime}-x64.zip"
+            cudart_url = release.assets.get(cudart_name)
+            if cudart_url and cudart_url != asset_url:
+                runtime_archive_name = cudart_name
+                runtime_archive_url = cudart_url
+        attempt_log = list(selection_log) + [
+            "windows_cuda_selection: selected published asset "
+            f"{artifact.asset_name} for runtime_line={runtime_line}"
+        ]
+        if runtime_archive_name:
+            attempt_log.append(
+                f"windows_cuda_selection: paired published runtime archive {runtime_archive_name}"
+            )
+        attempts.append(
+            AssetChoice(
+                repo = release.repo,
+                tag = release.release_tag,
+                name = artifact.asset_name,
+                url = asset_url,
+                source_label = "published",
+                install_kind = "windows-cuda",
+                runtime_line = runtime_line,
+                runtime_name = runtime_archive_name,
+                runtime_url = runtime_archive_url,
+                bundle_profile = artifact.bundle_profile,
+                coverage_class = artifact.coverage_class,
+                supported_sms = artifact.supported_sms,
+                min_sm = artifact.min_sm,
+                max_sm = artifact.max_sm,
+                selection_log = attempt_log,
+            )
+        )
     return attempts
 
 
@@ -3503,9 +3762,9 @@ def _detect_host_rocm_version() -> tuple[int, int] | None:
     """Return (major, minor) of the installed ROCm runtime, or None.
 
     Best-effort read from /opt/rocm/.info/version, amd-smi version, and
-    hipconfig --version. Used to pick a compatible upstream llama.cpp ROCm
-    prebuilt rather than the numerically newest one (which can be newer than
-    the host runtime).
+    hipconfig --version. Used to pick a compatible upstream llama.cpp
+    ROCm prebuilt rather than always taking the numerically newest one
+    (which can be newer than the host runtime).
     """
     rocm_root = os.environ.get("ROCM_PATH") or "/opt/rocm"
     for path in (
@@ -3515,23 +3774,19 @@ def _detect_host_rocm_version() -> tuple[int, int] | None:
         try:
             with open(path) as fh:
                 parts = fh.read().strip().split("-")[0].split(".")
-            # Explicit length guard so we don't rely on the broad except
-            # below to swallow IndexError when the version file has a single
-            # component (e.g. "6\n" on a partial install).
+            # Explicit length guard avoids relying on the broad except
+            # below to swallow IndexError when the version file contains
+            # a single component (e.g. "6\n" on a partial install).
             if len(parts) >= 2:
                 return int(parts[0]), int(parts[1])
         except Exception:
             pass
-    amd_smi = shutil.which("amd-smi")
+    amd_smi = shutil.which("amd-smi") if _amd_smi_allowed() else None
     if amd_smi:
         try:
-            result = subprocess.run(
-                [amd_smi, "version"],
-                stdout = subprocess.PIPE,
-                stderr = subprocess.DEVNULL,
-                text = True,
-                timeout = 5,
-            )
+            # Off on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt);
+            # hipconfig below and the version-file reads above cover that case.
+            result = run_capture([amd_smi, "version"], timeout = 5)
             if result.returncode == 0:
                 m = re.search(r"ROCm version:\s*(\d+)\.(\d+)", result.stdout)
                 if m:
@@ -3558,8 +3813,8 @@ def _detect_host_rocm_version() -> tuple[int, int] | None:
 
     # Distro package-manager fallbacks. Mirrors install.sh::get_torch_index_url
     # and _detect_rocm_version() in install_python_stack.py so package-managed
-    # ROCm hosts without /opt/rocm/.info/version still report a usable version,
-    # letting the <= host version filter in resolve_upstream_asset_choice pick
+    # ROCm hosts without /opt/rocm/.info/version still report a usable version
+    # and the <= host version filter in resolve_upstream_asset_choice picks
     # the correct upstream prebuilt instead of the newest-regardless fallback.
     for _cmd in (
         ["dpkg-query", "-W", "-f=${Version}\n", "rocm-core"],
@@ -3581,7 +3836,7 @@ def _detect_host_rocm_version() -> tuple[int, int] | None:
         if _result.returncode != 0 or not _result.stdout.strip():
             continue
         _raw = _result.stdout.strip()
-        # dpkg can prepend an epoch ("1:6.3.0-1"); strip it first.
+        # dpkg can prepend an epoch ("1:6.3.0-1"); strip it before parsing.
         _raw = re.sub(r"^\d+:", "", _raw)
         _m = re.match(r"(\d+)[.-](\d+)", _raw)
         if _m:
@@ -3590,7 +3845,7 @@ def _detect_host_rocm_version() -> tuple[int, int] | None:
 
 
 # Map detected gfx IDs to lemonade-sdk asset family suffixes.
-# More-specific prefixes must precede shorter ones (e.g. gfx1151 before gfx110).
+# More-specific prefixes must come before shorter ones (e.g. gfx1151 before gfx110).
 _LEMONADE_GFX_FAMILIES: list[tuple[str, str]] = [
     ("gfx1151", "gfx1151"),
     ("gfx1150", "gfx1150"),
@@ -3608,14 +3863,56 @@ def _lemonade_gfx_family(gfx_id: str) -> str | None:
     return None
 
 
+def published_rocm_choice_for_host(
+    release: PublishedReleaseBundle, host: HostInfo, install_kind: str
+) -> AssetChoice | None:
+    """Select the published ROCm bundle whose gfx target covers the host GPU.
+
+    The manifest's gfx_target uses the same umbrella family labels that
+    _lemonade_gfx_family produces (gfx110X, gfx120X, ...), so the host's detected
+    gfx is matched either to that family or to the bundle's concrete
+    mapped_targets list. Returns None when no published bundle covers the GPU, so
+    the caller can fall back (lemonade / upstream HIP)."""
+    if not host.rocm_gfx_target:
+        return None
+    gfx = host.rocm_gfx_target.lower().strip()
+    for artifact in release.artifacts:
+        if artifact.install_kind != install_kind:
+            continue
+        # Match on the concrete built-arch list, not the family prefix: an
+        # in-generation-but-unbuilt arch (e.g. gfx1033 in the gfx103 prefix) must
+        # NOT be served the family bundle. None makes the caller fall back to a
+        # source build for that GPU.
+        if gfx not in {target.lower() for target in artifact.mapped_targets}:
+            continue
+        asset_url = release.assets.get(artifact.asset_name)
+        if not asset_url:
+            continue
+        return AssetChoice(
+            repo = release.repo,
+            tag = release.release_tag,
+            name = artifact.asset_name,
+            url = asset_url,
+            source_label = "published",
+            install_kind = install_kind,
+            selection_log = list(release.selection_log)
+            + [
+                f"rocm_selection: gpu={host.rocm_gfx_target} "
+                f"selected published {artifact.asset_name}"
+            ],
+        )
+    return None
+
+
 def _is_trusted_github_release_url(url: str, expected_repo: str) -> bool:
     """Validate a release asset URL points at GitHub's expected hosts.
 
     Accepts:
       https://github.com/{expected_repo}/releases/download/...
       https://objects.githubusercontent.com/...   (GitHub's release CDN)
-    Anything else (http://, raw.githubusercontent.com, gist, etc.) is rejected
-    so a malicious API response can't redirect downloads to an attacker host.
+    Anything else (including http://, raw.githubusercontent.com, gist, etc.)
+    is rejected so a malicious API response cannot redirect downloads to an
+    attacker-chosen host.
     """
     if not isinstance(url, str) or not url:
         return False
@@ -3628,8 +3925,8 @@ def _is_trusted_github_release_url(url: str, expected_repo: str) -> bool:
     host = (parsed.netloc or "").lower()
     if host == "objects.githubusercontent.com":
         # GitHub's release CDN. Restrict to release-asset paths so a tampered
-        # API response pointing at an arbitrary CDN object is rejected. Real
-        # release asset URLs carry the "/github-production-release-asset-"
+        # API response pointing at an arbitrary CDN object is still rejected.
+        # Real release asset URLs carry the "/github-production-release-asset-"
         # prefix; gist / raw / avatar CDN paths do not.
         return parsed.path.startswith("/github-production-release-asset-")
     if host == "github.com":
@@ -3637,16 +3934,22 @@ def _is_trusted_github_release_url(url: str, expected_repo: str) -> bool:
     return False
 
 
+# (gfx_target, asset_name) pairs already logged. resolve_lemonade_rocm_choice()
+# runs twice per install (direct planner + resolve_upstream_asset_choice), so
+# this stops its selection banner and hash-manifest NOTE printing twice.
+_lemonade_selection_logged: "set[tuple[str, str]]" = set()
+
+
 @functools.lru_cache(maxsize = 8)
 def _fetch_lemonade_release_cached(api_url: str, llama_tag: str) -> "dict | None":
     """Cached wrapper around fetch_json for lemonade release lookups.
 
-    resolve_lemonade_rocm_choice() is called twice per install (direct planner
-    + resolve_upstream_asset_choice) with identical arguments. Without
-    memoisation each install hits api.github.com twice, doubling the
-    rate-limit failure surface on busy CI runners. Cache is process-scoped;
-    tests that vary fetch_json's return value across calls should call
-    cache_clear().
+    resolve_lemonade_rocm_choice() is called twice per install (once from the
+    direct planner, once from resolve_upstream_asset_choice) with identical
+    arguments. Without memoisation, each install hits api.github.com twice,
+    doubling the rate-limit failure surface on busy CI runners. Cache is
+    process-scoped; tests that need to vary fetch_json's return value across
+    invocations should call cache_clear().
     """
     try:
         return fetch_json(api_url)
@@ -3672,21 +3975,23 @@ def resolve_lemonade_rocm_choice(
 
     os_prefix:   lemonade's asset filename label, NOT a host-distro filter.
                  Pass "ubuntu" for any Linux host (Arch, Fedora, openSUSE,
-                 Debian, ...) -- lemonade publishes one Linux variant, a
-                 manylinux-style glibc build that runs on any distro with a
-                 recent-enough glibc. Pass "windows" for Windows hosts.
+                 Debian, ...) -- lemonade only publishes one Linux variant
+                 and it is a manylinux-style glibc build that runs on any
+                 distro with a recent-enough glibc. Pass "windows" for
+                 Windows hosts.
     install_kind: "linux-rocm" or "windows-hip"
     llama_tag:   the requested upstream llama.cpp tag ("latest" or a pinned
-                 release like "b1260"). When pinned, fetch the matching
-                 lemonade release; if lemonade hasn't published that tag, skip
-                 silently (caller falls through to upstream) rather than drift
-                 to whatever lemonade ships as latest.
+                 release like "b1260"). When pinned, the resolver fetches
+                 the matching lemonade release. When the pinned tag is not
+                 published by lemonade we skip silently (and the caller
+                 falls through to upstream) rather than drift to whatever
+                 lemonade ships as latest.
     """
     if not host.rocm_gfx_target:
         return None
     # Opt-out for users who want the upstream HIP build path only -- lemonade
-    # binaries lack approved-hash manifest entries, so their integrity gate is
-    # functional validation only.
+    # binaries are downloaded without entries in the approved-hash manifest, so
+    # the integrity gate is functional validation only.
     if os.environ.get("UNSLOTH_DISABLE_LEMONADE_ROCM", "").strip().lower() in (
         "1",
         "true",
@@ -3719,7 +4024,7 @@ def resolve_lemonade_rocm_choice(
         return None
     asset_url = assets[asset_name]
     if not asset_url:
-        # release_asset_map defaults to "" when an asset row lacks
+        # release_asset_map defaults to "" when an asset row is missing
         # browser_download_url; skip cleanly instead of letting
         # download_file("") raise a less obvious error downstream.
         log(
@@ -3728,9 +4033,9 @@ def resolve_lemonade_rocm_choice(
         )
         return None
     # Defence in depth: lemonade browser_download_url should be on github.com
-    # or githubusercontent.com. A compromised GitHub API response redirecting
-    # to an attacker host would otherwise be honoured silently (lemonade
-    # assets are not in the approved-hash manifest).
+    # or githubusercontent.com. A compromised GitHub API response that
+    # redirects to an attacker-chosen host would otherwise be honoured
+    # silently (lemonade assets are not in the approved-hash manifest).
     if not _is_trusted_github_release_url(asset_url, LEMONADE_ROCM_REPO):
         log(
             f"{LEMONADE_ROCM_REPO}@{release_tag} asset {asset_name!r} points "
@@ -3738,21 +4043,25 @@ def resolve_lemonade_rocm_choice(
             "lemonade prebuilt"
         )
         return None
-    # Note: lemonade tags Linux assets "ubuntu" but the binary is a generic
-    # glibc build that runs on any distro (Arch, Fedora, ...), so this attempt
-    # is selected for all Linux ROCm hosts, not just Ubuntu.
-    log(
-        f"AMD GPU {host.rocm_gfx_target!r} ({gfx_family}) -- "
-        f"trying lemonade-sdk ROCm prebuilt {asset_name} "
-        f"(works on any glibc Linux, not just Ubuntu)"
-    )
-    log(
-        f"NOTE: lemonade-sdk/llamacpp-rocm releases are not covered by the "
-        f"Unsloth approved-hash manifest; download integrity relies on "
-        f"functional validation (llama-bench / llama-server smoke tests) "
-        f"after extraction. Set UNSLOTH_DISABLE_LEMONADE_ROCM=1 to skip "
-        f"lemonade and fall back to the upstream HIP build path."
-    )
+    # Note: lemonade tags Linux assets with "ubuntu" but the binary is a
+    # generic glibc build that runs on any distro (Arch, Fedora, ...), so
+    # this attempt is selected for all Linux ROCm hosts, not just Ubuntu.
+    # Log once per (gfx_target, asset); see _lemonade_selection_logged.
+    log_key = (host.rocm_gfx_target, asset_name)
+    if log_key not in _lemonade_selection_logged:
+        _lemonade_selection_logged.add(log_key)
+        log(
+            f"AMD GPU {host.rocm_gfx_target!r} ({gfx_family}) -- "
+            f"trying lemonade-sdk ROCm prebuilt {asset_name} "
+            f"(works on any glibc Linux, not just Ubuntu)"
+        )
+        log(
+            f"NOTE: lemonade-sdk/llamacpp-rocm releases are not covered by the "
+            f"Unsloth approved-hash manifest; download integrity relies on "
+            f"functional validation (llama-bench / llama-server smoke tests) "
+            f"after extraction. Set UNSLOTH_DISABLE_LEMONADE_ROCM=1 to skip "
+            f"lemonade and fall back to the upstream HIP build path."
+        )
     return AssetChoice(
         repo = LEMONADE_ROCM_REPO,
         tag = release_tag,
@@ -3763,30 +4072,40 @@ def resolve_lemonade_rocm_choice(
     )
 
 
-def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice:
+def resolve_upstream_asset_choice(
+    host: HostInfo,
+    llama_tag: str,
+    lemonade_tag: "str | None" = None,
+) -> AssetChoice:
+    # lemonade_tag: tag for the lemonade lookup only. The release scan pins
+    # llama_tag to per-release upstream tags (b9518, ...) that lemonade's own
+    # tag series (b1292, ...) never contains, so pinning lemonade to them 404s
+    # on every scanned release. Scan callers pass the original request
+    # (normally "latest") here; upstream asset names keep the pinned tag.
     upstream_assets = github_release_assets(UPSTREAM_REPO, llama_tag)
     if host.is_linux and host.is_x86_64:
-        # AMD ROCm: try upstream ROCm prebuilt first, then a source build. The
-        # source build (via setup.sh) compiles with -DGGML_HIP=ON and
-        # auto-detects the exact GPU target via rocminfo, more reliable for
-        # consumer GPUs (e.g. gfx1151) that may not be in the prebuilt.
+        # AMD ROCm: try upstream ROCm prebuilt first, then fall back to source build.
+        # Source build (via setup.sh) compiles with -DGGML_HIP=ON and auto-detects
+        # the exact GPU target via rocminfo, which is more reliable for consumer
+        # GPUs (e.g. gfx1151) that may not be in the prebuilt.
         if host.has_rocm and not host.has_usable_nvidia:
-            # Try lemonade-sdk per-GPU prebuilt first: built against specific
-            # gfx targets and bundle all required ROCm runtime libs.
+            # Try lemonade-sdk per-GPU prebuilt first: these are built against
+            # specific gfx targets and bundle all required ROCm runtime libs.
             lemonade_choice = resolve_lemonade_rocm_choice(
-                host, "ubuntu", "linux-rocm", llama_tag = llama_tag
+                host, "ubuntu", "linux-rocm", llama_tag = lemonade_tag or llama_tag
             )
             if lemonade_choice is not None:
                 return lemonade_choice
 
-            # Fall back to the upstream combined ROCm tarball. Scan for any
-            # rocm-<version> prebuilt; when the host ROCm version is known,
-            # pick the newest candidate whose major.minor is <= host version
-            # -- otherwise a ROCm 6.4 host downloads the rocm-7.2 tarball,
-            # fails preflight, and source-builds even though a 6.4 prebuilt
-            # exists. If none is compatible (host older than every published
-            # prebuilt), fall back to the numerically newest so we try
-            # something.
+            # Fall back to upstream combined ROCm tarball.
+            # Scan upstream assets for any rocm-<version> prebuilt. When the
+            # host ROCm runtime version is known, pick the newest candidate
+            # whose major.minor is <= host version -- otherwise a ROCm 6.4
+            # host would download the rocm-7.2 tarball, fail preflight, and
+            # fall back to a source build even though a compatible 6.4
+            # prebuilt exists. If no compatible candidate matches (e.g. host
+            # runtime is older than every published prebuilt), fall back to
+            # the numerically newest so we at least try something.
             _rocm_pattern = re.compile(
                 rf"llama-{re.escape(llama_tag)}-bin-ubuntu-rocm-([0-9]+(?:\.[0-9]+)*)-x64\.tar\.gz"
             )
@@ -3805,10 +4124,10 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                     item for item in rocm_candidates if item[0][:2] <= _host_rocm_version
                 ]
             if rocm_candidates and not _compatible:
-                # Fall back to the newest candidate so we don't force a source
-                # build when the host runtime is older than every published
-                # prebuilt: preflight still catches a true incompatibility and
-                # triggers a fallback.
+                # Fall back to the newest candidate so a source build is
+                # not forced when the host runtime is older than every
+                # published prebuilt: preflight will still catch a true
+                # incompatibility and trigger a fallback.
                 _compatible = rocm_candidates[:1]
             if _compatible:
                 rocm_name = _compatible[0][1]
@@ -3831,7 +4150,7 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                     source_label = "upstream",
                     install_kind = "linux-rocm",
                 )
-            # No ROCm prebuilt available -- fall back to a source build
+            # No ROCm prebuilt available -- fall back to source build
             raise PrebuiltFallback(
                 "AMD ROCm detected but no upstream ROCm prebuilt found; "
                 "falling back to source build with HIP support"
@@ -3859,7 +4178,7 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
         # AMD ROCm on Windows: try lemonade per-GPU prebuilt first, then upstream HIP
         if host.has_rocm:
             lemonade_choice = resolve_lemonade_rocm_choice(
-                host, "windows", "windows-hip", llama_tag = llama_tag
+                host, "windows", "windows-hip", llama_tag = lemonade_tag or llama_tag
             )
             if lemonade_choice is not None:
                 return lemonade_choice
@@ -3918,12 +4237,16 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
     raise PrebuiltFallback(f"no prebuilt policy exists for {host.system} {host.machine}")
 
 
-def resolve_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice:
+def resolve_asset_choice(
+    host: HostInfo,
+    llama_tag: str,
+    lemonade_tag: "str | None" = None,
+) -> AssetChoice:
     if host.is_linux and host.is_x86_64 and host.has_usable_nvidia:
         raise PrebuiltFallback(
             "Linux CUDA installs require a compatible published bundle; upstream fallback is not available"
         )
-    return resolve_upstream_asset_choice(host, llama_tag)
+    return resolve_upstream_asset_choice(host, llama_tag, lemonade_tag = lemonade_tag)
 
 
 def resolve_release_asset_choice(
@@ -3931,6 +4254,7 @@ def resolve_release_asset_choice(
     llama_tag: str,
     release: PublishedReleaseBundle,
     checksums: ApprovedReleaseChecksums,
+    requested_tag: "str | None" = None,
 ) -> list[AssetChoice]:
     if host.is_windows and host.is_x86_64 and host.has_usable_nvidia:
         torch_preference = detect_torch_cuda_runtime_preference(host)
@@ -3961,13 +4285,15 @@ def resolve_release_asset_choice(
 
     published_choice: AssetChoice | None = None
     if host.is_windows and host.is_x86_64:
-        # AMD Windows hosts prefer a hash-approved published Windows HIP
-        # bundle when one exists, otherwise fall through to
-        # resolve_asset_choice() so the upstream HIP prebuilt is tried before
-        # the CPU fallback. Hard-pinning the published windows-cpu bundle here
-        # would make the HIP path unreachable.
+        # AMD Windows hosts prefer the fork's per-gfx windows-rocm bundle when one
+        # covers the GPU; otherwise fall through to resolve_asset_choice(). Note
+        # that on the fork repo the upstream win-hip archive has no approved hash,
+        # so apply_approved_hashes drops it and an uncovered gfx lands on a HIP
+        # source build (auto-detecting its exact gfx) rather than an upstream
+        # prebuilt. We still avoid hard-pinning windows-cpu here so a CPU bundle
+        # never shadows that ROCm path.
         if host.has_rocm:
-            published_choice = published_asset_choice_for_kind(release, "windows-hip")
+            published_choice = published_rocm_choice_for_host(release, host, "windows-rocm")
         else:
             published_choice = published_asset_choice_for_kind(release, "windows-cpu")
     elif host.is_macos and host.is_arm64:
@@ -3984,7 +4310,10 @@ def resolve_release_asset_choice(
                 f"{release.repo}@{release.release_tag} {published_choice.name} ({exc})"
             )
 
-    return apply_approved_hashes([resolve_asset_choice(host, llama_tag)], checksums)
+    return apply_approved_hashes(
+        [resolve_asset_choice(host, llama_tag, lemonade_tag = requested_tag)],
+        checksums,
+    )
 
 
 def extract_archive(archive_path: Path, destination: Path) -> None:
@@ -4239,13 +4568,17 @@ def hydrate_source_tree(
     expected_sha256: str | None,
     source_label: str | None = None,
     exact_source: bool = False,
+    asset_url: str | None = None,
 ) -> None:
     archive_path = work_dir / f"llama.cpp-source-{source_ref}.tar.gz"
-    source_urls = (
+    repo_urls = (
         commit_source_archive_urls(source_repo, source_ref)
         if exact_source
         else upstream_source_archive_urls(source_ref)
     )
+    # Prefer the published release asset (the only copy of a mix build's merged
+    # tree); fall back to codeload/archive for vanilla builds whose commit is real.
+    source_urls = ([asset_url] if asset_url else []) + repo_urls
     label = source_label or f"llama.cpp source tree for {source_ref}"
     extract_dir = Path(tempfile.mkdtemp(prefix = "source-extract-", dir = work_dir))
 
@@ -4369,7 +4702,13 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
     # libraries between b9279 and b9283) without us re-enumerating
     # every new file. Studio only invokes llama-server and llama-quantize;
     # other CLIs upstream ships (llama-cli, llama-bench, ...) are skipped.
-    if choice.install_kind in {"linux-cpu", "linux-cuda", "linux-rocm", "linux-arm64"}:
+    if choice.install_kind in {
+        "linux-cpu",
+        "linux-cuda",
+        "linux-arm64-cuda",
+        "linux-rocm",
+        "linux-arm64",
+    }:
         return ["llama-server", "llama-quantize", "lib*.so*"]
     if choice.install_kind in {"macos-arm64", "macos-x64"}:
         return ["llama-server", "llama-quantize", "lib*.dylib"]
@@ -4377,6 +4716,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "windows-cpu",
         "windows-cuda",
         "windows-hip",
+        "windows-rocm",
         "windows-arm64",
     }:
         return ["llama-server.exe", "llama-quantize.exe", "*.dll"]
@@ -4391,10 +4731,7 @@ def runtime_subdirs_for_choice(choice: AssetChoice) -> list[str]:
     (hipblaslt/library/<gfx>/ and rocblas/library/<gfx>/) to sit next to
     their shared libraries at runtime.  These trees are multi-level and
     cannot be handled by copy_globs (filename-only matching, flat copy)."""
-    if choice.source_label == "lemonade" and choice.install_kind in {
-        "linux-rocm",
-        "windows-hip",
-    }:
+    if choice.install_kind in {"linux-rocm", "windows-rocm", "windows-hip"}:
         return ["hipblaslt", "rocblas"]
     return []
 
@@ -4592,6 +4929,35 @@ def confirm_install_tree(install_dir: Path, host: HostInfo) -> None:
         raise RuntimeError("activated install was missing expected files: " + ", ".join(missing))
 
 
+def activate_staged_dir(staging_dir: Path, dst: Path) -> None:
+    """Move a freshly extracted ``staging_dir`` onto ``dst``.
+
+    ``os.replace`` is attempted first as the fast path. On Windows ARM64 the
+    antivirus scanner can transiently hold a freshly extracted DLL open at the
+    moment ``MoveFileEx`` runs, surfacing as ``[WinError 5] Access is denied``;
+    a file-by-file copy bypasses the rename entirely.
+
+    This fallback is intentionally limited to staging trees we just extracted.
+    It must not be used to move an existing/active install aside: there an
+    ``os.replace`` failure means the directory is genuinely in use, and a
+    silent copy + ``rmtree`` could partially delete a live install.
+
+    Only busy/lock errors (``is_busy_lock_error``) trigger the copy; anything
+    else (disk full, cross-device, missing path) re-raises so it cannot leave
+    a partially copied install behind. A copy is preferred over retrying the
+    rename because antivirus scans of large DLLs can outlast any reasonable
+    retry window.
+    """
+    try:
+        os.replace(staging_dir, dst)
+    except OSError as exc:
+        if not is_busy_lock_error(exc):
+            raise
+        log(f"os.replace failed ({exc!r}); falling back to file-by-file copy of staging tree")
+        shutil.copytree(staging_dir, dst, dirs_exist_ok = True)
+        remove_tree(staging_dir)
+
+
 def activate_install_tree(staging_dir: Path, install_dir: Path, host: HostInfo) -> None:
     rollback_dir: Path | None = None
     failed_dir: Path | None = None
@@ -4603,7 +4969,7 @@ def activate_install_tree(staging_dir: Path, install_dir: Path, host: HostInfo)
             log(f"moved existing install to rollback path {rollback_dir.name}")
 
         log(f"activating staged install {staging_dir} -> {install_dir}")
-        os.replace(staging_dir, install_dir)
+        activate_staged_dir(staging_dir, install_dir)
         log(f"activated staged install at {install_dir}")
         log(f"confirming activated install tree at {install_dir}")
         confirm_install_tree(install_dir, host)
@@ -4809,6 +5175,46 @@ def validated_validation_model_bytes(data: bytes) -> bytes:
     return data
 
 
+def _hf_resolve_url_parts(url: str) -> tuple[str, str, str] | None:
+    """Parse a huggingface.co .../resolve/<rev>/<path> URL into
+    (repo_id, revision, filename); None if it is not such a URL."""
+    try:
+        parsed = urllib.parse.urlparse(url)
+    except Exception:
+        return None
+    if (parsed.netloc or "").lower() not in ("huggingface.co", "www.huggingface.co"):
+        return None
+    parts = parsed.path.strip("/").split("/")
+    # <owner>/<name>/resolve/<rev>/<path...>
+    if len(parts) >= 5 and parts[2] == "resolve":
+        return f"{parts[0]}/{parts[1]}", parts[3], "/".join(parts[4:])
+    return None
+
+
+def _fetch_validation_model_bytes() -> bytes:
+    """Fetch the tiny GGUF validation model. Prefer huggingface_hub (completes
+    TLS chains via AIA fetching that bare urllib can't on some Windows/proxy
+    setups); fall back to the direct URL when hf_hub is unavailable or fails."""
+    parts = _hf_resolve_url_parts(TEST_MODEL_URL)
+    if parts is not None:
+        repo_id, revision, filename = parts
+        try:
+            from huggingface_hub import hf_hub_download
+            local = hf_hub_download(repo_id = repo_id, filename = filename, revision = revision)
+            return validated_validation_model_bytes(Path(local).read_bytes())
+        except Exception as exc:
+            log(
+                f"huggingface_hub fetch of validation model failed ({exc}); "
+                "falling back to direct URL"
+            )
+    return validated_validation_model_bytes(
+        download_bytes(
+            TEST_MODEL_URL,
+            progress_label = f"Downloading {download_label_from_url(TEST_MODEL_URL)}",
+        )
+    )
+
+
 def download_validation_model(path: Path, cache_path: Path | None = None) -> None:
     try:
         data: bytes | None = None
@@ -4821,12 +5227,7 @@ def download_validation_model(path: Path, cache_path: Path | None = None) -> Non
                 data = None
         if data is None:
             log("downloading tiny GGUF validation model")
-            data = validated_validation_model_bytes(
-                download_bytes(
-                    TEST_MODEL_URL,
-                    progress_label = f"Downloading {download_label_from_url(TEST_MODEL_URL)}",
-                )
-            )
+            data = _fetch_validation_model_bytes()
             if cache_path is not None:
                 atomic_write_bytes(cache_path, data)
         atomic_write_bytes(path, data)
@@ -5356,9 +5757,11 @@ def validate_server(
         # pass one (keeps backwards compatibility with older call sites).
         _gpu_kinds = {
             "linux-cuda",
+            "linux-arm64-cuda",
             "linux-rocm",
             "windows-cuda",
             "windows-hip",
+            "windows-rocm",
             "macos-arm64",
         }
         if install_kind is not None:
@@ -5677,7 +6080,7 @@ def selected_source_archive_metadata(
 def resolve_install_attempts(
     llama_tag: str, host: HostInfo, published_repo: str, published_release_tag: str
 ) -> tuple[str, str, list[AssetChoice], ApprovedReleaseChecksums]:
-    requested_tag, plans = resolve_install_release_plans(
+    requested_tag, plans = _fork_manifest_release_plans(
         llama_tag,
         host,
         published_repo,
@@ -5689,7 +6092,50 @@ def resolve_install_attempts(
     return requested_tag, plan.llama_tag, plan.attempts, plan.approved_checksums
 
 
-def resolve_install_release_plans(
+def _linux_published_attempts(
+    host: HostInfo, bundle: PublishedReleaseBundle, requested_tag: str
+) -> list[AssetChoice]:
+    """Build the install attempts for a fork Linux host from a manifest-described
+    bundle: CUDA (with a CPU fallback), per-gfx ROCm (with a lemonade fallback),
+    or CPU. Same selection the upstream filename path used, just sourced from the
+    manifest instead of reconstructed from asset names."""
+    attempts: list[AssetChoice] = []
+    if host.has_usable_nvidia:
+        # Prefer the cudart major Studio loads at runtime (torch's bundled
+        # libcudart), not the newest detected on disk. Without this a stray
+        # cuda13 runtime outranks the torch cuda12 the binary links against.
+        torch_preference = detect_torch_cuda_runtime_preference(host)
+        selection = linux_cuda_choice_from_release(
+            host,
+            bundle,
+            preferred_runtime_line = torch_preference.runtime_line,
+            selection_preamble = torch_preference.selection_log,
+        )
+        if selection is not None:
+            attempts.extend(selection.attempts)
+    if host.has_rocm and not host.has_usable_nvidia:
+        # Prefer the fork's own per-gfx ROCm bundle (hash-approved, ships the
+        # full ROCm runtime) and fall back to the external lemonade prebuilt.
+        # Do NOT append the CPU asset for ROCm-only hosts: if lemonade fails
+        # validation we want validate_prebuilt_attempts to raise PrebuiltFallback
+        # so the caller triggers the HIP source build, not silently install a
+        # CPU-only binary.
+        published_rocm = published_rocm_choice_for_host(bundle, host, "linux-rocm")
+        if published_rocm is not None:
+            attempts.append(published_rocm)
+        lemonade_choice = resolve_lemonade_rocm_choice(
+            host, "ubuntu", "linux-rocm", llama_tag = requested_tag
+        )
+        if lemonade_choice is not None:
+            attempts.append(lemonade_choice)
+    else:
+        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
+        if cpu_choice is not None:
+            attempts.append(cpu_choice)
+    return attempts
+
+
+def _fork_manifest_release_plans(
     llama_tag: str,
     host: HostInfo,
     published_repo: str,
@@ -5697,6 +6143,10 @@ def resolve_install_release_plans(
     *,
     max_release_fallbacks: int = DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS,
 ) -> tuple[str, list[InstallReleasePlan]]:
+    """Manifest-reading branch of resolve_simple_install_release_plans, used for
+    the fork's bundles whose GPU/arch coverage lives in
+    llama-prebuilt-manifest.json rather than in the filename: arm64 CUDA, Windows
+    CUDA, per-gfx ROCm, and macOS. Linux x64 takes the faster filename path."""
     requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
     release_limit = max(1, max_release_fallbacks)
@@ -5716,18 +6166,22 @@ def resolve_install_release_plans(
         checksums = resolved_release.checksums
         resolved_tag = bundle.upstream_tag
         try:
-            if host.is_linux and host.is_x86_64 and host.has_usable_nvidia:
-                linux_cuda_selection = resolve_linux_cuda_choice(host, bundle)
-                attempts = apply_approved_hashes(linux_cuda_selection.attempts, checksums)
+            if host.is_linux:
+                linux_attempts = _linux_published_attempts(host, bundle, requested_tag)
+                if not linux_attempts:
+                    raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
+                attempts = apply_approved_hashes(linux_attempts, checksums)
                 if not attempts:
-                    raise PrebuiltFallback("no compatible Linux CUDA asset was found")
-                log_lines(linux_cuda_selection.selection_log)
+                    raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
+                if attempts[0].selection_log:
+                    log_lines(attempts[0].selection_log)
             else:
                 attempts = resolve_release_asset_choice(
                     host,
                     resolved_tag,
                     bundle,
                     checksums,
+                    requested_tag = requested_tag,
                 )
                 if not attempts:
                     raise PrebuiltFallback("no compatible prebuilt asset was found")
@@ -5883,7 +6337,7 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
             ["libggml-cpu*.so*"],
             ["libmtmd.so*"],
         ]
-    if choice.install_kind == "linux-cuda":
+    if choice.install_kind in {"linux-cuda", "linux-arm64-cuda"}:
         return [
             ["libllama-common.so*"],
             ["libllama.so*"],
@@ -5926,7 +6380,7 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
             groups.append(["cublas64_*.dll"])
             groups.append(["cublasLt64_*.dll"])
         return groups
-    if choice.install_kind == "windows-hip":
+    if choice.install_kind in {"windows-hip", "windows-rocm"}:
         return [["llama.dll"], ["*hip*.dll"]]
     return []
 
@@ -6056,6 +6510,15 @@ def validate_prebuilt_choice(
     source_repo, source_ref, source_archive, exact_source = preferred_source_archive(
         approved_checksums, llama_tag
     )
+    # For an exact (mix) source the merge commit lives only in the release asset,
+    # not in any repo, so fetch the asset directly; codeload stays the fallback.
+    asset_url = (
+        release_asset_download_url(
+            approved_checksums.repo, approved_checksums.release_tag, source_archive.asset_name
+        )
+        if exact_source and source_archive is not None
+        else None
+    )
     if exact_source:
         log(f"hydrating exact llama.cpp source for {source_repo}@{source_ref} into {install_dir}")
     else:
@@ -6072,6 +6535,7 @@ def validate_prebuilt_choice(
             else f"llama.cpp source tree for {llama_tag}"
         ),
         exact_source = exact_source,
+        asset_url = asset_url,
     )
     log(f"overlaying prebuilt bundle {choice.name} into {install_dir}")
     server_path, quantize_path = install_from_archives(choice, host, install_dir, work_dir)
@@ -6195,7 +6659,6 @@ def install_prebuilt(
     published_repo: str,
     published_release_tag: str,
     *,
-    simple_policy: bool = False,
     override_has_rocm: bool = False,
     override_rocm_gfx: str | None = None,
     force_cpu: bool = False,
@@ -6218,20 +6681,14 @@ def install_prebuilt(
                 log(
                     f"no existing llama.cpp install detected at {install_dir}; performing fresh prebuilt install"
                 )
-            if simple_policy:
-                requested_tag, release_plans = resolve_simple_install_release_plans(
-                    llama_tag,
-                    host,
-                    published_repo,
-                    published_release_tag,
-                )
-            else:
-                requested_tag, release_plans = resolve_install_release_plans(
-                    llama_tag,
-                    host,
-                    published_repo,
-                    published_release_tag,
-                )
+            # Single resolver: linux-x64 takes the fast filename path internally,
+            # every other fork host reads the manifest.
+            requested_tag, release_plans = resolve_simple_install_release_plans(
+                llama_tag,
+                host,
+                published_repo,
+                published_release_tag,
+            )
             if release_plans and existing_install_matches_plan(install_dir, host, release_plans[0]):
                 current = release_plans[0]
                 log(
@@ -6331,11 +6788,6 @@ def parse_args() -> argparse.Namespace:
             "until a usable published llama.cpp release bundle is found."
         ),
     )
-    parser.add_argument(
-        "--simple-policy",
-        action = "store_true",
-        help = "Use the simplified platform-specific prebuilt selection policy.",
-    )
     parser.add_argument(
         "--has-rocm",
         action = "store_true",
@@ -6389,6 +6841,16 @@ def parse_args() -> argparse.Namespace:
         const = "latest",
         help = ("Resolve the source-build fallback plan."),
     )
+    resolve_group.add_argument(
+        "--resolve-prebuilt",
+        nargs = "?",
+        const = "latest",
+        help = (
+            "Report whether an official prebuilt exists for this host without "
+            "downloading. Picks the host's published repo when --published-repo "
+            "is left at the default. Use --output-format json."
+        ),
+    )
     parser.add_argument(
         "--output-format",
         choices = ("plain", "json"),
@@ -6473,16 +6935,58 @@ def main() -> int:
         )
         return EXIT_SUCCESS
 
+    if args.resolve_prebuilt is not None:
+        # Host-aware "is a prebuilt available" probe, no download. A default repo
+        # means "pick the repo for this host"; PrebuiltFallback == source build.
+        host = _apply_host_overrides(
+            detect_host(),
+            override_has_rocm = args.has_rocm,
+            override_rocm_gfx = args.rocm_gfx,
+            force_cpu = args.cpu_fallback,
+        )
+        # setup.sh routes Linux hosts with AMD tooling to the fork even when no GPU
+        # is probed; mirror that so a HIP source build is not offered a CPU prebuilt.
+        amd_tooling = host.is_linux and any(
+            shutil.which(t) for t in ("rocminfo", "amd-smi", "hipconfig", "hipinfo")
+        )
+        repo = (
+            published_repo_for_host(host, linux_amd_tooling_present = amd_tooling)
+            if args.published_repo == DEFAULT_PUBLISHED_REPO
+            else args.published_repo
+        )
+        try:
+            _requested, plans = resolve_simple_install_release_plans(
+                args.resolve_prebuilt, host, repo, args.published_release_tag or ""
+            )
+            choice = plans[0].attempts[0] if plans and plans[0].attempts else None
+            if choice is None:
+                payload = {"prebuilt_available": False, "repo": repo}
+            else:
+                payload = {
+                    "prebuilt_available": True,
+                    "repo": repo,
+                    "release_tag": plans[0].release_tag,
+                    "llama_tag": plans[0].llama_tag,
+                    "asset": choice.name,
+                    "install_kind": choice.install_kind,
+                }
+        except PrebuiltFallback:
+            payload = {"prebuilt_available": False, "repo": repo}
+        emit_resolver_output(payload, output_format = args.output_format)
+        return EXIT_SUCCESS
+
     if not args.install_dir:
         raise SystemExit(
             "install_llama_prebuilt.py: --install-dir is required unless --resolve-llama-tag, --resolve-install-tag, or --resolve-source-build is used"
         )
+    # Install path only: route status logs to stdout (see _LOG_TO_STDOUT note).
+    global _LOG_TO_STDOUT
+    _LOG_TO_STDOUT = True
     install_prebuilt(
         install_dir = Path(args.install_dir).expanduser().resolve(),
         llama_tag = args.llama_tag,
         published_repo = args.published_repo,
         published_release_tag = args.published_release_tag or "",
-        simple_policy = args.simple_policy,
         override_has_rocm = args.has_rocm,
         override_rocm_gfx = args.rocm_gfx,
         force_cpu = args.cpu_fallback,
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 2c8f6228d4..e540aac305 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -18,6 +18,7 @@ import re
 import shutil
 import subprocess
 import sys
+import sysconfig
 import tempfile
 import urllib.request
 from pathlib import Path
@@ -40,6 +41,14 @@ IS_MACOS = sys.platform == "darwin"
 IS_MAC_INTEL = IS_MACOS and platform.machine() == "x86_64"
 IS_MAC_ARM = IS_MACOS and platform.machine() == "arm64"
 IS_LINUX = sys.platform.startswith("linux")
+
+# DiskPart-prompt suppression: amd-smi auto-elevates on Windows, popping a
+# UAC/DiskPart prompt mid-install. This installer only spawns probes and pip/uv
+# (none need elevation), so set __COMPAT_LAYER=RunAsInvoker process-wide -- every
+# amd-smi subprocess then runs un-elevated, no per-call guard needed. setup.ps1
+# keeps per-call guards since it ALSO spawns winget installers that need elevation.
+if IS_WINDOWS:
+    os.environ.setdefault("__COMPAT_LAYER", "RunAsInvoker")
 # torchcodec ships wheels only for manylinux_2_28_x86_64, macosx_12_0_arm64,
 # and win_amd64. On other hosts the audio extras must be filtered out (the
 # extras-no-deps step would otherwise fail), regardless of NO_TORCH.
@@ -81,6 +90,18 @@ _PYTORCH_WHL_BASE = (
     os.environ.get("UNSLOTH_PYTORCH_MIRROR") or "https://download.pytorch.org/whl"
 ).rstrip("/")
 
+# CUDA torch repair specs (see _ensure_cuda_torch). torchvision/torchaudio are
+# pinned to the torch<2.11 family rather than left bare: the install uses an
+# exclusive --index-url (no PyPI fallback), so a bare name could resolve a
+# torchvision built against a different torch major (e.g. 0.27 for torch 2.12)
+# and fail at runtime with an ABI mismatch. Same bounds as the _default ROCm
+# spec above, which targets the same torch family.
+_CUDA_TORCH_PKG_SPEC: tuple[str, str, str] = (
+    "torch>=2.4,<2.11.0",
+    "torchvision>=0.19,<0.26.0",
+    "torchaudio>=2.4,<2.11.0",
+)
+
 # AMD Windows ROCm wheels (repo.amd.com/rocm/whl/{arch_family}/).
 # Override with UNSLOTH_ROCM_WINDOWS_MIRROR for air-gapped/mirror installs.
 _ROCM_WINDOWS_INDEX_BASE = (
@@ -137,6 +158,39 @@ def _bnb_rocm_prerelease_url() -> str | None:
     return _BNB_ROCM_PRERELEASE_URLS.get(arch)
 
 
+def _amd_smi_env() -> dict[str, str] | None:
+    """On Windows, env with __COMPAT_LAYER=RunAsInvoker; None elsewhere.
+    NB: RunAsInvoker doesn't stop amd-smi's runtime elevation (its manifest is
+    asInvoker -- it elevates a child via ShellExecute). The real guard is
+    _amd_smi_allowed() below; this is harmless belt-and-suspenders."""
+    if platform.system() != "Windows":
+        return None
+    return {**os.environ, "__COMPAT_LAYER": "RunAsInvoker"}
+
+
+def _amd_smi_allowed() -> bool:
+    """Whether it is safe to spawn amd-smi here.
+
+    On Windows w/o a working HIP runtime, amd-smi elevates a child and pops a
+    UAC/DiskPart prompt RunAsInvoker can't suppress. Only call it on Windows with
+    a HIP SDK (hipinfo present) or UNSLOTH_ENABLE_AMD_SMI=1; Linux/macOS always.
+    """
+    if platform.system() != "Windows":
+        return True
+    flag = os.environ.get("UNSLOTH_ENABLE_AMD_SMI", "").strip().lower()
+    if flag in ("1", "true", "yes", "on"):
+        return True
+    if flag in ("0", "false", "no", "off"):
+        return False
+    if shutil.which("hipinfo"):
+        return True
+    for _var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
+        _root = os.environ.get(_var)
+        if _root and os.path.isfile(os.path.join(_root, "bin", "hipinfo.exe")):
+            return True
+    return False
+
+
 def _detect_rocm_version() -> tuple[int, int] | None:
     """Return (major, minor) of the installed ROCm stack, or None."""
     rocm_root = os.environ.get("ROCM_PATH") or "/opt/rocm"
@@ -155,8 +209,10 @@ def _detect_rocm_version() -> tuple[int, int] | None:
         except Exception:
             pass
 
-    # Try amd-smi version (outputs "... | ROCm version: X.Y.Z")
-    amd_smi = shutil.which("amd-smi")
+    # Try amd-smi version (outputs "... | ROCm version: X.Y.Z").
+    # Gated off on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt);
+    # hipconfig below covers that case.
+    amd_smi = shutil.which("amd-smi") if _amd_smi_allowed() else None
     if amd_smi:
         try:
             result = subprocess.run(
@@ -165,9 +221,9 @@ def _detect_rocm_version() -> tuple[int, int] | None:
                 stderr = subprocess.DEVNULL,
                 text = True,
                 timeout = 5,
+                env = _amd_smi_env(),
             )
             if result.returncode == 0:
-                import re
                 m = re.search(r"ROCm version:\s*(\d+)\.(\d+)", result.stdout)
                 if m:
                     return int(m.group(1)), int(m.group(2))
@@ -282,6 +338,14 @@ def _detect_windows_gfx_arch() -> str | None:
                 if os.path.isfile(_candidate):
                     hipinfo = _candidate
                     break
+    if not hipinfo:
+        # 2b. AMD torch wheels ship hipInfo.exe into the venv Scripts dir
+        # (next to python.exe); resolvable even on driver-only hosts with no
+        # SDK install at all. Lets `studio update` re-detect the arch on a
+        # venv that already has the AMD wheel.
+        _venv_hipinfo = os.path.join(os.path.dirname(sys.executable), "hipInfo.exe")
+        if os.path.isfile(_venv_hipinfo):
+            hipinfo = _venv_hipinfo
     if hipinfo:
         try:
             result = subprocess.run(
@@ -304,7 +368,9 @@ def _detect_windows_gfx_arch() -> str | None:
             pass
 
     # 3. amd-smi fallback -- runtime-only Radeon installs ship amd-smi but no hipinfo.
-    amd_smi = shutil.which("amd-smi")
+    # Gated off on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt); the arch
+    # arrives via --rocm-gfx / name inference there, so this is only needed when safe.
+    amd_smi = shutil.which("amd-smi") if _amd_smi_allowed() else None
     if amd_smi:
         for _args in (("static", "--asic"), ("list",)):
             try:
@@ -313,6 +379,7 @@ def _detect_windows_gfx_arch() -> str | None:
                     stdout = subprocess.PIPE,
                     stderr = subprocess.DEVNULL,
                     timeout = 10,
+                    env = _amd_smi_env(),
                 )
                 if result.returncode != 0:
                     continue
@@ -330,6 +397,74 @@ def _detect_windows_gfx_arch() -> str | None:
                     return _pick
             except Exception:
                 continue
+
+    # 4. Last resort: GPU marketing name via WMI → arch table. Driver-only
+    #    hosts (Adrenalin, no HIP SDK) have neither hipinfo nor amd-smi
+    #    (amd-smi does not exist on Windows at all), but the display driver
+    #    always knows the GPU name. Mirrors setup.ps1's $nameArchTable so a
+    #    standalone `studio update` can repair a CPU-only venv on such hosts.
+    try:
+        result = subprocess.run(
+            [
+                "powershell.exe",
+                "-NoProfile",
+                "-NonInteractive",
+                "-Command",
+                "(Get-CimInstance Win32_VideoController).Name",
+            ],
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            timeout = 30,
+            creationflags = getattr(subprocess, "CREATE_NO_WINDOW", 0),
+        )
+        if result.returncode == 0:
+            _tokens = []
+            for _name in result.stdout.decode(errors = "replace").splitlines():
+                _arch = _gfx_arch_from_gpu_name(_name.strip())
+                if _arch:
+                    _tokens.append(_arch)
+            _pick = _dedup_pick(_tokens)
+            if _pick:
+                print(f"   gfx arch inferred from GPU name (WMI): {_pick}")
+                return _pick
+    except Exception:
+        pass
+    return None
+
+
+# GPU marketing-name → gfx arch table, mirroring setup.ps1's $nameArchTable.
+# Most-specific first; first match wins. Covers only arches the lemonade-sdk
+# prebuilts / AMD Windows torch indexes support; unknown names return None
+# (callers then fall back cleanly to CPU).
+_WIN_GPU_NAME_ARCH_TABLE: "list[tuple[str, str]]" = [
+    (r"9070 XT|9080", "gfx1201"),  # RDNA 4 (Radeon RX 9070 XT / 9080)
+    (r"9070|9060", "gfx1200"),  # RDNA 4 (Radeon RX 9070 / 9060)
+    # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+    (r"8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max", "gfx1151"),
+    # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+    (
+        r"890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]"
+        r"|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33",
+        "gfx1150",
+    ),
+    # RDNA 3 desktop / workstation (Navi 31)
+    (r"RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700", "gfx1100"),
+    (r"RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710", "gfx1102"),  # Navi 33
+    # RDNA 3 iGPU (Phoenix / Hawk Point)
+    (r"780M|760M|740M|Phoenix|Hawk Point|Z1 Extreme|Z2 Extreme", "gfx1103"),
+    (r"RX 6900|RX 6800|RX 6750|RX 6700|PRO W6800|PRO W6900", "gfx1030"),  # Navi 21
+    (r"RX 6650|RX 6600|PRO W6600|PRO W6650", "gfx1032"),  # Navi 23
+    (r"RX 6500|RX 6400|RX 6300|PRO W6400|PRO W6500", "gfx1034"),  # Navi 24
+]
+
+
+def _gfx_arch_from_gpu_name(name: str) -> "str | None":
+    """Map a GPU marketing name to its gfx arch via _WIN_GPU_NAME_ARCH_TABLE."""
+    if not name:
+        return None
+    for _pat, _arch in _WIN_GPU_NAME_ARCH_TABLE:
+        if re.search(_pat, name, re.IGNORECASE):
+            return _arch
     return None
 
 
@@ -366,8 +501,83 @@ def _detect_bnb_rocm_dll_ver() -> str | None:
     return max(all_vers, key = lambda v: int(v)) if all_vers else None
 
 
+_BNB_ROCM_SITECUSTOMIZE_BEGIN = "# BEGIN Unsloth BNB_ROCM_VERSION"
+_BNB_ROCM_SITECUSTOMIZE_END = "# END Unsloth BNB_ROCM_VERSION"
+_BNB_ROCM_VERSION_SOURCE_ENV = "UNSLOTH_BNB_ROCM_VERSION_SOURCE"
+_BNB_ROCM_VERSION_SOURCE_SITECUSTOMIZE = "sitecustomize"
+_BNB_ROCM_VERSION_SOURCE_DETECTED = "detected"
+
+
+def _persist_bnb_rocm_version(version: str) -> bool:
+    """Persist BNB_ROCM_VERSION for future Python processes in this venv."""
+    version = str(version).strip()
+    if not version:
+        return False
+
+    site_packages = sysconfig.get_path("purelib")
+    if not site_packages:
+        return False
+
+    sitecustomize_path = Path(site_packages) / "sitecustomize.py"
+    block = (
+        f"{_BNB_ROCM_SITECUSTOMIZE_BEGIN}\n"
+        "import os as _unsloth_os\n"
+        "_unsloth_existing_bnb_rocm = _unsloth_os.environ.get('BNB_ROCM_VERSION')\n"
+        f"_unsloth_os.environ.setdefault('BNB_ROCM_VERSION', {version!r})\n"
+        "if _unsloth_existing_bnb_rocm is None and "
+        f"_unsloth_os.environ.get('BNB_ROCM_VERSION') == {version!r}:\n"
+        "    _unsloth_os.environ.setdefault("
+        f"{_BNB_ROCM_VERSION_SOURCE_ENV!r}, "
+        f"{_BNB_ROCM_VERSION_SOURCE_SITECUSTOMIZE!r})\n"
+        "del _unsloth_existing_bnb_rocm\n"
+        f"{_BNB_ROCM_SITECUSTOMIZE_END}\n"
+    )
+
+    try:
+        sitecustomize_path.parent.mkdir(parents = True, exist_ok = True)
+        existing = (
+            sitecustomize_path.read_text(encoding = "utf-8") if sitecustomize_path.exists() else ""
+        )
+        # Strip all managed regions, including one whose END marker was lost to
+        # an interrupted write, then append exactly one fresh block.
+        pattern = re.compile(
+            rf"{re.escape(_BNB_ROCM_SITECUSTOMIZE_BEGIN)}.*?"
+            rf"(?:{re.escape(_BNB_ROCM_SITECUSTOMIZE_END)}\n?|\Z)",
+            re.DOTALL,
+        )
+        remainder = pattern.sub("", existing)
+        separator = "" if not remainder or remainder.endswith("\n") else "\n"
+        updated = f"{remainder}{separator}{block}"
+        tmp_path = sitecustomize_path.with_name(
+            f"{sitecustomize_path.name}.unsloth-tmp{os.getpid()}"
+        )
+        try:
+            tmp_path.write_text(updated, encoding = "utf-8")
+            if sitecustomize_path.exists():
+                shutil.copymode(sitecustomize_path, tmp_path)
+            os.replace(tmp_path, sitecustomize_path)
+        finally:
+            tmp_path.unlink(missing_ok = True)
+    except (OSError, UnicodeDecodeError) as exc:
+        print(
+            f"   Warning: could not persist BNB_ROCM_VERSION={version} "
+            f"to {sitecustomize_path}: {exc}"
+        )
+        return False
+
+    return True
+
+
 def _has_rocm_gpu() -> bool:
-    """Return True only if an actual AMD GPU is visible (not just ROCm tools installed)."""
+    """Return True only if an actual AMD GPU is visible (not just ROCm tools installed).
+
+    Always returns False when an NVIDIA GPU is present -- NVIDIA takes
+    priority on mixed hosts and prevents every detection path below
+    (rocminfo, amd-smi, KFD sysfs) from producing a false positive even
+    if ROCm tools are installed alongside the NVIDIA driver.
+    """
+    if _has_usable_nvidia_gpu():
+        return False
     for cmd, check_fn in (
         # rocminfo: look for a real gfx GPU id (3-4 chars, nonzero first digit).
         # gfx000 is the CPU agent; ROCm 6.1+ also emits generic ISA lines like
@@ -386,6 +596,10 @@ def _has_rocm_gpu() -> bool:
         exe = shutil.which(cmd[0])
         if not exe:
             continue
+        # Skip amd-smi on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt);
+        # rely on rocminfo / the sysfs fallback there.
+        if cmd[0] == "amd-smi" and not _amd_smi_allowed():
+            continue
         try:
             result = subprocess.run(
                 [exe, *cmd[1:]],
@@ -393,6 +607,7 @@ def _has_rocm_gpu() -> bool:
                 stderr = subprocess.DEVNULL,
                 text = True,
                 timeout = 10,
+                env = _amd_smi_env() if cmd[0] == "amd-smi" else None,
             )
         except Exception:
             continue
@@ -403,6 +618,13 @@ def _has_rocm_gpu() -> bool:
     # runtime-only detection. On minimal package-managed installs (no
     # rocminfo / no amd-smi tools), the kernel exposes AMD GPUs via
     # /sys/class/kfd so `studio update` can still detect and repair.
+    #
+    # Guard: reject any KFD node whose properties file reports a non-AMD
+    # vendor. With the NVIDIA open kernel module (driver 560+), NVIDIA GPUs
+    # can register KFD topology nodes with a non-zero gpu_id; those nodes
+    # have vendor_id 4318 (0x10DE) rather than the AMD value 4098 (0x1002).
+    # Without this check the fallback returns True on NVIDIA-only systems,
+    # causing _ensure_rocm_torch to install ROCm wheels on NVIDIA hardware.
     if sys.platform != "win32":
         try:
             kfd_nodes = "/sys/class/kfd/kfd/topology/nodes"
@@ -414,29 +636,69 @@ def _has_rocm_gpu() -> bool:
                             gpu_id = fh.read().strip()
                     except OSError:
                         continue
-                    if gpu_id and gpu_id != "0":  # gpu_id 0 = CPU node
-                        return True
+                    if not gpu_id or gpu_id == "0":  # gpu_id 0 = CPU node
+                        continue
+                    # Require AMD vendor_id 4098 (0x1002) in the properties file.
+                    # KFD properties files exist on every kernel that exposes
+                    # /sys/class/kfd, so absence of the file means we cannot
+                    # confirm AMD ownership -- skip the node rather than risk a
+                    # false positive (e.g. NVIDIA open driver KFD nodes that
+                    # lack a properties file on some kernel versions).
+                    props_path = os.path.join(kfd_nodes, entry, "properties")
+                    try:
+                        with open(props_path) as fh:
+                            props = fh.read()
+                    except OSError:
+                        continue  # can't confirm vendor -- skip
+                    if not re.search(r"\bvendor_id\s+4098\b", props):
+                        continue
+                    return True
         except OSError:
             pass
     return False
 
 
 def _has_usable_nvidia_gpu() -> bool:
-    """Return True only when nvidia-smi exists AND reports at least one GPU."""
+    """Return True when an NVIDIA GPU is present and usable.
+
+    Primary probe: nvidia-smi -L (subprocess).
+    Fallback: /proc/driver/nvidia/gpus/ sysfs (Linux only) -- handles the
+    case where nvidia-smi is present but the subprocess fails (PATH gap,
+    timeout, driver initialisation race). If either probe confirms an
+    NVIDIA GPU the function returns True so _has_rocm_gpu() is blocked.
+
+    CUDA_VISIBLE_DEVICES set to "" or "-1" hides every NVIDIA device (mixed
+    AMD+NVIDIA hosts steering work to the AMD card); neither probe honours
+    that env var, so check it first and report the GPU as not usable. Unset
+    means all devices visible.
+    """
+    cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if cvd is not None and cvd.strip() in ("", "-1"):
+        return False
     exe = shutil.which("nvidia-smi")
-    if not exe:
-        return False
-    try:
-        result = subprocess.run(
-            [exe, "-L"],
-            stdout = subprocess.PIPE,
-            stderr = subprocess.DEVNULL,
-            text = True,
-            timeout = 10,
-        )
-    except Exception:
-        return False
-    return result.returncode == 0 and "GPU " in result.stdout
+    if exe:
+        try:
+            result = subprocess.run(
+                [exe, "-L"],
+                stdout = subprocess.PIPE,
+                stderr = subprocess.DEVNULL,
+                text = True,
+                timeout = 10,
+            )
+            if result.returncode == 0 and "GPU " in result.stdout:
+                return True
+        except Exception:
+            pass
+    # Fallback: the NVIDIA driver exposes one subdirectory per GPU under
+    # /proc/driver/nvidia/gpus/ on Linux regardless of nvidia-smi state.
+    if sys.platform != "win32":
+        try:
+            gpu_dir = "/proc/driver/nvidia/gpus"
+            if os.path.isdir(gpu_dir) and os.listdir(gpu_dir):
+                return True
+        except OSError:
+            pass
+    return False
 
 
 def _detect_amd_gfx_codes() -> list[str]:
@@ -454,7 +716,8 @@ def _detect_amd_gfx_codes() -> list[str]:
     probes: list[list[str]] = []
     if shutil.which("rocminfo"):
         probes.append(["rocminfo"])
-    if shutil.which("amd-smi"):
+    # Gate amd-smi off on Windows w/o a HIP SDK (avoids the UAC/DiskPart prompt).
+    if shutil.which("amd-smi") and _amd_smi_allowed():
         probes.append(["amd-smi", "list"])
         probes.append(["amd-smi", "static", "--asic"])
     for cmd in probes:
@@ -465,6 +728,7 @@ def _detect_amd_gfx_codes() -> list[str]:
                 stderr = subprocess.DEVNULL,
                 text = True,
                 timeout = 15,
+                env = _amd_smi_env() if cmd[0] == "amd-smi" else None,
             )
         except Exception:
             continue
@@ -506,18 +770,175 @@ def _install_bnb_windows_rocm() -> bool:
     )
     if not _ok:
         return False
-    # After install: detect the actual ROCm DLL suffix in the wheel and set
-    # BNB_ROCM_VERSION so bitsandbytes loads the correct DLL regardless of what
-    # torch.version.hip reports. The wheel may ship an older suffix (e.g. "72")
-    # while torch reports a newer HIP version (e.g. 7.13); the override stops
-    # bitsandbytes from failing on a non-existent DLL. The worker subprocess
-    # inherits this env var. Fall back to "72" if detection fails (no-op/dry-run).
-    if "BNB_ROCM_VERSION" not in os.environ:
+    # After install: detect the actual ROCm DLL suffix shipped in the wheel and
+    # set BNB_ROCM_VERSION so bitsandbytes loads the correct DLL regardless of
+    # what torch.version.hip reports.  The wheel may ship an older suffix (e.g.
+    # "72") while torch reports a newer HIP version (e.g. 7.13); the env var
+    # override ensures bitsandbytes does not fail looking for a non-existent DLL.
+    # The worker subprocess inherits this env var automatically.
+    # Fall back to "72" if detection fails (e.g. install was a no-op / dry-run).
+    _env_ver = os.environ.get("BNB_ROCM_VERSION")
+    _env_is_persisted_default = (
+        os.environ.get(_BNB_ROCM_VERSION_SOURCE_ENV) == _BNB_ROCM_VERSION_SOURCE_SITECUSTOMIZE
+    )
+    _persist_detected_version = False
+    if _env_ver and not _env_is_persisted_default:
+        _ver = _env_ver
+    else:
         _ver = _detect_bnb_rocm_dll_ver() or "72"
         os.environ["BNB_ROCM_VERSION"] = _ver
+        os.environ[_BNB_ROCM_VERSION_SOURCE_ENV] = _BNB_ROCM_VERSION_SOURCE_DETECTED
+        _persist_detected_version = True
+    if _persist_detected_version:
+        _persist_bnb_rocm_version(_ver)
+    # Make hipInfo.exe (shipped into the venv Scripts dir by the AMD torch
+    # wheel) resolvable via PATH for this process and every child python the
+    # installer spawns (import checks, precompile): bitsandbytes runs
+    # `hipinfo.exe` at import time to detect the GPU arch and logs a scary
+    # (harmless) ERROR + WARNING on every import when it is missing. The venv
+    # Scripts dir is on PATH only when the venv is activated, which neither
+    # Studio nor the installer's child processes ever do.
+    _scripts_dir = os.path.dirname(sys.executable)
+    if os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe")) and not shutil.which(
+        "hipinfo.exe"
+    ):
+        os.environ["PATH"] = _scripts_dir + os.pathsep + os.environ.get("PATH", "")
     return True
 
 
+def _detect_cuda_torch_index_url() -> str:
+    """Return the pytorch.org CUDA wheel index URL for the host's NVIDIA driver.
+
+    Mirrors install.sh::get_torch_index_url's CUDA ladder so `studio update`
+    repairs to the same wheel family a fresh `curl | sh` install would pick.
+    Probes nvidia-smi (PATH, then /usr/bin/nvidia-smi) and parses both the
+    legacy "CUDA Version:" and the newer "CUDA UMD Version:" spellings.
+    Defaults to cu126 when nvidia-smi is missing or the version is unreadable
+    (e.g. NVIDIA detected only via the /proc/driver/nvidia/gpus fallback).
+    """
+    exe = shutil.which("nvidia-smi")
+    if not exe and os.path.isfile("/usr/bin/nvidia-smi"):
+        exe = "/usr/bin/nvidia-smi"
+    tag = "cu126"  # default when the driver CUDA version cannot be read
+    if exe:
+        try:
+            result = subprocess.run(
+                [exe],
+                stdout = subprocess.PIPE,
+                stderr = subprocess.DEVNULL,
+                text = True,
+                timeout = 10,
+            )
+            if result.returncode == 0:
+                m = re.search(r"CUDA(?: UMD)? Version:\s*(\d+)\.(\d+)", result.stdout)
+                if m:
+                    major, minor = int(m.group(1)), int(m.group(2))
+                    if major >= 13:
+                        tag = "cu130"
+                    elif major == 12 and minor >= 8:
+                        tag = "cu128"
+                    elif major == 12 and minor >= 6:
+                        tag = "cu126"
+                    elif major >= 12:
+                        tag = "cu124"
+                    elif major >= 11:
+                        tag = "cu118"
+                    else:
+                        tag = "cpu"  # ancient driver: no usable CUDA wheels
+        except Exception:
+            pass
+    return f"{_PYTORCH_WHL_BASE}/{tag}"
+
+
+def _ensure_cuda_torch() -> None:
+    """Repair a venv whose torch is a ROCm build on an NVIDIA host.
+
+    Counterpart to _ensure_rocm_torch. A venv poisoned by the pre-fix KFD
+    gpu_id false positive (ROCm torch installed on an NVIDIA-only machine)
+    keeps that broken torch on `studio update`, because a torch+rocm wheel
+    satisfies the version constraint and nothing force-reinstalls it. This
+    detects that exact case and reinstalls CUDA torch.
+
+    Only repairs when torch actually links against HIP/ROCm. Healthy CUDA
+    torch and deliberate CPU-only torch are left untouched.
+    """
+    # Respect an explicit backend choice from install.sh: only "" (standalone
+    # `studio update`) or "cuda" should ever force CUDA wheels. "rocm"/"cpu"
+    # (or any unrecognised value) are deliberate and must not be overridden.
+    if _TORCH_BACKEND not in ("", "cuda"):
+        return
+    # No CUDA torch on macOS; Windows venv/torch lifecycle is owned by
+    # install.ps1 (and the KFD poisoning bug is Linux-only), so skip both.
+    if IS_MACOS or IS_WINDOWS or NO_TORCH:
+        return
+    # Never undo a deliberate ROCm install (setup.ps1 sets this marker).
+    if os.environ.get("UNSLOTH_ROCM_TORCH_INSTALLED") == "1":
+        return
+    # CUDA_VISIBLE_DEVICES="" / "-1" deliberately hides the NVIDIA GPU (for
+    # example a mixed AMD+NVIDIA host that runs ROCm torch on the AMD card);
+    # never force CUDA wheels over that choice.
+    _cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if _cvd is not None and _cvd.strip() in ("", "-1"):
+        return
+    # Only NVIDIA hosts should carry CUDA torch. _has_usable_nvidia_gpu()
+    # covers the /proc/driver/nvidia/gpus fallback when nvidia-smi is absent.
+    if not _has_usable_nvidia_gpu():
+        return
+
+    # Classify the installed torch: "hip" (ROCm build -- the poisoning
+    # signature), "cuda" (healthy), or "cpu" (deliberate CPU wheel). A
+    # non-zero exit means torch is missing or un-importable; the base install
+    # step handles that, so leave it alone.
+    try:
+        probe = subprocess.run(
+            [
+                sys.executable,
+                "-c",
+                (
+                    "import torch; "
+                    "hip = getattr(torch.version, 'hip', '') or ''; "
+                    "cuda = getattr(torch.version, 'cuda', '') or ''; "
+                    "ver = getattr(torch, '__version__', '').lower(); "
+                    "print('hip' if (hip or 'rocm' in ver) else ('cuda' if cuda else 'cpu'))"
+                ),
+            ],
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            timeout = 90,
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        return
+    if probe.returncode != 0:
+        return
+    # Take the last non-empty stdout line: stray output from sitecustomize or
+    # an import hook must not mask the marker (fail-closed either way).
+    _marker_lines = [
+        line.strip() for line in probe.stdout.decode(errors = "replace").splitlines() if line.strip()
+    ]
+    if not _marker_lines or _marker_lines[-1] != "hip":
+        return  # healthy CUDA torch, or a deliberate CPU wheel -- leave as-is
+
+    index_url = _detect_cuda_torch_index_url()
+    _torch_pkg, _vision_pkg, _audio_pkg = _CUDA_TORCH_PKG_SPEC
+    print(
+        f"   torch is a ROCm build on an NVIDIA host -- reinstalling "
+        f"CUDA torch from {index_url}\n"
+        f"   (set UNSLOTH_TORCH_BACKEND=rocm to keep a deliberate ROCm torch "
+        f"on a mixed AMD+NVIDIA host)"
+    )
+    pip_install(
+        "CUDA torch repair",
+        "--force-reinstall",
+        "--no-cache-dir",
+        _torch_pkg,
+        _vision_pkg,
+        _audio_pkg,
+        "--index-url",
+        index_url,
+        constrain = False,
+    )
+
+
 def _ensure_rocm_torch() -> None:
     """Reinstall torch with ROCm wheels when the venv received CPU-only torch.
 
@@ -528,6 +949,13 @@ def _ensure_rocm_torch() -> None:
     Uses pip_install() to respect uv, constraints, and --python targeting.
     """
     global _rocm_windows_torch_installed
+    # install.sh sets UNSLOTH_TORCH_BACKEND to the resolved wheel family
+    # ("cuda", "rocm", "cpu"). Skip ROCm operations entirely when install.sh
+    # already selected a non-ROCm backend -- this is the authoritative signal
+    # and avoids re-running GPU detection in a subprocess that may see a
+    # different environment (different PATH, CUDA_VISIBLE_DEVICES, etc.).
+    if _TORCH_BACKEND in ("cuda", "cpu"):
+        return
     # setup.ps1 sets this after installing AMD wheels; skip the probe only when
     # torch is actually importable as ROCm. If the venv was wiped between runs,
     # the stale env-var would suppress a needed reinstall.
@@ -867,6 +1295,29 @@ def _infer_no_torch() -> bool:
 
 NO_TORCH = _infer_no_torch()
 
+# UNSLOTH_TORCH_BACKEND is set by install.sh after get_torch_index_url() so
+# that this script knows which torch variant was selected without re-running
+# GPU detection. Values: "cuda", "rocm", or "cpu". Empty means unknown
+# (standalone `unsloth studio update` runs, where we re-detect normally).
+_TORCH_BACKEND: str = os.environ.get("UNSLOTH_TORCH_BACKEND", "").lower()
+
+
+def _torch_step_label(suffix: str) -> str:
+    """Return a progress label like 'torch check (cuda)' using the known backend.
+
+    Falls back to GPU detection when UNSLOTH_TORCH_BACKEND is not set (e.g.
+    standalone `unsloth studio update` runs that bypass install.sh).
+    """
+    backend = _TORCH_BACKEND
+    if not backend:
+        if _has_usable_nvidia_gpu():
+            backend = "cuda"
+        elif _has_rocm_gpu():
+            backend = "rocm"
+        else:
+            backend = "cpu"
+    return f"torch {suffix} ({backend})"
+
 
 # -- Verbosity control ----------------------------------------------------------
 # By default the installer shows a minimal in-place one-line progress bar.
@@ -1549,7 +2000,8 @@ def install_python_stack() -> int:
     #     venv got CPU-only torch (common when pip resolves torch from PyPI).
     #     Must follow base packages so torch is present for inspection.
     if not IS_MACOS and not NO_TORCH:
-        _progress("ROCm torch check")
+        _progress(_torch_step_label("check"))
+        _ensure_cuda_torch()
         _ensure_rocm_torch()
 
     # Windows + AMD GPU: warn if ROCm torch was not installed (wrong Python
@@ -1569,6 +2021,12 @@ def install_python_stack() -> int:
             _wexe = shutil.which(_wcmd[0])
             if not _wexe:
                 continue
+            # Skip amd-smi on Windows w/o a HIP SDK (avoids the UAC/DiskPart
+            # prompt), as _has_rocm_gpu()/_detect_amd_gfx_codes do. The only loss
+            # is the best-effort "AMD GPU detected" note; ROCm-torch state below
+            # comes from the install itself.
+            if _wcmd[0] == "amd-smi" and not _amd_smi_allowed():
+                continue
             try:
                 _wr = subprocess.run(
                     [_wexe, *_wcmd[1:]],
@@ -1576,6 +2034,7 @@ def install_python_stack() -> int:
                     stderr = subprocess.DEVNULL,
                     text = True,
                     timeout = 10,
+                    env = _amd_smi_env() if _wcmd[0] == "amd-smi" else None,
                 )
             except Exception:
                 continue
@@ -1727,7 +2186,8 @@ def install_python_stack() -> int:
     #     Running the repair last ensures ROCm torch is in place at runtime,
     #     whichever intermediate step clobbered it.
     if not IS_WINDOWS and not IS_MACOS and not NO_TORCH:
-        _progress("ROCm torch (final)")
+        _progress(_torch_step_label("final"))
+        _ensure_cuda_torch()
         _ensure_rocm_torch()
 
     # 14. Final check (silent; third-party conflicts are expected)
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 4d47010136..94817b561a 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -220,14 +220,10 @@ function Get-InstalledLlamaPrebuiltRelease {
 function Find-Nvcc {
     param([string]$MaxVersion = "")
 
-    # If MaxVersion is set, we need to find a toolkit <= that version.
-    # CUDA toolkits install side-by-side under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.Y\
-
     $toolkitBase = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA'
 
     if ($MaxVersion -and (Test-Path $toolkitBase)) {
         $drMajor = [int]$MaxVersion.Split('.')[0]
-        $drMinor = [int]$MaxVersion.Split('.')[1]
 
         # Get all installed CUDA dirs, sorted descending (highest first)
         $cudaDirs = Get-ChildItem -Directory $toolkitBase | Where-Object {
@@ -236,8 +232,8 @@ function Find-Nvcc {
 
         foreach ($dir in $cudaDirs) {
             if ($dir.Name -match '^v(\d+)\.(\d+)') {
-                $tkMajor = [int]$Matches[1]; $tkMinor = [int]$Matches[2]
-                $compatible = ($tkMajor -lt $drMajor) -or ($tkMajor -eq $drMajor -and $tkMinor -le $drMinor)
+                $tkMajor = [int]$Matches[1]
+                $compatible = ($tkMajor -le $drMajor)
                 if ($compatible) {
                     $nvcc = Join-Path $dir.FullName 'bin\nvcc.exe'
                     if (Test-Path $nvcc) {
@@ -278,6 +274,19 @@ function Find-Nvcc {
     return $null
 }
 
+function Write-CudaDriverToolkitMismatch {
+    param(
+        [Parameter(Mandatory = $true)][string]$ToolkitVersion,
+        [Parameter(Mandatory = $true)][string]$DriverMaxCuda,
+        [string]$Color = "Yellow"
+    )
+    $toolkitMajor = $ToolkitVersion.Split('.')[0]
+    $driverMajor = $DriverMaxCuda.Split('.')[0]
+    substep "CUDA Toolkit $ToolkitVersion is a major-version mismatch: toolkit major $toolkitMajor exceeds driver CUDA major $driverMajor ($DriverMaxCuda)." $Color
+    substep "Update the NVIDIA GPU driver to run CUDA Toolkit $ToolkitVersion, or install a CUDA $driverMajor.x toolkit." $Color
+    substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." $Color
+}
+
 # Detect CUDA Compute Capability via nvidia-smi.
 # Returns e.g. "80" for A100 (8.0), "89" for RTX 4090 (8.9), etc.
 # Returns $null if detection fails.
@@ -290,7 +299,10 @@ function Get-CudaComputeCapability {
     if (-not $smiExe) { return $null }
 
     try {
-        $raw = & $smiExe --query-gpu=compute_cap --format=csv,noheader 2>$null
+        # Bounded: a wedged nvidia-smi must not hang setup after the initial
+        # -L probe succeeded (the helper merges stderr after stdout, so the
+        # first line is still the compute_cap value).
+        $raw = Invoke-NvidiaSmiBounded $smiExe @('--query-gpu=compute_cap', '--format=csv,noheader')
         if ($LASTEXITCODE -ne 0 -or -not $raw) { return $null }
 
         # nvidia-smi may return multiple GPUs; take the first one
@@ -354,10 +366,10 @@ function Get-PytorchCudaTag {
     if (-not $smiExe) { return "cu126" }
 
     try {
-        # 2>&1 | Out-String merges stderr into stdout then converts to a single
-        # string.  Plain 2>$null doesn't fully suppress stderr in PS 5.1 --
-        # ErrorRecord objects leak into $output and break the -match.
-        $output = & $smiExe 2>&1 | Out-String
+        # Bounded: a wedged nvidia-smi must not hang setup. The helper merges
+        # stderr into the returned string, matching the old 2>&1 | Out-String
+        # shape (plain 2>$null leaks ErrorRecord objects in PS 5.1).
+        $output = Invoke-NvidiaSmiBounded $smiExe
         # Newer NVIDIA drivers (e.g. 610.x on Windows) print
         # "CUDA UMD Version: X.Y" instead of the legacy "CUDA Version: X.Y".
         # Accept both spellings so we don't fall through to the cu126 default.
@@ -658,16 +670,58 @@ try {
 # ============================================
 # 1a. GPU detection
 # ============================================
+# ── Helper: run nvidia-smi under a timeout ──
+# A wedged NVIDIA driver can make nvidia-smi block during init or after a reset;
+# WaitForExit bounds it (mirrors Invoke-AmdSmiNoElevate below) so detection
+# cannot hang setup. No RunAsInvoker compat layer: nvidia-smi does not
+# auto-elevate. Returns combined stdout+stderr; "" on timeout/failure.
+function Invoke-NvidiaSmiBounded {
+    param(
+        [Parameter(Mandatory = $true, Position = 0)][string]$Exe,
+        [Parameter(Position = 1)][string[]]$SmiArgs = @(),
+        [int]$TimeoutSec = 10
+    )
+    try {
+        $psi = New-Object System.Diagnostics.ProcessStartInfo
+        $psi.FileName = $Exe
+        $psi.Arguments = ($SmiArgs -join ' ')
+        $psi.UseShellExecute = $false
+        $psi.RedirectStandardOutput = $true
+        $psi.RedirectStandardError = $true
+        $psi.CreateNoWindow = $true
+        $proc = [System.Diagnostics.Process]::Start($psi)
+        $outTask = $proc.StandardOutput.ReadToEndAsync()
+        $errTask = $proc.StandardError.ReadToEndAsync()
+        if (-not $proc.WaitForExit($TimeoutSec * 1000)) {
+            try { $proc.Kill() } catch {}
+            $global:LASTEXITCODE = 124
+            return ""
+        }
+        $global:LASTEXITCODE = $proc.ExitCode
+        return ($outTask.Result + "`n" + $errTask.Result)
+    } catch {
+        $global:LASTEXITCODE = 1
+        return ""
+    }
+}
+
+# ── Helper: nvidia-smi -L lists at least one real GPU ──
+# Exit code 0 alone is not enough: a stale/driverless nvidia-smi can exit 0
+# while listing no GPU, which would mark an AMD host NVIDIA and suppress ROCm
+# detection. Require a "GPU <n>:" data row.
+function Test-NvidiaSmiHasGpu {
+    param([Parameter(Mandatory = $true)][string]$Exe)
+    $out = Invoke-NvidiaSmiBounded $Exe @('-L')
+    return ($LASTEXITCODE -eq 0 -and $out -match '(?m)^GPU\s+\d+:')
+}
+
 $HasNvidiaSmi = $false
 $NvidiaSmiExe = $null  # Absolute path -- survives Refresh-Environment
 try {
     $nvSmiCmd = Get-Command nvidia-smi -ErrorAction SilentlyContinue
-    if ($nvSmiCmd) {
-        & $nvSmiCmd.Source *> $null
-        if ($LASTEXITCODE -eq 0) {
-            $HasNvidiaSmi = $true
-            $NvidiaSmiExe = $nvSmiCmd.Source
-        }
+    if ($nvSmiCmd -and (Test-NvidiaSmiHasGpu $nvSmiCmd.Source)) {
+        $HasNvidiaSmi = $true
+        $NvidiaSmiExe = $nvSmiCmd.Source
     }
 } catch {}
 # Fallback: nvidia-smi may not be on PATH even though a GPU + driver exist.
@@ -680,8 +734,7 @@ if (-not $HasNvidiaSmi) {
     foreach ($p in $nvSmiDefaults) {
         if (Test-Path $p) {
             try {
-                & $p *> $null
-                if ($LASTEXITCODE -eq 0) {
+                if (Test-NvidiaSmiHasGpu $p) {
                     $HasNvidiaSmi = $true
                     $NvidiaSmiExe = $p
                     Write-Host "   Found nvidia-smi at $(Split-Path $p -Parent)" -ForegroundColor Gray
@@ -691,6 +744,55 @@ if (-not $HasNvidiaSmi) {
         }
     }
 }
+# ── Helper: run amd-smi without triggering a UAC elevation prompt ──
+# amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing
+# DiskPart UAC prompt mid-install (Studio backend amd.py hits the same). RunAsInvoker
+# forces it (and helpers it spawns) to run un-elevated; on failure the WMI name ->
+# gfx fallback still resolves the arch.
+function Invoke-AmdSmiNoElevate {
+    param(
+        [Parameter(Mandatory = $true, Position = 0)][string]$Exe,
+        [Parameter(Position = 1)][string[]]$SmiArgs = @(),
+        [int]$TimeoutSec = 30
+    )
+    # RunAsInvoker blocks the auto-elevation/UAC prompt; the timeout bounds a flaky
+    # amd-smi that can otherwise spin for minutes (30s mirrors the backend amd.py).
+    $prevCompat = [Environment]::GetEnvironmentVariable('__COMPAT_LAYER', 'Process')
+    $env:__COMPAT_LAYER = 'RunAsInvoker'
+    try {
+        # [Process]::Start, NOT Start-Process -PassThru: the latter leaves .ExitCode
+        # $null after WaitForExit on PS 5.1, so $LASTEXITCODE (checked by callers)
+        # reads non-zero and kills detection. Async reads drain the pipes (no
+        # deadlock); amd-smi args have no spaces so a plain join is safe.
+        $psi = New-Object System.Diagnostics.ProcessStartInfo
+        $psi.FileName = $Exe
+        $psi.Arguments = ($SmiArgs -join ' ')
+        $psi.UseShellExecute = $false
+        $psi.RedirectStandardOutput = $true
+        $psi.RedirectStandardError = $true
+        $psi.CreateNoWindow = $true
+        $proc = [System.Diagnostics.Process]::Start($psi)
+        $outTask = $proc.StandardOutput.ReadToEndAsync()
+        $errTask = $proc.StandardError.ReadToEndAsync()
+        if (-not $proc.WaitForExit($TimeoutSec * 1000)) {
+            try { $proc.Kill() } catch {}
+            $global:LASTEXITCODE = 124
+            return ""
+        }
+        $global:LASTEXITCODE = $proc.ExitCode
+        return ($outTask.Result + "`n" + $errTask.Result)
+    } catch {
+        $global:LASTEXITCODE = 1
+        return ""
+    } finally {
+        if ($null -eq $prevCompat) {
+            Remove-Item Env:__COMPAT_LAYER -ErrorAction SilentlyContinue
+        } else {
+            $env:__COMPAT_LAYER = $prevCompat
+        }
+    }
+}
+
 # ── AMD ROCm detection (Windows): probe hipinfo/amd-smi for actual GPU ──
 $HasROCm = $false
 $HipSdkInstalled = $false   # HIP SDK binary found (independent of device accessibility)
@@ -741,14 +843,24 @@ if (-not $HasNvidiaSmi) {
         } catch {}
     }
     # amd-smi fallback: HIP runtime present but hipinfo unavailable (no full HIP SDK).
-    # Confirms GPU visibility via 'list', then attempts 'static --asic' to extract
-    # the gfx arch that hipinfo would have provided.  Critical for Strix Halo
-    # (gfx1151) and other iGPUs where only the HIP runtime is installed.
-    if (-not $HasROCm) {
+    # 'list' confirms GPU visibility, 'static --asic' extracts the gfx arch hipinfo
+    # would give. Critical for Strix Halo (gfx1151) and other HIP-runtime-only iGPUs.
+    #
+    # BUT on hosts without a working HIP runtime amd-smi elevates a child at runtime,
+    # popping a UAC/DiskPart prompt RunAsInvoker can't suppress (its manifest is
+    # asInvoker; even 'amd-smi version' hangs). So only probe when a HIP SDK is present
+    # (hipinfo found -> un-elevated) or the user opts in; else fall through to WMI name
+    # inference (enough to pick ROCm wheels + lemonade llama.cpp).
+    # An explicit opt-out (UNSLOTH_ENABLE_AMD_SMI=0/false/no/off) wins over the HIP-SDK
+    # heuristic: a HIP SDK binary with a broken runtime can still pop the prompt, so
+    # $HipSdkInstalled must NOT silently re-enable it.
+    $amdSmiOptOut = $env:UNSLOTH_ENABLE_AMD_SMI -match '^(?i)(0|false|no|off)$'
+    $amdSmiAllowed = (-not $amdSmiOptOut) -and ($HipSdkInstalled -or ($env:UNSLOTH_ENABLE_AMD_SMI -match '^(?i)(1|true|yes|on)$'))
+    if (-not $HasROCm -and $amdSmiAllowed) {
         $amdSmiExe = Get-Command "amd-smi" -ErrorAction SilentlyContinue
         if ($amdSmiExe) {
             try {
-                $smiOut = & $amdSmiExe.Source list 2>&1 | Out-String
+                $smiOut = Invoke-AmdSmiNoElevate $amdSmiExe.Source @('list')
                 if ($LASTEXITCODE -eq 0 -and $smiOut -match "(?im)^GPU\s*[:\[]\s*\d") {
                     $HasROCm = $true
                     # Attempt 1: newer amd-smi versions embed the gfx arch in list output.
@@ -782,7 +894,7 @@ if (-not $HasNvidiaSmi) {
                         # Attempt 2: 'static --asic' exposes ASIC details on ROCm 6+,
                         # including the GFX target needed for wheel index selection.
                         $smiAsicOut = ""
-                        try { $smiAsicOut = & $amdSmiExe.Source static --asic 2>&1 | Out-String } catch {}
+                        try { $smiAsicOut = Invoke-AmdSmiNoElevate $amdSmiExe.Source @('static','--asic') } catch {}
                         if ($smiAsicOut -match "(?i)\b(gfx\d+[a-z]?)\b") {
                             $script:ROCmGfxArch = $Matches[1].ToLower()
                             $ROCmGpuLabel = "AMD ROCm ($script:ROCmGfxArch)"
@@ -810,27 +922,34 @@ if (-not $HasNvidiaSmi) {
         } catch {}
     }
     # ── Arch resolution: env-var override → name inference ──────────────────
-    # Runs after all probe methods.  Covers users whose amd-smi version is too
-    # old to report the GFX target and who don't have hipinfo (HIP-runtime-only
-    # installs, common on Strix Halo / iGPU systems).
-    if ($HasROCm -and -not $script:ROCmGfxArch) {
+    # Runs after all probes, even when none confirmed a ROCm runtime ($HasROCm false):
+    # the Adrenalin driver alone runs the lemonade-sdk llama.cpp prebuilt (bundles its
+    # own runtime), and all it needs is the gfx arch, inferable from the WMI GPU name.
+    # Resolving it here lets setup.ps1 forward --rocm-gfx so a GPU llama.cpp is pulled
+    # instead of CPU. (PyTorch ROCm wheels still require a HIP SDK -- gated on $HasROCm
+    # below -- so this only affects llama.cpp / inference.)
+    if (-not $script:ROCmGfxArch) {
         # 1. Manual override: set UNSLOTH_ROCM_GFX_ARCH=gfx1151 before running.
         if ($env:UNSLOTH_ROCM_GFX_ARCH) {
             $script:ROCmGfxArch = $env:UNSLOTH_ROCM_GFX_ARCH.Trim().ToLower()
             $ROCmGpuLabel = "AMD ROCm ($script:ROCmGfxArch)"
             substep "gfx arch from UNSLOTH_ROCM_GFX_ARCH env override: $script:ROCmGfxArch" "Cyan"
         }
-        # 2. Best-effort name → arch lookup from marketing name (amd-smi / WMI).
-        #    Ordered most-specific first; first match wins.
+        # 2. Best-effort name → arch lookup (amd-smi / WMI). Most-specific first,
+        #    first match wins. Covers only arches the lemonade-sdk prebuilts support
+        #    (gfx120X/110X/1151/1150/103X); unknown names fall back cleanly to CPU.
         elseif ($ROCmGpuLabel) {
             $nameArchTable = @(
-                @{ P = "9070 XT|9080";                                         A = "gfx1201" }  # RDNA 4
-                @{ P = "9070|9060";                                            A = "gfx1200" }  # RDNA 4
-                @{ P = "8060S|890M|Strix Halo|HX 37[05]|HX 38[05]|AI 9 HX";    A = "gfx1151" }  # RDNA 3.5 iGPU (Strix Halo / Radeon 8060S retail)
-                @{ P = "880M|Strix Point|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]"; A = "gfx1150" }  # RDNA 3.5 iGPU (Strix Point)
-                @{ P = "RX 7900|RX 7800|RX 7700(?! S)";                        A = "gfx1100" }  # RDNA 3 desktop
-                @{ P = "RX 7600";                                              A = "gfx1102" }  # RDNA 3
-                @{ P = "780M|760M|740M|Phoenix";                               A = "gfx1103" }  # RDNA 3 iGPU (Phoenix)
+                @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4 (Radeon RX 9070 XT / 9080)
+                @{ P = "9070|9060";                                           A = "gfx1200" }  # RDNA 4 (Radeon RX 9070 / 9060)
+                @{ P = "8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+                @{ P = "890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1150" }  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+                @{ P = "RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700"; A = "gfx1100" }  # RDNA 3 desktop / workstation (Navi 31)
+                @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710"; A = "gfx1102" }  # RDNA 3 (Navi 33)
+                @{ P = "780M|760M|740M|Phoenix|Hawk Point|Z1 Extreme|Z2 Extreme"; A = "gfx1103" }  # RDNA 3 iGPU (Phoenix / Hawk Point)
+                @{ P = "RX 6900|RX 6800|RX 6750|RX 6700|PRO W6800|PRO W6900";  A = "gfx1030" }  # RDNA 2 (Navi 21) -- lemonade gfx103X
+                @{ P = "RX 6650|RX 6600|PRO W6600|PRO W6650";                  A = "gfx1032" }  # RDNA 2 (Navi 23) -- lemonade gfx103X
+                @{ P = "RX 6500|RX 6400|RX 6300|PRO W6400|PRO W6500";          A = "gfx1034" }  # RDNA 2 (Navi 24) -- lemonade gfx103X
             )
             foreach ($row in $nameArchTable) {
                 if ($ROCmGpuLabel -match $row.P) {
@@ -872,11 +991,11 @@ if (-not $HasNvidiaSmi) {
                 }
             } catch {}
         }
-        if (-not $script:ROCmVersion) {
+        if (-not $script:ROCmVersion -and $amdSmiAllowed) {
             $amdSmiVer = Get-Command "amd-smi" -ErrorAction SilentlyContinue
             if ($amdSmiVer) {
                 try {
-                    $smiVerOut = & $amdSmiVer.Source version 2>&1 | Out-String
+                    $smiVerOut = Invoke-AmdSmiNoElevate $amdSmiVer.Source @('version')
                     if ($LASTEXITCODE -eq 0 -and $smiVerOut -match 'ROCm version:\s*(\d+\.\d+)') { $script:ROCmVersion = $Matches[1] }
                 } catch {}
             }
@@ -901,11 +1020,20 @@ if ($HasNvidiaSmi) {
     substep "       This is a driver issue, not an SDK issue." "Yellow"
     substep "       Ensure the ROCm compute driver is installed alongside the display driver:" "Yellow"
     substep "       https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
+} elseif ($script:ROCmGfxArch) {
+    # Known arch: PyTorch comes from AMD's bundled-runtime ROCm wheels (repo.amd.com),
+    # which ship their own runtime -- HIP SDK optional (only adds the system toolchain).
+    Write-Host ""
+    step "gpu" "AMD ROCm ($script:ROCmGfxArch)" "Cyan"
+    substep "Detected: $ROCmGpuLabel" "Cyan"
+    substep "GPU PyTorch uses AMD's bundled-runtime ROCm wheels -- HIP SDK not required (optional)." "Cyan"
+    Write-Host ""
 } elseif ($ROCmGpuLabel) {
     Write-Host ""
-    step "gpu" "AMD GPU detected -- HIP SDK not found" "Yellow"
+    step "gpu" "AMD GPU detected -- arch unknown" "Yellow"
     substep "Detected: $ROCmGpuLabel" "Yellow"
-    substep "Install the HIP SDK for ROCm GPU inference:" "Yellow"
+    substep "Could not determine the GPU arch (gfx...). Install the HIP SDK or set" "Yellow"
+    substep "UNSLOTH_ROCM_GFX_ARCH to enable GPU ROCm PyTorch:" "Yellow"
     substep "https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
     Write-Host ""
 } else {
@@ -1062,17 +1190,22 @@ if ($vsResult) {
 # or installed. Without it, detection is best-effort and only sets the flag.
 function Resolve-CudaToolkit {
     param([switch]$RequireOrExit)
-# IMPORTANT: The CUDA Toolkit version must be <= the max CUDA version the
-# NVIDIA driver supports.  nvidia-smi reports this as "CUDA Version: X.Y".
-# If we install a toolkit newer than the driver supports, llama-server will
-# fail at runtime with "ggml_cuda_init: failed to initialize CUDA: (null)".
+# Toolkit major must be <= the driver's max CUDA major (nvidia-smi "CUDA Version: X.Y");
+# a newer-major toolkit fails at runtime ("ggml_cuda_init: failed to initialize CUDA").
 
-# -- Detect max CUDA version the driver supports --
 $DriverMaxCuda = $null
 try {
-    $smiOut = & $NvidiaSmiExe 2>&1 | Out-String
-    # Newer NVIDIA drivers (e.g. 610.x) report the driver max CUDA as
-    # "CUDA UMD Version: X.Y" rather than "CUDA Version: X.Y"; accept both.
+    # Bounded: source-build toolkit resolution must not hang on a wedged smi.
+    # test_resolve_cuda_toolkit.ps1 extracts this function alone into a child
+    # pwsh (no Invoke-NvidiaSmiBounded in scope) and stubs nvidia-smi with a
+    # .ps1 script, so fall back to direct invocation when the bounded runner
+    # is unavailable; production setup.ps1 always has it defined.
+    $smiOut = if (Get-Command Invoke-NvidiaSmiBounded -ErrorAction SilentlyContinue) {
+        Invoke-NvidiaSmiBounded $NvidiaSmiExe
+    } else {
+        & $NvidiaSmiExe 2>&1 | Out-String
+    }
+    # Newer drivers report "CUDA UMD Version: X.Y" instead of "CUDA Version: X.Y"; accept both.
     if ($smiOut -match "CUDA(?: UMD)? Version:\s+([\d]+)\.([\d]+)") {
         $DriverMaxCuda = "$($Matches[1]).$($Matches[2])"
         substep "driver supports up to CUDA $DriverMaxCuda"
@@ -1096,7 +1229,6 @@ $NvccPath = $null
 
 if ($DriverMaxCuda) {
     $drMajorCuda = [int]$DriverMaxCuda.Split('.')[0]
-    $drMinorCuda = [int]$DriverMaxCuda.Split('.')[1]
 
     # --- Step 1: Check existing CUDA_PATH first ---
     $existingCudaPath = [Environment]::GetEnvironmentVariable('CUDA_PATH', 'Machine')
@@ -1108,7 +1240,7 @@ if ($DriverMaxCuda) {
         $verOut = & $candidateNvcc --version 2>&1 | Out-String
         if ($verOut -match 'release\s+(\d+)\.(\d+)') {
             $tkMaj = [int]$Matches[1]; $tkMin = [int]$Matches[2]
-            $isCompat = ($tkMaj -lt $drMajorCuda) -or ($tkMaj -eq $drMajorCuda -and $tkMin -le $drMinorCuda)
+            $isCompat = ($tkMaj -le $drMajorCuda)
             if ($isCompat) {
                 # Also verify the toolkit supports our GPU architecture
                 $archOk = $true
@@ -1124,7 +1256,7 @@ if ($DriverMaxCuda) {
                     substep "using existing CUDA Toolkit at CUDA_PATH (nvcc: $NvccPath)"
                 }
             } else {
-                substep "CUDA_PATH ($existingCudaPath) has CUDA $tkMaj.$tkMin which exceeds driver max $DriverMaxCuda" "Yellow"
+                substep "CUDA_PATH ($existingCudaPath) has CUDA $tkMaj.$tkMin with major $tkMaj, which exceeds driver CUDA major $drMajorCuda ($DriverMaxCuda)" "Yellow"
             }
         }
     }
@@ -1141,12 +1273,19 @@ if ($DriverMaxCuda) {
                 }
             }
         } else {
-            # Check if there's an incompatible (too new) toolkit installed
+            # No side-by-side match: a major-compatible toolkit may still be on
+            # PATH/CUDA_PATH/a custom dir; use it, else record it as too-new.
             $AnyNvcc = Find-Nvcc
             if ($AnyNvcc) {
                 $NvccOut = & $AnyNvcc --version 2>&1 | Out-String
-                if ($NvccOut -match "release\s+([\d]+\.[\d]+)") {
-                    $IncompatibleToolkit = $Matches[1]
+                if ($NvccOut -match "release\s+(\d+)\.(\d+)") {
+                    $tkMaj = [int]$Matches[1]; $tkMin = [int]$Matches[2]
+                    if ($tkMaj -le $drMajorCuda) {
+                        $NvccPath = $AnyNvcc
+                        substep "found compatible CUDA Toolkit (nvcc: $NvccPath)"
+                    } else {
+                        $IncompatibleToolkit = "$tkMaj.$tkMin"
+                    }
                 }
             }
         }
@@ -1155,26 +1294,18 @@ if ($DriverMaxCuda) {
     $NvccPath = Find-Nvcc
 }
 
-# -- If incompatible toolkit is blocking, tell user to uninstall it --
+# A newer-major toolkit blocked by the driver: explain the mismatch.
 if (-not $NvccPath -and $IncompatibleToolkit) {
+    Write-CudaDriverToolkitMismatch -ToolkitVersion $IncompatibleToolkit -DriverMaxCuda $DriverMaxCuda
     if (-not $RequireOrExit) {
-        substep "CUDA Toolkit $IncompatibleToolkit exceeds driver max $DriverMaxCuda -- skipping; prebuilt llama.cpp needs no local toolkit" "Yellow"
         $script:CudaToolkitReady = $false
         return
     }
+    # Reached only by a source build (forced, or after a prebuilt-install failure);
+    # with no compatible toolkit it must fail (setup.sh degrades to CPU instead).
     Write-Host "" -ForegroundColor Red
     Write-Host "========================================================================" -ForegroundColor Red
-    Write-Host "[ERROR] CUDA Toolkit $IncompatibleToolkit is installed but INCOMPATIBLE" -ForegroundColor Red
-    Write-Host "        with your NVIDIA driver (which supports up to CUDA $DriverMaxCuda)." -ForegroundColor Red
-    Write-Host "" -ForegroundColor Red
-    Write-Host "  This will cause 'failed to initialize CUDA' errors at runtime." -ForegroundColor Red
-    Write-Host "" -ForegroundColor Red
-    Write-Host "  To fix:" -ForegroundColor Yellow
-    Write-Host "    1. Open Control Panel -> Programs -> Uninstall a program" -ForegroundColor Yellow
-    Write-Host "    2. Uninstall 'NVIDIA CUDA Toolkit $IncompatibleToolkit'" -ForegroundColor Yellow
-    Write-Host "    3. Re-run setup.bat (it will install CUDA $DriverMaxCuda automatically)" -ForegroundColor Yellow
-    Write-Host "" -ForegroundColor Yellow
-    Write-Host "  Alternatively, update your NVIDIA driver to one that supports CUDA $IncompatibleToolkit." -ForegroundColor Gray
+    Write-Host "[ERROR] CUDA source build cannot use the installed toolkit with this driver." -ForegroundColor Red
     Write-Host "========================================================================" -ForegroundColor Red
     exit 1
 }
@@ -1187,10 +1318,9 @@ if (-not $NvccPath -and $RequireOrExit) {
         if ($DriverMaxCuda) {
             # Query winget for available CUDA Toolkit versions
             $drMajor = [int]$DriverMaxCuda.Split('.')[0]
-            $drMinor = [int]$DriverMaxCuda.Split('.')[1]
             $AvailableVersions = @()
             try {
-                $rawOutput = winget show Nvidia.CUDA --versions --accept-source-agreements 2>&1 | Out-String
+                $rawOutput = winget show Nvidia.CUDA --versions --source winget --accept-source-agreements 2>&1 | Out-String
                 # Parse version lines (e.g. "12.6", "12.5", "11.8")
                 foreach ($line in $rawOutput -split "`n") {
                     $line = $line.Trim()
@@ -1200,13 +1330,12 @@ if (-not $NvccPath -and $RequireOrExit) {
                 }
             } catch {}
 
-            # Filter to compatible versions (<= driver max) and pick the highest
+            # Filter to compatible major versions and pick the highest
             $BestVersion = $null
             foreach ($ver in $AvailableVersions) {
                 $parts = $ver.Split('.')
                 $vMajor = [int]$parts[0]
-                $vMinor = [int]$parts[1]
-                if ($vMajor -lt $drMajor -or ($vMajor -eq $drMajor -and $vMinor -le $drMinor)) {
+                if ($vMajor -le $drMajor) {
                     $BestVersion = $ver
                     break  # list is descending, first match is highest compatible
                 }
@@ -1224,7 +1353,7 @@ if (-not $NvccPath -and $RequireOrExit) {
                     substep "CUDA Toolkit $BestVersion installed (nvcc: $NvccPath)"
                 }
             } else {
-                substep "no compatible CUDA Toolkit version found in winget (need <= $DriverMaxCuda)" "Yellow"
+                substep "no compatible CUDA Toolkit version found in winget (need CUDA major <= $drMajor)" "Yellow"
             }
         } else {
             substep "Installing CUDA Toolkit (latest) via winget..."
@@ -1246,7 +1375,7 @@ if (-not $NvccPath) {
     }
     Write-Host "[ERROR] CUDA Toolkit (nvcc) is required but could not be found or installed." -ForegroundColor Red
     if ($DriverMaxCuda) {
-        Write-Host "        Install CUDA Toolkit $DriverMaxCuda from https://developer.nvidia.com/cuda-toolkit-archive" -ForegroundColor Yellow
+        Write-Host "        Install a CUDA Toolkit with major version $($DriverMaxCuda.Split('.')[0]) from https://developer.nvidia.com/cuda-toolkit-archive" -ForegroundColor Yellow
     } else {
         Write-Host "        Install CUDA Toolkit from https://developer.nvidia.com/cuda-downloads" -ForegroundColor Yellow
     }
@@ -1344,8 +1473,12 @@ $script:CudaToolkitReady = $true
 if ($HasROCm) {
     $rocmVerLabel = if ($script:ROCmVersionFull) { "ROCm $script:ROCmVersionFull" } elseif ($script:ROCmVersion) { "ROCm $script:ROCmVersion" } else { "ROCm (version unknown)" }
     step "rocm" $rocmVerLabel
+} elseif ($script:ROCmGfxArch) {
+    # GPU training/inference works via AMD's bundled-runtime ROCm PyTorch wheels;
+    # the HIP SDK is optional (only the system ROCm toolchain).
+    step "rocm" "GPU via bundled ROCm wheels ($script:ROCmGfxArch) -- HIP SDK optional" "Cyan"
 } elseif ($ROCmGpuLabel) {
-    step "rocm" "HIP SDK not found -- GPU-accelerated training unavailable" "Yellow"
+    step "rocm" "AMD GPU detected -- arch unknown; HIP SDK not found" "Yellow"
 }
 
 # ============================================
@@ -1404,7 +1537,9 @@ if ($IsPipInstall) {
         substep "installing bun (faster frontend package installs)..."
         $prevEAP_bun = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
-        Invoke-SetupCommand { npm install -g bun } | Out-Null
+        # --allow-scripts=bun: npm >=11.16 gates install scripts and bun's
+        # postinstall fetches its binary; without it the install is a broken stub.
+        Invoke-SetupCommand { npm install -g bun --allow-scripts=bun } | Out-Null
         $ErrorActionPreference = $prevEAP_bun
         Refresh-Environment
         if (Get-Command bun -ErrorAction SilentlyContinue) {
@@ -1417,13 +1552,94 @@ if ($IsPipInstall) {
     }
 }
 
-# 1g. Python (>= 3.11 and < 3.14). Prefer py.exe so a 3.14 ahead of 3.13 on PATH does not trip the gate.
+# Conda CPython ships modified DLL search paths that break torch's c10.dll
+# loading on Windows; a venv made from conda Python inherits its base_prefix,
+# so check the executable path AND sys.base_prefix.
+$CondaSkipPattern = '(?i)(conda|miniconda|anaconda|miniforge|mambaforge)'
+function Test-IsConda {
+    param([string]$Exe)
+    if ($Exe -match $CondaSkipPattern) { return $true }
+    try {
+        $basePrefix = (& $Exe -c "import sys; print(sys.base_prefix)" 2>$null | Out-String).Trim()
+        if ($basePrefix -match $CondaSkipPattern) { return $true }
+    } catch { }
+    return $false
+}
+
+# 1g. Python (>= 3.11 and < 3.14). Prefer the interpreter install.ps1 already
+# resolved and built the venv with (UNSLOTH_SETUP_PYTHON), or the existing
+# venv python, before re-probing a system where a 3.14 or a WindowsApps stub
+# ahead on PATH would trip the gate. setup.ps1 only updates packages in that
+# venv, so the handoff is safe to reuse once validated.
+function Resolve-ReusedSetupPython {
+    if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_SETUP_PYTHON) -and
+        (Test-Path -LiteralPath $env:UNSLOTH_SETUP_PYTHON)) {
+        return $env:UNSLOTH_SETUP_PYTHON
+    }
+    # Standalone `unsloth studio setup/update` (install.ps1 did not run): derive
+    # the venv python from the studio root, mirroring the resolver below.
+    $root = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() }
+            elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() }
+            else { Join-Path $env:USERPROFILE ".unsloth\studio" }
+    if ($root -eq "~") {
+        # Join-Path with an empty child throws on Windows PowerShell 5.1.
+        $root = $env:USERPROFILE
+    } elseif ($root -like "~/*" -or $root -like "~\*") {
+        $root = Join-Path $env:USERPROFILE $root.Substring(1).TrimStart('/', '\')
+    }
+    $venvPy = Join-Path $root "unsloth_studio\Scripts\python.exe"
+    if (Test-Path -LiteralPath $venvPy) { return $venvPy }
+    return $null
+}
+$ReusedSetupPython = Resolve-ReusedSetupPython
+
 $HasPython = $null -ne (Get-Command python -ErrorAction SilentlyContinue)
-$PyLauncher = Get-Command py -CommandType Application -ErrorAction SilentlyContinue
 $PythonOk = $false
 $DetectedPyVer = $null
 
-if ($PyLauncher) {
+function Get-CompatiblePythonVersion {
+    param([string]$PythonExe)
+    try {
+        $out = & $PythonExe --version 2>&1 | Out-String
+        if ($out -match 'Python (3\.(11|12|13)(\.\d+)?)') {
+            return $Matches[1]
+        }
+    } catch { }
+    return $null
+}
+
+function Add-PythonDirToProcessPath {
+    param([string]$PythonExe)
+    try {
+        if ($PythonExe -and (Test-Path -LiteralPath $PythonExe)) {
+            $resolvedDir = Split-Path -Parent $PythonExe
+            $alreadyOnPath = ($env:PATH -split ';' | Where-Object { $_.TrimEnd('\') -ieq $resolvedDir.TrimEnd('\') }).Count -gt 0
+            if (-not $alreadyOnPath) {
+                $env:PATH = "$resolvedDir;$env:PATH"
+            }
+            $script:HasPython = $true
+        }
+    } catch { }
+}
+
+# Reuse the install.ps1 / venv interpreter before any system probe.
+if ($ReusedSetupPython) {
+    $_reusedVer = Get-CompatiblePythonVersion $ReusedSetupPython
+    if ($_reusedVer -and -not (Test-IsConda $ReusedSetupPython)) {
+        $DetectedPyVer = $_reusedVer
+        Add-PythonDirToProcessPath $ReusedSetupPython
+        $PythonOk = $true
+    }
+}
+
+# Fall back to every py.exe on PATH (all-users and per-user launchers can both
+# register). -All is required: Windows PowerShell 5.1 returns only the first
+# launcher without it, and the PowerShell 7 multi-match array breaks the call
+# operator if used directly.
+$PyLaunchers = if ($PythonOk) { @() } else { @(Get-Command py -All -CommandType Application -ErrorAction SilentlyContinue) }
+
+foreach ($PyLauncher in $PyLaunchers) {
+    if ($PyLauncher.Source -match $CondaSkipPattern) { continue }
     foreach ($minor in @("3.13", "3.12", "3.11")) {
         try {
             $out = & $PyLauncher.Source "-$minor" --version 2>&1 | Out-String
@@ -1435,12 +1651,7 @@ if ($PyLauncher) {
                 try {
                     $resolvedExe = (& $PyLauncher.Source "-$minor" -c "import sys; print(sys.executable)" 2>$null | Select-Object -First 1)
                     if ($resolvedExe -and (Test-Path $resolvedExe)) {
-                        $resolvedDir = Split-Path -Parent $resolvedExe
-                        $alreadyOnPath = ($env:PATH -split ';' | Where-Object { $_.TrimEnd('\') -ieq $resolvedDir.TrimEnd('\') }).Count -gt 0
-                        if (-not $alreadyOnPath) {
-                            $env:PATH = "$resolvedDir;$env:PATH"
-                        }
-                        $HasPython = $true
+                        Add-PythonDirToProcessPath $resolvedExe
                     }
                 } catch { }
                 $PythonOk = $true
@@ -1448,6 +1659,7 @@ if ($PyLauncher) {
             }
         } catch { }
     }
+    if ($PythonOk) { break }
 }
 
 if (-not $PythonOk -and $HasPython) {
@@ -1680,36 +1892,33 @@ if (Test-Path $OxcValidatorDir) {
 Write-Host ""
 substep "setting up Python environment..."
 
-# Find Python -- skip Anaconda/Miniconda distributions.
-# Conda-bundled CPython ships modified DLL search paths that break
-# torch's c10.dll loading on Windows. Standalone CPython (python.org,
-# winget, uv) does not have this issue.
-# Uses Get-Command -All to look past conda entries that shadow a valid
-# standalone Python further down PATH, and probes py.exe (the Python
-# Launcher) which reliably finds python.org installs.
-#
-# NOTE: A venv created from conda Python inherits conda's base_prefix
-# even though the venv path itself does not contain "conda". We check
-# both the executable path AND sys.base_prefix to catch this case.
-$CondaSkipPattern = '(?i)(conda|miniconda|anaconda|miniforge|mambaforge)'
+# Find Python -- skip Anaconda/Miniconda distributions ($CondaSkipPattern and
+# Test-IsConda are defined above the 1g gate). Standalone CPython (python.org,
+# winget, uv) does not have conda's torch c10.dll loading issue.
 $PythonCmd = $null
 
-# Helper: check if a Python executable is conda-based by inspecting
-# both the path and sys.base_prefix (catches venvs created from conda).
-function Test-IsConda {
-    param([string]$Exe)
-    if ($Exe -match $CondaSkipPattern) { return $true }
+# 0. Reuse the interpreter install.ps1 already resolved and built the venv with
+#    (UNSLOTH_SETUP_PYTHON, or the existing venv python) before probing the
+#    system -- it is already validated as supported and non-conda.
+if ($ReusedSetupPython) {
     try {
-        $basePrefix = (& $Exe -c "import sys; print(sys.base_prefix)" 2>$null | Out-String).Trim()
-        if ($basePrefix -match $CondaSkipPattern) { return $true }
+        $out = & $ReusedSetupPython --version 2>&1 | Out-String
+        if ($out -match 'Python 3\.(\d+)') {
+            $pyMinor = [int]$Matches[1]
+            if ($pyMinor -ge 11 -and $pyMinor -le 13 -and -not (Test-IsConda $ReusedSetupPython)) {
+                $PythonCmd = $ReusedSetupPython
+            }
+        }
     } catch { }
-    return $false
 }
 
 # 1. Try the Python Launcher (py.exe) first -- most reliable on Windows.
-#    py.exe is installed by python.org and resolves to standalone CPython.
-$pyLauncher = Get-Command py -CommandType Application -ErrorAction SilentlyContinue
-if ($pyLauncher -and $pyLauncher.Source -notmatch $CondaSkipPattern) {
+#    Enumerate every launcher with -All (Windows PowerShell 5.1 returns only
+#    the first match without it) and search each for a supported, non-conda
+#    interpreter.
+$PyLaunchersResolve = if ($PythonCmd) { @() } else { @(Get-Command py -All -CommandType Application -ErrorAction SilentlyContinue) }
+foreach ($pyLauncher in $PyLaunchersResolve) {
+    if ($pyLauncher.Source -match $CondaSkipPattern) { continue }
     foreach ($minor in @("3.13", "3.12", "3.11")) {
         try {
             $out = & $pyLauncher.Source "-$minor" --version 2>&1 | Out-String
@@ -1727,6 +1936,7 @@ if ($pyLauncher -and $pyLauncher.Source -notmatch $CondaSkipPattern) {
             }
         } catch { }
     }
+    if ($PythonCmd) { break }
 }
 
 # 2. Fall back to scanning python3.x / python3 / python on PATH.
@@ -1994,6 +2204,21 @@ if ($env:SKIP_STUDIO_BASE -ne "1" -and $env:STUDIO_LOCAL_INSTALL -ne "1") {
     if ($InstalledVer -and $LatestVer -and ($InstalledVer -eq $LatestVer)) {
         step "python" "$_PkgName $InstalledVer is up to date"
         $SkipPythonDeps = $true
+        # ...but not if an AMD GPU is present and installed PyTorch is CPU-only
+        # (host predates ROCm-wheel support, or GPU added later): the fast "up to
+        # date" path would leave the user on CPU torch with Train/Export disabled.
+        # Force the dependency pass so the ROCm wheels get installed.
+        if ($script:ROCmGfxArch) {
+            $_torchIsCpu = $true
+            try {
+                & python -c "import torch, sys; sys.exit(0 if torch.cuda.is_available() else 1)" 2>$null
+                if ($LASTEXITCODE -eq 0) { $_torchIsCpu = $false }
+            } catch {}
+            if ($_torchIsCpu) {
+                substep "AMD GPU ($script:ROCmGfxArch) detected but installed PyTorch is CPU-only -- reinstalling ROCm PyTorch" "Cyan"
+                $SkipPythonDeps = $false
+            }
+        }
     } elseif ($InstalledVer -and $LatestVer) {
         substep "$_PkgName $InstalledVer -> $LatestVer available, updating..."
     } elseif (-not $LatestVer) {
@@ -2057,7 +2282,13 @@ if ($HasNvidiaSmi) {
 # Wheels bundle their own ROCm runtime; HIP SDK version is irrelevant.
 $ROCmGfxArch = $script:ROCmGfxArch
 $ROCmIndexUrl = $null
-if ($HasROCm -and $CuTag -eq "cpu") {
+# Install AMD ROCm PyTorch wheels when ROCm is confirmed OR a gfx arch is known
+# (name-inferred on Adrenalin-only hosts). The per-arch wheels bundle the runtime
+# (rocm-sdk-libraries-<gfx>), so torch.cuda.is_available() is True without a HIP
+# SDK -- which flips Studio out of chat-only (CHAT_ONLY) and enables Train/Export.
+# Gating on $HasROCm alone left Strix Halo / Radeon 8060S on CPU torch; a failed
+# ROCm install still falls back to CPU below, so this is safe.
+if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
     $amdIndexBase = if ($env:UNSLOTH_ROCM_WINDOWS_MIRROR) { $env:UNSLOTH_ROCM_WINDOWS_MIRROR.TrimEnd('/') } else { "https://repo.amd.com/rocm/whl" }
     $archFamilyMap = @{
         "gfx1201" = "gfx120X-all"; "gfx1200" = "gfx120X-all"  # RDNA 4
@@ -2237,14 +2468,35 @@ if ($stackExit -ne 0) {
     $ErrorActionPreference = $prevEAP
 }
 
-# ── Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+# ── Pre-install transformers 5.x into .venv_t5_530/, .venv_t5_550/, and .venv_t5_510/ ──
 # Runs outside the deps fast-path gate so that upgrades from the legacy
 # single .venv_t5 are always migrated to the tiered layout.
 # T5 sidecar venvs live under the resolved $StudioHome so custom installs are self-contained.
 $VenvT5_530Dir = Join-Path $StudioHome ".venv_t5_530"
 $VenvT5_550Dir = Join-Path $StudioHome ".venv_t5_550"
+$VenvT5_510Dir = Join-Path $StudioHome ".venv_t5_510"
 $VenvT5Legacy = Join-Path $StudioHome ".venv_t5"
 
+function Test-TargetPackageVersion {
+    param(
+        [Parameter(Mandatory = $true)][string]$TargetDir,
+        [Parameter(Mandatory = $true)][string]$PackageName,
+        [Parameter(Mandatory = $true)][string]$ExpectedVersion
+    )
+    if (-not (Test-Path -LiteralPath $TargetDir -PathType Container)) { return $false }
+    $packageNorm = $PackageName.Replace("-", "_")
+    foreach ($pattern in @("$packageNorm-*.dist-info", "$PackageName-*.dist-info")) {
+        foreach ($distInfo in @(Get-ChildItem -LiteralPath $TargetDir -Directory -Filter $pattern -ErrorAction SilentlyContinue)) {
+            $metadata = Join-Path $distInfo.FullName "METADATA"
+            if (-not (Test-Path -LiteralPath $metadata -PathType Leaf)) { continue }
+            foreach ($line in (Get-Content -LiteralPath $metadata -ErrorAction SilentlyContinue)) {
+                if ($line -eq "Version: $ExpectedVersion") { return $true }
+            }
+        }
+    }
+    return $false
+}
+
 $_NeedT5Install = $false
 if (Test-Path -LiteralPath $VenvT5Legacy) {
     Assert-StudioOwnedOrAbsent -Path $VenvT5Legacy -Label "legacy transformers sidecar venv"
@@ -2253,6 +2505,10 @@ if (Test-Path -LiteralPath $VenvT5Legacy) {
 }
 if (-not (Test-Path -LiteralPath $VenvT5_530Dir)) { $_NeedT5Install = $true }
 if (-not (Test-Path -LiteralPath $VenvT5_550Dir)) { $_NeedT5Install = $true }
+if (-not (Test-Path -LiteralPath $VenvT5_510Dir)) { $_NeedT5Install = $true }
+if (-not (Test-TargetPackageVersion -TargetDir $VenvT5_530Dir -PackageName "transformers" -ExpectedVersion "5.3.0")) { $_NeedT5Install = $true }
+if (-not (Test-TargetPackageVersion -TargetDir $VenvT5_550Dir -PackageName "transformers" -ExpectedVersion "5.5.0")) { $_NeedT5Install = $true }
+if (-not (Test-TargetPackageVersion -TargetDir $VenvT5_510Dir -PackageName "transformers" -ExpectedVersion "5.10.2")) { $_NeedT5Install = $true }
 # Also reinstall when python deps were updated
 if (-not $SkipPythonDeps) { $_NeedT5Install = $true }
 
@@ -2330,9 +2586,44 @@ if ($script:UnslothVerbose) {
 if ($tiktokenInstallExit -ne 0) {
     substep "Could not install tiktoken into .venv_t5_550/ -- Qwen tokenizers may fail" "Yellow"
 }
-$ErrorActionPreference = $prevEAP_t5
 step "transformers" "5.5.0 pre-installed"
 
+# --- .venv_t5_510 (transformers 5.10.2) ---
+substep "pre-installing transformers 5.10.2 for Gemma 4 Unified support..."
+Assert-StudioOwnedOrAbsent -Path $VenvT5_510Dir -Label "transformers 5.10 sidecar venv"
+if (Test-Path -LiteralPath $VenvT5_510Dir) { Remove-Item -LiteralPath $VenvT5_510Dir -Recurse -Force }
+[System.IO.Directory]::CreateDirectory($VenvT5_510Dir) | Out-Null
+Mark-StudioOwned -Path $VenvT5_510Dir
+foreach ($pkg in @("transformers==5.10.2", "huggingface_hub==1.8.0", "hf_xet==1.4.2")) {
+    if ($script:UnslothVerbose) {
+        Fast-Install --target $VenvT5_510Dir --no-deps $pkg
+        $t5PkgExit = $LASTEXITCODE
+        $output = ""
+    } else {
+        $output = Fast-Install --target $VenvT5_510Dir --no-deps $pkg | Out-String
+        $t5PkgExit = $LASTEXITCODE
+    }
+    if ($t5PkgExit -ne 0) {
+        Write-Host "[FAIL] Could not install $pkg into .venv_t5_510/" -ForegroundColor Red
+        Write-Host $output -ForegroundColor Red
+        $ErrorActionPreference = $prevEAP_t5
+        exit 1
+    }
+}
+if ($script:UnslothVerbose) {
+    Fast-Install --target $VenvT5_510Dir tiktoken
+    $tiktokenInstallExit = $LASTEXITCODE
+    $output = ""
+} else {
+    $output = Fast-Install --target $VenvT5_510Dir tiktoken | Out-String
+    $tiktokenInstallExit = $LASTEXITCODE
+}
+if ($tiktokenInstallExit -ne 0) {
+    substep "Could not install tiktoken into .venv_t5_510/ -- Qwen tokenizers may fail" "Yellow"
+}
+$ErrorActionPreference = $prevEAP_t5
+step "transformers" "5.10.2 pre-installed"
+
 } # end $_NeedT5Install
 
 # ==========================================================================
@@ -2351,7 +2642,9 @@ $LlamaCppDir = Join-Path $UnslothHome "llama.cpp"
 $NeedLlamaSourceBuild = $false
 $SkipPrebuiltInstall = $false
 $RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
-$HelperReleaseRepo = "ggml-org/llama.cpp"
+# GPU Windows (CUDA / ROCm) installs the fork's app-* prebuilts; CPU-only stays
+# on ggml-org (the fork ships no windows-cpu bundle). Mirrors setup.sh's routing.
+$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm) { "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }
 $LlamaPr = if ($env:UNSLOTH_LLAMA_PR) { $env:UNSLOTH_LLAMA_PR.Trim() } else { "" }
 
 $LlamaPrForce = if ($env:UNSLOTH_LLAMA_PR_FORCE) { $env:UNSLOTH_LLAMA_PR_FORCE.Trim() } else { $DefaultLlamaPrForce }
@@ -2447,20 +2740,28 @@ if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
     substep "Skipping prebuilt install -- falling back to source build" "Yellow"
 } else {
     Write-Host ""
-    substep "installing prebuilt llama.cpp bundle (preferred path)..."
     if (Test-Path -LiteralPath $LlamaCppDir) {
         substep "Existing llama.cpp install detected -- validating staged prebuilt update before replacement"
         # If the existing install is the wrong kind (e.g. windows-cpu on a ROCm
-        # machine that should have windows-hip), remove it so the installer is
+        # machine that should have windows-rocm), remove it so the installer is
         # forced to download the correct variant rather than skipping on tag match.
         $existingMetaPath = Join-Path $LlamaCppDir "UNSLOTH_PREBUILT_INFO.json"
         if (Test-Path $existingMetaPath) {
             try {
                 $existingMeta = Get-Content $existingMetaPath -Raw | ConvertFrom-Json
                 $existingKind = $existingMeta.install_kind
-                $expectedKind = if ($HasROCm) { "windows-hip" } elseif ($HasNvidiaSmi) { "windows-cuda" } else { "windows-cpu" }
-                if ($existingKind -and $existingKind -ne $expectedKind) {
-                    substep "Removing mismatched llama.cpp install (found '$existingKind', need '$expectedKind')..."
+                # A ROCm host may legitimately carry the fork's windows-rocm bundle
+                # or the upstream windows-hip fallback, so accept either and never
+                # treat a valid ROCm install as mismatched. A name-inferred gfx
+                # arch (Adrenalin-only, no confirmed runtime) still counts as
+                # ROCm-capable -- the lemonade prebuilt bundles its own runtime,
+                # mirroring the --rocm-gfx forward below. NOTE: this block is
+                # currently inert -- write_prebuilt_metadata does not persist an
+                # install_kind key, so $existingKind is always null. If that changes,
+                # add the remaining host kinds (e.g. windows-arm64) before relying on it.
+                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu") }
+                if ($existingKind -and ($existingKind -notin $expectedKinds)) {
+                    substep "Removing mismatched llama.cpp install (found '$existingKind', need one of: $($expectedKinds -join ', '))..."
                     Remove-Item -Recurse -Force -LiteralPath $LlamaCppDir -ErrorAction SilentlyContinue
                 }
             } catch {
@@ -2468,6 +2769,7 @@ if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
             }
         }
     }
+    substep "installing prebuilt llama.cpp bundle (preferred path)..."
     # why: install_llama_prebuilt.py uses os.replace(), which would displace
     # an unrelated $env:UNSLOTH_STUDIO_HOME\llama.cpp before the source-build
     # ownership check below ever runs.
@@ -2478,17 +2780,18 @@ if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
             "$PSScriptRoot\install_llama_prebuilt.py",
             "--install-dir", $LlamaCppDir,
             "--llama-tag", $RequestedLlamaTag,
-            "--published-repo", $HelperReleaseRepo,
-            "--simple-policy"
+            "--published-repo", $HelperReleaseRepo
         )
         if ($HasROCm) {
             $prebuiltArgs += "--has-rocm"
-            # Forward the resolved gfx arch so the lemonade HIP prebuilt is picked
-            # even when the installer's own probe cannot report it (amd-smi-only
-            # hosts, name-inferred arch).
-            if ($script:ROCmGfxArch) {
-                $prebuiltArgs += @("--rocm-gfx", $script:ROCmGfxArch)
-            }
+        }
+        # Forward the resolved gfx arch so the lemonade HIP prebuilt is picked even
+        # when the installer's probe can't confirm the runtime (amd-smi-only /
+        # Adrenalin-only, name-inferred arch). --rocm-gfx is authoritative and
+        # implies ROCm in install_llama_prebuilt.py, so the GPU prebuilt is selected
+        # even with $HasROCm false. Gating on $HasROCm gave Strix Halo / 8060S CPU.
+        if ($script:ROCmGfxArch) {
+            $prebuiltArgs += @("--rocm-gfx", $script:ROCmGfxArch)
         }
         if ($env:UNSLOTH_LLAMA_RELEASE_TAG) {
             $prebuiltArgs += @("--published-release-tag", $env:UNSLOTH_LLAMA_RELEASE_TAG)
@@ -2583,7 +2886,7 @@ if ($NeedLlamaSourceBuild) {
         substep "installing OpenSSL dev (for HTTPS in llama-server)..."
         $HasWinget = $null -ne (Get-Command winget -ErrorAction SilentlyContinue)
         if ($HasWinget) {
-            winget install -e --id ShiningLight.OpenSSL.Dev --accept-package-agreements --accept-source-agreements
+            winget install -e --id ShiningLight.OpenSSL.Dev --source winget --accept-package-agreements --accept-source-agreements
             # Re-check after install
             foreach ($root in $OpenSslRoots) {
                 if (Test-Path (Join-Path $root 'include\openssl\ssl.h')) {
@@ -2665,6 +2968,18 @@ if (-not $NeedLlamaSourceBuild) {
     Write-Host ""
     if ($HasNvidiaSmi) {
         substep "building llama.cpp with CUDA support..."
+    } elseif ($HasROCm -or $script:ROCmGfxArch) {
+        # AMD GPU present but in the CPU-only source-build fallback: a HIP source
+        # build needs the full HIP SDK + ROCm clang toolchain. AMD GPU acceleration
+        # comes from the lemonade prebuilt (bundles the runtime, no SDK) -- reaching
+        # here means it couldn't be installed. Warn loudly, don't ship a slow CPU build.
+        $_amdArch = if ($script:ROCmGfxArch) { $script:ROCmGfxArch } else { "ROCm" }
+        substep "[WARN] AMD GPU ($_amdArch) detected, but the GPU-accelerated lemonade" "Yellow"
+        substep "       llama.cpp prebuilt could not be installed -- falling back to a CPU build." "Yellow"
+        substep "       The prebuilt is the AMD GPU path (no HIP SDK required). To restore GPU" "Yellow"
+        substep "       acceleration: re-run the installer (check your network / proxy), or set" "Yellow"
+        substep "       UNSLOTH_LLAMA_RELEASE_TAG to a tag with a gfx prebuilt for your GPU." "Yellow"
+        substep "building llama.cpp (CPU-only fallback)..." "Yellow"
     } else {
         substep "building llama.cpp (CPU-only, no NVIDIA GPU detected)..."
     }
diff --git a/studio/setup.sh b/studio/setup.sh
index fcec6f72fa..0e32289819 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -59,8 +59,9 @@ fi
 # ── Output helpers ──
 # Consistent column layout: 2-space indent, 15-char label (fits llama-quantize), then value.
 # Usage: step <label> <message> [color]   (color defaults to C_OK)
+# Usage: substep <message> [color]         (color defaults to C_DIM)
 step()    { printf "  ${C_DIM}%-15.15s${C_RST}${3:-$C_OK}%s${C_RST}\n" "$1" "$2"; }
-substep() { printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "$1"; }
+substep() { printf "  %-15s${2:-$C_DIM}%s${C_RST}\n" "" "$1"; }
 
 _is_verbose() {
     [ "${UNSLOTH_VERBOSE:-0}" = "1" ]
@@ -154,6 +155,158 @@ _nvcc_meets_llama_minimum() {
     echo "$_raw"
 }
 
+# Run a GPU probe under a 10s timeout when `timeout` is available so a wedged
+# NVIDIA driver cannot hang setup; fall back to a bare call where it is not.
+_setup_run_smi() {
+    if command -v timeout >/dev/null 2>&1; then
+        timeout 10 "$@"
+    else
+        "$@"
+    fi
+}
+
+# Returns 0 when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every NVIDIA
+# device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to the
+# AMD card). Unset means all devices visible. nvidia-smi ignores this env var,
+# so the probes below cannot see the distinction on their own.
+_setup_cvd_hides_nvidia() {
+    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
+    _setup_cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
+    [ -z "$_setup_cvd_trim" ] || [ "$_setup_cvd_trim" = "-1" ]
+}
+
+# Returns 0 when an NVIDIA GPU is present and usable. Primary probe is
+# `nvidia-smi -L` (timeout-bounded). Fallback is /proc/driver/nvidia/gpus,
+# which the driver populates per GPU regardless of nvidia-smi state -- handles
+# PATH gaps and driver init races. Mirrors install.sh _has_usable_nvidia_gpu
+# (PR 6174) so setup routes the same way as the torch installer. A GPU hidden
+# via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
+# install_llama_prebuilt.py has_usable_nvidia), so the AMD probes still run
+# and a mixed host steered to its AMD card keeps the ROCm route.
+_setup_has_usable_nvidia_gpu() {
+    if _setup_cvd_hides_nvidia; then
+        return 1
+    fi
+    _setup_nvsmi=""
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        _setup_nvsmi="nvidia-smi"
+    elif [ -x "/usr/bin/nvidia-smi" ]; then
+        _setup_nvsmi="/usr/bin/nvidia-smi"
+    fi
+    if [ -n "$_setup_nvsmi" ]; then
+        if _setup_run_smi "$_setup_nvsmi" -L 2>/dev/null \
+           | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+            return 0
+        fi
+    fi
+    if [ -d /proc/driver/nvidia/gpus ] && \
+       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
+        return 0
+    fi
+    return 1
+}
+
+_cuda_driver_max_version() {
+    command -v nvidia-smi >/dev/null 2>&1 || return 0
+    _setup_run_smi nvidia-smi 2>/dev/null \
+        | sed -nE 's/.*CUDA( UMD)? Version:[[:space:]]*([0-9]+)\.([0-9]+).*/\2.\3/p' \
+        | head -1 || true
+}
+
+_cuda_version_gt() {
+    local _left=${1:-}
+    local _right=${2:-}
+    if ! [[ "$_left" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+        return 1
+    fi
+    local _left_major=$((10#${BASH_REMATCH[1]}))
+    local _left_minor=$((10#${BASH_REMATCH[2]}))
+    if ! [[ "$_right" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+        return 1
+    fi
+    local _right_major=$((10#${BASH_REMATCH[1]}))
+    local _right_minor=$((10#${BASH_REMATCH[2]}))
+
+    if [ "$_left_major" -gt "$_right_major" ]; then
+        return 0
+    fi
+    if [ "$_left_major" -eq "$_right_major" ] && [ "$_left_minor" -gt "$_right_minor" ]; then
+        return 0
+    fi
+    return 1
+}
+
+_cuda_toolkit_major_gt_driver() {
+    local _toolkit_version=${1:-}
+    local _driver_version=${2:-}
+    if ! [[ "$_toolkit_version" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+        return 1
+    fi
+    local _toolkit_major=$((10#${BASH_REMATCH[1]}))
+    if ! [[ "$_driver_version" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+        return 1
+    fi
+    local _driver_major=$((10#${BASH_REMATCH[1]}))
+    [ "$_toolkit_major" -gt "$_driver_major" ]
+}
+
+_cuda_nvcc_candidate_paths() {
+    if command -v nvcc >/dev/null 2>&1; then
+        command -v nvcc
+    fi
+    if [ -x /usr/local/cuda/bin/nvcc ]; then
+        printf '%s\n' "/usr/local/cuda/bin/nvcc"
+    fi
+    ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V -r 2>/dev/null || true
+}
+
+_cuda_find_compatible_nvcc_for_driver() {
+    local _driver_version=$1
+    local _exclude_path=${2:-}
+    local _candidate _seen _check _status _version
+    local _best_path="" _best_version=""
+    _seen="
+"
+    while IFS= read -r _candidate; do
+        [ -n "$_candidate" ] || continue
+        [ "$_candidate" != "$_exclude_path" ] || continue
+        [ -x "$_candidate" ] || continue
+        case "$_seen" in
+            *"
+$_candidate
+"*) continue ;;
+        esac
+        _seen="${_seen}${_candidate}
+"
+        _check="$(_nvcc_meets_llama_minimum "$_candidate")"
+        _status="$(printf '%s\n' "$_check" | sed -n '1p')"
+        _version="$(printf '%s\n' "$_check" | sed -n '2p')"
+        [ "$_status" = "ok" ] || continue
+        [ -n "$_version" ] || continue
+        if _cuda_toolkit_major_gt_driver "$_version" "$_driver_version"; then
+            continue
+        fi
+        if [ -z "$_best_version" ] || _cuda_version_gt "$_version" "$_best_version"; then
+            _best_path="$_candidate"
+            _best_version="$_version"
+        fi
+    done <<EOF
+$(_cuda_nvcc_candidate_paths)
+EOF
+    [ -n "$_best_path" ] || return 1
+    printf '%s\n%s\n' "$_best_path" "$_best_version"
+}
+
+_print_cuda_driver_toolkit_mismatch() {
+    local _toolkit_version=$1
+    local _driver_version=$2
+    local _toolkit_major=${_toolkit_version%%.*}
+    local _driver_major=${_driver_version%%.*}
+    substep "CUDA Toolkit $_toolkit_version is a major-version mismatch: toolkit major $_toolkit_major exceeds driver CUDA major $_driver_major ($_driver_version)." "$C_WARN"
+    substep "Update the NVIDIA GPU driver to run CUDA Toolkit $_toolkit_version, or install a CUDA $_driver_major.x toolkit." "$C_WARN"
+    substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." "$C_WARN"
+}
+
 print_llama_error_log() {
     local log_file=$1
     [ -s "$log_file" ] || return 0
@@ -333,7 +486,9 @@ verbose_substep "node check: NEED_NODE=$NEED_NODE NODE_OK=${NODE_OK:-unknown} NP
 # avoids platform-specific installers, PATH issues, and admin requirements.
 if ! command -v bun &>/dev/null; then
     substep "installing bun..."
-    if run_maybe_quiet npm install -g bun && command -v bun &>/dev/null; then
+    # --allow-scripts=bun: npm >=11.16 gates install scripts and bun's
+    # postinstall fetches its binary; without it the install is a broken stub.
+    if run_maybe_quiet npm install -g bun --allow-scripts=bun && command -v bun &>/dev/null; then
         substep "bun installed ($(bun --version))"
     else
         substep "bun install skipped (npm will be used instead)"
@@ -483,6 +638,7 @@ fi
 VENV_DIR="$STUDIO_HOME/unsloth_studio"
 VENV_T5_530_DIR="$STUDIO_HOME/.venv_t5_530"
 VENV_T5_550_DIR="$STUDIO_HOME/.venv_t5_550"
+VENV_T5_510_DIR="$STUDIO_HOME/.venv_t5_510"
 
 [ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
 [ -d "$REPO_ROOT/.venv_overlay" ] && rm -rf "$REPO_ROOT/.venv_overlay"
@@ -596,9 +752,9 @@ else
     verbose_substep "python deps check: installed=$_PKG_NAME@${INSTALLED_VER:-unknown} latest=${LATEST_VER:-unknown}"
 fi
 
-# ── 6b. Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+# ── 6b. Pre-install transformers 5.x into .venv_t5_530/, .venv_t5_550/, and .venv_t5_510/ ──
 # Models like GLM-4.7-Flash, Qwen3 MoE need transformers>=5.3.0.
-# Gemma 4 models need transformers>=5.5.0.
+# Gemma 4 models need transformers>=5.5.0; Gemma 4 Unified needs 5.10.x.
 # Pre-install into separate directories to avoid runtime pip overhead.
 # The training subprocess prepends the appropriate dir to sys.path.
 #
@@ -633,6 +789,21 @@ _assert_studio_owned_or_absent() {
         exit 1
     fi
 }
+_target_has_pkg_version() {
+    _thpv_dir="$1"
+    _thpv_pkg="$2"
+    _thpv_version="$3"
+    [ -d "$_thpv_dir" ] || return 1
+    _thpv_pkg_norm=$(printf '%s' "$_thpv_pkg" | tr '-' '_')
+    for _thpv_metadata in \
+        "$_thpv_dir"/"$_thpv_pkg_norm"-*.dist-info/METADATA \
+        "$_thpv_dir"/"$_thpv_pkg"-*.dist-info/METADATA
+    do
+        [ -f "$_thpv_metadata" ] || continue
+        grep -qx "Version: $_thpv_version" "$_thpv_metadata" && return 0
+    done
+    return 1
+}
 _NEED_T5_INSTALL=false
 if [ -d "$STUDIO_HOME/.venv_t5" ]; then
     # Legacy layout — migrate
@@ -642,6 +813,10 @@ if [ -d "$STUDIO_HOME/.venv_t5" ]; then
 fi
 [ ! -d "$VENV_T5_530_DIR" ] && _NEED_T5_INSTALL=true
 [ ! -d "$VENV_T5_550_DIR" ] && _NEED_T5_INSTALL=true
+[ ! -d "$VENV_T5_510_DIR" ] && _NEED_T5_INSTALL=true
+_target_has_pkg_version "$VENV_T5_530_DIR" "transformers" "5.3.0" || _NEED_T5_INSTALL=true
+_target_has_pkg_version "$VENV_T5_550_DIR" "transformers" "5.5.0" || _NEED_T5_INSTALL=true
+_target_has_pkg_version "$VENV_T5_510_DIR" "transformers" "5.10.2" || _NEED_T5_INSTALL=true
 # Also reinstall when python deps were updated (packages may need rebuild)
 [ "$_SKIP_PYTHON_DEPS" = false ] && _NEED_T5_INSTALL=true
 
@@ -665,31 +840,68 @@ if [ "$_NEED_T5_INSTALL" = true ]; then
     run_quiet "install hf_xet for t5_550" fast_install --target "$VENV_T5_550_DIR" --no-deps "hf_xet==1.4.2"
     run_quiet "install tiktoken for t5_550" fast_install --target "$VENV_T5_550_DIR" "tiktoken"
     step "transformers" "5.5.0 pre-installed"
+
+    _assert_studio_owned_or_absent "$VENV_T5_510_DIR" "transformers 5.10 sidecar venv"
+    [ -d "$VENV_T5_510_DIR" ] && rm -rf "$VENV_T5_510_DIR"
+    mkdir -p "$VENV_T5_510_DIR"
+    : > "$VENV_T5_510_DIR/$_STUDIO_OWNED_MARKER" 2>/dev/null || true
+    run_quiet "install transformers 5.10.2" fast_install --target "$VENV_T5_510_DIR" --no-deps "transformers==5.10.2"
+    run_quiet "install huggingface_hub for t5_510" fast_install --target "$VENV_T5_510_DIR" --no-deps "huggingface_hub==1.8.0"
+    run_quiet "install hf_xet for t5_510" fast_install --target "$VENV_T5_510_DIR" --no-deps "hf_xet==1.4.2"
+    run_quiet "install tiktoken for t5_510" fast_install --target "$VENV_T5_510_DIR" "tiktoken"
+    step "transformers" "5.10.2 pre-installed"
 fi
 fi
 
 # ── GPU detection summary (mirrors setup.ps1 step "gpu" block) ──
+# WSL2 ROCDXG: the system rocminfo enumerates the GPU over /dev/dxg only when
+# HSA_ENABLE_DXG_DETECTION=1 (a no-op on bare metal), and /opt/rocm/bin can be
+# off PATH outside login shells (the profile.d drop-in). Seed both before the
+# probes or a ROCDXG WSL host is misdetected as CPU-only.
+export HSA_ENABLE_DXG_DETECTION="${HSA_ENABLE_DXG_DETECTION:-1}"
+if ! command -v rocminfo >/dev/null 2>&1 && [ -x /opt/rocm/bin/rocminfo ]; then
+    PATH="$PATH:/opt/rocm/bin"
+fi
 _setup_amd_detected=false
+_setup_nvidia_usable=false
 _setup_gfx_all=""
 _setup_mkt=""
-if command -v rocminfo >/dev/null 2>&1 && \
-   rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9][0-9]/{found=1} END{exit !found}'; then
-    _setup_amd_detected=true
-    _setup_gfx_all=$(rocminfo 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
-    _setup_mkt=$(rocminfo 2>/dev/null | awk -F': ' \
-        '/Marketing Name:/{gsub(/^[[:space:]]+|[[:space:]]+$/,"", $2); if($2){print $2; exit}}' || true)
-elif command -v amd-smi >/dev/null 2>&1 && \
-     amd-smi list 2>/dev/null | awk '/^GPU[[:space:]]*[:\[][[:space:]]*[0-9]/{ found=1 } END{ exit !found }'; then
-    _setup_amd_detected=true
-    _setup_gfx_all=$(amd-smi list 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
-    [ -z "$_setup_gfx_all" ] && \
-        _setup_gfx_all=$(amd-smi static --asic 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
-    _setup_mkt=$(amd-smi static --asic 2>/dev/null | awk -F'[:|]' \
-        '/[Mm]arket.?[Nn]ame/{gsub(/^[[:space:]]+|[[:space:]]+$/,"", $2); if($2){print $2; exit}}' || true)
+# NVIDIA priority: classify NVIDIA first and skip the AMD probes entirely on
+# a usable-NVIDIA host (mirrors _has_rocm_gpu in install_python_stack.py).
+# This also keeps a wedged rocminfo/amd-smi from hanging setup before the
+# host is classified; the AMD probes themselves run under _setup_run_smi.
+if _setup_has_usable_nvidia_gpu; then
+    _setup_nvidia_usable=true
+fi
+if [ "$_setup_nvidia_usable" != true ]; then
+    if command -v rocminfo >/dev/null 2>&1 && \
+       _setup_run_smi rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9][0-9]/{found=1} END{exit !found}'; then
+        _setup_amd_detected=true
+        _setup_gfx_all=$(_setup_run_smi rocminfo 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+        _setup_mkt=$(_setup_run_smi rocminfo 2>/dev/null | awk -F': ' \
+            '/Marketing Name:/{gsub(/^[[:space:]]+|[[:space:]]+$/,"", $2); if($2){print $2; exit}}' || true)
+    elif command -v amd-smi >/dev/null 2>&1 && \
+         _setup_run_smi amd-smi list 2>/dev/null | awk '/^GPU[[:space:]]*[:\[][[:space:]]*[0-9]/{ found=1 } END{ exit !found }'; then
+        _setup_amd_detected=true
+        _setup_gfx_all=$(_setup_run_smi amd-smi list 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+        [ -z "$_setup_gfx_all" ] && \
+            _setup_gfx_all=$(_setup_run_smi amd-smi static --asic 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+        _setup_mkt=$(_setup_run_smi amd-smi static --asic 2>/dev/null | awk -F'[:|]' \
+            '/[Mm]arket.?[Nn]ame/{gsub(/^[[:space:]]+|[[:space:]]+$/,"", $2); if($2){print $2; exit}}' || true)
+    elif [ -e /dev/kfd ] && \
+         awk 'FNR==1{ gpu=0; amd=0 } /gpu_id/{ gpu=($2+0>0) } /vendor_id/{ amd=($2==4098) } \
+              gpu && amd { found=1 } END{ exit !found }' \
+             /sys/class/kfd/kfd/topology/nodes/*/properties 2>/dev/null; then
+        # KFD sysfs fallback, AMD vendor_id 4098 only (mirrors install.sh
+        # _has_amd_rocm_gpu): covers AMD hosts where rocminfo/amd-smi are
+        # missing but the kernel exposes the GPU, so the source-build gate
+        # below does not drop them to a CPU llama.cpp build. No gfx arch is
+        # available from this path; name-based inference handles it.
+        _setup_amd_detected=true
+    fi
 fi
 
-if command -v nvidia-smi >/dev/null 2>&1 && \
-   nvidia-smi -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+if [ "$_setup_nvidia_usable" = true ]; then
     step "gpu" "NVIDIA GPU detected"
 elif [ "$_setup_amd_detected" = true ]; then
     _setup_vis="${HIP_VISIBLE_DEVICES:-${ROCR_VISIBLE_DEVICES:-}}"
@@ -706,14 +918,20 @@ elif [ "$_setup_amd_detected" = true ]; then
         substep "gfx arch from UNSLOTH_ROCM_GFX_ARCH env override: $_setup_gfx"
     # Name-based arch inference when tools don't report gfx (mirrors setup.ps1 nameArchTable)
     elif [ -z "$_setup_gfx" ] && [ -n "$_setup_mkt" ]; then
+        # Kept in sync with the table in install.sh (and the PS nameArchTable).
+        # gfx1102 matched BEFORE gfx1100 so the spaceless "RX 7700S" lands on
+        # gfx1102 (bash case has no negative lookahead like the PS tables).
         case "$_setup_mkt" in
-            *"9070 XT"*|*9080*)                                                 _setup_gfx="gfx1201" ;;  # RDNA 4
-            *9070*|*9060*)                                                      _setup_gfx="gfx1200" ;;  # RDNA 4
-            *"8060S"*|*"890M"*|*"Strix Halo"*|*"HX 37"*|*"HX 38"*|*"AI 9 HX"*)  _setup_gfx="gfx1151" ;;  # RDNA 3.5 iGPU
-            *"880M"*|*"Strix Point"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*)       _setup_gfx="gfx1150" ;;  # RDNA 3.5 iGPU
-            *"RX 7900"*|*"RX 7800"*|*"RX 7700"*)                                _setup_gfx="gfx1100" ;;  # RDNA 3 desktop
-            *"RX 7600"*)                                                        _setup_gfx="gfx1102" ;;  # RDNA 3
-            *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*)                             _setup_gfx="gfx1103" ;;  # RDNA 3 iGPU
+            *"9070 XT"*|*9080*)                                                                            _setup_gfx="gfx1201" ;;  # RDNA 4
+            *9070*|*9060*)                                                                                 _setup_gfx="gfx1200" ;;  # RDNA 4
+            *"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _setup_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+            *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _setup_gfx="gfx1150" ;;  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+            *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*)                  _setup_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
+            *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*)                  _setup_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
+            *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*|*"Hawk Point"*|*"Z1 Extreme"*|*"Z2 Extreme"*)            _setup_gfx="gfx1103" ;;  # RDNA 3 iGPU (Phoenix / Hawk Point)
+            *"RX 6900"*|*"RX 6800"*|*"RX 6750"*|*"RX 6700"*|*"PRO W6800"*|*"PRO W6900"*)                    _setup_gfx="gfx1030" ;;  # RDNA 2 (Navi 21)
+            *"RX 6650"*|*"RX 6600"*|*"PRO W6600"*|*"PRO W6650"*)                                            _setup_gfx="gfx1032" ;;  # RDNA 2 (Navi 23)
+            *"RX 6500"*|*"RX 6400"*|*"RX 6300"*|*"PRO W6400"*|*"PRO W6500"*)                                _setup_gfx="gfx1034" ;;  # RDNA 2 (Navi 24)
         esac
         if [ -n "$_setup_gfx" ]; then
             substep "gfx arch inferred from GPU name: $_setup_gfx"
@@ -766,34 +984,46 @@ _HOST_SYSTEM="$(uname -s 2>/dev/null || true)"
 _HOST_MACHINE="$(uname -m 2>/dev/null || true)"
 
 # Pick the release repo install_llama_prebuilt.py plans against.
-# unslothai/llama.cpp ships only Linux CUDA bundles, so CPU-only Linux
-# x86_64 routes to ggml-org for bin-ubuntu-x64.tar.gz. Anything with a
-# GPU tool installed stays on unslothai (CUDA bundle / ROCm source build).
+# The fork ships CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and
+# macOS bundles. Only the plain CPU/Vulkan bundles still come from ggml-org, so
+# CPU-only Linux (x86_64 and arm64) routes there; GPU Linux, Windows and macOS
+# use unslothai.
 _LINUX_HAS_GPU=false
-for _GPU_TOOL in nvidia-smi rocminfo amd-smi hipconfig hipinfo; do
-    if command -v "$_GPU_TOOL" >/dev/null 2>&1; then
-        _LINUX_HAS_GPU=true
-        break
-    fi
-done
+# Route to the fork only for a usable GPU. NVIDIA counts only when a device is
+# actually enumerated and not hidden via CUDA_VISIBLE_DEVICES=""/-1
+# (_setup_nvidia_usable, from _setup_has_usable_nvidia_gpu above) -- mirroring
+# install_llama_prebuilt.py's has_usable_nvidia. Mere nvidia-smi presence
+# (CPU-only CUDA-toolkit containers, broken drivers) or a hidden GPU therefore
+# takes the ggml-org CPU prebuilt instead of a slow source build. AMD is
+# deliberately left on tooling presence, not usability: an unusable NVIDIA host
+# has a good CPU prebuilt to fall back to, whereas tightening AMD would regress
+# ROCm hosts exposing only hipconfig/hipinfo into an unnecessary CPU build.
+if [ "$_setup_nvidia_usable" = true ]; then
+    _LINUX_HAS_GPU=true
+else
+    for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo; do
+        if command -v "$_GPU_TOOL" >/dev/null 2>&1; then
+            _LINUX_HAS_GPU=true
+            break
+        fi
+    done
+fi
 
-if [ "$_HOST_SYSTEM" = "Darwin" ]; then
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-elif [ "$_HOST_SYSTEM" = "Linux" ] \
+if [ "$_HOST_SYSTEM" = "Linux" ] \
         && [ "$_HOST_MACHINE" = "x86_64" ] \
         && [ "$_LINUX_HAS_GPU" = false ]; then
     _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
 elif [ "$_HOST_SYSTEM" = "Linux" ] \
         && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; } \
         && [ "$_LINUX_HAS_GPU" = false ]; then
-    # Linux ARM64 (Ampere Altra, Raspberry Pi 5, GitHub `ubuntu-24.04-arm`,
-    # CPU-only Jetson rescue mode, ...). unslothai/llama.cpp only ships
-    # the Linux CUDA bundles, so without this branch the prebuilt
-    # resolver returns 0 attempts on every release and the installer
-    # falls all the way back to a source build. Upstream ggml-org ships
+    # CPU-only Linux ARM64 (Ampere Altra, Raspberry Pi 5, GitHub
+    # `ubuntu-24.04-arm`, CPU-only Jetson rescue mode, ...). The fork ships no
+    # arm64 CPU bundle, so without this branch the prebuilt resolver returns 0
+    # attempts and the installer falls back to a source build. ggml-org ships
     # llama-bNNNN-bin-ubuntu-arm64.tar.gz from at least b9072 onward.
     _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
 else
+    # GPU Linux (x64 CUDA/ROCm, arm64 CUDA), Windows (CUDA/ROCm), and macOS.
     _HELPER_RELEASE_REPO="unslothai/llama.cpp"
 fi
 unset _GPU_TOOL
@@ -856,7 +1086,6 @@ else
         --install-dir "$LLAMA_CPP_DIR"
         --llama-tag "$_REQUESTED_LLAMA_TAG"
         --published-repo "$_HELPER_RELEASE_REPO"
-        --simple-policy
     )
     if [ -n "${UNSLOTH_LLAMA_RELEASE_TAG:-}" ]; then
         _PREBUILT_CMD+=(--published-release-tag "$UNSLOTH_LLAMA_RELEASE_TAG")
@@ -1137,23 +1366,35 @@ else
 
             GPU_BACKEND=""
             NVCC_PATH=""
-            if command -v nvcc &>/dev/null; then
-                NVCC_PATH="$(command -v nvcc)"
-                GPU_BACKEND="cuda"
-            elif [ -x /usr/local/cuda/bin/nvcc ]; then
-                NVCC_PATH="/usr/local/cuda/bin/nvcc"
-                export PATH="/usr/local/cuda/bin:$PATH"
-                GPU_BACKEND="cuda"
-            elif ls /usr/local/cuda-*/bin/nvcc &>/dev/null 2>&1; then
-                # Pick the newest cuda-XX.X directory
-                NVCC_PATH="$(ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V | tail -1)"
-                export PATH="$(dirname "$NVCC_PATH"):$PATH"
-                GPU_BACKEND="cuda"
+            # Gate the CUDA toolkit search on an actually-usable NVIDIA GPU
+            # (_setup_nvidia_usable, computed in the GPU summary block above;
+            # already false when hidden via CUDA_VISIBLE_DEVICES=""/-1).
+            # A CUDA toolkit alone (CPU-only build container, leftover packages)
+            # is not proof of a GPU: building with -DGGML_CUDA=ON there yields a
+            # binary that fails at runtime, so fall through to the CPU build.
+            if [ "$_setup_nvidia_usable" = true ]; then
+                if command -v nvcc &>/dev/null; then
+                    NVCC_PATH="$(command -v nvcc)"
+                    GPU_BACKEND="cuda"
+                elif [ -x /usr/local/cuda/bin/nvcc ]; then
+                    NVCC_PATH="/usr/local/cuda/bin/nvcc"
+                    export PATH="/usr/local/cuda/bin:$PATH"
+                    GPU_BACKEND="cuda"
+                elif ls /usr/local/cuda-*/bin/nvcc &>/dev/null 2>&1; then
+                    # Pick the newest cuda-XX.X directory
+                    NVCC_PATH="$(ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V | tail -1)"
+                    export PATH="$(dirname "$NVCC_PATH"):$PATH"
+                    GPU_BACKEND="cuda"
+                fi
             fi
 
-            # Check for ROCm (AMD) only if CUDA was not already selected
+            # Check for ROCm (AMD) only if CUDA was not already selected, and
+            # only when an AMD GPU was actually detected (_setup_amd_detected).
+            # hipcc presence alone (HIP SDK, no GPU) must not select a HIP build.
+            # NVIDIA-usable hosts never build HIP (defense in depth: the AMD
+            # probes above are already skipped when NVIDIA is usable).
             ROCM_HIPCC=""
-            if [ -z "$GPU_BACKEND" ]; then
+            if [ -z "$GPU_BACKEND" ] && [ "$_setup_nvidia_usable" != true ] && [ "$_setup_amd_detected" = true ]; then
                 if command -v hipcc &>/dev/null; then
                     ROCM_HIPCC="$(command -v hipcc)"
                     GPU_BACKEND="rocm"
@@ -1189,6 +1430,7 @@ else
                     GPU_BACKEND=""
                     _BUILD_DESC="building (CPU, CUDA toolkit < 12.4)"
                 else
+<<<<<<< ours
                     CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
 
                     # glibc >= 2.41 + CUDA < 13.3: rsqrt/rsqrtf header clash fails every .cu
@@ -1222,22 +1464,60 @@ else
                                 esac
                             fi
                         done <<< "$_raw_caps"
+=======
+                    _DRIVER_MAX_CUDA="$(_cuda_driver_max_version)"
+                    _CUDA_TOOLKIT_ALLOWED=true
+                    if [ -n "$_NVCC_VER" ] && [ -n "$_DRIVER_MAX_CUDA" ] && \
+                       _cuda_toolkit_major_gt_driver "$_NVCC_VER" "$_DRIVER_MAX_CUDA"; then
+                        _BLOCKED_NVCC_VER="$_NVCC_VER"
+                        if _ALT_NVCC_CHECK="$(_cuda_find_compatible_nvcc_for_driver "$_DRIVER_MAX_CUDA" "$NVCC_PATH")"; then
+                            NVCC_PATH="$(printf '%s\n' "$_ALT_NVCC_CHECK" | sed -n '1p')"
+                            _NVCC_VER="$(printf '%s\n' "$_ALT_NVCC_CHECK" | sed -n '2p')"
+                            GPU_BACKEND="cuda"
+                            export PATH="$(dirname "$NVCC_PATH"):$PATH"
+                            substep "CUDA Toolkit $_BLOCKED_NVCC_VER is a major-version mismatch with driver CUDA $_DRIVER_MAX_CUDA; using compatible CUDA Toolkit $_NVCC_VER at $NVCC_PATH." "$C_WARN"
+                        else
+                            _print_cuda_driver_toolkit_mismatch "$_NVCC_VER" "$_DRIVER_MAX_CUDA"
+                            substep "falling back to CPU llama.cpp build for this run." "$C_WARN"
+                            NVCC_PATH=""
+                            GPU_BACKEND=""
+                            _BUILD_DESC="building (CPU, CUDA toolkit major > driver)"
+                            _CUDA_TOOLKIT_ALLOWED=false
+                        fi
+>>>>>>> theirs
                     fi
 
-                    if [ -n "$CUDA_ARCHS" ]; then
-                        CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
-                        _BUILD_DESC="building (CUDA, sm_${CUDA_ARCHS//;/+sm_})"
-                    else
-                        _BUILD_DESC="building (CUDA)"
+                    if [ "$_CUDA_TOOLKIT_ALLOWED" = true ]; then
+                        CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
+
+                        CUDA_ARCHS=""
+                        if command -v nvidia-smi &>/dev/null; then
+                            _raw_caps=$(_setup_run_smi nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
+                            while IFS= read -r _cap; do
+                                _cap=$(echo "$_cap" | tr -d '[:space:]')
+                                if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+                                    _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
+                                    case ";$CUDA_ARCHS;" in
+                                        *";$_arch;"*) ;;
+                                        *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
+                                    esac
+                                fi
+                            done <<< "$_raw_caps"
+                        fi
+
+                        if [ -n "$CUDA_ARCHS" ]; then
+                            CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+                            _BUILD_DESC="building (CUDA, sm_${CUDA_ARCHS//;/+sm_})"
+                        else
+                            _BUILD_DESC="building (CUDA)"
+                        fi
+
+                        CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
+
+                        # Allow a host gcc/clang newer than nvcc's whitelist (else a fresh
+                        # toolkit aborts with "unsupported GNU version"); via env to avoid word-splitting.
+                        export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                     fi
-
-                    CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
-
-                    # Accept a host gcc/clang newer than nvcc's whitelist; a fresh
-                    # toolkit (e.g. CUDA 13.3) otherwise aborts with "#error --
-                    # unsupported GNU version". Via env, not CMAKE_ARGS, to avoid
-                    # word-splitting.
-                    export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                 fi
             elif [ "$GPU_BACKEND" = "rocm" ]; then
                 # Resolve hipcc symlinks to find the real ROCm root
@@ -1318,7 +1598,7 @@ else
                     CMAKE_ARGS="$CMAKE_ARGS -DGPU_TARGETS=${GPU_TARGETS}"
                     _BUILD_DESC="building (ROCm, ${GPU_TARGETS//;/+})"
                 fi
-            elif [ -d /usr/local/cuda ] || nvidia-smi &>/dev/null; then
+            elif [ -d /usr/local/cuda ] || _setup_run_smi nvidia-smi &>/dev/null; then
                 _BUILD_DESC="building (CPU, CUDA driver found but nvcc missing)"
             elif [ -d /opt/rocm ] || command -v rocm-smi &>/dev/null; then
                 _BUILD_DESC="building (CPU, ROCm driver found but hipcc missing)"
@@ -1495,7 +1775,6 @@ if [ "$_LLAMA_CPP_DEGRADED" = true ] \
         --install-dir "$LLAMA_CPP_DIR"
         --llama-tag "$_REQUESTED_LLAMA_TAG"
         --published-repo "ggml-org/llama.cpp"
-        --simple-policy
         --cpu-fallback
     )
     # Trust the installer's exit code: it validates the server before exiting 0,
diff --git a/tests/sh/test_get_torch_index_url.sh b/tests/sh/test_get_torch_index_url.sh
index e20fd1ca86..3dece1f69a 100755
--- a/tests/sh/test_get_torch_index_url.sh
+++ b/tests/sh/test_get_torch_index_url.sh
@@ -13,6 +13,10 @@ FAIL=0
 _FUNC_FILE=$(mktemp)
 _FAKE_SMI_DIR=$(mktemp -d)
 {
+    sed -n '/^_run_bounded()/,/^}/p' "$INSTALL_SH"
+    echo ""
+    sed -n '/^_cvd_hides_nvidia()/,/^}/p' "$INSTALL_SH"
+    echo ""
     sed -n '/^_has_amd_rocm_gpu()/,/^}/p' "$INSTALL_SH"
     echo ""
     sed -n '/^_has_usable_nvidia_gpu()/,/^}/p' "$INSTALL_SH"
@@ -107,7 +111,7 @@ MOCK
 # Build a minimal tools directory with symlinks to essential commands
 # (uname, grep, head, etc.) but WITHOUT nvidia-smi or amd-smi.
 _TOOLS_DIR=$(mktemp -d)
-for _cmd in uname grep sed head sh bash cat awk printf; do
+for _cmd in uname grep sed head sh bash cat awk printf tr; do
     _real=$(command -v "$_cmd" 2>/dev/null || true)
     [ -n "$_real" ] && ln -sf "$_real" "$_TOOLS_DIR/$_cmd"
 done
@@ -116,12 +120,19 @@ done
 # $1 = directory with mock nvidia-smi (prepended to PATH), or "none" for no-GPU test
 run_func() {
     _mock_dir="$1"
+    # Default: strip CUDA_VISIBLE_DEVICES so the host environment cannot leak
+    # in; a second argument sets it explicitly (hidden-GPU scenarios).
+    if [ "$#" -ge 2 ]; then
+        _cvd_setup="export CUDA_VISIBLE_DEVICES='$2'"
+    else
+        _cvd_setup="unset CUDA_VISIBLE_DEVICES"
+    fi
     if [ "$_mock_dir" = "none" ]; then
         # Minimal PATH with only basic tools, no nvidia-smi anywhere
-        PATH="$_TOOLS_DIR" bash -c ". '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
+        PATH="$_TOOLS_DIR" bash -c "$_cvd_setup; . '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
     else
         # Put mock nvidia-smi dir first, then basic tools
-        PATH="$_mock_dir:$_TOOLS_DIR" bash -c ". '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
+        PATH="$_mock_dir:$_TOOLS_DIR" bash -c "$_cvd_setup; . '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
     fi
 }
 
@@ -332,6 +343,40 @@ _result=$(run_func "$_dir")
 assert_eq "CUDA Version 13.7 -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
 rm -rf "$_dir"
 
+# 34) CUDA_VISIBLE_DEVICES="" hides the NVIDIA GPU -> cpu (no AMD present)
+_dir=$(make_mock_smi "12.8")
+_result=$(run_func "$_dir" "")
+assert_eq "CVD='' hides NVIDIA -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
+rm -rf "$_dir"
+
+# 35) CUDA_VISIBLE_DEVICES=-1 hides the NVIDIA GPU -> cpu (no AMD present)
+_dir=$(make_mock_smi "12.8")
+_result=$(run_func "$_dir" "-1")
+assert_eq "CVD=-1 hides NVIDIA -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
+rm -rf "$_dir"
+
+# 36) Mixed AMD+NVIDIA host with NVIDIA hidden -> ROCm route is restored
+_cuda_dir=$(make_mock_smi "12.6")
+_amd_dir=$(make_mock_amd_smi "6.4")
+_combined_dir=$(mktemp -d)
+ln -sf "$_cuda_dir/nvidia-smi" "$_combined_dir/nvidia-smi"
+ln -sf "$_amd_dir/amd-smi" "$_combined_dir/amd-smi"
+_result=$(run_func "$_combined_dir" "-1")
+assert_eq "CUDA+ROCm with CVD=-1 -> rocm6.4" "https://download.pytorch.org/whl/rocm6.4" "$_result"
+rm -rf "$_cuda_dir" "$_amd_dir" "$_combined_dir"
+
+# 37) CUDA_VISIBLE_DEVICES=0 (a visible device) must NOT hide the GPU
+_dir=$(make_mock_smi "12.8")
+_result=$(run_func "$_dir" "0")
+assert_eq "CVD=0 keeps NVIDIA -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
+rm -rf "$_dir"
+
+# 38) Whitespace-padded "-1" still hides the GPU
+_dir=$(make_mock_smi "12.8")
+_result=$(run_func "$_dir" " -1 ")
+assert_eq "CVD=' -1 ' hides NVIDIA -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
+rm -rf "$_dir"
+
 rm -f "$_FUNC_FILE"
 rm -rf "$_FAKE_SMI_DIR"
 rm -rf "$_TOOLS_DIR"
diff --git a/tests/studio/install/test_cuda_repair.py b/tests/studio/install/test_cuda_repair.py
new file mode 100644
index 0000000000..83c2d962e8
--- /dev/null
+++ b/tests/studio/install/test_cuda_repair.py
@@ -0,0 +1,248 @@
+"""Tests for CUDA torch repair on poisoned NVIDIA venvs.
+
+Verifies _ensure_cuda_torch (studio/install_python_stack.py) reinstalls CUDA
+torch when a venv on an NVIDIA host carries a ROCm torch build (the pre-fix KFD
+gpu_id false positive), without touching healthy CUDA, deliberate CPU wheels,
+ROCm hosts, macOS, or Windows. All tests use mocks -- no GPU required.
+"""
+
+import importlib.util
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+# ── Load module under test (mirrors test_rocm_support.py) ────────────────────
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+
+_STACK_PATH = PACKAGE_ROOT / "studio" / "install_python_stack.py"
+_STACK_SPEC = importlib.util.spec_from_file_location("studio_install_python_stack", _STACK_PATH)
+assert _STACK_SPEC is not None and _STACK_SPEC.loader is not None
+stack_mod = importlib.util.module_from_spec(_STACK_SPEC)
+sys.modules[_STACK_SPEC.name] = stack_mod
+_STACK_SPEC.loader.exec_module(stack_mod)
+
+_ensure_cuda_torch = stack_mod._ensure_cuda_torch
+_detect_cuda_torch_index_url = stack_mod._detect_cuda_torch_index_url
+
+
+# ── Helpers ──────────────────────────────────────────────────────────────────
+
+
+def _make_run(
+    torch_state = "hip",
+    cuda_version = "12.8",
+    torch_rc = 0,
+    smi_rc = 0,
+):
+    """Build a subprocess.run side_effect.
+
+    The torch-classify probe runs sys.executable and reads bytes stdout; the
+    nvidia-smi version probe runs the smi path with text=True. Distinguish by
+    the executable.
+    """
+
+    def _run(cmd, *args, **kwargs):
+        result = MagicMock()
+        exe = str(cmd[0]) if cmd else ""
+        if exe == sys.executable:
+            result.returncode = torch_rc
+            result.stdout = (torch_state + "\n").encode()
+            return result
+        # nvidia-smi version probe (text = True)
+        result.returncode = smi_rc
+        out = f"CUDA Version: {cuda_version}\n" if cuda_version else "No devices found\n"
+        result.stdout = out if kwargs.get("text") else out.encode()
+        return result
+
+    return _run
+
+
+def _run_cuda_repair(
+    *,
+    backend = "",
+    nvidia = True,
+    torch_state = "hip",
+    cuda_version = "12.8",
+    torch_rc = 0,
+    smi_rc = 0,
+    is_macos = False,
+    is_windows = False,
+    no_torch = False,
+    rocm_marker = False,
+    smi_path = "/usr/bin/nvidia-smi",
+    cvd = None,
+):
+    """Invoke _ensure_cuda_torch under a fully mocked host; return the pip mock.
+
+    cvd controls CUDA_VISIBLE_DEVICES: None removes it from the environment
+    (the host machine may export one), any string sets it explicitly.
+    """
+    env = {}
+    if rocm_marker:
+        env["UNSLOTH_ROCM_TORCH_INSTALLED"] = "1"
+    if cvd is not None:
+        env["CUDA_VISIBLE_DEVICES"] = cvd
+
+    def _which(name, *a, **k):
+        if name == "nvidia-smi":
+            return smi_path
+        return None
+
+    with (
+        patch.object(stack_mod, "_TORCH_BACKEND", backend),
+        patch.object(stack_mod, "IS_MACOS", is_macos),
+        patch.object(stack_mod, "IS_WINDOWS", is_windows),
+        patch.object(stack_mod, "NO_TORCH", no_torch),
+        patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = nvidia),
+        patch.object(stack_mod.shutil, "which", side_effect = _which),
+        patch.object(stack_mod.os.path, "isfile", return_value = bool(smi_path)),
+        patch.object(stack_mod, "pip_install") as mock_pip,
+        patch.object(
+            stack_mod.subprocess,
+            "run",
+            side_effect = _make_run(torch_state, cuda_version, torch_rc, smi_rc),
+        ),
+        patch.dict(stack_mod.os.environ, env, clear = False),
+    ):
+        if not rocm_marker:
+            stack_mod.os.environ.pop("UNSLOTH_ROCM_TORCH_INSTALLED", None)
+        if cvd is None:
+            stack_mod.os.environ.pop("CUDA_VISIBLE_DEVICES", None)
+        _ensure_cuda_torch()
+    return mock_pip
+
+
+def _index_url(mock_pip) -> str:
+    """Return the --index-url value from the recorded pip_install call."""
+    args = [str(a) for a in mock_pip.call_args.args]
+    return args[args.index("--index-url") + 1]
+
+
+# ── Repair fires only on the poisoning signature ─────────────────────────────
+
+
+class TestCudaRepairFires:
+    def test_hip_build_on_nvidia_triggers_repair(self):
+        mock_pip = _run_cuda_repair(torch_state = "hip", cuda_version = "12.8")
+        assert mock_pip.call_count == 1
+        call_args = [str(a) for a in mock_pip.call_args.args]
+        assert "--force-reinstall" in call_args
+        assert "--no-cache-dir" in call_args
+        assert "cu128" in _index_url(mock_pip)
+        assert mock_pip.call_args.kwargs["constrain"] is False
+
+    def test_rocm_in_version_string_triggers_repair(self):
+        # AMD SDK / Radeon wheels may not set torch.version.hip but encode
+        # rocm in __version__; the probe prints "hip" for both.
+        mock_pip = _run_cuda_repair(torch_state = "hip")
+        assert mock_pip.call_count == 1
+
+
+# ── No-op cases ──────────────────────────────────────────────────────────────
+
+
+class TestCudaRepairSkips:
+    def test_healthy_cuda_torch_no_repair(self):
+        mock_pip = _run_cuda_repair(torch_state = "cuda")
+        mock_pip.assert_not_called()
+
+    def test_deliberate_cpu_wheel_no_repair(self):
+        mock_pip = _run_cuda_repair(torch_state = "cpu")
+        mock_pip.assert_not_called()
+
+    def test_backend_rocm_skips(self):
+        mock_pip = _run_cuda_repair(backend = "rocm", torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_backend_cpu_skips(self):
+        mock_pip = _run_cuda_repair(backend = "cpu", torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_unknown_backend_skips(self):
+        mock_pip = _run_cuda_repair(backend = "auto", torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_no_nvidia_gpu_skips(self):
+        mock_pip = _run_cuda_repair(nvidia = False, torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_torch_missing_skips(self):
+        # Non-zero probe exit = torch missing / un-importable.
+        mock_pip = _run_cuda_repair(torch_state = "hip", torch_rc = 1)
+        mock_pip.assert_not_called()
+
+    def test_macos_skips(self):
+        mock_pip = _run_cuda_repair(is_macos = True, torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_windows_skips(self):
+        mock_pip = _run_cuda_repair(is_windows = True, torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_no_torch_mode_skips(self):
+        mock_pip = _run_cuda_repair(no_torch = True, torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_rocm_install_marker_skips(self):
+        mock_pip = _run_cuda_repair(rocm_marker = True, torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_cvd_minus_one_skips(self):
+        # CUDA_VISIBLE_DEVICES=-1 deliberately hides the NVIDIA GPU (mixed
+        # AMD+NVIDIA host running ROCm torch on the AMD card).
+        mock_pip = _run_cuda_repair(cvd = "-1", torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_cvd_empty_skips(self):
+        mock_pip = _run_cuda_repair(cvd = "", torch_state = "hip")
+        mock_pip.assert_not_called()
+
+    def test_cvd_explicit_device_still_repairs(self):
+        mock_pip = _run_cuda_repair(cvd = "0", torch_state = "hip")
+        assert mock_pip.call_count == 1
+
+
+# ── CUDA index ladder ────────────────────────────────────────────────────────
+
+
+class TestCudaIndexResolution:
+    def test_cuda_128_selects_cu128(self):
+        assert "cu128" in _index_url(_run_cuda_repair(cuda_version = "12.8"))
+
+    def test_cuda_130_selects_cu130(self):
+        assert "cu130" in _index_url(_run_cuda_repair(cuda_version = "13.0"))
+
+    def test_cuda_126_selects_cu126(self):
+        assert "cu126" in _index_url(_run_cuda_repair(cuda_version = "12.6"))
+
+    def test_cuda_124_selects_cu124(self):
+        assert "cu124" in _index_url(_run_cuda_repair(cuda_version = "12.4"))
+
+    def test_cuda_118_selects_cu118(self):
+        assert "cu118" in _index_url(_run_cuda_repair(cuda_version = "11.8"))
+
+    def test_unreadable_version_defaults_cu126(self):
+        # nvidia-smi runs but prints no CUDA version line (or fails).
+        mock_pip = _run_cuda_repair(cuda_version = "", smi_rc = 1)
+        assert "cu126" in _index_url(mock_pip)
+
+    def test_proc_fallback_no_smi_defaults_cu126(self):
+        # NVIDIA usable via /proc fallback, nvidia-smi absent entirely.
+        mock_pip = _run_cuda_repair(smi_path = None)
+        assert "cu126" in _index_url(mock_pip)
+
+    def test_detect_index_url_uses_pytorch_base(self):
+        with (
+            patch.object(stack_mod.shutil, "which", return_value = None),
+            patch.object(stack_mod.os.path, "isfile", return_value = False),
+        ):
+            url = _detect_cuda_torch_index_url()
+        assert url == f"{stack_mod._PYTORCH_WHL_BASE}/cu126"
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/tests/studio/install/test_gpu_detection_followups.py b/tests/studio/install/test_gpu_detection_followups.py
new file mode 100644
index 0000000000..983969a4d3
--- /dev/null
+++ b/tests/studio/install/test_gpu_detection_followups.py
@@ -0,0 +1,480 @@
+"""Tests for the GPU-detection follow-ups to PR 6174.
+
+PR 6174 made NVIDIA take precedence and added a /proc/driver/nvidia/gpus
+fallback in install.sh and studio/install_python_stack.py. These tests cover the
+same hardening ported to the llama.cpp prebuilt installer
+(studio/install_llama_prebuilt.py) and the Studio shell setup (studio/setup.sh):
+
+  * detect_host() recognises NVIDIA via /proc/driver/nvidia/gpus when nvidia-smi
+    is unavailable, and skips ROCm probing when NVIDIA is usable.
+  * setup.sh routes through a timeout-bounded NVIDIA probe with a /proc fallback
+    and only selects a CUDA/ROCm source build when the matching GPU is detected.
+
+All tests use mocks or source-level assertions -- no GPU, network, or real
+nvidia-smi/rocminfo invocation.
+"""
+
+import importlib.util
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+
+# Load studio/install_llama_prebuilt.py the same way the sibling suite does.
+_MODULE_PATH = PACKAGE_ROOT / "studio" / "install_llama_prebuilt.py"
+_SPEC = importlib.util.spec_from_file_location(
+    "studio_install_llama_prebuilt_followups", _MODULE_PATH
+)
+assert _SPEC is not None and _SPEC.loader is not None
+prebuilt_mod = importlib.util.module_from_spec(_SPEC)
+sys.modules[_SPEC.name] = prebuilt_mod
+_SPEC.loader.exec_module(prebuilt_mod)
+
+detect_host = prebuilt_mod.detect_host
+_apply_host_overrides = prebuilt_mod._apply_host_overrides
+
+SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
+
+
+def _make_run_capture(rocminfo_stdout: str = ""):
+    """Return a fake run_capture: rocminfo reports rocminfo_stdout, everything
+    else (nvidia-smi, amd-smi) returns empty so only the patched probes matter."""
+
+    def _run_capture(cmd, *args, **kwargs):
+        exe = str(cmd[0]) if cmd else ""
+        result = MagicMock()
+        if exe.endswith("rocminfo"):
+            result.returncode = 0
+            result.stdout = rocminfo_stdout
+        else:
+            result.returncode = 1
+            result.stdout = ""
+        result.stderr = ""
+        return result
+
+    return _run_capture
+
+
+def _run_detect_host(
+    *,
+    machine: str = "x86_64",
+    system: str = "Linux",
+    which_map: dict | None = None,
+    proc_dir_entries: list | None = None,
+    rocminfo_stdout: str = "",
+    env: dict | None = None,
+):
+    """Drive detect_host() against a fully synthetic host."""
+    which_map = which_map or {}
+    proc_dir_entries = proc_dir_entries if proc_dir_entries is not None else []
+
+    real_isdir = prebuilt_mod.os.path.isdir
+    real_listdir = prebuilt_mod.os.listdir
+    proc_path = "/proc/driver/nvidia/gpus"
+
+    def fake_isdir(p):
+        if str(p) == proc_path:
+            return bool(proc_dir_entries)
+        return real_isdir(p)
+
+    def fake_listdir(p):
+        if str(p) == proc_path:
+            if not proc_dir_entries:
+                raise OSError("no such dir")
+            return list(proc_dir_entries)
+        return real_listdir(p)
+
+    patches = [
+        patch.object(prebuilt_mod.platform, "system", return_value = system),
+        patch.object(prebuilt_mod.platform, "machine", return_value = machine),
+        patch.object(prebuilt_mod.platform, "mac_ver", return_value = ("", ("", "", ""), "")),
+        patch.object(prebuilt_mod.shutil, "which", side_effect = lambda n: which_map.get(n)),
+        patch.object(prebuilt_mod, "run_capture", side_effect = _make_run_capture(rocminfo_stdout)),
+        patch.object(prebuilt_mod.os.path, "isdir", side_effect = fake_isdir),
+        patch.object(prebuilt_mod.os, "listdir", side_effect = fake_listdir),
+        patch.object(prebuilt_mod.os, "access", return_value = False),
+        patch.dict(prebuilt_mod.os.environ, env or {}, clear = False),
+    ]
+    for p in patches:
+        p.start()
+    try:
+        # Ensure CUDA_VISIBLE_DEVICES does not leak in from the test host unless
+        # the scenario sets it explicitly.
+        if env is None or "CUDA_VISIBLE_DEVICES" not in env:
+            prebuilt_mod.os.environ.pop("CUDA_VISIBLE_DEVICES", None)
+        return detect_host()
+    finally:
+        for p in patches:
+            p.stop()
+
+
+# ── install_llama_prebuilt.detect_host(): /proc NVIDIA fallback ──────────────
+
+
+class TestDetectHostProcFallback:
+    def test_proc_fallback_marks_physical_nvidia_when_smi_absent(self):
+        """No nvidia-smi, but /proc/driver/nvidia/gpus is populated -> NVIDIA."""
+        host = _run_detect_host(
+            which_map = {},  # nvidia-smi resolves to None
+            proc_dir_entries = ["0000:01:00.0"],
+        )
+        assert host.has_physical_nvidia is True
+
+    def test_proc_fallback_has_usable_nvidia_when_devices_visible(self):
+        """Default CUDA_VISIBLE_DEVICES (unset) -> visible tokens non-empty -> usable."""
+        host = _run_detect_host(
+            which_map = {},
+            proc_dir_entries = ["0000:01:00.0"],
+        )
+        assert host.has_usable_nvidia is True
+
+    def test_proc_fallback_not_usable_when_devices_hidden(self):
+        """CUDA_VISIBLE_DEVICES='' hides all GPUs -> physical yes, usable no."""
+        host = _run_detect_host(
+            which_map = {},
+            proc_dir_entries = ["0000:01:00.0"],
+            env = {"CUDA_VISIBLE_DEVICES": ""},
+        )
+        assert host.has_physical_nvidia is True
+        assert host.has_usable_nvidia is False
+
+    def test_empty_proc_dir_does_not_mark_nvidia(self):
+        """A driver dir that exists but is empty must not assert a GPU."""
+        host = _run_detect_host(which_map = {}, proc_dir_entries = [])
+        assert host.has_physical_nvidia is False
+
+    def test_proc_fallback_is_linux_only(self):
+        """The /proc fallback must not run on Windows (path is Linux-only)."""
+        host = _run_detect_host(
+            system = "Windows",
+            machine = "amd64",
+            which_map = {},
+            proc_dir_entries = ["0000:01:00.0"],
+        )
+        assert host.has_physical_nvidia is False
+
+
+# ── install_llama_prebuilt.detect_host(): NVIDIA precedence over ROCm ────────
+
+
+class TestDetectHostNvidiaPrecedence:
+    def test_rocm_probe_skipped_when_proc_nvidia_present(self):
+        """rocminfo reports gfx1100, but a proc-detected NVIDIA GPU wins."""
+        host = _run_detect_host(
+            which_map = {"rocminfo": "/usr/bin/rocminfo"},
+            proc_dir_entries = ["0000:01:00.0"],
+            rocminfo_stdout = "  Name:                    gfx1100\n",
+        )
+        assert host.has_usable_nvidia is True
+        assert host.has_rocm is False
+
+    def test_rocm_detected_when_no_nvidia(self):
+        """With no NVIDIA signal at all, rocminfo gfx1100 -> has_rocm True."""
+        host = _run_detect_host(
+            which_map = {"rocminfo": "/usr/bin/rocminfo"},
+            proc_dir_entries = [],
+            rocminfo_stdout = "  Name:                    gfx1100\n",
+        )
+        assert host.has_usable_nvidia is False
+        assert host.has_rocm is True
+
+
+# ── _apply_host_overrides: forwarded --rocm-gfx / --has-rocm still win ───────
+
+
+class TestOverridesStillWin:
+    def test_forwarded_gfx_forces_rocm_on_non_nvidia_host(self):
+        host = _run_detect_host(which_map = {}, proc_dir_entries = [])
+        assert host.has_rocm is False
+        overridden = _apply_host_overrides(host, override_rocm_gfx = "gfx1100")
+        assert overridden.has_rocm is True
+        assert overridden.rocm_gfx_target == "gfx1100"
+
+    def test_override_has_rocm_forces_rocm(self):
+        host = _run_detect_host(which_map = {}, proc_dir_entries = [])
+        overridden = _apply_host_overrides(host, override_has_rocm = True)
+        assert overridden.has_rocm is True
+
+    def test_force_cpu_drops_nvidia_attributes(self):
+        host = _run_detect_host(which_map = {}, proc_dir_entries = ["0000:01:00.0"])
+        assert host.has_usable_nvidia is True
+        overridden = _apply_host_overrides(host, force_cpu = True)
+        assert overridden.has_usable_nvidia is False
+        assert overridden.has_physical_nvidia is False
+        assert overridden.has_rocm is False
+
+
+# ── setup.sh source-level guarantees ────────────────────────────────────────
+
+
+class TestSetupShHardening:
+    @pytest.fixture(scope = "class")
+    def setup_src(self) -> str:
+        return SETUP_SH.read_text(encoding = "utf-8")
+
+    def test_has_usable_nvidia_helper_exists(self, setup_src):
+        assert "_setup_has_usable_nvidia_gpu()" in setup_src
+
+    def test_helper_uses_proc_fallback(self, setup_src):
+        start = setup_src.find("_setup_has_usable_nvidia_gpu()")
+        end = setup_src.find("\n}", start)
+        body = setup_src[start:end]
+        assert (
+            "/proc/driver/nvidia/gpus" in body
+        ), "_setup_has_usable_nvidia_gpu must fall back to /proc/driver/nvidia/gpus"
+
+    def test_gpu_summary_uses_helper(self, setup_src):
+        assert "if _setup_has_usable_nvidia_gpu; then" in setup_src
+
+    def test_timeout_wrapper_exists(self, setup_src):
+        start = setup_src.find("_setup_run_smi()")
+        assert start >= 0, "_setup_run_smi timeout wrapper must exist"
+        end = setup_src.find("\n}", start)
+        body = setup_src[start:end]
+        assert "timeout 10" in body
+        assert "command -v timeout" in body
+
+    def test_cuda_source_build_gated_on_usable_nvidia(self, setup_src):
+        """The nvcc source-build search must be gated on _setup_nvidia_usable.
+
+        The hidden-GPU policy (CUDA_VISIBLE_DEVICES=""/-1) lives inside
+        _setup_has_usable_nvidia_gpu, so the gate itself only needs the flag.
+        """
+        anchor = setup_src.find('NVCC_PATH=""\n')
+        assert anchor >= 0
+        window = setup_src[anchor : anchor + 700]
+        assert (
+            'if [ "$_setup_nvidia_usable" = true ]' in window
+        ), "CUDA toolkit search must require a usable NVIDIA GPU, not just nvcc"
+
+    def test_nvidia_helper_honours_hidden_cvd(self, setup_src):
+        """_setup_has_usable_nvidia_gpu must consult the hidden-CVD helper so
+        CUDA_VISIBLE_DEVICES=""/-1 suppresses NVIDIA before the AMD probes are
+        gated (mixed hosts steered to the AMD card keep the ROCm route)."""
+        assert "_setup_cvd_hides_nvidia()" in setup_src
+        start = setup_src.find("_setup_has_usable_nvidia_gpu() {")
+        end = setup_src.find("\n}", start)
+        body = setup_src[start:end]
+        assert "_setup_cvd_hides_nvidia" in body
+
+    def test_rocm_source_build_gated_on_amd_detected(self, setup_src):
+        """The hipcc source-build search must be gated on _setup_amd_detected."""
+        anchor = setup_src.find('ROCM_HIPCC=""')
+        assert anchor >= 0
+        window = setup_src[anchor : anchor + 400]
+        assert (
+            '[ "$_setup_amd_detected" = true ]' in window
+        ), "ROCm toolkit search must require a detected AMD GPU, not just hipcc"
+
+    def test_compute_cap_probe_timeout_wrapped(self, setup_src):
+        assert "_setup_run_smi nvidia-smi --query-gpu=compute_cap" in setup_src
+
+    def test_driver_version_probe_timeout_wrapped(self, setup_src):
+        start = setup_src.find("_cuda_driver_max_version()")
+        end = setup_src.find("\n}", start)
+        body = setup_src[start:end]
+        assert "_setup_run_smi nvidia-smi" in body
+
+
+# TEST: install.sh -- UNSLOTH_TORCH_BACKEND classified on the final path segment
+
+
+class TestBackendExportLeafClassification:
+    """A custom UNSLOTH_PYTORCH_MIRROR whose base path contains "rocm" or
+    "gfx" must not mislabel a cu*/cpu index as ROCm; classification uses the
+    final path segment of TORCH_INDEX_URL only."""
+
+    @pytest.fixture(scope = "class")
+    def install_src(self) -> str:
+        return (PACKAGE_ROOT / "install.sh").read_text(encoding = "utf-8")
+
+    def test_export_block_uses_leaf(self, install_src):
+        anchor = install_src.find("_torch_index_leaf=")
+        assert anchor >= 0, "backend export must classify on the final path segment"
+        window = install_src[anchor : anchor + 500]
+        assert 'export UNSLOTH_TORCH_BACKEND="rocm"' in window
+        assert 'export UNSLOTH_TORCH_BACKEND="cpu"' in window
+        assert 'export UNSLOTH_TORCH_BACKEND="cuda"' in window
+
+    def test_leaf_classification_behaviour(self, tmp_path):
+        import subprocess as sp
+
+        script = tmp_path / "leaf.sh"
+        src = (PACKAGE_ROOT / "install.sh").read_text(encoding = "utf-8")
+        anchor = src.find("_torch_index_leaf=")
+        block = src[anchor : src.find("esac", anchor) + 4]
+        # Drive the extracted block with adversarial mirror URLs.
+        script.write_text(
+            "#!/bin/sh\n"
+            'TORCH_INDEX_URL="$1"\n' + block + "\n"
+            'printf "%s" "$UNSLOTH_TORCH_BACKEND"\n'
+        )
+        cases = {
+            "https://download.pytorch.org/whl/cu128": "cuda",
+            "https://download.pytorch.org/whl/cpu": "cpu",
+            "https://download.pytorch.org/whl/rocm6.4": "rocm",
+            "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.1/": "rocm",
+            "https://repo.amd.com/rocm/whl/gfx1151/": "rocm",
+            "https://mirror.local/rocm-cache/cu128": "cuda",
+            "https://mirror.local/gfx-cache/cpu": "cpu",
+        }
+        for url, expected in cases.items():
+            out = sp.run(
+                ["sh", str(script), url], capture_output = True, text = True, timeout = 30
+            ).stdout.strip()
+            assert out == expected, f"{url} classified as {out!r}, expected {expected!r}"
+
+
+# TEST: CUDA_VISIBLE_DEVICES=""/-1 hides NVIDIA in every usable-GPU helper
+
+
+_STACK_PATH = PACKAGE_ROOT / "studio" / "install_python_stack.py"
+_STACK_SPEC = importlib.util.spec_from_file_location(
+    "studio_install_python_stack_followups", _STACK_PATH
+)
+assert _STACK_SPEC is not None and _STACK_SPEC.loader is not None
+stack_mod = importlib.util.module_from_spec(_STACK_SPEC)
+sys.modules[_STACK_SPEC.name] = stack_mod
+_STACK_SPEC.loader.exec_module(stack_mod)
+
+
+def _stack_nvidia_usable(cvd):
+    """Drive install_python_stack._has_usable_nvidia_gpu with a mocked
+    nvidia-smi that always reports a GPU; cvd = None removes the env var."""
+
+    def fake_run(cmd, *args, **kwargs):
+        result = MagicMock()
+        result.returncode = 0
+        result.stdout = "GPU 0: NVIDIA Fake (UUID: GPU-x)\n"
+        return result
+
+    env = {} if cvd is None else {"CUDA_VISIBLE_DEVICES": cvd}
+    with (
+        patch.object(
+            stack_mod.shutil,
+            "which",
+            side_effect = lambda n: "/usr/bin/nvidia-smi" if n == "nvidia-smi" else None,
+        ),
+        patch.object(stack_mod.subprocess, "run", side_effect = fake_run),
+        patch.dict(stack_mod.os.environ, env, clear = False),
+    ):
+        if cvd is None:
+            stack_mod.os.environ.pop("CUDA_VISIBLE_DEVICES", None)
+        return stack_mod._has_usable_nvidia_gpu()
+
+
+class TestHiddenCvdNotUsable:
+    """CUDA_VISIBLE_DEVICES set to "" or "-1" deliberately hides every NVIDIA
+    device (mixed AMD+NVIDIA hosts steering work to the AMD card). All three
+    _has_usable_nvidia_gpu implementations (install_python_stack.py, install.sh,
+    setup.sh) must report the GPU as not usable so the AMD/CPU routes run,
+    matching install_llama_prebuilt.py's has_usable_nvidia."""
+
+    def test_python_unset_cvd_is_usable(self):
+        assert _stack_nvidia_usable(None) is True
+
+    def test_python_empty_cvd_not_usable(self):
+        assert _stack_nvidia_usable("") is False
+
+    def test_python_minus_one_not_usable(self):
+        assert _stack_nvidia_usable("-1") is False
+
+    def test_python_padded_minus_one_not_usable(self):
+        assert _stack_nvidia_usable(" -1 ") is False
+
+    def test_python_explicit_device_is_usable(self):
+        assert _stack_nvidia_usable("0") is True
+
+    def test_python_device_list_is_usable(self):
+        assert _stack_nvidia_usable("0,1") is True
+
+    def test_hidden_nvidia_restores_rocm_detection(self):
+        """Mixed host, NVIDIA hidden via CVD=-1, rocminfo reports gfx1100:
+        _has_rocm_gpu must proceed past the NVIDIA guard and return True
+        (before this fix the guard ignored CVD and blocked ROCm)."""
+
+        def fake_run(cmd, *args, **kwargs):
+            result = MagicMock()
+            result.returncode = 0
+            exe = str(cmd[0])
+            if exe.endswith("rocminfo"):
+                result.stdout = "  Name:                    gfx1100\n"
+            else:
+                result.stdout = "GPU 0: NVIDIA Fake (UUID: GPU-x)\n"
+            return result
+
+        which_map = {
+            "rocminfo": "/usr/bin/rocminfo",
+            "nvidia-smi": "/usr/bin/nvidia-smi",
+        }
+        with (
+            patch.object(stack_mod.shutil, "which", side_effect = which_map.get),
+            patch.object(stack_mod.subprocess, "run", side_effect = fake_run),
+            patch.dict(stack_mod.os.environ, {"CUDA_VISIBLE_DEVICES": "-1"}, clear = False),
+        ):
+            assert stack_mod._has_rocm_gpu() is True
+
+    @staticmethod
+    def _run_sh_helper(tmp_path, src: str, fn_names: list, cvd):
+        """Extract shell functions, run the usable-GPU one against a fake
+        nvidia-smi, and return "usable"/"not_usable"."""
+        import os as _os
+        import subprocess as sp
+
+        blocks = []
+        for name in fn_names:
+            start = src.find(f"{name}() {{")
+            assert start >= 0, f"{name} missing"
+            end = src.find("\n}", start) + 2
+            blocks.append(src[start:end])
+        fake_bin = tmp_path / "bin"
+        fake_bin.mkdir(exist_ok = True)
+        smi = fake_bin / "nvidia-smi"
+        smi.write_text("#!/bin/sh\necho 'GPU 0: NVIDIA Fake (UUID: GPU-x)'\n")
+        smi.chmod(0o755)
+        script = tmp_path / "probe.sh"
+        script.write_text(
+            "#!/bin/sh\n" + "\n".join(blocks) + "\n"
+            f"if {fn_names[-1]}; then echo usable; else echo not_usable; fi\n"
+        )
+        env = dict(_os.environ)
+        env["PATH"] = f"{fake_bin}:{env['PATH']}"
+        if cvd is None:
+            env.pop("CUDA_VISIBLE_DEVICES", None)
+        else:
+            env["CUDA_VISIBLE_DEVICES"] = cvd
+        return sp.run(
+            ["sh", str(script)], capture_output = True, text = True, timeout = 30, env = env
+        ).stdout.strip()
+
+    @pytest.mark.parametrize(
+        "cvd, expected",
+        [(None, "usable"), ("", "not_usable"), ("-1", "not_usable"), ("0", "usable")],
+    )
+    def test_install_sh_helper_cvd(self, tmp_path, cvd, expected):
+        src = (PACKAGE_ROOT / "install.sh").read_text(encoding = "utf-8")
+        out = self._run_sh_helper(
+            tmp_path,
+            src,
+            ["_run_bounded", "_cvd_hides_nvidia", "_has_usable_nvidia_gpu"],
+            cvd,
+        )
+        assert out == expected
+
+    @pytest.mark.parametrize(
+        "cvd, expected",
+        [(None, "usable"), ("", "not_usable"), ("-1", "not_usable"), ("0", "usable")],
+    )
+    def test_setup_sh_helper_cvd(self, tmp_path, cvd, expected):
+        src = SETUP_SH.read_text(encoding = "utf-8")
+        out = self._run_sh_helper(
+            tmp_path,
+            src,
+            ["_setup_run_smi", "_setup_cvd_hides_nvidia", "_setup_has_usable_nvidia_gpu"],
+            cvd,
+        )
+        assert out == expected
diff --git a/tests/studio/install/test_hf_auth.py b/tests/studio/install/test_hf_auth.py
new file mode 100644
index 0000000000..7c3296d4fd
--- /dev/null
+++ b/tests/studio/install/test_hf_auth.py
@@ -0,0 +1,114 @@
+"""Tests for Hugging Face auth on the llama.cpp prebuilt installer's fetches.
+
+Anonymous huggingface.co downloads (tiny GGUF validation model) share a
+per-IP rate limit that CI fleets exhaust (HTTP 429), forcing the prebuilt
+path into a source build. auth_headers now sends HF_TOKEN to huggingface.co
+hosts, and a redirect handler strips Authorization when the download is
+redirected to a different host (CDN signed URLs). All tests are offline.
+"""
+
+import importlib.util
+import sys
+import urllib.request
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+
+_MODULE_PATH = PACKAGE_ROOT / "studio" / "install_llama_prebuilt.py"
+_SPEC = importlib.util.spec_from_file_location(
+    "studio_install_llama_prebuilt_hf_auth", _MODULE_PATH
+)
+assert _SPEC is not None and _SPEC.loader is not None
+mod = importlib.util.module_from_spec(_SPEC)
+sys.modules[_SPEC.name] = mod
+_SPEC.loader.exec_module(mod)
+
+_TOKEN_VARS = ("GH_TOKEN", "GITHUB_TOKEN", "HF_TOKEN", "HUGGING_FACE_HUB_TOKEN")
+HF_URL = "https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf"
+GH_URL = "https://api.github.com/repos/unslothai/llama.cpp/releases"
+
+
+def _headers(url, env):
+    """auth_headers under a fully controlled token environment."""
+    with patch.dict(mod.os.environ, env, clear = False):
+        for var in _TOKEN_VARS:
+            if var not in env:
+                mod.os.environ.pop(var, None)
+        return mod.auth_headers(url)
+
+
+class TestAuthHeaderRouting:
+    def test_hf_token_sent_to_huggingface(self):
+        headers = _headers(HF_URL, {"HF_TOKEN": "hf_x"})
+        assert headers.get("Authorization") == "Bearer hf_x"
+
+    def test_hub_token_fallback(self):
+        headers = _headers(HF_URL, {"HUGGING_FACE_HUB_TOKEN": "hf_y"})
+        assert headers.get("Authorization") == "Bearer hf_y"
+
+    def test_hf_token_not_sent_to_github(self):
+        headers = _headers(GH_URL, {"HF_TOKEN": "hf_x"})
+        assert "Authorization" not in headers
+
+    def test_hf_token_not_sent_to_other_hosts(self):
+        headers = _headers("https://cdn-lfs.huggingface.co/x", {"HF_TOKEN": "hf_x"})
+        assert "Authorization" not in headers
+
+    def test_gh_token_not_sent_to_huggingface(self):
+        headers = _headers(HF_URL, {"GH_TOKEN": "gh_x"})
+        assert "Authorization" not in headers
+
+    def test_gh_token_still_wins_on_github(self):
+        headers = _headers(GH_URL, {"GH_TOKEN": "gh_x", "HF_TOKEN": "hf_x"})
+        assert headers.get("Authorization") == "Bearer gh_x"
+
+    def test_no_tokens_no_auth(self):
+        assert "Authorization" not in _headers(HF_URL, {})
+
+    def test_validation_model_url_is_hf(self):
+        assert mod.should_send_hf_auth(mod.TEST_MODEL_URL) is True
+
+
+class TestCrossHostRedirectStripsAuth:
+    def _redirect(self, newurl):
+        req = urllib.request.Request(HF_URL, headers = {"Authorization": "Bearer hf_x"})
+        handler = mod._CrossHostAuthStrippingRedirectHandler()
+        return handler.redirect_request(req, None, 302, "Found", {}, newurl)
+
+    def test_cross_host_redirect_drops_authorization(self):
+        new_request = self._redirect("https://cdn-lfs.huggingface.co/signed/blob")
+        assert new_request is not None
+        assert "Authorization" not in new_request.headers
+        assert "Authorization" not in new_request.unredirected_hdrs
+
+    def test_same_host_redirect_keeps_authorization(self):
+        new_request = self._redirect("https://huggingface.co/elsewhere/blob")
+        assert new_request is not None
+        assert new_request.headers.get("Authorization") == "Bearer hf_x"
+
+
+class TestDownloadBytesWiring:
+    def test_download_bytes_sends_hf_auth(self):
+        response = MagicMock()
+        response.__enter__ = lambda s: s
+        response.__exit__ = lambda s, *a: False
+        response.headers.get.return_value = None
+        response.read.side_effect = [b"data", b""]
+        with (
+            patch.object(mod._URL_OPENER, "open", return_value = response) as opened,
+            patch.dict(mod.os.environ, {"HF_TOKEN": "hf_x"}, clear = False),
+        ):
+            for var in ("GH_TOKEN", "GITHUB_TOKEN"):
+                mod.os.environ.pop(var, None)
+            data = mod.download_bytes(HF_URL)
+        assert data == b"data"
+        request = opened.call_args.args[0]
+        assert request.headers.get("Authorization") == "Bearer hf_x"
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/tests/studio/install/test_install_llama_prebuilt_logic.py b/tests/studio/install/test_install_llama_prebuilt_logic.py
index de3808469c..000c5a7ace 100644
--- a/tests/studio/install/test_install_llama_prebuilt_logic.py
+++ b/tests/studio/install/test_install_llama_prebuilt_logic.py
@@ -1,3 +1,4 @@
+import errno
 import importlib.util
 import io
 import json
@@ -28,6 +29,7 @@ ApprovedReleaseChecksums = INSTALL_LLAMA_PREBUILT.ApprovedReleaseChecksums
 hydrate_source_tree = INSTALL_LLAMA_PREBUILT.hydrate_source_tree
 validate_prebuilt_choice = INSTALL_LLAMA_PREBUILT.validate_prebuilt_choice
 activate_install_tree = INSTALL_LLAMA_PREBUILT.activate_install_tree
+activate_staged_dir = INSTALL_LLAMA_PREBUILT.activate_staged_dir
 create_install_staging_dir = INSTALL_LLAMA_PREBUILT.create_install_staging_dir
 sha256_file = INSTALL_LLAMA_PREBUILT.sha256_file
 source_archive_logical_name = INSTALL_LLAMA_PREBUILT.source_archive_logical_name
@@ -206,6 +208,110 @@ def test_hydrate_source_tree_extracts_upstream_archive_contents(
     assert not (install_dir / f"llama.cpp-{upstream_tag}").exists()
 
 
+def test_release_asset_download_url():
+    fn = INSTALL_LLAMA_PREBUILT.release_asset_download_url
+    assert fn(
+        "unslothai/llama.cpp", "b9000-mix-abc1234", "llama.cpp-source-commit-deadbeef.tar.gz"
+    ) == (
+        "https://github.com/unslothai/llama.cpp/releases/download/"
+        "b9000-mix-abc1234/llama.cpp-source-commit-deadbeef.tar.gz"
+    )
+    # Any missing component -> None (no asset url, caller falls back to codeload).
+    assert fn(None, "b9000", "x.tar.gz") is None
+    assert fn("unslothai/llama.cpp", None, "x.tar.gz") is None
+    assert fn("unslothai/llama.cpp", "b9000", None) is None
+
+
+def _mk_source_tarball(path: Path, tag: str) -> None:
+    with tarfile.open(path, "w:gz") as archive:
+        add_bytes_to_tar(
+            archive, f"llama.cpp-{tag}/CMakeLists.txt", b"cmake_minimum_required(VERSION 3.14)\n"
+        )
+        add_bytes_to_tar(
+            archive,
+            f"llama.cpp-{tag}/convert_hf_to_gguf.py",
+            b"#!/usr/bin/env python3\nimport gguf\n",
+        )
+        add_bytes_to_tar(archive, f"llama.cpp-{tag}/gguf-py/gguf/__init__.py", b"__all__ = []\n")
+
+
+def test_hydrate_source_tree_prefers_release_asset_for_mix(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    # A mix build's merge commit is in no repo, so the codeload/archive URLs 404.
+    # hydrate must fetch the release asset and never touch codeload.
+    commit = "a" * 40
+    archive_path = tmp_path / "merged-source.tar.gz"
+    _mk_source_tarball(archive_path, f"b9000-mix-{commit[:7]}")
+    asset_url = INSTALL_LLAMA_PREBUILT.release_asset_download_url(
+        "unslothai/llama.cpp", "b9000-mix-abc1234", f"llama.cpp-source-commit-{commit}.tar.gz"
+    )
+    codeload_urls = set(
+        INSTALL_LLAMA_PREBUILT.commit_source_archive_urls("unslothai/llama.cpp", commit)
+    )
+    seen = []
+
+    def fake_download_file(url: str, destination: Path) -> None:
+        seen.append(url)
+        if url in codeload_urls:
+            raise AssertionError("codeload was hit even though the release asset was available")
+        assert url == asset_url
+        destination.write_bytes(archive_path.read_bytes())
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "download_file", fake_download_file)
+
+    install_dir = tmp_path / "install"
+    work_dir = tmp_path / "work"
+    work_dir.mkdir()
+    hydrate_source_tree(
+        commit,
+        install_dir,
+        work_dir,
+        source_repo = "unslothai/llama.cpp",
+        expected_sha256 = sha256_file(archive_path),
+        exact_source = True,
+        asset_url = asset_url,
+    )
+    assert seen == [asset_url]
+    assert (install_dir / "CMakeLists.txt").exists()
+    assert (install_dir / "convert_hf_to_gguf.py").exists()
+
+
+def test_hydrate_source_tree_falls_back_to_codeload_when_asset_missing(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    # If the release asset 404s, fall back to codeload/archive (vanilla path).
+    commit = "b" * 40
+    archive_path = tmp_path / "vanilla-source.tar.gz"
+    _mk_source_tarball(archive_path, f"commit-{commit[:7]}")
+    asset_url = INSTALL_LLAMA_PREBUILT.release_asset_download_url(
+        "unslothai/llama.cpp", "b9000", f"llama.cpp-source-commit-{commit}.tar.gz"
+    )
+    codeload_urls = INSTALL_LLAMA_PREBUILT.commit_source_archive_urls("unslothai/llama.cpp", commit)
+
+    def fake_download_file(url: str, destination: Path) -> None:
+        if url == asset_url:
+            raise RuntimeError("404 Not Found")
+        assert url in codeload_urls
+        destination.write_bytes(archive_path.read_bytes())
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "download_file", fake_download_file)
+
+    install_dir = tmp_path / "install"
+    work_dir = tmp_path / "work"
+    work_dir.mkdir()
+    hydrate_source_tree(
+        commit,
+        install_dir,
+        work_dir,
+        source_repo = "unslothai/llama.cpp",
+        expected_sha256 = sha256_file(archive_path),
+        exact_source = True,
+        asset_url = asset_url,
+    )
+    assert (install_dir / "CMakeLists.txt").exists()
+
+
 def test_validate_prebuilt_choice_creates_repo_shaped_linux_install(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):
@@ -338,336 +444,6 @@ def test_validate_prebuilt_choice_creates_repo_shaped_linux_install(
     assert (install_dir / "BUILD_INFO.txt").exists()
 
 
-def test_simple_linux_direct_release_uses_published_source_checksums_for_branch(
-    monkeypatch: pytest.MonkeyPatch,
-):
-    source_commit = "25b1bc9c2f9aa0a390b968ee1ffd9ff01340a3fe"
-    release = {
-        "tag_name": "llama-prebuilt-master-3a92bc9",
-        "assets": [
-            {
-                "name": "app-master-linux-x64-cuda13-newer.tar.gz",
-                "browser_download_url": "https://example.test/app-master-linux-x64-cuda13-newer.tar.gz",
-            },
-            {
-                "name": "llama-prebuilt-sha256.json",
-                "browser_download_url": "https://example.test/llama-prebuilt-sha256.json",
-            },
-        ],
-    }
-    checksums = ApprovedReleaseChecksums(
-        repo = "unslothai/llama.cpp",
-        release_tag = "llama-prebuilt-master-3a92bc9",
-        upstream_tag = "b9174",
-        source_commit = source_commit,
-        source_repo = "ggml-org/llama.cpp",
-        source_repo_url = "https://github.com/ggml-org/llama.cpp",
-        source_ref_kind = "branch",
-        requested_source_ref = "master",
-        resolved_source_ref = "master",
-        artifacts = {
-            "app-master-linux-x64-cuda13-newer.tar.gz": ApprovedArtifactHash(
-                asset_name = "app-master-linux-x64-cuda13-newer.tar.gz",
-                sha256 = "a" * 64,
-                repo = "unslothai/llama.cpp",
-                kind = "linux-cuda-app",
-            ),
-            INSTALL_LLAMA_PREBUILT.exact_source_archive_logical_name(
-                source_commit
-            ): ApprovedArtifactHash(
-                asset_name = INSTALL_LLAMA_PREBUILT.exact_source_archive_logical_name(source_commit),
-                sha256 = "b" * 64,
-                repo = "ggml-org/llama.cpp",
-                kind = "exact-source",
-            ),
-        },
-    )
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "load_approved_release_checksums",
-        lambda repo, release_tag: checksums,
-    )
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "detected_linux_runtime_lines",
-        lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-    )
-    host = HostInfo(
-        system = "Linux",
-        machine = "x86_64",
-        is_windows = False,
-        is_linux = True,
-        is_macos = False,
-        is_x86_64 = True,
-        is_arm64 = False,
-        nvidia_smi = None,
-        driver_cuda_version = (13, 1),
-        compute_caps = ["100"],
-        visible_cuda_devices = None,
-        has_physical_nvidia = True,
-        has_usable_nvidia = True,
-    )
-
-    plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-        release,
-        host,
-        "unslothai/llama.cpp",
-        "latest",
-    )
-
-    assert plan is not None
-    assert plan.llama_tag == "master"
-    assert plan.approved_checksums.upstream_tag == "b9174"
-    assert plan.approved_checksums.source_commit == source_commit
-    assert plan.attempts[0].expected_sha256 == "a" * 64
-    source_repo, source_ref, _source_archive, exact_source = (
-        INSTALL_LLAMA_PREBUILT.preferred_source_archive(plan.approved_checksums, plan.llama_tag)
-    )
-    assert source_repo == "ggml-org/llama.cpp"
-    assert source_ref == source_commit
-    assert exact_source is True
-
-
-def test_simple_linux_direct_release_honors_torch_cudart_preference(
-    monkeypatch: pytest.MonkeyPatch,
-):
-    # Regression: a Blackwell host (sm_120, driver 13.0) with BOTH cudart majors
-    # visible -- a stray cuda13 wheel plus torch's cuda12 -- must install the
-    # cuda12 build that matches the runtime torch, not the newest-major cuda13
-    # build (which loads no GPU and silently falls back to CPU).
-    release = {
-        "tag_name": "b9334",
-        "assets": [
-            {
-                "name": f"app-b9334-linux-x64-{profile}.tar.gz",
-                "browser_download_url": f"https://example.test/app-b9334-linux-x64-{profile}.tar.gz",
-            }
-            for profile in (
-                "cuda12-newer",
-                "cuda12-portable",
-                "cuda13-newer",
-                "cuda13-portable",
-            )
-        ],
-    }
-    # cuda13 detected first (newest-major order); both compatible with driver 13.0.
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "detected_linux_runtime_lines",
-        lambda: (
-            ["cuda13", "cuda12"],
-            {
-                "cuda13": ["/usr/local/lib/python3.13/site-packages/nvidia/cu13/lib"],
-                "cuda12": ["/venv/lib/python3.13/site-packages/nvidia/cuda_runtime/lib"],
-            },
-        ),
-    )
-    host = HostInfo(
-        system = "Linux",
-        machine = "x86_64",
-        is_windows = False,
-        is_linux = True,
-        is_macos = False,
-        is_x86_64 = True,
-        is_arm64 = False,
-        nvidia_smi = "nvidia-smi",
-        driver_cuda_version = (13, 0),
-        compute_caps = ["120"],
-        visible_cuda_devices = None,
-        has_physical_nvidia = True,
-        has_usable_nvidia = True,
-    )
-
-    def first_asset_for_torch(line):
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detect_torch_cuda_runtime_preference",
-            lambda h: INSTALL_LLAMA_PREBUILT.CudaRuntimePreference(
-                runtime_line = line, selection_log = []
-            ),
-        )
-        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-            release, host, "unslothai/llama.cpp", "latest"
-        )
-        return plan.attempts[0]
-
-    # torch reports cuda12 (the cu128 runtime) -> install the cuda12 build.
-    primary = first_asset_for_torch("cuda12")
-    assert primary.name == "app-b9334-linux-x64-cuda12-newer.tar.gz"
-    assert primary.runtime_line == "cuda12"
-
-    # torch unavailable -> unchanged newest-major fallback (documents the residual).
-    assert first_asset_for_torch(None).name == "app-b9334-linux-x64-cuda13-newer.tar.gz"
-
-
-@pytest.mark.parametrize(
-    "mutate, expected_match",
-    [
-        # Missing source_commit.
-        (
-            lambda c: setattr(c, "source_commit", None) or setattr(c, "source_commit_short", None),
-            "exact source provenance",
-        ),
-        # source_commit present, but no exact-source archive hash.
-        (
-            lambda c: c.artifacts.pop(
-                INSTALL_LLAMA_PREBUILT.exact_source_archive_logical_name(c.source_commit),
-                None,
-            ),
-            "exact source provenance",
-        ),
-        # source_commit + exact-source archive present, but no source_repo.
-        (
-            lambda c: setattr(c, "source_repo", None) or setattr(c, "source_repo_url", None),
-            "exact source provenance",
-        ),
-    ],
-    ids = [
-        "missing_source_commit",
-        "missing_exact_source_artifact",
-        "missing_source_repo",
-    ],
-)
-def test_simple_linux_direct_release_rejects_branch_without_exact_source_metadata(
-    monkeypatch: pytest.MonkeyPatch, mutate, expected_match
-):
-    source_commit = "25b1bc9c2f9aa0a390b968ee1ffd9ff01340a3fe"
-    release = {
-        "tag_name": "llama-prebuilt-master-3a92bc9",
-        "assets": [
-            {
-                "name": "app-master-linux-x64-cuda13-newer.tar.gz",
-                "browser_download_url": "https://example.test/app-master-linux-x64-cuda13-newer.tar.gz",
-            },
-            {
-                "name": "llama-prebuilt-sha256.json",
-                "browser_download_url": "https://example.test/llama-prebuilt-sha256.json",
-            },
-        ],
-    }
-    checksums = ApprovedReleaseChecksums(
-        repo = "unslothai/llama.cpp",
-        release_tag = "llama-prebuilt-master-3a92bc9",
-        upstream_tag = "b9174",
-        source_commit = source_commit,
-        source_repo = "ggml-org/llama.cpp",
-        source_repo_url = "https://github.com/ggml-org/llama.cpp",
-        source_ref_kind = "branch",
-        requested_source_ref = "master",
-        resolved_source_ref = "master",
-        artifacts = {
-            "app-master-linux-x64-cuda13-newer.tar.gz": ApprovedArtifactHash(
-                asset_name = "app-master-linux-x64-cuda13-newer.tar.gz",
-                sha256 = "a" * 64,
-                repo = "unslothai/llama.cpp",
-                kind = "linux-cuda-app",
-            ),
-            INSTALL_LLAMA_PREBUILT.exact_source_archive_logical_name(
-                source_commit
-            ): ApprovedArtifactHash(
-                asset_name = INSTALL_LLAMA_PREBUILT.exact_source_archive_logical_name(source_commit),
-                sha256 = "b" * 64,
-                repo = "ggml-org/llama.cpp",
-                kind = "exact-source",
-            ),
-        },
-    )
-    mutate(checksums)
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "load_approved_release_checksums",
-        lambda repo, release_tag: checksums,
-    )
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "detected_linux_runtime_lines",
-        lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-    )
-    host = HostInfo(
-        system = "Linux",
-        machine = "x86_64",
-        is_windows = False,
-        is_linux = True,
-        is_macos = False,
-        is_x86_64 = True,
-        is_arm64 = False,
-        nvidia_smi = None,
-        driver_cuda_version = (13, 1),
-        compute_caps = ["100"],
-        visible_cuda_devices = None,
-        has_physical_nvidia = True,
-        has_usable_nvidia = True,
-    )
-
-    with pytest.raises(PrebuiltFallback, match = expected_match):
-        INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-            release,
-            host,
-            "unslothai/llama.cpp",
-            "latest",
-        )
-
-
-def test_simple_linux_direct_release_keeps_legacy_b_tag_path_without_checksums(
-    monkeypatch: pytest.MonkeyPatch,
-):
-    release = {
-        "tag_name": "b9999",
-        "assets": [
-            {
-                "name": "app-b9999-linux-x64-cuda13-newer.tar.gz",
-                "browser_download_url": "https://example.test/app-b9999-linux-x64-cuda13-newer.tar.gz",
-            },
-            {
-                "name": "llama-prebuilt-sha256.json",
-                "browser_download_url": "https://example.test/llama-prebuilt-sha256.json",
-            },
-        ],
-    }
-
-    def unexpected_checksum_load(repo: str, release_tag: str):
-        raise AssertionError("legacy b-tag direct releases should not require checksum metadata")
-
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "load_approved_release_checksums",
-        unexpected_checksum_load,
-    )
-    monkeypatch.setattr(
-        INSTALL_LLAMA_PREBUILT,
-        "detected_linux_runtime_lines",
-        lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-    )
-    host = HostInfo(
-        system = "Linux",
-        machine = "x86_64",
-        is_windows = False,
-        is_linux = True,
-        is_macos = False,
-        is_x86_64 = True,
-        is_arm64 = False,
-        nvidia_smi = None,
-        driver_cuda_version = (13, 1),
-        compute_caps = ["100"],
-        visible_cuda_devices = None,
-        has_physical_nvidia = True,
-        has_usable_nvidia = True,
-    )
-
-    plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-        release,
-        host,
-        "unslothai/llama.cpp",
-        "latest",
-    )
-
-    assert plan is not None
-    assert plan.llama_tag == "b9999"
-    assert plan.release_tag == "b9999"
-    assert plan.approved_checksums.source_commit is None
-    assert plan.attempts[0].expected_sha256 is None
-
-
 def test_validate_prebuilt_choice_creates_repo_shaped_windows_install(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):
@@ -892,6 +668,48 @@ def test_activate_install_tree_cleans_all_paths_when_rollback_restore_fails(
     assert "removing rollback path" in output
 
 
+def test_activate_staged_dir_copies_when_replace_hits_busy_lock(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+):
+    staging_dir = tmp_path / "llama.cpp.staging-test"
+    (staging_dir / "bin").mkdir(parents = True)
+    (staging_dir / "bin" / "ggml-base.dll").write_bytes(b"fake dll")
+    dst = tmp_path / "llama.cpp"
+
+    def denied_replace(src, dst_arg):
+        raise PermissionError(errno.EACCES, "Access is denied", str(src))
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT.os, "replace", denied_replace)
+
+    activate_staged_dir(staging_dir, dst)
+
+    assert (dst / "bin" / "ggml-base.dll").read_bytes() == b"fake dll"
+    assert not staging_dir.exists()
+
+    captured = capsys.readouterr()
+    assert "falling back to file-by-file copy" in captured.out + captured.err
+
+
+def test_activate_staged_dir_reraises_non_busy_errors(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    staging_dir = tmp_path / "llama.cpp.staging-test"
+    staging_dir.mkdir()
+    (staging_dir / "new.txt").write_text("new install\n")
+    dst = tmp_path / "llama.cpp"
+
+    def out_of_space_replace(src, dst_arg):
+        raise OSError(errno.ENOSPC, "No space left on device", str(src))
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT.os, "replace", out_of_space_replace)
+
+    with pytest.raises(OSError, match = "No space left on device"):
+        activate_staged_dir(staging_dir, dst)
+
+    assert not dst.exists()
+    assert (staging_dir / "new.txt").read_text() == "new install\n"
+
+
 def test_binary_env_linux_includes_binary_parent_in_ld_library_path(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):
@@ -993,7 +811,7 @@ def test_install_prebuilt_falls_back_to_older_release_plan(
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [first_plan, second_plan],
@@ -1921,7 +1739,7 @@ def test_install_prebuilt_skips_download_when_existing_install_matches(
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [plan],
@@ -2011,7 +1829,7 @@ def test_install_prebuilt_does_not_skip_unhealthy_existing_install(
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [plan],
@@ -2139,7 +1957,7 @@ def test_install_prebuilt_skips_when_older_release_fallback_matches_existing_ins
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [latest_plan, fallback_plan],
@@ -2286,7 +2104,7 @@ def test_install_prebuilt_skips_same_release_fallback_attempt_when_installed(
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [plan],
@@ -2405,7 +2223,7 @@ def test_install_prebuilt_same_tag_upstream_failure_uses_older_unsloth_release_p
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
         INSTALL_LLAMA_PREBUILT,
-        "resolve_install_release_plans",
+        "resolve_simple_install_release_plans",
         lambda llama_tag, host, published_repo, published_release_tag: (
             "latest",
             [latest_plan, older_plan],
diff --git a/tests/studio/install/test_llama_pr_force_and_source.py b/tests/studio/install/test_llama_pr_force_and_source.py
index 8e2a56ae39..6c780057ad 100644
--- a/tests/studio/install/test_llama_pr_force_and_source.py
+++ b/tests/studio/install/test_llama_pr_force_and_source.py
@@ -463,8 +463,13 @@ class TestSourcePatternsPs1:
         assert "$LlamaSource = $DefaultLlamaSource" in self.content
 
     def test_release_repo_override_removed(self):
+        # No env-based release-repo override; the repo is chosen by GPU detection
+        # (GPU -> fork, CPU -> ggml-org), mirroring setup.sh.
         assert "$HelperReleaseRepo = if ($env:UNSLOTH_LLAMA_RELEASE_REPO)" not in self.content
-        assert '$HelperReleaseRepo = "ggml-org/llama.cpp"' in self.content
+        assert (
+            "$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm) "
+            '{ "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }' in self.content
+        )
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {' in self.content
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index 5b1c4a44e5..9dfa4e0005 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -686,14 +686,26 @@ class TestSourceCodePatterns:
         assert "_RESOLVED_SOURCE_REF_KIND" in content
         assert "_RESOLVED_SOURCE_REF" in content
 
-    def test_setup_sh_prebuilt_install_uses_simple_policy_only(self):
-        """Shell prebuilt path should use the simplified helper install entrypoint."""
+    def test_setup_sh_prebuilt_install_entrypoint(self):
+        """Shell prebuilt path should call the helper install entrypoint, not the
+        old tag-resolution / releases-latest flow."""
         content = SETUP_SH.read_text()
-        assert "--simple-policy" in content
         assert "--resolve-install-tag" not in content
         assert "_HELPER_RELEASE_REPO}/releases/latest" not in content
         assert "ggml-org/llama.cpp/releases/latest" not in content
 
+    def test_setup_sh_routes_to_fork_only_on_usable_gpu(self):
+        """Linux fork-vs-ggml routing must gate NVIDIA on actual GPU usability,
+        not mere nvidia-smi presence, so CPU-only / hidden-GPU hosts (e.g.
+        CUDA_VISIBLE_DEVICES=-1) get the ggml CPU prebuilt instead of a source
+        build. Guards against a silent revert to the old presence-only loop."""
+        content = SETUP_SH.read_text()
+        assert '[ "$_setup_nvidia_usable" = true ]' in content
+        assert "CUDA_VISIBLE_DEVICES" in content
+        # nvidia-smi must NOT be back in the bare presence loop.
+        assert "for _GPU_TOOL in nvidia-smi" not in content
+        assert "for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo" in content
+
     def test_setup_sh_reports_installed_prebuilt_release(self):
         """Shell wrapper should report the installed prebuilt release from metadata."""
         content = SETUP_SH.read_text()
@@ -832,10 +844,10 @@ class TestSourceCodePatterns:
                 if "LlamaCppDir" in context:
                     pytest.fail(f"Found 'git pull' in llama.cpp build section at line {i+1}")
 
-    def test_setup_ps1_prebuilt_install_uses_simple_policy_only(self):
-        """PS1 prebuilt path should use the simplified helper install entrypoint."""
+    def test_setup_ps1_prebuilt_install_entrypoint(self):
+        """PS1 prebuilt path should call the helper install entrypoint, not the
+        old tag-resolution / releases-latest flow."""
         content = SETUP_PS1.read_text()
-        assert '"--simple-policy"' in content
         assert "--resolve-install-tag" not in content
         assert "$HelperReleaseRepo/releases/latest" not in content
         assert "ggml-org/llama.cpp/releases/latest" not in content
diff --git a/tests/studio/install/test_pr5940_followups.py b/tests/studio/install/test_pr5940_followups.py
new file mode 100644
index 0000000000..583532eb74
--- /dev/null
+++ b/tests/studio/install/test_pr5940_followups.py
@@ -0,0 +1,334 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Tests for the AMD-Windows installer follow-ups (PR #5940):
+
+  * the huggingface_hub validation-model fetch + its urllib fallback,
+  * run_capture's Windows-only amd-smi __COMPAT_LAYER=RunAsInvoker injection,
+  * parity of the name->arch table between install.ps1 and setup.ps1.
+
+Mock-only; no AMD hardware or network required.
+"""
+
+import importlib.util
+import re
+import subprocess
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+
+# install_llama_prebuilt.py is self-contained (stdlib + optional filelock), so it
+# loads without the studio backend on sys.path.
+_PREBUILT_PATH = PACKAGE_ROOT / "studio" / "install_llama_prebuilt.py"
+_SPEC = importlib.util.spec_from_file_location(
+    "studio_install_llama_prebuilt_pr5940", _PREBUILT_PATH
+)
+assert _SPEC is not None and _SPEC.loader is not None
+prebuilt = importlib.util.module_from_spec(_SPEC)
+sys.modules[_SPEC.name] = prebuilt
+_SPEC.loader.exec_module(prebuilt)
+
+_INSTALL_PS1 = PACKAGE_ROOT / "install.ps1"
+_SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1"
+_INSTALL_SH = PACKAGE_ROOT / "install.sh"
+
+
+# ── _hf_resolve_url_parts ────────────────────────────────────────────────────
+
+
+def test_hf_resolve_url_parts_valid():
+    assert prebuilt._hf_resolve_url_parts(
+        "https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf"
+    ) == ("ggml-org/models", "main", "tinyllamas/stories260K.gguf")
+
+
+@pytest.mark.parametrize(
+    "url",
+    [
+        "https://github.com/owner/repo/releases/download/x.gguf",  # not huggingface
+        "https://huggingface.co/owner/repo",  # no /resolve/<rev>/
+        "https://huggingface.co/owner/repo/blob/main/x.gguf",  # /blob/ not /resolve/
+        "not even a url",
+    ],
+)
+def test_hf_resolve_url_parts_non_hf_returns_none(url):
+    assert prebuilt._hf_resolve_url_parts(url) is None
+
+
+# ── _fetch_validation_model_bytes ────────────────────────────────────────────
+
+
+def test_fetch_validation_model_prefers_huggingface_hub(tmp_path):
+    model = tmp_path / "stories260K.gguf"
+    model.write_bytes(b"GGUF-via-hf")
+    fake_hf = MagicMock(return_value = str(model))
+    with (
+        patch.object(prebuilt, "validated_validation_model_bytes", side_effect = lambda b: b),
+        patch.dict(sys.modules, {"huggingface_hub": MagicMock(hf_hub_download = fake_hf)}),
+    ):
+        assert prebuilt._fetch_validation_model_bytes() == b"GGUF-via-hf"
+    assert fake_hf.called  # hf path was taken, urllib not needed
+
+
+def test_fetch_validation_model_falls_back_to_urllib_on_hf_failure():
+    fake_hf = MagicMock(side_effect = RuntimeError("hf unreachable"))
+    with (
+        patch.object(prebuilt, "validated_validation_model_bytes", side_effect = lambda b: b),
+        patch.dict(sys.modules, {"huggingface_hub": MagicMock(hf_hub_download = fake_hf)}),
+        patch.object(prebuilt, "download_bytes", return_value = b"GGUF-via-urllib") as dl,
+    ):
+        assert prebuilt._fetch_validation_model_bytes() == b"GGUF-via-urllib"
+    assert dl.called  # fell back to the direct URL download
+
+
+# ── run_capture amd-smi RunAsInvoker injection ───────────────────────────────
+
+
+def _capture_env(command, system):
+    captured = {"env": "sentinel"}
+
+    def fake_run(cmd, **kwargs):
+        captured["env"] = kwargs.get("env")
+        return subprocess.CompletedProcess(cmd, 0, "", "")
+
+    with (
+        patch.object(prebuilt.subprocess, "run", side_effect = fake_run),
+        patch.object(prebuilt.platform, "system", return_value = system),
+    ):
+        prebuilt.run_capture(command)
+    return captured["env"]
+
+
+def test_run_capture_injects_runasinvoker_for_amd_smi_on_windows():
+    env = _capture_env(["amd-smi", "list"], "Windows")
+    assert env is not None and env.get("__COMPAT_LAYER") == "RunAsInvoker"
+
+
+def test_run_capture_injects_for_full_path_amd_smi_exe_on_windows():
+    env = _capture_env(["amd-smi.exe", "version"], "Windows")
+    assert env is not None and env.get("__COMPAT_LAYER") == "RunAsInvoker"
+
+
+def test_run_capture_no_injection_for_non_amd_smi_on_windows():
+    assert _capture_env(["rocminfo"], "Windows") is None
+
+
+def test_run_capture_no_injection_on_linux():
+    # amd-smi does not auto-elevate on Linux, so no env override is applied.
+    assert _capture_env(["amd-smi", "list"], "Linux") is None
+
+
+# ── name->arch table parity (install.ps1 vs setup.ps1) ───────────────────────
+
+
+def _ps_name_arch_rows(text):
+    return re.findall(r'@\{\s*P\s*=\s*"([^"]*)"\s*;\s*A\s*=\s*"(gfx[0-9a-z]+)"', text)
+
+
+def test_ps_name_arch_tables_in_sync():
+    t1 = _ps_name_arch_rows(_INSTALL_PS1.read_text(encoding = "utf-8"))
+    t2 = _ps_name_arch_rows(_SETUP_PS1.read_text(encoding = "utf-8"))
+    assert t1, "no nameArchTable found in install.ps1"
+    assert t1 == t2, f"name->arch tables drifted:\ninstall.ps1={t1}\nsetup.ps1={t2}"
+
+
+def test_rx_7700s_resolves_to_gfx1102_not_gfx1100():
+    rows = _ps_name_arch_rows(_INSTALL_PS1.read_text(encoding = "utf-8"))
+    name = "AMD Radeon RX 7700S"
+    matched = next((arch for pattern, arch in rows if re.search(pattern, name)), None)
+    assert matched == "gfx1102", f"RX 7700S matched {matched!r}, expected gfx1102"
+
+
+def test_radeon_8060s_resolves_to_gfx1151():
+    rows = _ps_name_arch_rows(_INSTALL_PS1.read_text(encoding = "utf-8"))
+    name = "AMD Radeon(TM) 8060S Graphics"
+    matched = next((arch for pattern, arch in rows if re.search(pattern, name)), None)
+    assert matched == "gfx1151"
+
+
+def _sh_name_arch_rows(text, var = "_gpu_disp_gfx"):
+    """Parse a bash `case "$..._mkt" in ... ) <var>="gfxNNNN"` name->arch
+    table into [(substr_tokens, arch), ...] preserving order."""
+    rows = []
+    for line in text.splitlines():
+        m = re.search(var + r'="(gfx[0-9a-z]+)"', line)
+        if not m or '*"' not in line:
+            continue
+        tokens = re.findall(r'\*"([^"]+)"\*', line)
+        if tokens:
+            rows.append((tokens, m.group(1)))
+    return rows
+
+
+def _sh_resolve(rows, name):
+    for tokens, arch in rows:
+        if any(tok in name for tok in tokens):  # bash *"X"* == substring
+            return arch
+    return None
+
+
+def test_install_sh_name_arch_agrees_with_ps_for_strix_and_non_amd():
+    """The bash install.sh name->arch table must agree with the PowerShell
+    source-of-truth for the Strix Halo (gfx1151) vs Strix Point (gfx1150)
+    split, and must never misclassify NVIDIA/Intel as an AMD gfx."""
+    sh_rows = _sh_name_arch_rows(_INSTALL_SH.read_text(encoding = "utf-8"))
+    ps_rows = _ps_name_arch_rows(_INSTALL_PS1.read_text(encoding = "utf-8"))
+    assert sh_rows, "no name->arch case table found in install.sh"
+    cases = {
+        "AMD Radeon(TM) 8060S Graphics": "gfx1151",  # Strix Halo
+        "AMD Ryzen AI Max+ PRO 395 w/ Radeon 8060S": "gfx1151",
+        "AMD Radeon 890M Graphics": "gfx1150",  # Strix Point (NOT gfx1151)
+        "AMD Ryzen AI 9 HX 370 w/ Radeon 890M": "gfx1150",
+        "AMD Radeon RX 7700S": "gfx1102",
+        "NVIDIA GeForce RTX 4090": None,
+        "Intel(R) Arc A770 Graphics": None,
+    }
+    for name, expect in cases.items():
+        sh = _sh_resolve(sh_rows, name)
+        assert sh == expect, f"install.sh: {name!r} -> {sh!r}, expected {expect!r}"
+        if expect is not None:  # cross-check bash agrees with the PowerShell table
+            ps = next((a for p, a in ps_rows if re.search(p, name)), None)
+            assert sh == ps, f"install.sh/install.ps1 drift for {name!r}: {sh!r} vs {ps!r}"
+
+
+def test_setup_sh_name_arch_table_in_sync_with_install_sh():
+    """studio/setup.sh keeps its own copy of the bash name->arch table (over
+    `_setup_gfx`); it must stay row-for-row identical to install.sh's, both in
+    tokens and in match order (order carries the RX 7700S -> gfx1102 rule)."""
+    install_rows = _sh_name_arch_rows(_INSTALL_SH.read_text(encoding = "utf-8"))
+    setup_rows = _sh_name_arch_rows(
+        (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8"),
+        var = "_setup_gfx",
+    )
+    assert setup_rows, "no name->arch case table found in studio/setup.sh"
+    assert install_rows == setup_rows, (
+        "bash name->arch tables drifted:\n"
+        f"install.sh={install_rows}\nstudio/setup.sh={setup_rows}"
+    )
+    # The historical drift this guards against: Strix Point SKUs must be
+    # gfx1150, and the spaceless RX 7700S must match gfx1102 before gfx1100.
+    for name, expect in {
+        "AMD Radeon 890M Graphics": "gfx1150",
+        "AMD Ryzen AI 9 HX 370 w/ Radeon 890M": "gfx1150",
+        "AMD Radeon(TM) 8060S Graphics": "gfx1151",
+        "AMD Radeon RX 7700S": "gfx1102",
+    }.items():
+        got = _sh_resolve(setup_rows, name)
+        assert got == expect, f"setup.sh: {name!r} -> {got!r}, expected {expect!r}"
+
+
+# ── amd-smi gating (DiskPart UAC-prompt avoidance) ───────────────────────────
+# On Windows w/o a HIP SDK, amd-smi elevates and pops a UAC/DiskPart prompt
+# RunAsInvoker can't suppress, so _amd_smi_allowed() skips it by default;
+# HIP-SDK hosts and an explicit opt-in keep it.
+
+
+def _amd_smi_allowed_under(system, hipinfo_present, env):
+    which = (
+        (lambda name: r"C:\hip\bin\hipinfo.exe" if name == "hipinfo" else None)
+        if hipinfo_present
+        else (lambda name: None)
+    )
+    with (
+        patch.object(prebuilt.platform, "system", return_value = system),
+        patch.object(prebuilt.shutil, "which", side_effect = which),
+        patch.dict(prebuilt.os.environ, env, clear = True),
+    ):
+        return prebuilt._amd_smi_allowed()
+
+
+def test_amd_smi_allowed_on_linux_regardless():
+    # Linux amd-smi does not elevate -> always allowed (no regression on Linux).
+    assert _amd_smi_allowed_under("Linux", hipinfo_present = False, env = {}) is True
+
+
+def test_amd_smi_skipped_on_windows_without_hip_sdk():
+    # The DiskPart fix: no HIP SDK + no opt-in -> do not spawn amd-smi.
+    assert _amd_smi_allowed_under("Windows", hipinfo_present = False, env = {}) is False
+
+
+def test_amd_smi_allowed_on_windows_with_hip_sdk():
+    # hipinfo present => amd-smi runs un-elevated, so it is allowed (no regression
+    # for HIP-SDK Windows users, who never saw the prompt).
+    assert _amd_smi_allowed_under("Windows", hipinfo_present = True, env = {}) is True
+
+
+def test_amd_smi_opt_in_forces_on_windows_no_sdk():
+    assert (
+        _amd_smi_allowed_under(
+            "Windows", hipinfo_present = False, env = {"UNSLOTH_ENABLE_AMD_SMI": "1"}
+        )
+        is True
+    )
+
+
+def test_amd_smi_opt_out_overrides_hip_sdk():
+    assert (
+        _amd_smi_allowed_under("Windows", hipinfo_present = True, env = {"UNSLOTH_ENABLE_AMD_SMI": "0"})
+        is False
+    )
+
+
+def test_ps_installers_gate_amd_smi_on_windows():
+    # Both PowerShell installers must gate amd-smi behind HIP-SDK presence + the
+    # UNSLOTH_ENABLE_AMD_SMI opt-in, mirroring _amd_smi_allowed().
+    for ps in (_INSTALL_PS1, _SETUP_PS1):
+        text = ps.read_text(encoding = "utf-8")
+        assert "UNSLOTH_ENABLE_AMD_SMI" in text, f"{ps.name} missing amd-smi opt-in gate"
+        assert "amdSmiAllowed" in text, f"{ps.name} missing amd-smi gate variable"
+
+
+def test_install_python_stack_gates_every_amd_smi_spawn():
+    # Regression for the DiskPart UAC prompt: every function that both names the
+    # `amd-smi` command AND spawns a subprocess must gate it behind
+    # _amd_smi_allowed(). The "ROCm torch missing" probe once spawned `amd-smi
+    # list` ungated on Adrenalin-only hosts; not-spawning is the only fix.
+    import ast
+
+    src = (PACKAGE_ROOT / "studio" / "install_python_stack.py").read_text(encoding = "utf-8")
+    tree = ast.parse(src)
+
+    def _names_amd_smi_command(node):
+        # Exact "amd-smi"/"amd-smi.exe" constant, not a substring in a log message.
+        return any(
+            isinstance(n, ast.Constant)
+            and isinstance(n.value, str)
+            and n.value.lower() in ("amd-smi", "amd-smi.exe")
+            for n in ast.walk(node)
+        )
+
+    def _spawns_subprocess(node):
+        for n in ast.walk(node):
+            if (
+                isinstance(n, ast.Call)
+                and isinstance(n.func, ast.Attribute)
+                and isinstance(n.func.value, ast.Name)
+                and n.func.value.id == "subprocess"
+            ):
+                return True
+        return False
+
+    def _references_gate(node):
+        return any(isinstance(n, ast.Name) and n.id == "_amd_smi_allowed" for n in ast.walk(node))
+
+    offenders = [
+        node.name
+        for node in ast.walk(tree)
+        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))
+        and _names_amd_smi_command(node)
+        and _spawns_subprocess(node)
+        and not _references_gate(node)
+    ]
+    assert not offenders, (
+        "install_python_stack.py spawns amd-smi without an _amd_smi_allowed() "
+        f"gate in: {offenders} -- this pops the Windows UAC/DiskPart prompt on "
+        "Adrenalin-only (no HIP SDK) hosts."
+    )
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])
diff --git a/tests/studio/install/test_probe_timeouts.py b/tests/studio/install/test_probe_timeouts.py
new file mode 100644
index 0000000000..acea0ed34d
--- /dev/null
+++ b/tests/studio/install/test_probe_timeouts.py
@@ -0,0 +1,202 @@
+"""Tests that NVIDIA probes in the installers are bounded by a timeout.
+
+Covers audit findings 5 and 6: a wedged nvidia-smi must not hang the installer,
+and the Windows probe must require a real GPU listing (not just exit code 0).
+
+Source-level assertions verify the guards are present in install.sh / install.ps1
+/ setup.ps1; one behavioral shell test confirms the bash helper actually returns
+within the timeout when nvidia-smi hangs.
+"""
+
+import os
+import shutil
+import stat
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+import pytest
+
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+INSTALL_SH = PACKAGE_ROOT / "install.sh"
+INSTALL_PS1 = PACKAGE_ROOT / "install.ps1"
+SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1"
+
+
+def _extract_sh_function_body(source: str, name: str) -> str:
+    """Return a shell function body from `source` by brace matching."""
+    needle = f"{name}() {{"
+    start = source.find(needle)
+    if start < 0:
+        return ""
+    depth = 0
+    i = start + len(needle) - 1
+    n = len(source)
+    while i < n:
+        ch = source[i]
+        if ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return source[start : i + 1]
+        i += 1
+    return source[start:]
+
+
+# ── install.sh: _run_bounded helper and its use at every nvidia-smi call ──
+
+
+class TestInstallShBoundedProbe:
+    def _src(self) -> str:
+        return INSTALL_SH.read_text(encoding = "utf-8")
+
+    def test_run_bounded_helper_defined(self):
+        body = _extract_sh_function_body(self._src(), "_run_bounded")
+        assert body, "install.sh must define a _run_bounded helper"
+        assert (
+            "command -v timeout" in body
+        ), "_run_bounded must check for the `timeout` binary before using it"
+        assert "timeout 10" in body, "_run_bounded must apply a 10s timeout"
+        # Must fall back to running unbounded when `timeout` is unavailable
+        # (e.g. macOS) so semantics are unchanged there.
+        assert (
+            "else" in body and '"$@"' in body
+        ), "_run_bounded must run the command unbounded when `timeout` is absent"
+
+    def test_nvidia_smi_dash_l_probe_is_bounded(self):
+        body = _extract_sh_function_body(self._src(), "_has_usable_nvidia_gpu")
+        assert body, "install.sh must define _has_usable_nvidia_gpu"
+        # The -L probe must go through the bounded runner, not call nvidia-smi raw.
+        assert (
+            '_run_bounded "$_nvsmi" -L' in body
+        ), "_has_usable_nvidia_gpu must run nvidia-smi -L through _run_bounded"
+        # The /proc fallback from PR 6174 must still be present.
+        assert "/proc/driver/nvidia" in body
+
+    def test_cuda_version_parse_is_bounded(self):
+        body = _extract_sh_function_body(self._src(), "get_torch_index_url")
+        assert body, "install.sh must define get_torch_index_url"
+        assert (
+            "_run_bounded" in body
+        ), "get_torch_index_url CUDA-version parse must run nvidia-smi through _run_bounded"
+        # The locale must be forced without depending on `env` being on PATH.
+        assert "LC_ALL=C" in body
+        # _nvidia_detected gating from PR 6174 must remain.
+        assert "_nvidia_detected" in body
+
+    def test_no_unbounded_nvidia_smi_invocation_remains(self):
+        """Every nvidia-smi *execution* in install.sh goes through _run_bounded.
+
+        `command -v nvidia-smi` and `-x /usr/bin/nvidia-smi` are resolution
+        checks, not executions, and are allowed. An execution looks like
+        `"$_nvsmi" ...` / `$_smi ...` / `nvidia-smi -L`.
+        """
+        body_nvidia = _extract_sh_function_body(self._src(), "_has_usable_nvidia_gpu")
+        body_torch = _extract_sh_function_body(self._src(), "get_torch_index_url")
+        # In _has_usable_nvidia_gpu the only execution of $_nvsmi must be bounded.
+        assert '"$_nvsmi" -L' not in body_nvidia.replace(
+            '_run_bounded "$_nvsmi" -L', ""
+        ), "found an unbounded nvidia-smi -L execution in _has_usable_nvidia_gpu"
+        # In get_torch_index_url the $_smi execution must be bounded.
+        assert (
+            "LC_ALL=C $_smi" not in body_torch
+        ), "found an unbounded LC_ALL=C $_smi execution in get_torch_index_url"
+
+
+# ── install.ps1 / setup.ps1: bounded, GPU-row-validated Windows probe ──
+
+
+class TestPowerShellBoundedProbe:
+    @pytest.mark.parametrize("path", [INSTALL_PS1, SETUP_PS1])
+    def test_bounded_helper_present(self, path):
+        src = path.read_text(encoding = "utf-8")
+        assert (
+            "function Invoke-NvidiaSmiBounded" in src
+        ), f"{path.name} must define Invoke-NvidiaSmiBounded"
+        assert (
+            "WaitForExit($TimeoutSec * 1000)" in src
+        ), f"{path.name} bounded probe must use WaitForExit with a timeout"
+        # Kill + sentinel on timeout, mirroring Invoke-AmdSmiNoElevate.
+        assert (
+            "$proc.Kill()" in src and "124" in src
+        ), f"{path.name} must kill nvidia-smi and signal a timeout exit code"
+
+    @pytest.mark.parametrize("path", [INSTALL_PS1, SETUP_PS1])
+    def test_probe_requires_gpu_row(self, path):
+        src = path.read_text(encoding = "utf-8")
+        assert (
+            "function Test-NvidiaSmiHasGpu" in src
+        ), f"{path.name} must define Test-NvidiaSmiHasGpu"
+        assert "@('-L')" in src, f"{path.name} must probe nvidia-smi with -L"
+        assert (
+            "^GPU\\s+\\d+:" in src
+        ), f"{path.name} must require a 'GPU <n>:' data row, not just exit code 0"
+
+    @pytest.mark.parametrize("path", [INSTALL_PS1, SETUP_PS1])
+    def test_detection_uses_validated_probe(self, path):
+        src = path.read_text(encoding = "utf-8")
+        # The exit-code-only pattern must be gone from the detection block.
+        assert (
+            "& $nvSmiCmd.Source *> $null" not in src
+        ), f"{path.name} must not use the exit-code-only nvidia-smi probe"
+        assert (
+            "Test-NvidiaSmiHasGpu $nvSmiCmd.Source" in src
+        ), f"{path.name} PATH probe must use Test-NvidiaSmiHasGpu"
+        assert (
+            "Test-NvidiaSmiHasGpu $p" in src
+        ), f"{path.name} hardcoded-path fallback must use Test-NvidiaSmiHasGpu"
+
+
+# ── Behavioral: a hanging nvidia-smi must not hang _has_usable_nvidia_gpu ──
+
+
+def _have_timeout() -> bool:
+    return shutil.which("timeout") is not None
+
+
+@pytest.mark.skipif(not _have_timeout(), reason = "`timeout` binary not available")
+def test_has_usable_nvidia_gpu_returns_under_timeout():
+    """Extract _run_bounded + _has_usable_nvidia_gpu, point them at a fake
+    nvidia-smi that sleeps 30s, and assert the probe returns well under that.
+    """
+    src = INSTALL_SH.read_text(encoding = "utf-8")
+    helper = _extract_sh_function_body(src, "_run_bounded")
+    fn = _extract_sh_function_body(src, "_has_usable_nvidia_gpu")
+    assert helper and fn
+
+    workdir = tempfile.mkdtemp(prefix = "pr6174_timeout_", dir = str(PACKAGE_ROOT.parent))
+    try:
+        fake_dir = Path(workdir, "bin")
+        fake_dir.mkdir()
+        fake_smi = fake_dir / "nvidia-smi"
+        fake_smi.write_text("#!/bin/sh\nsleep 30\n")
+        fake_smi.chmod(fake_smi.stat().st_mode | stat.S_IEXEC | stat.S_IXGRP | stat.S_IXOTH)
+
+        # Build a minimal PATH that includes the fake nvidia-smi plus the real
+        # `timeout`/`awk`/`ls` it needs. Use the fake dir first so it wins.
+        real_bins = {Path(shutil.which(c)).parent for c in ("timeout", "awk", "ls", "sh")}
+        path_env = os.pathsep.join([str(fake_dir)] + [str(p) for p in real_bins])
+
+        # Force the /proc fallback off so the result depends only on the probe,
+        # and so a host with real NVIDIA does not mask the timeout behaviour.
+        script = (
+            f"{helper}\n{fn}\n"
+            "if _has_usable_nvidia_gpu; then echo DETECTED; else echo NONE; fi\n"
+        )
+        proc = subprocess.run(
+            ["sh", "-c", script],
+            env = {"PATH": path_env},
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            text = True,
+            timeout = 20,  # generous: the internal timeout is 10s, sleep is 30s
+        )
+        # The probe must have returned (not hung). On this CI host /proc/driver/
+        # nvidia/gpus is absent, so a timed-out smi yields NONE; on a real NVIDIA
+        # host the /proc fallback yields DETECTED. Either way it must not hang.
+        assert proc.stdout.strip() in {"NONE", "DETECTED"}
+    finally:
+        shutil.rmtree(workdir, ignore_errors = True)
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index e0309dbd35..9aa06b6ff7 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -485,6 +485,21 @@ class TestDetectRocmVersion:
                     result = _detect_rocm_version()
                     assert result == (6, 3)
 
+    def test_dpkg_fallback_without_hipconfig(self, tmp_path):
+        """dpkg rocm-core fallback works when amd-smi and hipconfig are absent
+        (regression: a shadowing local re import raised UnboundLocalError)."""
+
+        def which(cmd):
+            return "/usr/bin/dpkg-query" if cmd == "dpkg-query" else None
+
+        mock_result = MagicMock()
+        mock_result.returncode = 0
+        mock_result.stdout = "1:6.3.0-1\n"
+        with patch.dict(os.environ, {"ROCM_PATH": str(tmp_path / "nonexistent")}):
+            with patch("shutil.which", side_effect = which):
+                with patch("subprocess.run", return_value = mock_result):
+                    assert _detect_rocm_version() == (6, 3)
+
     def test_empty_version_file(self, tmp_path):
         """Empty version file should return None."""
         info_dir = tmp_path / ".info"
@@ -550,9 +565,13 @@ class TestEnsureRocmTorch:
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
     def test_no_rocm_skips(self, mock_nvidia, mock_pip):
         """No ROCm toolchain should skip entirely."""
-        with patch("os.path.isdir", return_value = False):
-            with patch("shutil.which", return_value = None):
-                _ensure_rocm_torch()
+        # _detect_windows_gfx_arch pinned to None: on a real AMD test host its
+        # WMI name fallback would otherwise answer and defeat the "no ROCm
+        # anywhere" premise of this test.
+        with patch.object(stack_mod, "_detect_windows_gfx_arch", return_value = None):
+            with patch("os.path.isdir", return_value = False):
+                with patch("shutil.which", return_value = None):
+                    _ensure_rocm_torch()
         mock_pip.assert_not_called()
 
     @patch.object(stack_mod, "pip_install")
@@ -692,10 +711,151 @@ class TestEnsureRocmTorch:
     @patch.object(stack_mod, "_has_rocm_gpu", return_value = False)
     def test_no_gpu_with_rocm_tools_skips(self, mock_gpu, mock_nvidia, mock_pip):
         """ROCm tools present but no actual AMD GPU should skip entirely."""
-        with patch("os.path.isdir", return_value = True):
-            _ensure_rocm_torch()
+        # Pin the Windows arch probe to None: on a real AMD test host the WMI
+        # name fallback would otherwise answer and defeat the "no actual GPU"
+        # premise (the Linux path under test uses _has_rocm_gpu, mocked False).
+        with patch.object(stack_mod, "_detect_windows_gfx_arch", return_value = None):
+            with patch("os.path.isdir", return_value = True):
+                _ensure_rocm_torch()
         mock_pip.assert_not_called()
 
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True)
+    def test_torch_backend_cuda_env_skips_entirely(self, mock_nvidia, mock_gpu, mock_pip):
+        """UNSLOTH_TORCH_BACKEND=cuda must short-circuit before any GPU probe."""
+        with patch.dict(os.environ, {"UNSLOTH_TORCH_BACKEND": "cuda"}):
+            # Reload _TORCH_BACKEND from the patched environment.
+            with patch.object(stack_mod, "_TORCH_BACKEND", "cuda"):
+                _ensure_rocm_torch()
+        mock_pip.assert_not_called()
+
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True)
+    def test_torch_backend_cpu_env_skips_entirely(self, mock_nvidia, mock_gpu, mock_pip):
+        """UNSLOTH_TORCH_BACKEND=cpu must short-circuit before any GPU probe."""
+        with patch.dict(os.environ, {"UNSLOTH_TORCH_BACKEND": "cpu"}):
+            with patch.object(stack_mod, "_TORCH_BACKEND", "cpu"):
+                _ensure_rocm_torch()
+        mock_pip.assert_not_called()
+
+
+# TEST: install_python_stack.py -- _has_rocm_gpu KFD sysfs vendor_id guard
+
+
+class TestHasRocmGpuKfdVendorGuard:
+    """Verify that the KFD sysfs fallback rejects non-AMD (NVIDIA) KFD nodes.
+
+    These tests are source-level: they verify the regex and logic present in
+    the _has_rocm_gpu implementation rather than running the sysfs traversal
+    (which requires Linux path conventions).
+    """
+
+    def _src(self) -> str:
+        """Return the source of _has_rocm_gpu from install_python_stack.py."""
+        import inspect
+        return inspect.getsource(stack_mod._has_rocm_gpu)
+
+    def test_vendor_id_check_present(self):
+        """_has_rocm_gpu sysfs fallback must check vendor_id 4098 (AMD 0x1002)."""
+        src = self._src()
+        assert "vendor_id" in src, (
+            "_has_rocm_gpu KFD sysfs fallback must read the properties file "
+            "to check vendor_id and exclude NVIDIA KFD nodes"
+        )
+        assert "4098" in src, (
+            "_has_rocm_gpu must require AMD vendor_id 4098 (0x1002) in the "
+            "KFD node properties to avoid false positives on NVIDIA systems"
+        )
+
+    def test_vendor_regex_pattern_anchored(self):
+        """The vendor_id regex must use a word boundary to avoid partial matches."""
+        import re as _re
+
+        src = self._src()
+        # The pattern should have a word boundary before and after the number
+        # so "vendor_id 41098" doesn't match "vendor_id 4098".
+        assert (
+            _re.search(r"\\b.*vendor_id.*\\b", src) or "\\bvendor_id" in src
+        ), "_has_rocm_gpu vendor_id check should use word boundary anchors"
+
+    def test_sysfs_fallback_guarded_by_non_win32(self):
+        """KFD sysfs fallback must be Linux-only (guarded by sys.platform != 'win32')."""
+        src = self._src()
+        assert "win32" in src, "_has_rocm_gpu sysfs fallback must be guarded by sys.platform check"
+
+    def test_cpu_node_excluded(self):
+        """gpu_id == '0' must be excluded (CPU topology nodes)."""
+        src = self._src()
+        assert (
+            '!= "0"' in src or "== '0'" in src or "!= '0'" in src or '"0"' in src
+        ), "_has_rocm_gpu must skip gpu_id 0 nodes (CPU nodes)"
+
+    def test_install_sh_has_vendor_check(self):
+        """_has_amd_rocm_gpu in install.sh sysfs fallback must also check vendor_id 4098."""
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("_has_amd_rocm_gpu()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert "vendor_id" in func_body, "_has_amd_rocm_gpu sysfs fallback must check vendor_id"
+        assert "4098" in func_body, "_has_amd_rocm_gpu must require AMD vendor_id 4098 (0x1002)"
+
+    def test_has_rocm_gpu_returns_false_when_nvidia_present(self):
+        """_has_rocm_gpu must return False immediately when _has_usable_nvidia_gpu is True.
+
+        This is the primary guard: even if rocminfo, amd-smi, or KFD sysfs
+        produce a false positive, an NVIDIA GPU always wins.
+        """
+        with patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True):
+            with patch("shutil.which", return_value = "/usr/bin/rocminfo"):
+                # Simulate rocminfo claiming an AMD GPU is present
+                mock_result = MagicMock()
+                mock_result.returncode = 0
+                mock_result.stdout = "Name: gfx1100\n"
+                with patch("subprocess.run", return_value = mock_result):
+                    assert not stack_mod._has_rocm_gpu(), (
+                        "_has_rocm_gpu must return False when NVIDIA GPU is detected, "
+                        "regardless of what rocminfo reports"
+                    )
+
+    def test_install_sh_has_rocm_gpu_nvidia_guard(self):
+        """_has_amd_rocm_gpu in install.sh must call _has_usable_nvidia_gpu and return 1 if true."""
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("_has_amd_rocm_gpu()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert (
+            "_has_usable_nvidia_gpu" in func_body
+        ), "_has_amd_rocm_gpu must call _has_usable_nvidia_gpu to block NVIDIA hosts"
+        assert (
+            "return 1" in func_body
+        ), "_has_amd_rocm_gpu must return 1 (false) when NVIDIA GPU is detected"
+
+    def test_has_usable_nvidia_gpu_proc_fallback_present(self):
+        """`_has_usable_nvidia_gpu` must have a /proc/driver/nvidia fallback."""
+        import inspect
+
+        src = inspect.getsource(stack_mod._has_usable_nvidia_gpu)
+        assert "/proc/driver/nvidia" in src, (
+            "_has_usable_nvidia_gpu must fall back to /proc/driver/nvidia/gpus when "
+            "nvidia-smi subprocess fails, to handle PATH gaps and driver init races"
+        )
+
+    def test_install_sh_has_usable_nvidia_gpu_proc_fallback(self):
+        """_has_usable_nvidia_gpu in install.sh must also have a /proc/driver/nvidia fallback."""
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("_has_usable_nvidia_gpu()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert "/proc/driver/nvidia" in func_body, (
+            "_has_usable_nvidia_gpu in install.sh must fall back to "
+            "/proc/driver/nvidia/gpus when nvidia-smi fails"
+        )
+
 
 # TEST: install_python_stack.py -- _ROCM_TORCH_INDEX mapping
 
@@ -862,7 +1022,14 @@ class TestInstallShStructure:
     """Verify install.sh structural properties without running it."""
 
     def test_no_here_strings(self):
-        """install.sh must not use <<< (not POSIX)."""
+        """install.sh must not use the bash-only `<<<` here-string operator.
+
+        `<<<` inside a quoted literal (e.g. a marker in a printf) is just data,
+        not a here-string, so strip quoted spans first: this still catches a
+        real `cmd <<< word` (outside quotes) without false-positiving on data.
+        """
+        import re
+
         sh_path = PACKAGE_ROOT / "install.sh"
         source = sh_path.read_text(encoding = "utf-8")
         # <<< is bash-only; breaks dash
@@ -870,7 +1037,11 @@ class TestInstallShStructure:
             stripped = line.lstrip()
             if stripped.startswith("#"):
                 continue
-            assert "<<<" not in line, f"install.sh:{i} uses non-POSIX <<< here-string"
+            # Remove quoted string literals so `<<<` inside them is ignored;
+            # a genuine here-string operator lives outside any quotes.
+            unquoted = re.sub(r"'[^']*'", "", line)
+            unquoted = re.sub(r'"[^"]*"', "", unquoted)
+            assert "<<<" not in unquoted, f"install.sh:{i} uses non-POSIX <<< here-string"
 
     def test_rocm_detection_present(self):
         """install.sh should have ROCm detection in get_torch_index_url."""
@@ -893,16 +1064,21 @@ class TestInstallShStructure:
         source = sh_path.read_text(encoding = "utf-8")
         body = _extract_sh_function_body(source, "get_torch_index_url")
         nvidia_call = body.find("_has_usable_nvidia_gpu")
-        no_nvidia_branch = body.find('if [ -z "$_smi" ]')
+        # Gate changed from [ -z "$_smi" ] to [ "$_nvidia_detected" -eq 0 ] to
+        # handle proc-only NVIDIA hosts where nvidia-smi is absent but _has_usable_nvidia_gpu
+        # returns true via /proc/driver/nvidia/gpus.
+        no_nvidia_branch = body.find('if [ "$_nvidia_detected" -eq 0 ]')
+        if no_nvidia_branch < 0:
+            no_nvidia_branch = body.find('if [ -z "$_smi" ]')
         rocm_call = body.find("_has_amd_rocm_gpu")
         assert nvidia_call >= 0, "get_torch_index_url should call _has_usable_nvidia_gpu"
-        assert no_nvidia_branch >= 0, "get_torch_index_url should gate ROCm on no-nvidia-smi"
+        assert no_nvidia_branch >= 0, "get_torch_index_url should gate ROCm on no-nvidia branch"
         assert (
             rocm_call > no_nvidia_branch
-        ), "ROCm detection should sit inside the 'no nvidia-smi' branch"
+        ), "ROCm detection should sit inside the 'no NVIDIA' branch"
         assert (
             nvidia_call < no_nvidia_branch
-        ), "NVIDIA detection should run before the no-nvidia-smi branch"
+        ), "NVIDIA detection should run before the no-NVIDIA branch"
 
     def test_bitsandbytes_amd_install(self):
         """install.sh should install bitsandbytes for AMD when ROCm detected."""
@@ -984,6 +1160,89 @@ class TestInstallShStructure:
         rocm_pos = func_body.find("amd-smi")
         assert darwin_pos < rocm_pos, "macOS check should come before ROCm detection"
 
+    def test_unsloth_torch_backend_exported_after_get_torch_index_url(self):
+        """install.sh must export UNSLOTH_TORCH_BACKEND after TORCH_INDEX_URL is set.
+
+        This lets install_python_stack.py skip ROCm torch operations on CUDA
+        and CPU hosts without re-running GPU detection in a subprocess.
+        """
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        torch_url_pos = source.find("TORCH_INDEX_URL=$(get_torch_index_url)")
+        backend_pos = source.find("UNSLOTH_TORCH_BACKEND")
+        assert backend_pos > 0, "UNSLOTH_TORCH_BACKEND must be set in install.sh"
+        assert (
+            backend_pos > torch_url_pos
+        ), "UNSLOTH_TORCH_BACKEND must be set AFTER TORCH_INDEX_URL is resolved"
+        # Verify all three cases are covered
+        assert '"cuda"' in source[backend_pos : backend_pos + 500]
+        assert '"rocm"' in source[backend_pos : backend_pos + 500]
+        assert '"cpu"' in source[backend_pos : backend_pos + 500]
+        # Must be exported so subprocesses (setup.sh, install_python_stack.py) see it
+        assert "export UNSLOTH_TORCH_BACKEND" in source
+
+    def test_kfd_sysfs_amd_vendor_check_in_has_amd_rocm_gpu(self):
+        """_has_amd_rocm_gpu sysfs fallback must require AMD vendor_id 4098.
+
+        NVIDIA open kernel module (560+) registers KFD nodes with vendor_id
+        4318 (0x10DE). Without the vendor check, _has_amd_rocm_gpu returns 0
+        (true) on NVIDIA-only hosts that have the nvidia-open driver, causing
+        get_torch_index_url to select a ROCm wheel index.
+        """
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("_has_amd_rocm_gpu()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert (
+            "vendor_id" in func_body
+        ), "_has_amd_rocm_gpu sysfs fallback must check vendor_id to exclude NVIDIA KFD nodes"
+        assert (
+            "4098" in func_body
+        ), "_has_amd_rocm_gpu sysfs fallback must require AMD vendor_id 4098 (0x1002)"
+
+    def test_kfd_awk_resets_state_per_file(self):
+        """KFD sysfs awk must reset gpu/amd state per file (FNR==1).
+
+        Without the reset, a Ryzen+NVIDIA host where node 0 is an AMD CPU
+        agent (vendor_id 4098, gpu_id 0) and node 1 is an NVIDIA GPU
+        (gpu_id > 0, vendor_id 4318) can produce a false positive: node 0
+        sets amd=1, node 1 sets gpu=1, and the combined state triggers found=1
+        before vendor_id 4318 is seen on node 1.
+        """
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("_has_amd_rocm_gpu()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert "FNR==1" in func_body, (
+            "_has_amd_rocm_gpu KFD awk must reset state per file with FNR==1 "
+            "to avoid false positives on Ryzen+NVIDIA hosts with multiple KFD nodes"
+        )
+
+    def test_get_torch_index_url_uses_nvidia_detected_flag(self):
+        """get_torch_index_url must track NVIDIA detection independently of _smi.
+
+        When _has_usable_nvidia_gpu returns true via /proc/driver/nvidia fallback
+        but nvidia-smi is not on PATH, _smi stays empty. Without a separate
+        _nvidia_detected flag, the function falls into the AMD/CPU branch even
+        though NVIDIA was confirmed, silently installing CPU wheels instead of CUDA.
+        """
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        func_start = source.find("get_torch_index_url()")
+        func_end = source.find("\n}", func_start)
+        func_body = source[func_start:func_end]
+        assert "_nvidia_detected" in func_body, (
+            "get_torch_index_url must use a _nvidia_detected flag (separate from "
+            "_smi) so that proc-only NVIDIA detection still selects CUDA wheels"
+        )
+        # The AMD/ROCm branch must be gated on _nvidia_detected being 0, not on
+        # _smi being empty.
+        assert (
+            '_nvidia_detected" -eq 0' in func_body or "_nvidia_detected" in func_body
+        ), "get_torch_index_url AMD branch must be skipped when _nvidia_detected=1"
+
 
 # TEST: Live regression on current host (NVIDIA B200 expected)
 
@@ -1202,6 +1461,10 @@ class TestAmdGpuMonitoring:
         ):
             monkeypatch.delenv(var, raising = False)
 
+        # amd-smi is gated off on Windows w/o a HIP SDK; this test mocks it as
+        # available, so opt in so the gate allows it on every platform.
+        monkeypatch.setenv("UNSLOTH_ENABLE_AMD_SMI", "1")
+
         mock_json = json.dumps(
             [
                 {
@@ -1216,8 +1479,12 @@ class TestAmdGpuMonitoring:
         mock_result.returncode = 0
         mock_result.stdout = mock_json
 
-        with patch.object(subprocess, "run", return_value = mock_result):
-            result = amd_mod.get_primary_gpu_utilization()
+        # The premise is "amd-smi exists and answers": the absence guard
+        # which()-checks before spawning, so hosts without a real amd-smi
+        # (Linux CI, driver-only Windows) need which mocked too.
+        with patch.object(amd_mod.shutil, "which", return_value = "/usr/bin/amd-smi"):
+            with patch.object(subprocess, "run", return_value = mock_result):
+                result = amd_mod.get_primary_gpu_utilization()
         assert result["available"] is True
         assert result["gpu_utilization_pct"] == 50.0
         assert result["temperature_c"] == 65.0
@@ -1238,7 +1505,12 @@ class TestAmdGpuMonitoring:
         except Exception:
             pytest.skip("Could not load amd module")
 
-        with patch.object(subprocess, "run", side_effect = OSError("amd-smi not found")):
+        # Opt in so the call reaches subprocess.run (gated off on Windows w/o a
+        # HIP SDK); testing the OSError handling here.
+        with (
+            patch.dict(os.environ, {"UNSLOTH_ENABLE_AMD_SMI": "1"}),
+            patch.object(subprocess, "run", side_effect = OSError("amd-smi not found")),
+        ):
             result = amd_mod.get_primary_gpu_utilization()
         assert result["available"] is False
 
@@ -1258,10 +1530,15 @@ class TestAmdGpuMonitoring:
         except Exception:
             pytest.skip("Could not load amd module")
 
-        with patch.object(
-            subprocess,
-            "run",
-            side_effect = subprocess.TimeoutExpired("amd-smi", 5),
+        # Opt in so the call reaches subprocess.run (gated off on Windows w/o a
+        # HIP SDK); testing the timeout handling here.
+        with (
+            patch.dict(os.environ, {"UNSLOTH_ENABLE_AMD_SMI": "1"}),
+            patch.object(
+                subprocess,
+                "run",
+                side_effect = subprocess.TimeoutExpired("amd-smi", 5),
+            ),
         ):
             result = amd_mod.get_primary_gpu_utilization()
         assert result["available"] is False
@@ -1337,14 +1614,18 @@ class TestApplyGpuIdsRocmFallback:
         func_body = source[func_start : source.find("\ndef ", func_start + 1)]
         assert 'getattr(_torch.version, "hip", None)' in func_body
 
-    def test_apply_gpu_ids_sets_hip_and_rocr_visible_devices(self):
-        """apply_gpu_ids should set both HIP_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES on ROCm."""
+    def test_apply_gpu_ids_sets_hip_but_not_rocr_visible_devices(self):
+        """apply_gpu_ids should set HIP_VISIBLE_DEVICES but leave ROCR_VISIBLE_DEVICES inherited.
+
+        ROCR_VISIBLE_DEVICES uses HSA agent-level indexing, not physical GPU indices.
+        Overwriting it breaks multi-GPU ROCm systems (see issue #6118).
+        """
         hw_path = PACKAGE_ROOT / "studio" / "backend" / "utils" / "hardware" / "hardware.py"
         source = hw_path.read_text(encoding = "utf-8")
         func_start = source.find("def apply_gpu_ids")
         func_body = source[func_start : source.find("\ndef ", func_start + 1)]
         assert 'os.environ["HIP_VISIBLE_DEVICES"] = value' in func_body
-        assert 'os.environ["ROCR_VISIBLE_DEVICES"] = value' in func_body
+        assert 'os.environ["ROCR_VISIBLE_DEVICES"] = value' not in func_body
 
     def test_apply_gpu_ids_rocm_fallback_is_guarded_by_try_except(self):
         """torch import in apply_gpu_ids must be wrapped in try/except so a missing torch never crashes."""
@@ -1508,8 +1789,13 @@ class TestDetectWindowsGfxArch:
     """Verify hipinfo parsing for GPU arch detection on Windows."""
 
     def test_returns_none_when_hipinfo_not_on_path(self):
+        # Also neutralise the venv-hipInfo and WMI-name fallbacks: this test
+        # pins "no probe source available -> None", and the suite may run on a
+        # real AMD host where WMI would legitimately answer.
         with patch("shutil.which", return_value = None):
-            result = stack_mod._detect_windows_gfx_arch()
+            with patch("os.path.isfile", return_value = False):
+                with patch("subprocess.run", side_effect = FileNotFoundError):
+                    result = stack_mod._detect_windows_gfx_arch()
         assert result is None
 
     def test_parses_gcnarchname_from_hipinfo_output(self):
@@ -1531,11 +1817,21 @@ class TestDetectWindowsGfxArch:
         assert result is None
 
     def test_returns_none_when_no_gcnarchname_in_output(self):
+        # hipinfo answers but without a gcnArchName line. Route only the
+        # hipinfo/amd-smi probes to the mock; the WMI fallback must get
+        # nothing (FileNotFoundError) -- otherwise the mocked device name
+        # would legitimately resolve via the name table.
         mock_result = MagicMock()
         mock_result.returncode = 0
-        mock_result.stdout = b"deviceName : Radeon RX 9060 XT\n"
+        mock_result.stdout = b"deviceName : SomeUnknownDevice\n"
+
+        def _run(cmd, **kwargs):
+            if cmd and "powershell" in str(cmd[0]).lower():
+                raise FileNotFoundError(cmd[0])
+            return mock_result
+
         with patch("shutil.which", return_value = "/usr/bin/hipinfo"):
-            with patch("subprocess.run", return_value = mock_result):
+            with patch("subprocess.run", side_effect = _run):
                 result = stack_mod._detect_windows_gfx_arch()
         assert result is None
 
@@ -1558,12 +1854,116 @@ class TestDetectWindowsGfxArch:
         assert result == "gfx1201"
 
 
+# TEST: install_python_stack.py -- GPU-name / WMI fallback (no amd-smi, no hipinfo)
+
+
+class TestGfxArchNameFallback:
+    """amd-smi does not exist on Windows (neither Adrenalin consistently nor
+    the HIP SDK ship a CLI) and driver-only hosts lack hipinfo too. The
+    detection chain must still resolve the arch from the GPU marketing name
+    (WMI), mirroring setup.ps1's $nameArchTable."""
+
+    @pytest.mark.parametrize(
+        "name, expected",
+        [
+            ("AMD Radeon(TM) 8060S Graphics", "gfx1151"),
+            ("AMD Ryzen AI MAX+ 395 w/ Radeon 8060S", "gfx1151"),
+            ("AMD Radeon(TM) 890M", "gfx1150"),
+            ("AMD Ryzen AI 9 HX 370 w/ Radeon 890M", "gfx1150"),
+            ("AMD Radeon RX 9070 XT", "gfx1201"),
+            ("AMD Radeon RX 9070", "gfx1200"),
+            ("AMD Radeon RX 7700S", "gfx1102"),  # (?!S) lookahead must not hit gfx1100
+            ("AMD Radeon RX 7700 XT", "gfx1100"),
+            ("AMD Radeon(TM) 780M", "gfx1103"),
+            ("NVIDIA GeForce RTX 4090", None),
+            ("Microsoft Basic Display Adapter", None),
+            ("", None),
+        ],
+    )
+    def test_name_to_arch_mapping(self, name, expected):
+        assert stack_mod._gfx_arch_from_gpu_name(name) == expected
+
+    def test_wmi_fallback_resolves_arch_without_any_tools(self):
+        """hipinfo absent everywhere + amd-smi absent -> WMI name fallback."""
+        ps_result = MagicMock()
+        ps_result.returncode = 0
+        ps_result.stdout = b"AMD Radeon(TM) 8060S Graphics\r\nMicrosoft Basic Display Adapter\r\n"
+
+        def _run(cmd, **kwargs):
+            if cmd and "powershell.exe" in str(cmd[0]).lower():
+                return ps_result
+            raise FileNotFoundError(cmd[0])
+
+        with patch.dict(os.environ, {}, clear = False):
+            for _v in (
+                "HIP_PATH",
+                "ROCM_PATH",
+                "UNSLOTH_ROCM_GFX_ARCH",
+                "UNSLOTH_ENABLE_AMD_SMI",
+            ):
+                os.environ.pop(_v, None)
+            with patch("shutil.which", return_value = None):
+                with patch("os.path.isfile", return_value = False):
+                    with patch("subprocess.run", side_effect = _run):
+                        result = stack_mod._detect_windows_gfx_arch()
+        assert result == "gfx1151"
+
+    def test_wmi_fallback_returns_none_for_non_amd_hosts(self):
+        ps_result = MagicMock()
+        ps_result.returncode = 0
+        ps_result.stdout = b"NVIDIA GeForce RTX 4090\r\n"
+
+        def _run(cmd, **kwargs):
+            if cmd and "powershell.exe" in str(cmd[0]).lower():
+                return ps_result
+            raise FileNotFoundError(cmd[0])
+
+        with patch.dict(os.environ, {}, clear = False):
+            for _v in ("HIP_PATH", "ROCM_PATH", "UNSLOTH_ROCM_GFX_ARCH"):
+                os.environ.pop(_v, None)
+            with patch("shutil.which", return_value = None):
+                with patch("os.path.isfile", return_value = False):
+                    with patch("subprocess.run", side_effect = _run):
+                        result = stack_mod._detect_windows_gfx_arch()
+        assert result is None
+
+    def test_stack_probes_venv_hipinfo(self):
+        """The venv Scripts dir hipInfo.exe (shipped by AMD torch wheels) must
+        be a probe candidate so `studio update` works on driver-only hosts."""
+        source = _STACK_PATH.read_text(encoding = "utf-8")
+        assert 'os.path.join(os.path.dirname(sys.executable), "hipInfo.exe")' in source
+
+    def test_prebuilt_resolve_exe_probes_venv_dir(self):
+        """install_llama_prebuilt's _resolve_exe must include the venv Scripts
+        candidate for the same driver-only standalone-rerun scenario."""
+        source = _PREBUILT_PATH.read_text(encoding = "utf-8")
+        assert "_venv_candidate" in source
+
+    def test_runtime_monitor_guards_amd_smi_absence(self):
+        """amd.py must which()-check amd-smi before spawning so absence
+        disables the poller in one step (no FileNotFoundError strikes)."""
+        amd_path = PACKAGE_ROOT / "studio" / "backend" / "utils" / "hardware" / "amd.py"
+        source = amd_path.read_text(encoding = "utf-8")
+        assert 'shutil.which("amd-smi") is None' in source
+
+
 # TEST: install_python_stack.py -- _install_bnb_windows_rocm
 
 
 class TestInstallBnbWindowsRocm:
     """Verify AMD Windows BNB wheel install helper."""
 
+    @pytest.fixture(autouse = True)
+    def _isolate_sitecustomize_persistence(self, monkeypatch, request):
+        """Keep helper tests from writing to the active interpreter site-packages."""
+        if request.node.name.startswith("test_persist"):
+            return
+        monkeypatch.setattr(
+            stack_mod,
+            "_persist_bnb_rocm_version",
+            lambda version: True,
+        )
+
     def test_calls_pip_install_try_with_win_amd64_url(self):
         """Should call pip_install_try with the win_amd64 wheel URL via plain pip."""
         with patch.object(stack_mod, "pip_install_try", return_value = True) as mock_pip:
@@ -1617,6 +2017,7 @@ class TestInstallBnbWindowsRocm:
         """BNB_ROCM_VERSION is set from the DLL detected after install."""
         with patch.dict(os.environ, {}, clear = False):
             os.environ.pop("BNB_ROCM_VERSION", None)
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
             with patch.object(stack_mod, "pip_install_try", return_value = True):
                 with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = "72"):
                     stack_mod._install_bnb_windows_rocm()
@@ -1626,6 +2027,7 @@ class TestInstallBnbWindowsRocm:
         """If AMD ships a newer DLL (e.g. rocm713.dll), that version is used."""
         with patch.dict(os.environ, {}, clear = False):
             os.environ.pop("BNB_ROCM_VERSION", None)
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
             with patch.object(stack_mod, "pip_install_try", return_value = True):
                 with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = "713"):
                     stack_mod._install_bnb_windows_rocm()
@@ -1635,6 +2037,7 @@ class TestInstallBnbWindowsRocm:
         """Falls back to '72' when DLL detection returns None."""
         with patch.dict(os.environ, {}, clear = False):
             os.environ.pop("BNB_ROCM_VERSION", None)
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
             with patch.object(stack_mod, "pip_install_try", return_value = True):
                 with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = None):
                     stack_mod._install_bnb_windows_rocm()
@@ -1643,10 +2046,222 @@ class TestInstallBnbWindowsRocm:
     def test_does_not_override_existing_bnb_rocm_version(self):
         """An explicit BNB_ROCM_VERSION in the caller's env must not be clobbered."""
         with patch.dict(os.environ, {"BNB_ROCM_VERSION": "60"}):
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
             with patch.object(stack_mod, "pip_install_try", return_value = True):
                 stack_mod._install_bnb_windows_rocm()
             assert os.environ.get("BNB_ROCM_VERSION") == "60"
 
+    def test_does_not_persist_existing_bnb_rocm_version(self):
+        """A caller override must not become the venv's managed default."""
+        with patch.dict(os.environ, {"BNB_ROCM_VERSION": "60"}):
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
+            with patch.object(stack_mod, "pip_install_try", return_value = True):
+                with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver") as mock_detect:
+                    with patch.object(
+                        stack_mod, "_persist_bnb_rocm_version", return_value = True
+                    ) as mock_persist:
+                        stack_mod._install_bnb_windows_rocm()
+
+            assert os.environ.get("BNB_ROCM_VERSION") == "60"
+            mock_detect.assert_not_called()
+            mock_persist.assert_not_called()
+
+    def test_redetects_when_bnb_rocm_version_came_from_sitecustomize(self):
+        """Persisted defaults should not mask a newer DLL suffix after reinstall."""
+        with patch.dict(
+            os.environ,
+            {
+                "BNB_ROCM_VERSION": "72",
+                stack_mod._BNB_ROCM_VERSION_SOURCE_ENV: (
+                    stack_mod._BNB_ROCM_VERSION_SOURCE_SITECUSTOMIZE
+                ),
+            },
+        ):
+            with patch.object(stack_mod, "pip_install_try", return_value = True):
+                with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = "713"):
+                    with patch.object(
+                        stack_mod, "_persist_bnb_rocm_version", return_value = True
+                    ) as mock_persist:
+                        stack_mod._install_bnb_windows_rocm()
+
+            assert os.environ.get("BNB_ROCM_VERSION") == "713"
+            assert (
+                os.environ.get(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV)
+                == stack_mod._BNB_ROCM_VERSION_SOURCE_DETECTED
+            )
+            mock_persist.assert_called_once_with("713")
+
+    def test_persists_bnb_rocm_version_for_direct_venv_python(self, tmp_path):
+        """BNB_ROCM_VERSION must apply to a fresh Python process in the venv."""
+        site_packages = tmp_path / "site-packages"
+
+        with patch.dict(os.environ, {}, clear = False):
+            os.environ.pop("BNB_ROCM_VERSION", None)
+            os.environ.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
+            with patch.object(stack_mod, "pip_install_try", return_value = True):
+                with patch.object(stack_mod, "_detect_bnb_rocm_dll_ver", return_value = "72"):
+                    with patch.object(
+                        stack_mod.sysconfig, "get_path", return_value = str(site_packages)
+                    ):
+                        stack_mod._install_bnb_windows_rocm()
+
+        sitecustomize = site_packages / "sitecustomize.py"
+        source = sitecustomize.read_text(encoding = "utf-8")
+        assert "BNB_ROCM_VERSION" in source
+        assert stack_mod._BNB_ROCM_VERSION_SOURCE_ENV in source
+        assert "'72'" in source
+
+        probe_env = os.environ.copy()
+        probe_env.pop("BNB_ROCM_VERSION", None)
+        probe_env.pop(stack_mod._BNB_ROCM_VERSION_SOURCE_ENV, None)
+        probe_env["PYTHONPATH"] = str(site_packages)
+        result = subprocess.run(
+            [
+                sys.executable,
+                "-c",
+                (
+                    "import os; "
+                    "print(os.environ.get('BNB_ROCM_VERSION', ''), "
+                    "os.environ.get('UNSLOTH_BNB_ROCM_VERSION_SOURCE', ''))"
+                ),
+            ],
+            env = probe_env,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.PIPE,
+            text = True,
+            check = True,
+        )
+        assert result.stdout.strip() == "72 sitecustomize"
+
+    def test_persist_bnb_rocm_version_replaces_existing_managed_block(self, tmp_path):
+        """Updating sitecustomize.py must not duplicate the managed BNB block."""
+        site_packages = tmp_path / "site-packages"
+        site_packages.mkdir()
+        sitecustomize = site_packages / "sitecustomize.py"
+        sitecustomize.write_text(
+            "EXISTING = True\n"
+            "# BEGIN Unsloth BNB_ROCM_VERSION\n"
+            "import os as _unsloth_os\n"
+            "_unsloth_os.environ.setdefault('BNB_ROCM_VERSION', '72')\n"
+            "# END Unsloth BNB_ROCM_VERSION\n",
+            encoding = "utf-8",
+        )
+
+        with patch.object(stack_mod.sysconfig, "get_path", return_value = str(site_packages)):
+            assert stack_mod._persist_bnb_rocm_version("713") is True
+
+        source = sitecustomize.read_text(encoding = "utf-8")
+        assert source.count("# BEGIN Unsloth BNB_ROCM_VERSION") == 1
+        assert "EXISTING = True" in source
+        assert "'713'" in source
+        assert "'72'" not in source
+
+    def test_persist_bnb_rocm_version_handles_non_utf8_sitecustomize(self, tmp_path):
+        """A legacy non-UTF-8 sitecustomize.py should not abort installation."""
+        site_packages = tmp_path / "site-packages"
+        site_packages.mkdir()
+        sitecustomize = site_packages / "sitecustomize.py"
+        sitecustomize.write_bytes(b"\xff\xfe\x00")
+
+        with patch.object(stack_mod.sysconfig, "get_path", return_value = str(site_packages)):
+            assert stack_mod._persist_bnb_rocm_version("72") is False
+
+    def test_persist_bnb_rocm_version_repairs_truncated_block(self, tmp_path):
+        """A managed block missing its END marker is replaced, not duplicated."""
+        site_packages = tmp_path / "site-packages"
+        site_packages.mkdir()
+        sitecustomize = site_packages / "sitecustomize.py"
+        sitecustomize.write_text(
+            "EXISTING = True\n"
+            "# BEGIN Unsloth BNB_ROCM_VERSION\n"
+            "import os as _unsloth_os\n"
+            "_unsloth_os.environ.setdefault('BNB_ROCM_VERSION', '72')\n",
+            encoding = "utf-8",
+        )
+
+        with patch.object(stack_mod.sysconfig, "get_path", return_value = str(site_packages)):
+            assert stack_mod._persist_bnb_rocm_version("713") is True
+
+        source = sitecustomize.read_text(encoding = "utf-8")
+        assert source.count("# BEGIN Unsloth BNB_ROCM_VERSION") == 1
+        assert source.count("# END Unsloth BNB_ROCM_VERSION") == 1
+        assert "EXISTING = True" in source
+        assert "'713'" in source
+        assert "'72'" not in source
+
+    def test_persist_bnb_rocm_version_dedupes_duplicate_blocks(self, tmp_path):
+        """Multiple managed blocks collapse to one while preserving user content."""
+        site_packages = tmp_path / "site-packages"
+        site_packages.mkdir()
+        sitecustomize = site_packages / "sitecustomize.py"
+        block = (
+            "# BEGIN Unsloth BNB_ROCM_VERSION\n"
+            "import os as _unsloth_os\n"
+            "_unsloth_os.environ.setdefault('BNB_ROCM_VERSION', '72')\n"
+            "# END Unsloth BNB_ROCM_VERSION\n"
+        )
+        sitecustomize.write_text(block + "USER_MID = 1\n" + block, encoding = "utf-8")
+
+        with patch.object(stack_mod.sysconfig, "get_path", return_value = str(site_packages)):
+            assert stack_mod._persist_bnb_rocm_version("713") is True
+
+        source = sitecustomize.read_text(encoding = "utf-8")
+        assert source.count("# BEGIN Unsloth BNB_ROCM_VERSION") == 1
+        assert source.count("# END Unsloth BNB_ROCM_VERSION") == 1
+        assert "USER_MID = 1" in source
+        assert "'713'" in source
+        assert "'72'" not in source
+
+    def test_persist_bnb_rocm_version_atomic_no_leftover_tmp(self, tmp_path):
+        """The write-then-rename path must not leave its temp file behind."""
+        site_packages = tmp_path / "site-packages"
+        site_packages.mkdir()
+
+        with patch.object(stack_mod.sysconfig, "get_path", return_value = str(site_packages)):
+            assert stack_mod._persist_bnb_rocm_version("72") is True
+
+        leftovers = [p.name for p in site_packages.iterdir() if "unsloth-tmp" in p.name]
+        assert leftovers == []
+        assert (site_packages / "sitecustomize.py").exists()
+
+
+class TestRuntimeBnbRocmSourceGuards:
+    """Runtime entrypoints redetect managed defaults but keep caller overrides."""
+
+    _MAIN_PATH = PACKAGE_ROOT / "studio" / "backend" / "main.py"
+    _TRAINING_WORKER_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "training" / "worker.py"
+
+    def test_main_gate_redetects_persisted_default(self):
+        source = self._MAIN_PATH.read_text(encoding = "utf-8")
+        assert 'os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") == "sitecustomize"' in source
+        assert 'os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] = "detected"' in source
+
+    def test_worker_gate_redetects_persisted_default(self):
+        source = self._TRAINING_WORKER_PATH.read_text(encoding = "utf-8")
+        assert 'os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") == "sitecustomize"' in source
+        assert 'os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] = "detected"' in source
+
+    def test_fallback_prefers_seeded_value_over_hardcoded_72(self):
+        """A failed redetect must not downgrade a persisted suffix to '72'."""
+        for path in (self._MAIN_PATH, self._TRAINING_WORKER_PATH):
+            source = path.read_text(encoding = "utf-8")
+            assert (
+                '_bnb_rocm_ver or os.environ.get("BNB_ROCM_VERSION") or "72"' in source
+            ), path.name
+
+    def test_main_requires_found_rocm_dll(self):
+        """HIP_PATH/ROCM_PATH alone (HIP SDK on a CUDA/CPU box) must not force
+        a ROCm backend onto a non-ROCm bitsandbytes."""
+        source = self._MAIN_PATH.read_text(encoding = "utf-8")
+        assert "if _found_rocm_bnb:" in source
+        assert "_hip_env" not in source
+
+    def test_worker_requires_found_rocm_dll(self):
+        """No DLL found: the worker must not write any override or touch the
+        seeded marker (later import fixes must still see sitecustomize)."""
+        source = self._TRAINING_WORKER_PATH.read_text(encoding = "utf-8")
+        assert "if _found_rocm_bnb:" in source
+
 
 class TestDetectBnbRocmDllVer:
     """Unit tests for _detect_bnb_rocm_dll_ver()."""
@@ -1847,8 +2462,9 @@ class TestWorkerWindowsRocmPatches:
         assert "BNB_ROCM_VERSION" in source
         # Detection helper must be used
         assert "_detect_bnb_rocm_dll_ver" in source or "libbitsandbytes_rocm" in source
-        # "72" must appear as the safe fallback
-        assert '"72"' in source or "'72'" in source
+        # Falls back to the seeded value, never a blind "72" (which would
+        # force a ROCm backend onto a non-ROCm bitsandbytes wheel)
+        assert '_bnb_rocm_ver or os.environ.get("BNB_ROCM_VERSION")' in source
 
     def test_bnb_rocm_version_set_before_ml_imports(self):
         """BNB_ROCM_VERSION must appear in section 1f, before section 2 ML imports."""
@@ -2350,6 +2966,38 @@ class TestServerStartupRocmFixes:
         source = _MAIN_PY_PATH.read_text(encoding = "utf-8")
         assert '"BNB_ROCM_VERSION" not in os.environ' in source
 
+    # ── hipInfo.exe PATH prepend (bitsandbytes arch-probe fix) ────────────────
+    # bitsandbytes' get_rocm_gpu_arch() runs `hipinfo.exe` via subprocess PATH
+    # at import time. The AMD torch wheel ships hipInfo.exe in the venv
+    # Scripts dir, which is on PATH only for activated venvs -- Studio and the
+    # installer launch python directly, so without the prepend every bnb
+    # import logs "Could not detect ROCm GPU architecture: [WinError 2]".
+
+    def test_main_py_prepends_hipinfo_dir_to_path(self):
+        """main.py must make hipInfo.exe resolvable before bnb imports."""
+        source = _MAIN_PY_PATH.read_text(encoding = "utf-8")
+        assert "hipInfo.exe" in source
+        # The prepend must come before the BNB_ROCM_VERSION block (both run
+        # pre-import; order documents that bnb sees the fixed PATH).
+        assert source.find("hipInfo.exe") < source.find("BNB_ROCM_VERSION")
+
+    def test_main_py_hipinfo_prepend_gated_on_file_presence(self):
+        """Only AMD ROCm wheels ship hipInfo.exe; NVIDIA/CPU hosts must be
+        untouched, so the prepend must check the file exists first."""
+        source = _MAIN_PY_PATH.read_text(encoding = "utf-8")
+        assert 'os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe"))' in source
+
+    def test_worker_py_prepends_hipinfo_dir_to_path(self):
+        """worker.py must mirror the prepend for standalone-spawned workers."""
+        source = _WORKER_PATH.read_text(encoding = "utf-8")
+        assert "hipInfo.exe" in source
+
+    def test_install_stack_prepends_hipinfo_dir_to_path(self):
+        """install_python_stack.py must prepend so the installer's child
+        import checks inherit a PATH where bnb's probe succeeds."""
+        source = _STACK_PATH.read_text(encoding = "utf-8")
+        assert "hipInfo.exe" in source
+
     # ── torch._C._distributed_c10d stubs in hardware.py ──────────────────────
 
     def test_hardware_py_injects_distributed_c10d_stub(self):
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 27cf30496f..a4fc09a183 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -11,8 +11,11 @@ No GPU, no network, no torch required -- all I/O is monkeypatched.
 """
 
 import importlib.util
+import os
 import socket
+import subprocess
 import sys
+import textwrap
 import types
 from pathlib import Path
 
@@ -51,12 +54,11 @@ compatible_windows_runtime_lines = INSTALL_LLAMA_PREBUILT.compatible_windows_run
 runtime_line_from_cuda_version = INSTALL_LLAMA_PREBUILT.runtime_line_from_cuda_version
 apply_approved_hashes = INSTALL_LLAMA_PREBUILT.apply_approved_hashes
 linux_cuda_choice_from_release = INSTALL_LLAMA_PREBUILT.linux_cuda_choice_from_release
-parse_direct_linux_release_bundle = INSTALL_LLAMA_PREBUILT.parse_direct_linux_release_bundle
 windows_cuda_attempts = INSTALL_LLAMA_PREBUILT.windows_cuda_attempts
 resolve_upstream_asset_choice = INSTALL_LLAMA_PREBUILT.resolve_upstream_asset_choice
 resolve_requested_install_tag = INSTALL_LLAMA_PREBUILT.resolve_requested_install_tag
 resolve_install_attempts = INSTALL_LLAMA_PREBUILT.resolve_install_attempts
-resolve_install_release_plans = INSTALL_LLAMA_PREBUILT.resolve_install_release_plans
+_fork_manifest_release_plans = INSTALL_LLAMA_PREBUILT._fork_manifest_release_plans
 resolve_published_release = INSTALL_LLAMA_PREBUILT.resolve_published_release
 resolve_source_build_plan = INSTALL_LLAMA_PREBUILT.resolve_source_build_plan
 validated_checksums_for_bundle = INSTALL_LLAMA_PREBUILT.validated_checksums_for_bundle
@@ -111,7 +113,9 @@ def load_studio_run_module(monkeypatch):
     return module
 
 
+# ---------------------------------------------------------------------------
 # Helper factories
+# ---------------------------------------------------------------------------
 
 
 def make_host(**overrides):
@@ -269,7 +273,9 @@ def mock_windows_runtime(monkeypatch, lines):
     )
 
 
+# ===========================================================================
 # Studio run.py localhost warning
+# ===========================================================================
 
 
 class TestStudioLocalhostIpv6Warning:
@@ -450,7 +456,9 @@ class TestStudioLocalhostIpv6Warning:
         assert calls["stop_hint"] == 1
 
 
+# ===========================================================================
 # A. normalize_compute_cap
+# ===========================================================================
 
 
 class TestNormalizeComputeCap:
@@ -482,7 +490,9 @@ class TestNormalizeComputeCap:
         assert normalize_compute_cap("9.0") == "90"
 
 
+# ===========================================================================
 # B. normalize_compute_caps
+# ===========================================================================
 
 
 class TestNormalizeComputeCaps:
@@ -499,7 +509,9 @@ class TestNormalizeComputeCaps:
         assert normalize_compute_caps([]) == []
 
 
+# ===========================================================================
 # C. parse_cuda_visible_devices
+# ===========================================================================
 
 
 class TestParseCudaVisibleDevices:
@@ -522,7 +534,9 @@ class TestParseCudaVisibleDevices:
         assert parse_cuda_visible_devices(" 0 , 1 ") == ["0", "1"]
 
 
+# ===========================================================================
 # D. supports_explicit_visible_device_matching
+# ===========================================================================
 
 
 class TestSupportsExplicitVisibleDeviceMatching:
@@ -542,7 +556,9 @@ class TestSupportsExplicitVisibleDeviceMatching:
         assert supports_explicit_visible_device_matching(["0", "MIG-device"]) is False
 
 
+# ===========================================================================
 # E. select_visible_gpu_rows
+# ===========================================================================
 
 
 class TestSelectVisibleGpuRows:
@@ -575,7 +591,9 @@ class TestSelectVisibleGpuRows:
         assert result == []
 
 
+# ===========================================================================
 # F. compatible_linux_runtime_lines
+# ===========================================================================
 
 
 class TestCompatibleLinuxRuntimeLines:
@@ -601,38 +619,9 @@ class TestCompatibleLinuxRuntimeLines:
         assert compatible_linux_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
 
 
-class TestParseDirectLinuxReleaseBundle:
-    def _release(self, *targets):
-        names = [f"app-bTEST-linux-x64-{t}.tar.gz" for t in targets]
-        return {
-            "tag_name": "bTEST",
-            "assets": [{"name": n, "browser_download_url": "https://x/" + n} for n in names],
-        }
-
-    def _cuda_artifact(self, bundle):
-        return [a for a in bundle.artifacts if a.install_kind == "linux-cuda"][0]
-
-    def test_parses_known_cuda13_bundle(self):
-        bundle = parse_direct_linux_release_bundle(
-            "unslothai/llama.cpp", self._release("cuda13-newer")
-        )
-        assert bundle is not None
-        assert self._cuda_artifact(bundle).runtime_line == "cuda13"
-
-    def test_parses_future_cuda_major_with_forward_profile(self):
-        # A future major name parses and inherits the newest known major's
-        # coverage for the same class as a forward default.
-        bundle = parse_direct_linux_release_bundle(
-            "unslothai/llama.cpp", self._release("cuda14-newer")
-        )
-        assert bundle is not None
-        art = self._cuda_artifact(bundle)
-        assert art.runtime_line == "cuda14"
-        assert art.coverage_class == "newer"
-        assert art.max_sm == 120  # inherited from cuda13-newer
-
-
+# ===========================================================================
 # G. pick_windows_cuda_runtime + compatible_windows_runtime_lines
+# ===========================================================================
 
 
 class TestPickWindowsCudaRuntime:
@@ -666,6 +655,14 @@ class TestCompatibleWindowsRuntimeLines:
         host = make_host(driver_cuda_version = (12, 4))
         assert compatible_windows_runtime_lines(host) == ["cuda12"]
 
+    @pytest.mark.parametrize("minor", [0, 1, 2, 3])
+    def test_cuda12_runs_on_any_12_x_driver(self, minor):
+        # cuda12 app bundles are toolkit-12.8 builds with bundled runtime; CUDA
+        # minor-version compatibility runs them on any 12.x driver, same as Linux.
+        # Previously Windows wrongly gated cuda12 below a 12.4 driver.
+        host = make_host(driver_cuda_version = (12, minor))
+        assert compatible_windows_runtime_lines(host) == ["cuda12"]
+
     def test_driver_13_1(self):
         host = make_host(driver_cuda_version = (13, 1))
         assert compatible_windows_runtime_lines(host) == ["cuda13", "cuda12"]
@@ -679,7 +676,9 @@ class TestCompatibleWindowsRuntimeLines:
         assert compatible_windows_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
 
 
+# ===========================================================================
 # H. runtime_line_from_cuda_version
+# ===========================================================================
 
 
 class TestRuntimeLineFromCudaVersion:
@@ -699,7 +698,9 @@ class TestRuntimeLineFromCudaVersion:
         assert runtime_line_from_cuda_version("") is None
 
 
+# ===========================================================================
 # I. apply_approved_hashes
+# ===========================================================================
 
 
 class TestApplyApprovedHashes:
@@ -813,7 +814,9 @@ class TestApplyApprovedHashes:
             apply_approved_hashes([], checksums)
 
 
+# ===========================================================================
 # J. published release resolution
+# ===========================================================================
 
 
 class TestPublishedReleaseResolution:
@@ -945,6 +948,7 @@ class TestPublishedReleaseResolution:
                 [],
                 release_tag = release_tag,
                 upstream_tag = "b9000",
+                source_repo = "example/custom-llama.cpp",
                 source_commit = commit,
             ),
         )
@@ -1146,8 +1150,55 @@ class TestValidatedChecksumsForBundle:
         with pytest.raises(PrebuiltFallback, match = "manifest checksum"):
             validated_checksums_for_bundle("unslothai/llama.cpp", bundle)
 
+    def test_rejects_exact_source_without_repo(self, monkeypatch):
+        # An exact source archive with no source repo to clone from would let
+        # preferred_source_archive silently fall back to upstream source at the
+        # tag, so validation must fail closed (clean source build instead).
+        bundle = make_release([], release_tag = "r1", upstream_tag = "b8508")
+        checksums = make_checksums_with_source(
+            [], release_tag = "r1", upstream_tag = "b8508", source_commit = "a" * 40
+        )  # exact source archive, but no source_repo
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "load_approved_release_checksums",
+            lambda repo, release_tag: checksums,
+        )
 
+        with pytest.raises(PrebuiltFallback, match = "exact source archive"):
+            validated_checksums_for_bundle("unslothai/llama.cpp", bundle)
+
+    def test_accepts_exact_source_when_only_bundle_has_repo(self, monkeypatch):
+        # The source repo can live only in the manifest bundle, not the checksum
+        # payload. source_build_plan_for_release coalesces checksums-or-bundle, so
+        # validation must accept the bundle's repo rather than failing closed.
+        bundle = make_release(
+            [],
+            release_tag = "r1",
+            upstream_tag = "b8508",
+            source_repo = "ggml-org/llama.cpp",
+            source_repo_url = "https://github.com/ggml-org/llama.cpp",
+        )
+        checksums = make_checksums_with_source(
+            [], release_tag = "r1", upstream_tag = "b8508", source_commit = "a" * 40
+        )  # exact source archive, repo only on the bundle
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "load_approved_release_checksums",
+            lambda repo, release_tag: checksums,
+        )
+
+        assert validated_checksums_for_bundle("unslothai/llama.cpp", bundle) is checksums
+        plan = INSTALL_LLAMA_PREBUILT.source_build_plan_for_release(
+            INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(bundle = bundle, checksums = checksums)
+        )
+        assert plan.source_url == "https://github.com/ggml-org/llama.cpp"
+        assert plan.source_ref_kind == "commit"
+        assert plan.source_ref == "a" * 40
+
+
+# ===========================================================================
 # K. linux_cuda_choice_from_release -- core selection
+# ===========================================================================
 
 
 class TestLinuxCudaChoiceFromRelease:
@@ -1197,6 +1248,36 @@ class TestLinuxCudaChoiceFromRelease:
         log_entries = result.selection_log
         assert any("unavailable_on_host" in entry for entry in log_entries)
 
+    def test_arm64_host_selects_linux_arm64_cuda_kind(self, monkeypatch):
+        # An arm64 CUDA host (DGX Spark / Grace Hopper) selects the
+        # linux-arm64-cuda bundle and ignores the x64 linux-cuda one.
+        mock_linux_runtime(monkeypatch, ["cuda13"])
+        host = make_host(
+            machine = "aarch64",
+            driver_cuda_version = (13, 0),
+            compute_caps = ["90"],
+        )
+        arm = make_artifact(
+            "app-b9457-linux-arm64-cuda13-portable.tar.gz",
+            install_kind = "linux-arm64-cuda",
+            runtime_line = "cuda13",
+            coverage_class = "portable",
+            supported_sms = ["90", "100", "120", "121"],
+            min_sm = 90,
+            max_sm = 121,
+            bundle_profile = "cuda13-portable",
+        )
+        x64 = make_artifact(
+            "app-b9457-linux-x64-cuda13-portable.tar.gz",
+            install_kind = "linux-cuda",
+            runtime_line = "cuda13",
+        )
+        release = make_release([arm, x64])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is not None
+        assert result.primary.install_kind == "linux-arm64-cuda"
+        assert result.primary.name == "app-b9457-linux-arm64-cuda13-portable.tar.gz"
+
     # --- SM matching ---
 
     def test_exact_sm_match(self, monkeypatch):
@@ -1401,80 +1482,9 @@ class TestLinuxCudaChoiceFromRelease:
         assert result is None
 
 
-def make_profile_artifact(asset_name, profile_name, **overrides):
-    profile = INSTALL_LLAMA_PREBUILT.DIRECT_LINUX_BUNDLE_PROFILES[profile_name]
-    defaults = dict(
-        runtime_line = profile["runtime_line"],
-        coverage_class = profile["coverage_class"],
-        supported_sms = [str(value) for value in profile["supported_sms"]],
-        min_sm = int(profile["min_sm"]),
-        max_sm = int(profile["max_sm"]),
-        bundle_profile = profile_name,
-        rank = int(profile["rank"]),
-    )
-    defaults.update(overrides)
-    return make_artifact(asset_name, **defaults)
-
-
-class TestBlackwellUltraSm103Coverage:
-    """sm_103 (B300 / GB300) runs on the bundled base compute_100 PTX via JIT."""
-
-    def test_profiles_list_sm103_wherever_sm100_is_shipped(self):
-        for (
-            name,
-            profile,
-        ) in INSTALL_LLAMA_PREBUILT.DIRECT_LINUX_BUNDLE_PROFILES.items():
-            sms = {str(value) for value in profile["supported_sms"]}
-            if "100" in sms:
-                assert "103" in sms, name
-            else:
-                assert "103" not in sms, name
-
-    def test_b300_selects_cuda13_newer_prebuilt(self, monkeypatch):
-        mock_linux_runtime(monkeypatch, ["cuda13"])
-        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
-        art = make_profile_artifact("cuda13-newer.tar.gz", "cuda13-newer")
-        release = make_release([art])
-        result = linux_cuda_choice_from_release(host, release)
-        assert result is not None
-        assert result.primary.name == "cuda13-newer.tar.gz"
-
-    def test_b300_selects_cuda12_newer_prebuilt(self, monkeypatch):
-        mock_linux_runtime(monkeypatch, ["cuda12"])
-        host = make_host(compute_caps = ["103"], driver_cuda_version = (12, 8))
-        art = make_profile_artifact("cuda12-newer.tar.gz", "cuda12-newer")
-        release = make_release([art])
-        result = linux_cuda_choice_from_release(host, release)
-        assert result is not None
-        assert result.primary.name == "cuda12-newer.tar.gz"
-
-    def test_b300_reported_as_decimal_normalizes_and_matches(self, monkeypatch):
-        mock_linux_runtime(monkeypatch, ["cuda13"])
-        host = make_host(compute_caps = ["10.3"], driver_cuda_version = (13, 0))
-        art = make_profile_artifact("cuda13-portable.tar.gz", "cuda13-portable")
-        release = make_release([art])
-        result = linux_cuda_choice_from_release(host, release)
-        assert result is not None
-
-    def test_b300_falls_back_to_portable_when_only_portable_present(self, monkeypatch):
-        mock_linux_runtime(monkeypatch, ["cuda13"])
-        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
-        art = make_profile_artifact("cuda13-portable.tar.gz", "cuda13-portable")
-        release = make_release([art])
-        result = linux_cuda_choice_from_release(host, release)
-        assert result is not None
-        assert result.primary.name == "cuda13-portable.tar.gz"
-
-    def test_older_bundle_still_rejects_b300(self, monkeypatch):
-        mock_linux_runtime(monkeypatch, ["cuda13"])
-        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
-        art = make_profile_artifact("cuda13-older.tar.gz", "cuda13-older")
-        release = make_release([art])
-        result = linux_cuda_choice_from_release(host, release)
-        assert result is None
-
-
+# ===========================================================================
 # L. resolve_install_attempts
+# ===========================================================================
 
 
 class TestResolveInstallAttempts:
@@ -1599,7 +1609,11 @@ class TestResolveInstallAttempts:
         assert attempts[0].expected_sha256 == "a" * 64
         assert approved.release_tag == "llama-prebuilt-latest"
 
-    def test_linux_cpu_uses_same_tag_upstream_asset(self, monkeypatch):
+    def test_linux_cpu_fork_without_bundle_raises_no_upstream_fallback(self, monkeypatch):
+        # A CPU-only Linux host on the fork no longer falls back to the ggml-org
+        # CPU asset: production routes CPU-only Linux to ggml-org, never the fork.
+        # With no fork CPU bundle in the manifest the resolver raises rather than
+        # quietly reaching for an upstream asset.
         host = make_host(
             has_usable_nvidia = False,
             has_physical_nvidia = False,
@@ -1607,7 +1621,7 @@ class TestResolveInstallAttempts:
         )
         release = make_release([], release_tag = "llama-prebuilt-latest", upstream_tag = "b9000")
         checksums = make_checksums_with_source(
-            ["llama-b9000-bin-ubuntu-x64.tar.gz"],
+            [],
             release_tag = release.release_tag,
             upstream_tag = "b9000",
         )
@@ -1627,22 +1641,13 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "github_release_assets",
-            lambda repo, tag: {
-                f"llama-{tag}-bin-ubuntu-x64.tar.gz": f"https://example.com/llama-{tag}-bin-ubuntu-x64.tar.gz"
-            },
+            lambda repo, tag: (_ for _ in ()).throw(
+                AssertionError("fork CPU host must not query upstream assets")
+            ),
         )
 
-        _requested_tag, resolved_tag, attempts, _approved = resolve_install_attempts(
-            "latest",
-            host,
-            "unslothai/llama.cpp",
-            "",
-        )
-
-        assert resolved_tag == "b9000"
-        assert attempts[0].name == "llama-b9000-bin-ubuntu-x64.tar.gz"
-        assert attempts[0].source_label == "upstream"
-        assert attempts[0].expected_sha256 == "a" * 64
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt asset was found"):
+            resolve_install_attempts("latest", host, "unslothai/llama.cpp", "")
 
     def test_linux_cuda_does_not_fall_back_to_upstream_cpu(self, monkeypatch):
         host = make_host(system = "Linux", machine = "x86_64", compute_caps = ["86"])
@@ -1667,7 +1672,7 @@ class TestResolveInstallAttempts:
         )
         mock_linux_runtime(monkeypatch, ["cuda12"])
 
-        with pytest.raises(PrebuiltFallback, match = "no compatible published Linux CUDA bundle"):
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt asset was found"):
             resolve_install_attempts("latest", host, "unslothai/llama.cpp", "")
 
     def test_windows_cpu_prefers_published_asset(self, monkeypatch):
@@ -1863,37 +1868,35 @@ class TestResolveInstallAttempts:
 
 
 class TestResolveInstallReleasePlans:
-    def test_latest_collects_multiple_older_release_plans_up_to_limit(self, monkeypatch):
-        host = make_host(
-            has_usable_nvidia = False,
-            has_physical_nvidia = False,
-            nvidia_smi = None,
+    def _cuda_bundle(self, asset_name, release_tag, upstream_tag):
+        # A fork CUDA bundle that covers the default NVIDIA host (sm 86,
+        # cuda12 runtime), so each release yields a plan via
+        # linux_cuda_choice_from_release.
+        art = make_artifact(
+            asset_name,
+            install_kind = "linux-cuda",
+            runtime_line = "cuda12",
+            coverage_class = "portable",
+            supported_sms = ["75", "80", "86", "89", "90"],
+            min_sm = 75,
+            max_sm = 90,
         )
+        return INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+            bundle = make_release([art], release_tag = release_tag, upstream_tag = upstream_tag),
+            checksums = make_checksums_with_source(
+                [asset_name],
+                release_tag = release_tag,
+                upstream_tag = upstream_tag,
+            ),
+        )
+
+    def test_latest_collects_multiple_older_release_plans_up_to_limit(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda12"])
+        host = make_host(system = "Linux", machine = "x86_64", compute_caps = ["86"])
         releases = [
-            INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
-                bundle = make_release([], release_tag = "r3", upstream_tag = "b9003"),
-                checksums = make_checksums_with_source(
-                    ["llama-b9003-bin-ubuntu-x64.tar.gz"],
-                    release_tag = "r3",
-                    upstream_tag = "b9003",
-                ),
-            ),
-            INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
-                bundle = make_release([], release_tag = "r2", upstream_tag = "b9002"),
-                checksums = make_checksums_with_source(
-                    ["llama-b9002-bin-ubuntu-x64.tar.gz"],
-                    release_tag = "r2",
-                    upstream_tag = "b9002",
-                ),
-            ),
-            INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
-                bundle = make_release([], release_tag = "r1", upstream_tag = "b9001"),
-                checksums = make_checksums_with_source(
-                    ["llama-b9001-bin-ubuntu-x64.tar.gz"],
-                    release_tag = "r1",
-                    upstream_tag = "b9001",
-                ),
-            ),
+            self._cuda_bundle("app-b9003-linux-x64-cuda12.tar.gz", "r3", "b9003"),
+            self._cuda_bundle("app-b9002-linux-x64-cuda12.tar.gz", "r2", "b9002"),
+            self._cuda_bundle("app-b9001-linux-x64-cuda12.tar.gz", "r1", "b9001"),
         ]
 
         monkeypatch.setattr(
@@ -1901,15 +1904,8 @@ class TestResolveInstallReleasePlans:
             "iter_resolved_published_releases",
             lambda requested_tag, published_repo, published_release_tag = "": iter(releases),
         )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "github_release_assets",
-            lambda repo, tag: {
-                f"llama-{tag}-bin-ubuntu-x64.tar.gz": f"https://example.com/llama-{tag}-bin-ubuntu-x64.tar.gz"
-            },
-        )
 
-        requested_tag, plans = resolve_install_release_plans(
+        requested_tag, plans = _fork_manifest_release_plans(
             "latest",
             host,
             "unslothai/llama.cpp",
@@ -1922,12 +1918,10 @@ class TestResolveInstallReleasePlans:
         assert [plan.llama_tag for plan in plans] == ["b9003", "b9002"]
 
     def test_latest_skips_non_installable_release_and_keeps_searching(self, monkeypatch):
-        host = make_host(
-            has_usable_nvidia = False,
-            has_physical_nvidia = False,
-            nvidia_smi = None,
-        )
+        mock_linux_runtime(monkeypatch, ["cuda12"])
+        host = make_host(system = "Linux", machine = "x86_64", compute_caps = ["86"])
         releases = [
+            # r2 ships no fork bundle, so it yields no plan and is skipped.
             INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                 bundle = make_release([], release_tag = "r2", upstream_tag = "b9002"),
                 checksums = make_checksums_with_source(
@@ -1936,14 +1930,7 @@ class TestResolveInstallReleasePlans:
                     upstream_tag = "b9002",
                 ),
             ),
-            INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
-                bundle = make_release([], release_tag = "r1", upstream_tag = "b9001"),
-                checksums = make_checksums_with_source(
-                    ["llama-b9001-bin-ubuntu-x64.tar.gz"],
-                    release_tag = "r1",
-                    upstream_tag = "b9001",
-                ),
-            ),
+            self._cuda_bundle("app-b9001-linux-x64-cuda12.tar.gz", "r1", "b9001"),
         ]
 
         monkeypatch.setattr(
@@ -1951,19 +1938,8 @@ class TestResolveInstallReleasePlans:
             "iter_resolved_published_releases",
             lambda requested_tag, published_repo, published_release_tag = "": iter(releases),
         )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "github_release_assets",
-            lambda repo, tag: (
-                {}
-                if tag == "b9002"
-                else {
-                    f"llama-{tag}-bin-ubuntu-x64.tar.gz": f"https://example.com/llama-{tag}-bin-ubuntu-x64.tar.gz"
-                }
-            ),
-        )
 
-        _requested_tag, plans = resolve_install_release_plans(
+        _requested_tag, plans = _fork_manifest_release_plans(
             "latest",
             host,
             "unslothai/llama.cpp",
@@ -1995,7 +1971,9 @@ class TestResolveInstallReleasePlans:
             sys.modules.pop(spec.name, None)
 
 
+# ===========================================================================
 # N. windows_cuda_attempts
+# ===========================================================================
 
 
 class TestWindowsCudaAttempts:
@@ -2213,7 +2191,9 @@ class TestWindowsCudaAttempts:
         assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
 
 
+# ===========================================================================
 # N.1b. _pinned_windows_cuda_fallback -- pinned b9360 cuda-13.1 Blackwell fallback
+# ===========================================================================
 
 
 class TestPinnedBlackwellCudaFallback:
@@ -2354,8 +2334,49 @@ class TestPinnedBlackwellCudaFallback:
         )
         assert _windows_cuda_attempt_covers_blackwell(cpu) is False
 
+    def _app_attempt(self, profile, runtime_line, max_sm):
+        # The fork's app-named windows-cuda bundle: no toolkit minor in the name,
+        # SM coverage declared directly (as published_windows_cuda_attempts sets it).
+        return AssetChoice(
+            repo = UPSTREAM_REPO,
+            tag = self.TAG,
+            name = f"app-{self.TAG}-windows-x64-{runtime_line}-{profile}.zip",
+            url = "https://example.com/x",
+            source_label = "published",
+            install_kind = "windows-cuda",
+            runtime_line = runtime_line,
+            coverage_class = "newer" if profile == "newer" else profile,
+            max_sm = max_sm,
+            min_sm = 80,
+            supported_sms = ["120"] if max_sm >= 120 else ["86", "89"],
+        )
 
+    @pytest.mark.parametrize(
+        "profile, runtime_line, max_sm, covers",
+        [
+            ("newer", "cuda13", 120, True),  # native Blackwell build
+            ("newer", "cuda12", 120, True),  # 12.8 toolkit app bundle reaches sm120
+            ("older", "cuda12", 89, False),  # 12.4 toolkit app bundle stops at Ada
+        ],
+    )
+    def test_attempt_covers_blackwell_app_bundle(self, profile, runtime_line, max_sm, covers):
+        # App-named bundles carry no toolkit minor; coverage is read from max_sm.
+        attempt = self._app_attempt(profile, runtime_line, max_sm)
+        assert _windows_cuda_attempt_covers_blackwell(attempt) is covers
+
+    def test_pin_dormant_when_app_bundle_covers_blackwell(self):
+        # Regression: the fork's app-named cuda13 bundle covers Blackwell, so the
+        # b9360 pin must retire instead of being prepended ahead of the native
+        # in-release build (previously the coverage check only matched legacy
+        # -bin-win-cuda-X.Y-x64.zip names, so the pin never went dormant).
+        host = self._win_host((13, 1), ["120"])
+        existing = [self._app_attempt("newer", "cuda13", 120)]
+        assert _pinned_windows_cuda_fallback(host, existing) is None
+
+
+# ===========================================================================
 # N.1c. direct_upstream_release_plan -- pinned Blackwell fallback ordering
+# ===========================================================================
 
 
 class TestDirectUpstreamBlackwellPin:
@@ -2398,7 +2419,10 @@ class TestDirectUpstreamBlackwellPin:
         )
         plan = direct_upstream_release_plan(self._release(), host, UPSTREAM_REPO, "latest")
         order = [(a.tag, a.runtime_line or a.install_kind) for a in plan.attempts]
-        assert order == [("b9360", "cuda13"), (self.TAG, "cuda12"), (self.TAG, "windows-cpu")]
+        # cuda-12.4 (toolkit 12.4, no sm_120) is dropped entirely on Blackwell:
+        # behind the pin it would still be attempted if the pin download failed,
+        # and the functional validator accepts its slow non-native path.
+        assert order == [("b9360", "cuda13"), (self.TAG, "windows-cpu")]
         assert plan.attempts[0].name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
         # Direct/upstream path stays unverified-by-manifest (no approved hashes).
         assert plan.approved_checksums.artifacts == {}
@@ -2419,7 +2443,177 @@ class TestDirectUpstreamBlackwellPin:
         assert plan.attempts[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
 
 
+# N.1c2. Blackwell never falls to a non-sm_120 windows-cuda attempt
+
+
+class TestBlackwellCuda124Exclusion:
+    """A Blackwell host must never have a windows-cuda attempt that cannot
+    offload sm_120 anywhere in its chain: behind the pin it is one failed
+    download away from a validated-but-7-tok/s install."""
+
+    def _bw_host(self):
+        return make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+
+    def _upstream_cuda(
+        self,
+        minor,
+        tag = "b9365",
+    ):
+        return AssetChoice(
+            repo = "ggml-org/llama.cpp",
+            tag = tag,
+            name = f"llama-{tag}-bin-win-cuda-{minor}-x64.zip",
+            url = f"https://example.com/{minor}",
+            source_label = "upstream",
+            install_kind = "windows-cuda",
+            runtime_line = "cuda" + minor.split(".")[0],
+        )
+
+    def test_drops_124_keeps_133_on_blackwell(self):
+        kept = INSTALL_LLAMA_PREBUILT._drop_blackwell_incapable_windows_cuda(
+            self._bw_host(),
+            [self._upstream_cuda("13.3"), self._upstream_cuda("12.4")],
+        )
+        assert [a.name for a in kept] == ["llama-b9365-bin-win-cuda-13.3-x64.zip"]
+
+    def test_keeps_manifest_cuda12_bundle_with_sm120(self):
+        # Published cuda12 app bundles are toolkit-12.8 builds that include
+        # sm_120; the manifest SM metadata must keep them on Blackwell.
+        bundle = AssetChoice(
+            repo = "unslothai/llama.cpp",
+            tag = "b9585",
+            name = "app-b9585-windows-x64-cuda12-portable.zip",
+            url = "https://example.com/app",
+            source_label = "published",
+            install_kind = "windows-cuda",
+            runtime_line = "cuda12",
+            supported_sms = ["70", "120"],
+            max_sm = 120,
+        )
+        kept = INSTALL_LLAMA_PREBUILT._drop_blackwell_incapable_windows_cuda(
+            self._bw_host(), [bundle]
+        )
+        assert kept == [bundle]
+        assert _windows_cuda_attempt_covers_blackwell(bundle)
+
+    def test_manifest_bundle_without_sm120_dropped(self):
+        bundle = AssetChoice(
+            repo = "unslothai/llama.cpp",
+            tag = "b9585",
+            name = "app-b9585-windows-x64-cuda12-older.zip",
+            url = "https://example.com/app",
+            source_label = "published",
+            install_kind = "windows-cuda",
+            runtime_line = "cuda12",
+            supported_sms = ["70", "75", "80"],
+            max_sm = 80,
+        )
+        assert (
+            INSTALL_LLAMA_PREBUILT._drop_blackwell_incapable_windows_cuda(self._bw_host(), [bundle])
+            == []
+        )
+
+    def test_non_blackwell_host_unfiltered(self):
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (12, 9),
+            compute_caps = ["89"],
+        )
+        attempts = [self._upstream_cuda("12.4")]
+        assert (
+            INSTALL_LLAMA_PREBUILT._drop_blackwell_incapable_windows_cuda(host, attempts)
+            == attempts
+        )
+
+    def test_non_cuda_attempts_pass_through(self):
+        cpu = AssetChoice(
+            repo = "ggml-org/llama.cpp",
+            tag = "b9365",
+            name = "llama-b9365-bin-win-cpu-x64.zip",
+            url = "https://example.com/cpu",
+            source_label = "upstream",
+            install_kind = "windows-cpu",
+        )
+        kept = INSTALL_LLAMA_PREBUILT._drop_blackwell_incapable_windows_cuda(
+            self._bw_host(), [self._upstream_cuda("12.4"), cpu]
+        )
+        assert kept == [cpu]
+
+
+# N.1c3. direct_linux_release_plan -- no silent CPU on NVIDIA hosts
+
+
+class TestDirectLinuxNvidiaCpuGate:
+    """When a release ships a linux-cpu bundle but no CUDA line this NVIDIA
+    host can use, the planner must raise (so the caller walks back to an older
+    release with a usable CUDA line) instead of silently planning a CPU
+    install on a GPU host. CPU-only hosts keep taking the CPU bundle."""
+
+    def _bundle_cpu_only(self):
+        return make_release(
+            [
+                make_artifact(
+                    "llama-b8508-bin-ubuntu-x64.tar.gz",
+                    install_kind = "linux-cpu",
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = None,
+                ),
+            ]
+        )
+
+    def _patch(self, monkeypatch):
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "parse_direct_linux_release_bundle",
+            lambda repo, release: self._bundle_cpu_only(),
+        )
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "detect_torch_cuda_runtime_preference",
+            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
+        )
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "detected_linux_runtime_lines",
+            lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
+        )
+
+    def test_nvidia_host_without_cuda_line_raises_for_walkback(self, monkeypatch):
+        self._patch(monkeypatch)
+        host = make_host(driver_cuda_version = (13, 1), compute_caps = ["100"])
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt"):
+            INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
+                {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
+            )
+
+    def test_cpu_host_still_gets_cpu_bundle(self, monkeypatch):
+        self._patch(monkeypatch)
+        host = make_host(
+            nvidia_smi = None,
+            driver_cuda_version = None,
+            compute_caps = [],
+            has_physical_nvidia = False,
+            has_usable_nvidia = False,
+        )
+        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
+            {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
+        )
+        assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
+
+
+# ===========================================================================
 # N.1d. published_windows_cuda_attempts -- version-dynamic ordering seed
+# ===========================================================================
 
 
 class TestPublishedWindowsCudaAttemptsDynamicMajor:
@@ -2434,6 +2628,7 @@ class TestPublishedWindowsCudaAttemptsDynamicMajor:
             f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip",
             install_kind = "windows-cuda",
             runtime_line = runtime_line,
+            supported_sms = ["75", "80", "86", "89", "90", "100", "120"],
             max_sm = 120,
         )
 
@@ -2485,12 +2680,14 @@ class TestPublishedWindowsCudaAttemptsDynamicMajor:
         assert result[0].runtime_line == "cuda12"
 
 
+# ===========================================================================
 # N.1e. resolve_release_asset_choice -- pin on the published install path
+# ===========================================================================
 
 
 class TestResolveReleaseAssetChoicePin:
-    """The published (non --simple-policy) install path reaches the same b9360
-    Blackwell pin as the simple path, with its verified hash threaded."""
+    """The manifest install path reaches the same b9360 Blackwell pin as the
+    filename path, with its verified hash threaded."""
 
     TAG = "b8508"
 
@@ -2500,6 +2697,7 @@ class TestResolveReleaseAssetChoicePin:
                 f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip",
                 install_kind = "windows-cuda",
                 runtime_line = line,
+                supported_sms = ["75", "80", "86", "89", "90", "100", "120"],
                 max_sm = 120,
             )
             for minor, line in minors_lines
@@ -2546,7 +2744,9 @@ class TestResolveReleaseAssetChoicePin:
         # augmented checksums (the pin survives the approved-hash gate).
         assert result[0].expected_sha256 and len(result[0].expected_sha256) == 64
         assert result[0].runtime_sha256 and len(result[0].runtime_sha256) == 64
-        assert any(a.runtime_line == "cuda12" for a in result)
+        # The sm_120-incapable upstream cuda-12.4 zip is excluded on Blackwell
+        # rather than left behind the pin as a slow-path fallback.
+        assert not any(a.runtime_line == "cuda12" for a in result)
 
     def test_pin_dormant_on_published_path_for_13_3(self, monkeypatch):
         mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
@@ -2578,7 +2778,243 @@ class TestResolveReleaseAssetChoicePin:
         assert "b9360" not in [a.tag for a in result]
 
 
+class TestPublishedWindowsCudaAppBundleSmSelection:
+    """app-named windows-cuda bundles carry no minor in the filename, so the
+    driver-minor gate is skipped. Selection must instead filter by SM coverage,
+    or every host gets the lowest-rank "older" bundle regardless of its GPU."""
+
+    TAG = "b9457"
+
+    def _app(self, klass, supported, min_sm, max_sm, rank):
+        return make_artifact(
+            f"app-{self.TAG}-windows-x64-cuda12-{klass}.zip",
+            install_kind = "windows-cuda",
+            runtime_line = "cuda12",
+            coverage_class = klass,
+            supported_sms = supported,
+            min_sm = min_sm,
+            max_sm = max_sm,
+            bundle_profile = f"cuda12-{klass}",
+            rank = rank,
+        )
+
+    def test_blackwell_sm120_skips_older_bundle(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda12"])
+        older = self._app("older", ["70", "75", "80", "86", "89"], 70, 89, 10)
+        newer = self._app("newer", ["86", "89", "90", "100", "120"], 86, 120, 20)
+        portable = self._app(
+            "portable", ["70", "75", "80", "86", "89", "90", "100", "120"], 70, 120, 30
+        )
+        release = make_release([older, newer, portable], upstream_tag = self.TAG)
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (12, 8),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, None)
+        assert result, "expected a windows-cuda attempt for an sm120 host"
+        # The lowest-rank "older" bundle (max_sm 89) must not be chosen, and the
+        # tightest covering bundle is cuda12-newer (range 86-120).
+        assert result[0].name == f"app-{self.TAG}-windows-x64-cuda12-newer.zip"
+
+    def _line(self, line, klass, rank):
+        return make_artifact(
+            f"app-{self.TAG}-windows-x64-{line}-{klass}.zip",
+            install_kind = "windows-cuda",
+            runtime_line = line,
+            coverage_class = klass,
+            supported_sms = ["86", "89", "90", "100", "120"],
+            min_sm = 86,
+            max_sm = 120,
+            bundle_profile = f"{line}-{klass}",
+            rank = rank,
+        )
+
+    def test_cuda13_reachable_on_driver_13_0(self, monkeypatch):
+        # app-named cuda13 bundles must be reachable on a 13.0 driver. The old
+        # synthetic '13.1' minor gate dropped the whole cuda13 line (13.1 > 13.0),
+        # so a cu13 host fell to cuda12. cuda13 is gated at the major level now.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        release = make_release(
+            [self._line("cuda12", "newer", 20), self._line("cuda13", "newer", 50)],
+            upstream_tag = self.TAG,
+        )
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 0),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, "cuda13")
+        assert result
+        assert result[0].runtime_line == "cuda13"
+        assert result[0].name == f"app-{self.TAG}-windows-x64-cuda13-newer.zip"
+
+    def test_app_bundle_offered_when_no_runtime_dll_detected(self, monkeypatch):
+        # Windows torch bundles cudart in torch/lib, which runtime-DLL probing
+        # misses, so detected_windows_runtime_lines() returns nothing. The app
+        # bundle ships its own runtime, so selection must fall back to the
+        # driver-derived order instead of yielding no attempt (which would drop
+        # the host to the upstream build).
+        mock_windows_runtime(monkeypatch, [])
+        release = make_release(
+            [self._line("cuda12", "newer", 20), self._line("cuda13", "newer", 50)],
+            upstream_tag = self.TAG,
+        )
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 0),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, "cuda13")
+        assert result, "torch-only host must still get the fork app bundle"
+        assert result[0].name == f"app-{self.TAG}-windows-x64-cuda13-newer.zip"
+
+
+class TestPublishedRocmGfxSelection:
+    """Published ROCm bundles are matched by the host's detected gfx family, not
+    by rank -- rank ties would alphabetically hand every AMD GPU the gfx103X
+    bundle (e.g. a gfx1151 Strix Halo host)."""
+
+    GFX = ["gfx103X", "gfx110X", "gfx120X", "gfx1150", "gfx1151"]
+    MEMBERS = {
+        "gfx103X": ["gfx1030", "gfx1031", "gfx1032", "gfx1034"],
+        "gfx110X": ["gfx1100", "gfx1101", "gfx1102", "gfx1103"],
+        "gfx120X": ["gfx1200", "gfx1201"],
+        "gfx1150": ["gfx1150"],
+        "gfx1151": ["gfx1151"],
+    }
+
+    def _release(self, install_kind, prefix):
+        artifacts = [
+            make_artifact(
+                f"{prefix}-{gfx}.{'zip' if 'windows' in install_kind else 'tar.gz'}",
+                install_kind = install_kind,
+                runtime_line = None,
+                coverage_class = None,
+                supported_sms = [],
+                min_sm = None,
+                max_sm = None,
+                bundle_profile = None,
+                rank = 1000,
+                gfx_target = gfx,
+                mapped_targets = self.MEMBERS[gfx],
+            )
+            for gfx in self.GFX
+        ]
+        return make_release(artifacts, upstream_tag = "b9457")
+
+    def _host(self, gfx):
+        return make_host(
+            machine = "x86_64",
+            nvidia_smi = None,
+            driver_cuda_version = None,
+            compute_caps = [],
+            has_physical_nvidia = False,
+            has_usable_nvidia = False,
+            has_rocm = True,
+            rocm_gfx_target = gfx,
+        )
+
+    def test_gfx1100_selects_gfx110X_family(self):
+        release = self._release("linux-rocm", "app-b9457-linux-x64-rocm")
+        choice = INSTALL_LLAMA_PREBUILT.published_rocm_choice_for_host(
+            release, self._host("gfx1100"), "linux-rocm"
+        )
+        assert choice is not None
+        assert choice.name == "app-b9457-linux-x64-rocm-gfx110X.tar.gz"
+
+    def test_gfx1151_strix_halo_not_handed_gfx103X(self):
+        release = self._release("linux-rocm", "app-b9457-linux-x64-rocm")
+        choice = INSTALL_LLAMA_PREBUILT.published_rocm_choice_for_host(
+            release, self._host("gfx1151"), "linux-rocm"
+        )
+        assert choice is not None
+        assert choice.name == "app-b9457-linux-x64-rocm-gfx1151.tar.gz"
+
+    def test_windows_rocm_gfx_match(self):
+        release = self._release("windows-rocm", "app-b9457-windows-x64-rocm")
+        choice = INSTALL_LLAMA_PREBUILT.published_rocm_choice_for_host(
+            release, self._host("gfx1201"), "windows-rocm"
+        )
+        assert choice is not None
+        assert choice.name == "app-b9457-windows-x64-rocm-gfx120X.zip"
+
+    def test_uncovered_gpu_returns_none(self):
+        release = self._release("linux-rocm", "app-b9457-linux-x64-rocm")
+        assert (
+            INSTALL_LLAMA_PREBUILT.published_rocm_choice_for_host(
+                release, self._host("gfx900"), "linux-rocm"
+            )
+            is None
+        )
+
+    def test_in_prefix_but_unbuilt_arch_returns_none(self):
+        # gfx1033 shares the gfx103 prefix but is not in any bundle's
+        # mapped_targets, so it must fall back to source, not be served gfx103X.
+        release = self._release("linux-rocm", "app-b9457-linux-x64-rocm")
+        for unbuilt in ("gfx1033", "gfx1035", "gfx1104", "gfx1202"):
+            assert (
+                INSTALL_LLAMA_PREBUILT.published_rocm_choice_for_host(
+                    release, self._host(unbuilt), "linux-rocm"
+                )
+                is None
+            ), unbuilt
+
+
+class TestPublishedMacosForkSelection:
+    """macOS now routes to the fork (setup.sh), which ships
+    llama-<tag>-bin-macos-<arch>.tar.gz with pinned deployment targets, selected
+    by install_kind."""
+
+    def _release(self):
+        arts = [
+            make_artifact(
+                "llama-b9457-bin-macos-arm64.tar.gz",
+                install_kind = "macos-arm64",
+                runtime_line = None,
+                coverage_class = None,
+                supported_sms = [],
+                min_sm = None,
+                max_sm = None,
+                bundle_profile = "macos-metal-arm64",
+                rank = 50,
+            ),
+            make_artifact(
+                "llama-b9457-bin-macos-x64.tar.gz",
+                install_kind = "macos-x64",
+                runtime_line = None,
+                coverage_class = None,
+                supported_sms = [],
+                min_sm = None,
+                max_sm = None,
+                bundle_profile = "macos-cpu-x64",
+                rank = 50,
+            ),
+        ]
+        return make_release(arts, upstream_tag = "b9457")
+
+    def test_macos_arm64_selects_fork_bundle(self):
+        choice = INSTALL_LLAMA_PREBUILT.published_asset_choice_for_kind(
+            self._release(), "macos-arm64"
+        )
+        assert choice is not None
+        assert choice.name == "llama-b9457-bin-macos-arm64.tar.gz"
+        assert choice.install_kind == "macos-arm64"
+
+    def test_macos_x64_selects_fork_bundle(self):
+        choice = INSTALL_LLAMA_PREBUILT.published_asset_choice_for_kind(
+            self._release(), "macos-x64"
+        )
+        assert choice is not None
+        assert choice.name == "llama-b9457-bin-macos-x64.tar.gz"
+
+
+# ===========================================================================
 # N.1. apply_approved_hashes -- runtime archive checksum threading
+# ===========================================================================
 
 
 class TestApplyApprovedHashesRuntimePair:
@@ -2650,7 +3086,9 @@ class TestApplyApprovedHashesRuntimePair:
         assert result[0].runtime_sha256 is None
 
 
+# ===========================================================================
 # O. resolve_upstream_asset_choice -- platform routing
+# ===========================================================================
 
 
 class TestResolveUpstreamAssetChoice:
@@ -2793,7 +3231,9 @@ class TestResolveUpstreamAssetChoice:
         assert result.name == cuda_name
 
 
+# ===========================================================================
 # N.2. Deterministic macOS prebuilt pin (b9415)
+# ===========================================================================
 
 
 def _macos_host(machine = "arm64", version = (15, 5)):
@@ -2916,37 +3356,47 @@ class TestResolveSimpleMacosPin:
         assert calls[0][2] == "latest"
 
 
+# ===========================================================================
 # Linux arm64 + GPU must not install the x64-only fork bundle
+# ===========================================================================
 
 
 class TestLinuxArm64ForkFallsBackToSource:
-    """The unslothai/llama.cpp fork ships only linux-x64 bundles. An arm64
-    Linux host with a GPU (GH200/GB200/DGX Spark) routes to the fork and must
-    fall back to a source build instead of selecting an x64 binary."""
+    """The fork now ships linux-arm64-cuda bundles (GH200/GB200/DGX Spark). An
+    arm64 Linux host on the fork no longer hard-fails on the simple path; it
+    delegates to the manifest-aware resolver, which selects the arm64 CUDA
+    bundle (or falls back to source only if none matches)."""
 
-    def test_arm64_nvidia_fork_raises_before_fetching_releases(self, monkeypatch):
-        # Guard fires before any release is fetched: poison the iterator to prove
-        # it is never called.
-        def _boom(*_a, **_k):
-            raise AssertionError("iterator must not run for arm64 fork hosts")
+    def test_arm64_nvidia_fork_delegates_to_manifest_resolver(self, monkeypatch):
+        # arm64 fork hosts are no longer blocked up front; the simple resolver
+        # hands them to the manifest-aware resolver instead.
+        called = {}
 
-        monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "iter_release_payloads_by_time", _boom)
+        def _full(llama_tag, host, repo, tag, **_kw):
+            called["args"] = (host.machine, repo)
+            return "b9457", ["plan"]
+
+        monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "_fork_manifest_release_plans", _full)
         host = make_host(system = "Linux", machine = "aarch64")
-        with pytest.raises(PrebuiltFallback, match = "linux-x64 prebuilts"):
-            resolve_simple_install_release_plans("latest", host, "unslothai/llama.cpp", "")
+        tag, plans = resolve_simple_install_release_plans("latest", host, "unslothai/llama.cpp", "")
+        assert called.get("args") == ("aarch64", "unslothai/llama.cpp")
+        assert plans == ["plan"]
 
-    def test_x86_64_fork_is_not_blocked_by_the_arch_guard(self, monkeypatch):
-        # x64 host must pass the guard and reach the iterator (here empty, so it
-        # raises the generic message, not the arch one).
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "iter_release_payloads_by_time",
-            lambda *_a, **_k: iter(()),
-        )
+    def test_x86_64_fork_delegates_to_manifest_resolver(self, monkeypatch):
+        # The old linux-x64 arch guard is gone: an x64 fork host is routed to the
+        # manifest resolver exactly like every other fork host, not down a
+        # separate filename-parsing path.
+        called = {}
+
+        def _full(llama_tag, host, repo, tag, **_kw):
+            called["args"] = (host.machine, repo)
+            return "b9457", ["plan"]
+
+        monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "_fork_manifest_release_plans", _full)
         host = make_host(system = "Linux", machine = "x86_64")
-        with pytest.raises(PrebuiltFallback) as exc:
-            resolve_simple_install_release_plans("latest", host, "unslothai/llama.cpp", "")
-        assert "linux-x64 prebuilts" not in str(exc.value)
+        tag, plans = resolve_simple_install_release_plans("latest", host, "unslothai/llama.cpp", "")
+        assert called.get("args") == ("x86_64", "unslothai/llama.cpp")
+        assert plans == ["plan"]
 
     def test_arm64_cpu_on_ggml_org_is_not_blocked(self, monkeypatch):
         # CPU-only arm64 routes to ggml-org (not the fork), so the guard must not
@@ -2970,7 +3420,9 @@ class TestLinuxArm64ForkFallsBackToSource:
         assert "linux-x64 prebuilts" not in str(exc.value)
 
 
+# ===========================================================================
 # arm64 Linux GPU: CPU prebuilt fallback after a failed source build (--cpu-fallback)
+# ===========================================================================
 
 
 class TestCpuFallback:
@@ -3011,7 +3463,6 @@ class TestCpuFallback:
                 llama_tag = "latest",
                 published_repo = "ggml-org/llama.cpp",
                 published_release_tag = "",
-                simple_policy = True,
                 force_cpu = True,
             )
         host = captured["host"]
@@ -3062,3 +3513,443 @@ class TestCpuFallback:
         # is gated on a degraded source build for arm64.
         assert "ggml-org/llama.cpp" in source
         assert "_LLAMA_CPP_DEGRADED" in source
+
+
+# ===========================================================================
+# setup.sh / setup.ps1: CUDA toolkit newer than driver diagnostics
+# ===========================================================================
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "bash-only Studio installer tests")
+class TestCudaDriverToolkitMismatchMessage:
+    _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
+    _SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1"
+
+    def _setup_sh_cuda_helper_fragment(self):
+        source = self._SETUP_SH.read_text(encoding = "utf-8")
+        start = source.index("_nvcc_meets_llama_minimum()")
+        end = source.index("print_llama_error_log()")
+        return source[start:end]
+
+    def _run_bash(
+        self,
+        script,
+        *,
+        env = None,
+    ):
+        proc = subprocess.run(
+            ["/bin/bash", "-c", script],
+            check = True,
+            text = True,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.PIPE,
+            env = {**os.environ, **(env or {})},
+        )
+        return proc.stdout + proc.stderr
+
+    def _fake_nvidia_smi(self, tmp_path, output):
+        mock_bin = tmp_path / "bin"
+        mock_bin.mkdir()
+        nvidia_smi = mock_bin / "nvidia-smi"
+        nvidia_smi.write_text(
+            textwrap.dedent(
+                f"""\
+                #!/usr/bin/env bash
+                cat <<'OUT'
+                {output}
+                OUT
+                """
+            ),
+            encoding = "utf-8",
+        )
+        nvidia_smi.chmod(0o755)
+        return mock_bin
+
+    def test_setup_sh_same_major_minor_mismatch_is_accepted(self, tmp_path):
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 580.95   Driver Version: 580.95   CUDA Version: 13.1 |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _driver="$(_cuda_driver_max_version)"
+            if _cuda_toolkit_major_gt_driver "13.3" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "13.3" "$_driver"
+            else
+                printf 'compatible:%s\\n' "$_driver"
+            fi
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert "compatible:13.1" in output
+        assert "major-version mismatch" not in output
+
+    def test_setup_sh_driver_major_too_old_message_names_major_mismatch(self, tmp_path):
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 570.95   Driver Version: 570.95   CUDA Version: 12.9 |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _driver="$(_cuda_driver_max_version)"
+            if _cuda_toolkit_major_gt_driver "13.3" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "13.3" "$_driver"
+            fi
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert (
+            "CUDA Toolkit 13.3 is a major-version mismatch: toolkit major 13 "
+            "exceeds driver CUDA major 12 (12.9)."
+        ) in output
+        assert (
+            "Update the NVIDIA GPU driver to run CUDA Toolkit 13.3, or install "
+            "a CUDA 12.x toolkit."
+        ) in output
+        assert "prebuilt CUDA bundle" in output
+
+    def test_setup_sh_happy_path_does_not_print_mismatch(self, tmp_path):
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 580.95   Driver Version: 580.95   CUDA Version: 13.3 |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _driver="$(_cuda_driver_max_version)"
+            if _cuda_toolkit_major_gt_driver "13.3" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "13.3" "$_driver"
+            else
+                printf 'compatible:%s\\n' "$_driver"
+            fi
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert "compatible:13.3" in output
+        assert "Unsloth supports CUDA Toolkit" not in output
+
+    def test_setup_sh_skips_check_without_nvidia_smi(self, tmp_path):
+        empty_bin = tmp_path / "empty-bin"
+        empty_bin.mkdir()
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _driver="$(_cuda_driver_max_version)"
+            if [ -n "$_driver" ] && _cuda_toolkit_major_gt_driver "13.3" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "13.3" "$_driver"
+            else
+                printf 'skipped\\n'
+            fi
+            """
+        )
+        output = self._run_bash(script, env = {"PATH": str(empty_bin)})
+        assert "skipped" in output
+        assert "Unsloth supports CUDA Toolkit" not in output
+
+    def test_setup_sh_unparsable_nvidia_smi_output_falls_back(self, tmp_path):
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 580.95   Driver Version: 580.95   CUDA Version: N/A |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _driver="$(_cuda_driver_max_version)"
+            if [ -n "$_driver" ] && _cuda_toolkit_major_gt_driver "13.3" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "13.3" "$_driver"
+            else
+                printf 'fallback:%s\\n' "${{_driver:-generic}}"
+            fi
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert "fallback:generic" in output
+        assert "Unsloth supports CUDA Toolkit" not in output
+
+    def test_setup_sh_cuda_version_gt_compares_numerically(self):
+        # 13.9 vs 13.10 is where a lexical compare goes wrong (9 > 1).
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            {self._setup_sh_cuda_helper_fragment()}
+            for pair in "13.9 13.10" "13.10 13.9" "14.0 13.9" "13.3 13.3"; do
+                if _cuda_version_gt $pair; then r=gt; else r=le; fi
+                printf '%s -> %s\\n' "$pair" "$r"
+            done
+            """
+        )
+        output = self._run_bash(script)
+        assert "13.9 13.10 -> le" in output
+        assert "13.10 13.9 -> gt" in output
+        assert "14.0 13.9 -> gt" in output
+        assert "13.3 13.3 -> le" in output
+
+    def test_setup_ps1_mirrors_driver_mismatch_guidance(self):
+        source = self._SETUP_PS1.read_text(encoding = "utf-8")
+        assert "Write-CudaDriverToolkitMismatch" in source
+        assert (
+            "CUDA Toolkit $ToolkitVersion is a major-version mismatch: toolkit "
+            "major $toolkitMajor exceeds driver CUDA major $driverMajor"
+        ) in source
+        assert (
+            "Update the NVIDIA GPU driver to run CUDA Toolkit $ToolkitVersion, "
+            "or install a CUDA $driverMajor.x toolkit." in source
+        )
+        assert (
+            "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit."
+        ) in source
+        assert (
+            "Write-CudaDriverToolkitMismatch -ToolkitVersion $IncompatibleToolkit "
+            "-DriverMaxCuda $DriverMaxCuda"
+        ) in source
+
+    def _fake_nvcc(self, tmp_path, release):
+        mock_bin = tmp_path / f"nvcc-bin-{release.replace('.', '-')}"
+        mock_bin.mkdir()
+        nvcc = mock_bin / "nvcc"
+        nvcc.write_text(
+            textwrap.dedent(
+                f"""\
+                #!/usr/bin/env bash
+                printf '%s\\n' 'Cuda compilation tools, release {release}, V{release}.0'
+                """
+            ),
+            encoding = "utf-8",
+        )
+        nvcc.chmod(0o755)
+        return nvcc
+
+    def test_setup_sh_major_mismatch_uses_newest_compatible_detected_toolkit(self, tmp_path):
+        blocked_nvcc = self._fake_nvcc(tmp_path, "13.3")
+        older_nvcc = self._fake_nvcc(tmp_path, "12.6")
+        compatible_nvcc = self._fake_nvcc(tmp_path, "12.8")
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _cuda_nvcc_candidate_paths() {{
+                printf '%s\\n' "{blocked_nvcc}" "{older_nvcc}" "{compatible_nvcc}"
+            }}
+            NVCC_PATH="{blocked_nvcc}"
+            GPU_BACKEND="cuda"
+            _BUILD_DESC="building"
+            _NVCC_CHECK="$(_nvcc_meets_llama_minimum "$NVCC_PATH")"
+            _NVCC_STATUS="$(printf '%s\\n' "$_NVCC_CHECK" | sed -n '1p')"
+            _NVCC_VER="$(printf '%s\\n' "$_NVCC_CHECK" | sed -n '2p')"
+            _DRIVER_MAX_CUDA="12.9"
+            _CUDA_TOOLKIT_ALLOWED=true
+            if [ "$_NVCC_STATUS" != "too_old" ] && \
+               [ -n "$_NVCC_VER" ] && \
+               _cuda_toolkit_major_gt_driver "$_NVCC_VER" "$_DRIVER_MAX_CUDA"; then
+                _BLOCKED_NVCC_VER="$_NVCC_VER"
+                if _ALT_NVCC_CHECK="$(_cuda_find_compatible_nvcc_for_driver "$_DRIVER_MAX_CUDA" "$NVCC_PATH")"; then
+                    NVCC_PATH="$(printf '%s\\n' "$_ALT_NVCC_CHECK" | sed -n '1p')"
+                    _NVCC_VER="$(printf '%s\\n' "$_ALT_NVCC_CHECK" | sed -n '2p')"
+                    GPU_BACKEND="cuda"
+                    substep "CUDA Toolkit $_BLOCKED_NVCC_VER is a major-version mismatch with driver CUDA $_DRIVER_MAX_CUDA; using compatible CUDA Toolkit $_NVCC_VER at $NVCC_PATH."
+                else
+                    NVCC_PATH=""
+                    GPU_BACKEND=""
+                    _BUILD_DESC="building (CPU, CUDA toolkit major > driver)"
+                    _CUDA_TOOLKIT_ALLOWED=false
+                fi
+            fi
+            printf 'NVCC_PATH=%s\\n' "$NVCC_PATH"
+            printf 'NVCC_VER=%s\\n' "$_NVCC_VER"
+            printf 'GPU_BACKEND=%s\\n' "$GPU_BACKEND"
+            printf 'BUILD_DESC=%s\\n' "$_BUILD_DESC"
+            printf 'ALLOWED=%s\\n' "$_CUDA_TOOLKIT_ALLOWED"
+            """
+        )
+        output = self._run_bash(script)
+        assert f"NVCC_PATH={compatible_nvcc}" in output
+        assert "NVCC_VER=12.8" in output
+        assert "GPU_BACKEND=cuda" in output
+        assert "ALLOWED=true" in output
+        assert "CPU" not in output
+
+    def test_setup_sh_parses_cuda_umd_version_variant(self, tmp_path):
+        # Newer drivers report "CUDA UMD Version"; the helper must read it too.
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 610.00   Driver Version: 610.00   CUDA UMD Version: 13.0 |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            {self._setup_sh_cuda_helper_fragment()}
+            printf '%s' "$(_cuda_driver_max_version)"
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert output.strip() == "13.0"
+
+    def test_setup_sh_empty_toolkit_version_skips_mismatch(self, tmp_path):
+        # The real guard requires a non-empty nvcc version before warning.
+        mock_bin = self._fake_nvidia_smi(
+            tmp_path,
+            "| NVIDIA-SMI 580.95   Driver Version: 580.95   CUDA Version: 13.0 |",
+        )
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            _nvcc=""
+            _driver="$(_cuda_driver_max_version)"
+            if [ -n "$_nvcc" ] && [ -n "$_driver" ] && _cuda_toolkit_major_gt_driver "$_nvcc" "$_driver"; then
+                _print_cuda_driver_toolkit_mismatch "$_nvcc" "$_driver"
+            else
+                printf 'skipped\\n'
+            fi
+            """
+        )
+        output = self._run_bash(
+            script,
+            env = {"PATH": f"{mock_bin}:{os.environ.get('PATH', '')}"},
+        )
+        assert "skipped" in output
+        assert "Unsloth supports CUDA Toolkit" not in output
+
+    def test_setup_sh_nvcc_below_minimum_is_too_old(self, tmp_path):
+        # CUDA toolkit < 12.4 short-circuits to the too_old branch (no mismatch).
+        nvcc = self._fake_nvcc(tmp_path, "12.0")
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            {self._setup_sh_cuda_helper_fragment()}
+            _nvcc_meets_llama_minimum "{nvcc}"
+            """
+        )
+        output = self._run_bash(script)
+        assert output.splitlines()[0] == "too_old"
+        assert "12.0" in output
+
+    def test_setup_sh_compatible_finder_rejects_too_old_only_candidate(self, tmp_path):
+        # Only candidate is below the llama minimum: finder must fail (CPU fallback), not pick 12.0.
+        blocked_nvcc = self._fake_nvcc(tmp_path, "13.3")
+        too_old_nvcc = self._fake_nvcc(tmp_path, "12.0")
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            {self._setup_sh_cuda_helper_fragment()}
+            _cuda_nvcc_candidate_paths() {{
+                printf '%s\\n' "{blocked_nvcc}" "{too_old_nvcc}"
+            }}
+            if _ALT="$(_cuda_find_compatible_nvcc_for_driver "12.9" "{blocked_nvcc}")"; then
+                printf 'FOUND:%s\\n' "$_ALT"
+            else
+                printf 'NONE\\n'
+            fi
+            """
+        )
+        output = self._run_bash(script)
+        assert "NONE" in output
+        assert "FOUND" not in output
+
+    def _cuda_build_decision_output(self, *, nvcc_path, driver):
+        # Mirror setup.sh's source-build decision: keep the toolkit, switch, or degrade to CPU.
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            C_WARN=
+            substep() {{ printf '%s\\n' "$1"; }}
+            {self._setup_sh_cuda_helper_fragment()}
+            NVCC_PATH="{nvcc_path}"
+            GPU_BACKEND="cuda"
+            _DRIVER_MAX_CUDA="{driver}"
+            _NVCC_CHECK="$(_nvcc_meets_llama_minimum "$NVCC_PATH")"
+            _NVCC_STATUS="$(printf '%s\\n' "$_NVCC_CHECK" | sed -n '1p')"
+            _NVCC_VER="$(printf '%s\\n' "$_NVCC_CHECK" | sed -n '2p')"
+            _CUDA_TOOLKIT_ALLOWED=true
+            if [ "$_NVCC_STATUS" = "too_old" ]; then
+                NVCC_PATH=""; GPU_BACKEND=""; _CUDA_TOOLKIT_ALLOWED=false
+            elif [ -n "$_NVCC_VER" ] && [ -n "$_DRIVER_MAX_CUDA" ] && _cuda_toolkit_major_gt_driver "$_NVCC_VER" "$_DRIVER_MAX_CUDA"; then
+                if _ALT="$(_cuda_find_compatible_nvcc_for_driver "$_DRIVER_MAX_CUDA" "$NVCC_PATH")"; then
+                    NVCC_PATH="$(printf '%s\\n' "$_ALT" | sed -n '1p')"
+                    _NVCC_VER="$(printf '%s\\n' "$_ALT" | sed -n '2p')"
+                else
+                    NVCC_PATH=""; GPU_BACKEND=""; _CUDA_TOOLKIT_ALLOWED=false
+                fi
+            fi
+            printf 'NVCC_PATH=%s\\n' "$NVCC_PATH"
+            printf 'NVCC_VER=%s\\n' "$_NVCC_VER"
+            printf 'GPU_BACKEND=%s\\n' "$GPU_BACKEND"
+            printf 'ALLOWED=%s\\n' "$_CUDA_TOOLKIT_ALLOWED"
+            """
+        )
+        return self._run_bash(script)
+
+    def test_setup_sh_same_major_newer_minor_keeps_original_toolkit(self, tmp_path):
+        # Same-major newer-minor (13.3 vs driver 13.0): build CUDA with it, never fall back.
+        toolkit = self._fake_nvcc(tmp_path, "13.3")
+        output = self._cuda_build_decision_output(nvcc_path = toolkit, driver = "13.0")
+        assert f"NVCC_PATH={toolkit}" in output
+        assert "NVCC_VER=13.3" in output
+        assert "GPU_BACKEND=cuda" in output
+        assert "ALLOWED=true" in output
+
+    def test_setup_sh_missing_driver_version_still_enables_cuda(self, tmp_path):
+        # No driver CUDA version from nvidia-smi: keep CUDA enabled (pre-fix behavior), not CPU.
+        toolkit = self._fake_nvcc(tmp_path, "13.3")
+        output = self._cuda_build_decision_output(nvcc_path = toolkit, driver = "")
+        assert f"NVCC_PATH={toolkit}" in output
+        assert "GPU_BACKEND=cuda" in output
+        assert "ALLOWED=true" in output
+
+    def test_setup_sh_compatible_finder_rejects_newer_major_only_candidate(self, tmp_path):
+        # Only alternative is still newer-major than the driver: finder must fail, not pick it.
+        blocked_nvcc = self._fake_nvcc(tmp_path, "13.3")
+        other_newer_nvcc = self._fake_nvcc(tmp_path, "13.1")
+        script = textwrap.dedent(
+            f"""\
+            set -euo pipefail
+            {self._setup_sh_cuda_helper_fragment()}
+            _cuda_nvcc_candidate_paths() {{
+                printf '%s\\n' "{blocked_nvcc}" "{other_newer_nvcc}"
+            }}
+            if _ALT="$(_cuda_find_compatible_nvcc_for_driver "12.9" "{blocked_nvcc}")"; then
+                printf 'FOUND:%s\\n' "$_ALT"
+            else
+                printf 'NONE\\n'
+            fi
+            """
+        )
+        output = self._run_bash(script)
+        assert "NONE" in output
+        assert "FOUND" not in output
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index dc62194be8..4739d519d9 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1100,6 +1100,22 @@ with sync_playwright() as p:
         plus_btn.click(force = True)
         page.wait_for_timeout(400)
         compare_item = page.get_by_role("menuitem", name = re.compile(r"Compare chat", re.I)).first
+        if compare_item.count() == 0:
+            # The plus menu was decluttered: Compare chat now lives in the
+            # "More" submenu; hover (then click as fallback) to open it.
+            more_trigger = page.get_by_role("menuitem", name = re.compile(r"^More$", re.I)).first
+            if more_trigger.count() > 0:
+                more_trigger.hover()
+                page.wait_for_timeout(400)
+                compare_item = page.get_by_role(
+                    "menuitem", name = re.compile(r"Compare chat", re.I)
+                ).first
+                if compare_item.count() == 0:
+                    more_trigger.click(force = True)
+                    page.wait_for_timeout(400)
+                    compare_item = page.get_by_role(
+                        "menuitem", name = re.compile(r"Compare chat", re.I)
+                    ).first
         if compare_item.count() > 0:
             compare_item.click(force = True)
             page.wait_for_timeout(800)
diff --git a/tests/studio/playwright_extra_ui.py b/tests/studio/playwright_extra_ui.py
index 20c7bda87c..26c3c244ca 100644
--- a/tests/studio/playwright_extra_ui.py
+++ b/tests/studio/playwright_extra_ui.py
@@ -332,6 +332,21 @@ with sync_playwright() as p:
         plus_btn.click(force = True)
         page.wait_for_timeout(400)
         compare_item = page.get_by_role("menuitem", name = re.compile(r"Compare chat", re.I)).first
+        if compare_item.count() == 0:
+            # Compare chat moved into the "More" submenu; hover, then click fallback.
+            more_trigger = page.get_by_role("menuitem", name = re.compile(r"^More$", re.I)).first
+            if more_trigger.count() > 0:
+                more_trigger.hover()
+                page.wait_for_timeout(400)
+                compare_item = page.get_by_role(
+                    "menuitem", name = re.compile(r"Compare chat", re.I)
+                ).first
+                if compare_item.count() == 0:
+                    more_trigger.click(force = True)
+                    page.wait_for_timeout(400)
+                    compare_item = page.get_by_role(
+                        "menuitem", name = re.compile(r"Compare chat", re.I)
+                    ).first
         if compare_item.count() > 0:
             compare_item.click(force = True)
             compare_opened = True
diff --git a/tests/studio/test_cli_studio_stop_windows.py b/tests/studio/test_cli_studio_stop_windows.py
new file mode 100644
index 0000000000..7ba3d12214
--- /dev/null
+++ b/tests/studio/test_cli_studio_stop_windows.py
@@ -0,0 +1,128 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression tests for `unsloth studio stop` on Windows (PR #5940).
+
+`stop` once used the POSIX `os.kill(pid, 0)` probe, which raises OSError
+(WinError 87) for every pid on Windows -- crashing before reaching taskkill.
+The fix adds a cross-platform `_pid_alive(pid)` (tasklist on Windows, signal-0
+elsewhere).
+
+AST + mock-only; no real process management, no Studio deps imported.
+"""
+
+import ast
+import os
+import subprocess
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+_STUDIO_CMD_PY = Path(__file__).resolve().parents[2] / "unsloth_cli" / "commands" / "studio.py"
+_SOURCE = _STUDIO_CMD_PY.read_text(encoding = "utf-8")
+
+
+def _func_source(name: str) -> str:
+    """Return the source of a top-level function `name` in studio.py."""
+    tree = ast.parse(_SOURCE)
+    for node in ast.walk(tree):
+        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)) and node.name == name:
+            return ast.get_source_segment(_SOURCE, node)
+    raise AssertionError(f"function {name!r} not found in studio.py")
+
+
+def _load_pid_alive(platform: str, fake_run = None):
+    """Exec just `_pid_alive` with injectable sys/subprocess, so we can drive
+    the win32 branch on any host without importing the full unsloth_cli."""
+    src = _func_source("_pid_alive")
+    fake_sys = types.SimpleNamespace(platform = platform)
+    fake_sub = types.SimpleNamespace(run = fake_run) if fake_run is not None else subprocess
+    ns = {"os": os, "sys": fake_sys, "subprocess": fake_sub}
+    exec(src, ns)
+    return ns["_pid_alive"]
+
+
+# ── AST: stop() must not use the broken bare liveness probe ──────────────────
+
+
+def test_stop_does_not_use_bare_oskill_liveness_probe():
+    """stop() must not call os.kill(pid, 0) -- it crashes on Windows."""
+    stop_src = _func_source("stop")
+    tree = ast.parse(stop_src)
+    for call in ast.walk(tree):
+        if not isinstance(call, ast.Call):
+            continue
+        f = call.func
+        is_os_kill = (
+            isinstance(f, ast.Attribute)
+            and f.attr == "kill"
+            and isinstance(f.value, ast.Name)
+            and f.value.id == "os"
+        )
+        if is_os_kill and len(call.args) == 2:
+            sig = call.args[1]
+            if isinstance(sig, ast.Constant) and sig.value == 0:
+                raise AssertionError(
+                    "stop() still uses os.kill(pid, 0); it raises WinError 87 on "
+                    "Windows. Use the cross-platform _pid_alive() helper instead."
+                )
+
+
+def test_pid_alive_helper_is_defined_and_used_by_stop():
+    assert "def _pid_alive(" in _SOURCE, "_pid_alive helper missing"
+    assert "_pid_alive(pid)" in _func_source("stop"), "stop() must use _pid_alive"
+    # The helper must special-case Windows via tasklist (os.kill(pid,0) is invalid there).
+    helper = _func_source("_pid_alive")
+    assert 'sys.platform == "win32"' in helper
+    assert "tasklist" in helper
+
+
+# ── Behavioral: the win32 tasklist branch ────────────────────────────────────
+
+
+def _fake_tasklist(returns_pid: int | None, *, raises: bool = False):
+    def _run(
+        cmd,
+        capture_output = False,
+        text = False,
+        timeout = None,
+    ):
+        assert cmd[0] == "tasklist"
+        assert "/FI" in cmd  # filtered by PID
+        if raises:
+            raise OSError("boom")
+        if returns_pid is None:
+            stdout = "INFO: No tasks are running which match the specified criteria.\n"
+        else:
+            stdout = f'"python.exe","{returns_pid}","Console","1","12,345 K"\n'
+        return types.SimpleNamespace(stdout = stdout, returncode = 0)
+
+    return _run
+
+
+def test_pid_alive_windows_true_when_tasklist_lists_pid():
+    pid_alive = _load_pid_alive("win32", fake_run = _fake_tasklist(4242))
+    assert pid_alive(4242) is True
+
+
+def test_pid_alive_windows_false_when_tasklist_empty():
+    pid_alive = _load_pid_alive("win32", fake_run = _fake_tasklist(None))
+    assert pid_alive(4242) is False
+
+
+def test_pid_alive_windows_assumes_alive_when_tasklist_errors():
+    # Can't determine -> assume alive; taskkill is the source of truth.
+    pid_alive = _load_pid_alive("win32", fake_run = _fake_tasklist(None, raises = True))
+    assert pid_alive(4242) is True
+
+
+# ── Behavioral: the POSIX signal-0 branch (skip on Windows runners) ───────────
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX os.kill(pid,0) branch")
+def test_pid_alive_posix_true_for_self_false_for_dead():
+    pid_alive = _load_pid_alive("linux")
+    assert pid_alive(os.getpid()) is True
+    assert pid_alive(2_000_000_000) is False
diff --git a/tests/studio/test_resolve_cuda_toolkit.ps1 b/tests/studio/test_resolve_cuda_toolkit.ps1
index 42076cdfe0..77fcdca59c 100644
--- a/tests/studio/test_resolve_cuda_toolkit.ps1
+++ b/tests/studio/test_resolve_cuda_toolkit.ps1
@@ -1,7 +1,7 @@
 #!/usr/bin/env pwsh
 # Unit test for Resolve-CudaToolkit in studio/setup.ps1. No GPU required: the
 # detection helpers (nvidia-smi, nvcc, Find-Nvcc, ...) are stubbed so the real
-# function logic runs against a spoofed Blackwell sm_120 / driver 13.2 host.
+# function logic runs against spoofed Blackwell sm_120 driver/toolkit scenarios.
 #
 # The function is extracted via AST and run in a child pwsh per scenario, because
 # the -RequireOrExit path calls `exit` (which would otherwise kill this harness).
@@ -22,13 +22,24 @@ $fn = $ast.FindAll({ param($n)
 if ($fn.Count -ne 1) { throw "expected exactly one Resolve-CudaToolkit, found $($fn.Count)" }
 $fnText = $fn[0].Extent.Text
 
-# --- Spoof executables: nvidia-smi reports driver max CUDA 13.2; nvcc 13.3 ---
+# Resolve-CudaToolkit calls Write-CudaDriverToolkitMismatch, so extract it too.
+$mismatchFn = $ast.FindAll({ param($n)
+    $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq "Write-CudaDriverToolkitMismatch"
+}, $true)
+if ($mismatchFn.Count -ne 1) { throw "expected exactly one Write-CudaDriverToolkitMismatch, found $($mismatchFn.Count)" }
+$mismatchText = $mismatchFn[0].Extent.Text
+
+# --- Spoof executables for driver/toolkit compatibility scenarios ---
 $work = Join-Path ([System.IO.Path]::GetTempPath()) ("rct_" + [guid]::NewGuid().ToString("N"))
 New-Item -ItemType Directory -Force -Path $work | Out-Null
-$smiFake  = Join-Path $work "nvidia-smi.ps1"
-$nvccFake = Join-Path $work "nvcc.ps1"
-Set-Content -LiteralPath $smiFake  -Value "'CUDA Version: 13.2'"
-Set-Content -LiteralPath $nvccFake -Value "'Cuda compilation tools, release 13.3, V13.3.0'"
+$smiMajorMismatchFake = Join-Path $work "nvidia-smi-12.9.ps1"
+$smiSameMajorFake     = Join-Path $work "nvidia-smi-13.2.ps1"
+$nvccIncompatibleFake = Join-Path $work "nvcc-13.3.ps1"
+$nvccCompatibleFake   = Join-Path $work "nvcc-12.8.ps1"
+Set-Content -LiteralPath $smiMajorMismatchFake -Value "'CUDA Version: 12.9'"
+Set-Content -LiteralPath $smiSameMajorFake     -Value "'CUDA Version: 13.2'"
+Set-Content -LiteralPath $nvccIncompatibleFake -Value "'Cuda compilation tools, release 13.3, V13.3.0'"
+Set-Content -LiteralPath $nvccCompatibleFake   -Value "'Cuda compilation tools, release 12.8, V12.8.0'"
 
 $failures = 0
 function Check($name, $cond) {
@@ -38,13 +49,15 @@ function Check($name, $cond) {
 
 # Build + run one scenario in a child pwsh; returns @{ Exit; Out }.
 function Run-Case {
-    param([string]$FindMode, [bool]$Require)
+    param([string]$FindMode, [bool]$Require, [string]$DriverMode = "major-mismatch")
     $requireLit = if ($Require) { '$true' } else { '$false' }
+    $smiForCase = if ($DriverMode -eq "same-major") { $smiSameMajorFake } else { $smiMajorMismatchFake }
     $child = @"
 `$ErrorActionPreference = 'Continue'
 [Environment]::SetEnvironmentVariable('CUDA_PATH', `$null, 'Process')
 `$FindNvccMode = '$FindMode'
-`$NvccFake = '$nvccFake'
+`$NvccIncompatibleFake = '$nvccIncompatibleFake'
+`$NvccCompatibleFake = '$nvccCompatibleFake'
 function substep { param(`$m, `$c) Write-Host "  `$m" }
 function step    { param(`$l, `$v, `$c) Write-Host "[`$l] `$v" }
 function Add-ToUserPath { param(`$Directory, `$Position) `$true }
@@ -57,17 +70,20 @@ function winget { `$script:WingetCalled = `$true; 'no matching versions' }
 function Find-Nvcc {
     param([string]`$MaxVersion = '')
     switch (`$FindNvccMode) {
-        'compatible'   { return `$NvccFake }
-        'incompatible' { if (`$MaxVersion) { return `$null } else { return `$NvccFake } }
+        'compatible'   { return `$NvccCompatibleFake }
+        'same-major'   { return `$NvccIncompatibleFake }
+        'incompatible' { if (`$MaxVersion) { return `$null } else { return `$NvccIncompatibleFake } }
         default        { return `$null }
     }
 }
-`$NvidiaSmiExe  = '$smiFake'
+`$NvidiaSmiExe  = '$smiForCase'
 `$VsInstallPath = `$null
 `$HasNvidiaSmi  = `$true
 `$script:CudaToolkitReady = `$false
 `$script:NvccPath = `$null; `$script:CudaToolkitRoot = `$null; `$script:CudaArch = `$null
 
+$mismatchText
+
 $fnText
 
 if ($requireLit) { Resolve-CudaToolkit -RequireOrExit } else { Resolve-CudaToolkit }
@@ -80,37 +96,54 @@ Write-Host ("RESULT ready={0} nvcc={1} winget={2}" -f `$script:CudaToolkitReady,
 }
 
 try {
-    Write-Host "Scenario 1: prebuilt path, too-new toolkit (no -RequireOrExit) -> defers, no exit"
+    Write-Host "Scenario 1: prebuilt path, newer-major toolkit (no -RequireOrExit) -> defers, no exit"
     $r = Run-Case -FindMode "incompatible" -Require $false
     Check "exits 0 (not blocked)"        ($r.Exit -eq 0)
     Check "CudaToolkitReady = false"     ($r.Out -match "ready=False")
     Check "winget NOT called"            ($r.Out -match "winget=False")
-    Check "no INCOMPATIBLE error text"   (-not ($r.Out -match "INCOMPATIBLE"))
+    Check "explains major mismatch"      ($r.Out -match "major-version mismatch")
+    Check "does not blame the toolkit"   (-not ($r.Out -match "INCOMPATIBLE"))
 
-    Write-Host "Scenario 2: forced source build, too-new toolkit (-RequireOrExit) -> hard exit"
+    Write-Host "Scenario 2: forced source build, newer-major toolkit (-RequireOrExit) -> hard exit"
     $r = Run-Case -FindMode "incompatible" -Require $true
     Check "exits non-zero"               ($r.Exit -ne 0)
-    Check "preserved INCOMPATIBLE error" ($r.Out -match "is installed but INCOMPATIBLE")
+    Check "explains major mismatch"      ($r.Out -match "major-version mismatch")
+    Check "one-line source-build error"  ($r.Out -match "CUDA source build cannot use the installed toolkit")
 
-    Write-Host "Scenario 3: compatible toolkit (-RequireOrExit) -> resolves, env set"
+    Write-Host "Scenario 3: same-major newer-minor toolkit (-RequireOrExit) -> resolves, env set"
+    $r = Run-Case -FindMode "same-major" -Require $true -DriverMode "same-major"
+    Check "exits 0"                      ($r.Exit -eq 0)
+    Check "CudaToolkitReady = true"      ($r.Out -match "ready=True")
+    Check "NvccPath published"           ($r.Out -match "nvcc=.*nvcc-13\.3")
+    Check "no mismatch warning"          (-not ($r.Out -match "major-version mismatch"))
+
+    Write-Host "Scenario 4: compatible older-major toolkit (-RequireOrExit) -> resolves, env set"
     $r = Run-Case -FindMode "compatible" -Require $true
     Check "exits 0"                      ($r.Exit -eq 0)
     Check "CudaToolkitReady = true"      ($r.Out -match "ready=True")
     Check "NvccPath published"           ($r.Out -match "nvcc=.*nvcc")
 
-    Write-Host "Scenario 4: no toolkit, prebuilt path (no -RequireOrExit) -> defers, no winget"
+    Write-Host "Scenario 5: no toolkit, prebuilt path (no -RequireOrExit) -> defers, no winget"
     $r = Run-Case -FindMode "none" -Require $false
     Check "exits 0"                      ($r.Exit -eq 0)
     Check "CudaToolkitReady = false"     ($r.Out -match "ready=False")
     Check "winget NOT called"            ($r.Out -match "winget=False")
 
-    Write-Host "Scenario 5: no toolkit, forced (-RequireOrExit) -> winget attempted then exit"
+    Write-Host "Scenario 6: no toolkit, forced (-RequireOrExit) -> winget attempted then exit"
     # The function exits before the RESULT line here, so assert on the winget-block
     # marker in output rather than the flag.
     $r = Run-Case -FindMode "none" -Require $true
     Check "winget attempted"             ($r.Out -match "installing via winget")
     Check "exits non-zero"               ($r.Exit -ne 0)
     Check "preserved nvcc-required error" ($r.Out -match "CUDA Toolkit \(nvcc\) is required")
+
+    Write-Host "Scenario 7: same-major toolkit only on PATH, missed by -MaxVersion (-RequireOrExit) -> accepted, not rejected"
+    # -MaxVersion misses it (not in side-by-side base) but plain Find-Nvcc finds it on PATH: must be used.
+    $r = Run-Case -FindMode "incompatible" -Require $true -DriverMode "same-major"
+    Check "exits 0"                      ($r.Exit -eq 0)
+    Check "CudaToolkitReady = true"      ($r.Out -match "ready=True")
+    Check "NvccPath published"           ($r.Out -match "nvcc=.*nvcc-13\.3")
+    Check "no mismatch warning"          (-not ($r.Out -match "major-version mismatch"))
 }
 finally {
     Remove-Item -Recurse -Force -LiteralPath $work -ErrorAction SilentlyContinue
diff --git a/tests/studio/test_sync_allow_scripts_pins.py b/tests/studio/test_sync_allow_scripts_pins.py
new file mode 100644
index 0000000000..2aa0e9368c
--- /dev/null
+++ b/tests/studio/test_sync_allow_scripts_pins.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+"""Offline tests for scripts/sync_allow_scripts_pins.py."""
+
+from __future__ import annotations
+
+import json
+import sys
+import tempfile
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(REPO_ROOT / "scripts"))
+
+import sync_allow_scripts_pins as sync  # noqa: E402
+
+
+def write_fixture(tmp: Path, policy: dict, lock_packages: dict) -> None:
+    (tmp / "package.json").write_text(
+        json.dumps(
+            {
+                "name": "fixture",
+                "dependencies": {"a": "^1.0.0"},
+                "allowScripts": policy,
+            },
+            indent = 2,
+        )
+        + "\n"
+    )
+    (tmp / "package-lock.json").write_text(
+        json.dumps(
+            {
+                "lockfileVersion": 3,
+                "packages": lock_packages,
+            }
+        )
+        + "\n"
+    )
+
+
+LOCK = {
+    "": {"name": "fixture"},
+    "node_modules/@biomejs/biome": {"version": "1.9.9", "hasInstallScript": True},
+    "node_modules/msw": {"version": "2.15.0", "hasInstallScript": True},
+    "node_modules/vite/node_modules/fsevents": {"version": "2.3.3", "hasInstallScript": True},
+    "node_modules/clean": {"version": "3.0.0"},
+}
+
+
+def test_split_spec():
+    assert sync.split_spec("fsevents") == ("fsevents", None)
+    assert sync.split_spec("@biomejs/biome") == ("@biomejs/biome", None)
+    assert sync.split_spec("@biomejs/biome@1.9.4") == ("@biomejs/biome", "1.9.4")
+    assert sync.split_spec("msw@2.14.3 || 2.15.0") == ("msw", "2.14.3 || 2.15.0")
+
+
+def test_in_sync_is_clean():
+    with tempfile.TemporaryDirectory() as d:
+        tmp = Path(d)
+        write_fixture(
+            tmp,
+            {
+                "@biomejs/biome@1.9.9": True,
+                "msw@2.15.0": True,
+                "fsevents": True,
+            },
+            LOCK,
+        )
+        assert sync.main(["--check", "--dir", str(tmp)]) == 0
+
+
+def test_stale_pin_fails_check_and_fix_repairs():
+    with tempfile.TemporaryDirectory() as d:
+        tmp = Path(d)
+        write_fixture(
+            tmp,
+            {
+                "@biomejs/biome@1.9.4": True,  # stale, bumped to 1.9.9
+                "msw@2.14.3": False,  # stale denial, bumped to 2.15.0
+                "fsevents": True,  # bare: never stale
+                "ghost@9.9.9": True,  # not in lockfile: left alone
+                "weird@*": True,  # non-exact spec: left alone
+            },
+            LOCK,
+        )
+        assert sync.main(["--check", "--dir", str(tmp)]) == 1
+        assert sync.main(["--fix", "--dir", str(tmp)]) == 0
+        got = json.loads((tmp / "package.json").read_text())["allowScripts"]
+        assert got == {
+            "@biomejs/biome@1.9.9": True,
+            "msw@2.15.0": False,  # value and key order preserved
+            "fsevents": True,
+            "ghost@9.9.9": True,
+            "weird@*": True,
+        }
+        assert list(got) == [
+            "@biomejs/biome@1.9.9",
+            "msw@2.15.0",
+            "fsevents",
+            "ghost@9.9.9",
+            "weird@*",
+        ]
+        assert sync.main(["--check", "--dir", str(tmp)]) == 0
+
+
+def test_multi_version_disjunction():
+    lock = dict(LOCK)
+    lock["node_modules/x/node_modules/msw"] = {"version": "2.14.3", "hasInstallScript": True}
+    with tempfile.TemporaryDirectory() as d:
+        tmp = Path(d)
+        write_fixture(tmp, {"msw@2.14.3": True}, lock)
+        assert sync.main(["--fix", "--dir", str(tmp)]) == 0
+        got = json.loads((tmp / "package.json").read_text())["allowScripts"]
+        assert got == {"msw@2.14.3 || 2.15.0": True}
+
+
+def test_no_policy_is_noop():
+    with tempfile.TemporaryDirectory() as d:
+        tmp = Path(d)
+        (tmp / "package.json").write_text(json.dumps({"name": "fixture"}) + "\n")
+        (tmp / "package-lock.json").write_text(json.dumps({"packages": {}}) + "\n")
+        before = (tmp / "package.json").read_text()
+        assert sync.main(["--fix", "--dir", str(tmp)]) == 0
+        assert (tmp / "package.json").read_text() == before
+
+
+def test_missing_files_is_noop():
+    with tempfile.TemporaryDirectory() as d:
+        assert sync.main(["--check", "--dir", d]) == 0
+
+
+if __name__ == "__main__":
+    failures = 0
+    for name, fn in sorted(globals().items()):
+        if name.startswith("test_") and callable(fn):
+            try:
+                fn()
+                print(f"PASS {name}")
+            except AssertionError as e:
+                failures += 1
+                print(f"FAIL {name}: {e}")
+    sys.exit(1 if failures else 0)
diff --git a/tests/test_import_fixes_drift.py b/tests/test_import_fixes_drift.py
index 373830ae4a..6c75a4b886 100644
--- a/tests/test_import_fixes_drift.py
+++ b/tests/test_import_fixes_drift.py
@@ -584,3 +584,54 @@ def test_accelerate_utils_imports_module_present():
         "accelerate.utils.imports.is_wandb_available is gone; "
         "disable_broken_wandb cannot patch the source module."
     )
+
+
+# ===========================================================================
+# bitsandbytes -- ROCm arch / warp-size detection shape
+# ===========================================================================
+
+
+def test_bitsandbytes_rocm_detection_helpers_recognizable():
+    """``fix_bitsandbytes_rocm_arch_detection`` swaps bnb's ROCm helpers
+    only when they shell out via subprocess and never consult torch device
+    props; a third shape is declined by design, silently restoring Windows
+    ROCm noise. Fail so the sniff gets updated. Reads source, no import."""
+    spec = importlib.util.find_spec("bitsandbytes")
+    if spec is None:
+        pytest.skip("bitsandbytes not installed -- nothing to drift-check.")
+    cuda_specs_path = None
+    for location in spec.submodule_search_locations or []:
+        candidate = os.path.join(location, "cuda_specs.py")
+        if os.path.isfile(candidate):
+            cuda_specs_path = candidate
+            break
+    if cuda_specs_path is None:
+        pytest.skip("bitsandbytes has no cuda_specs.py (pre-ROCm version).")
+
+    import ast
+
+    with open(cuda_specs_path, "r", encoding = "utf-8") as f:
+        source = f.read()
+    helpers = [
+        node
+        for node in ast.walk(ast.parse(source))
+        if isinstance(node, ast.FunctionDef)
+        and node.name in ("get_rocm_gpu_arch", "get_rocm_warpsize")
+    ]
+    if not helpers:
+        pytest.skip("bitsandbytes cuda_specs has no ROCm detection helpers.")
+    for node in helpers:
+        segment = ast.get_source_segment(source, node) or ""
+        recognized = (
+            "subprocess" in segment
+            or "get_device_properties" in segment
+            or "gcnArchName" in segment
+        )
+        if not recognized:
+            pytest.fail(
+                f"DRIFT DETECTED: bitsandbytes.cuda_specs.{node.name} uses "
+                "neither subprocess nor torch device properties; "
+                "fix_bitsandbytes_rocm_arch_detection's shape sniff will "
+                "decline to patch it and Windows ROCm import-time noise / "
+                "wrong ROCM_GPU_ARCH may return."
+            )
diff --git a/tests/test_video_path_validation.py b/tests/test_video_path_validation.py
new file mode 100644
index 0000000000..d1406230e3
--- /dev/null
+++ b/tests/test_video_path_validation.py
@@ -0,0 +1,570 @@
+"""
+Tests for check_dataset_for_missing_videos (issue #5085).
+
+Fixtures extract the function from vision.py via AST so the pure-Python logic
+tests run without the full unsloth import chain (triton/CUDA kernels).
+"""
+
+import ast
+import os
+import tempfile
+import warnings
+from pathlib import Path
+
+import pytest
+
+
+# ── Fixtures ──────────────────────────────────────────────────────────────────
+
+
+def _extract_fns_via_ast(
+    source_path,
+    fn_names,
+    extra_ns = None,
+):
+    """Exec a set of top-level functions out of a .py file so intra-module
+    references between them resolve."""
+    source = source_path.read_text(encoding = "utf-8")
+    tree = ast.parse(source, filename = str(source_path))
+    wanted = set(fn_names)
+    nodes = [n for n in tree.body if isinstance(n, ast.FunctionDef) and n.name in wanted]
+    missing = wanted - {n.name for n in nodes}
+    if missing:
+        pytest.fail(f"{sorted(missing)} not found in {source_path}")
+    mini = ast.Module(body = nodes, type_ignores = [])
+    ast.fix_missing_locations(mini)
+    ns = {"os": os, "warnings": warnings, "__name__": "_extracted"}
+    if extra_ns:
+        ns.update(extra_ns)
+    exec(compile(mini, str(source_path), "exec"), ns)
+    return {name: ns[name] for name in fn_names}
+
+
+def _extract_fn_via_ast(
+    source_path,
+    fn_name,
+    extra_ns = None,
+):
+    return _extract_fns_via_ast(source_path, [fn_name], extra_ns)[fn_name]
+
+
+@pytest.fixture(scope = "session")
+def check_dataset_for_missing_videos():
+    """Direct import when possible, else AST extraction from vision.py."""
+    try:
+        from unsloth.models.vision import check_dataset_for_missing_videos as fn
+        return fn
+    except Exception:
+        pass
+
+    vision_path = Path(__file__).parent.parent / "unsloth" / "models" / "vision.py"
+    fns = _extract_fns_via_ast(
+        vision_path,
+        [
+            "_looks_like_message_list",
+            "_iter_message_lists",
+            "_local_path_from_video_value",
+            "check_dataset_for_missing_videos",
+        ],
+    )
+    return fns["check_dataset_for_missing_videos"]
+
+
+@pytest.fixture(scope = "session")
+def make_auto_validating_collator(check_dataset_for_missing_videos):
+    """Factory for a minimal collator mirroring the trainer.py wrapper."""
+
+    class _FakeBase:
+        def __init__(self, formatting_func = None):
+            self.formatting_func = formatting_func
+
+        def __call__(self, examples):
+            if self.formatting_func is not None:
+                examples = [self.formatting_func(e) for e in examples]
+            return {"ok": True, "examples": examples}
+
+    class _AutoValidatingCollator(_FakeBase):
+        def __init__(self, formatting_func = None):
+            super().__init__(formatting_func = formatting_func)
+            self._checked_video_paths = set()
+
+        def __call__(self, examples):
+            formatting_func = self.formatting_func
+            if formatting_func is not None:
+                examples = [formatting_func(e) for e in examples]
+            check_dataset_for_missing_videos(
+                examples,
+                raise_error = True,
+                checked = self._checked_video_paths,
+            )
+            if formatting_func is None:
+                return super().__call__(examples)
+            self.formatting_func = None
+            try:
+                return super().__call__(examples)
+            finally:
+                self.formatting_func = formatting_func
+
+    return _AutoValidatingCollator
+
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+
+def _make_video_dataset(*video_paths):
+    return [
+        {"messages": [{"role": "user", "content": [{"type": "video", "video": p}]}]}
+        for p in video_paths
+    ]
+
+
+def _batch(*video_paths):
+    return _make_video_dataset(*video_paths)
+
+
+# ── Tests: check_dataset_for_missing_videos ───────────────────────────────────
+
+
+def test_missing_local_file_raises(check_dataset_for_missing_videos):
+    """Missing local path raises FileNotFoundError."""
+    ds = _make_video_dataset("/nonexistent/videos/clip.mp4")
+    with pytest.raises(FileNotFoundError):
+        check_dataset_for_missing_videos(ds)
+
+
+def test_remote_url_skipped(check_dataset_for_missing_videos):
+    """http/https URLs are not checked locally."""
+    ds = _make_video_dataset("https://example.com/video.mp4")
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_existing_file_accepted(check_dataset_for_missing_videos):
+    """Existing local file passes without error."""
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"fake video bytes")
+        tmp = f.name
+    try:
+        ds = _make_video_dataset(tmp)
+        assert check_dataset_for_missing_videos(ds) == []
+    finally:
+        os.unlink(tmp)
+
+
+def test_file_uri_scheme_stripped(check_dataset_for_missing_videos):
+    """file:// scheme is stripped before the path check."""
+    ds = _make_video_dataset("file:///nonexistent/clip.mp4")
+    with pytest.raises(FileNotFoundError):
+        check_dataset_for_missing_videos(ds)
+
+
+def test_warn_only_mode(check_dataset_for_missing_videos):
+    """raise_error=False warns and returns the missing paths."""
+    ds = _make_video_dataset("/nonexistent/videos/clip.mp4")
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        missing = check_dataset_for_missing_videos(ds, raise_error = False)
+
+    assert len(caught) == 1
+    assert "could not be found" in str(caught[0].message)
+    assert missing == ["/nonexistent/videos/clip.mp4"]
+
+
+def test_duplicate_paths_deduplicated(check_dataset_for_missing_videos):
+    """Repeated missing path is listed once."""
+    ds = _make_video_dataset("/nonexistent/clip.mp4", "/nonexistent/clip.mp4")
+    with pytest.raises(FileNotFoundError) as exc_info:
+        check_dataset_for_missing_videos(ds)
+    assert str(exc_info.value).count("/nonexistent/clip.mp4") == 1
+
+
+# ── Tests: UnslothVisionDataCollator auto-validation ─────────────────────────
+
+
+def test_collator_raises_on_first_batch_with_missing_video(make_auto_validating_collator):
+    """Collator raises on a missing path with no user action needed."""
+    collator = make_auto_validating_collator()
+    batch = _batch("/nonexistent/auto/clip.mp4")
+    with pytest.raises(FileNotFoundError):
+        collator(batch)
+
+
+def test_collator_passes_on_first_batch_with_valid_video(make_auto_validating_collator):
+    """Collator passes a valid batch through."""
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"fake video bytes")
+        tmp = f.name
+    try:
+        collator = make_auto_validating_collator()
+        batch = _batch(tmp)
+        result = collator(batch)
+        assert result["ok"] is True
+    finally:
+        os.unlink(tmp)
+
+
+def test_collator_validates_every_batch(make_auto_validating_collator):
+    """A missing video first appearing after batch 0 must still raise."""
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"fake video bytes")
+        tmp = f.name
+    try:
+        collator = make_auto_validating_collator()
+        collator(_batch(tmp))  # batch 0: valid
+        with pytest.raises(FileNotFoundError):
+            collator(_batch("/nonexistent/late.mp4"))
+    finally:
+        os.unlink(tmp)
+
+
+def test_collator_dedupes_across_batches(make_auto_validating_collator):
+    """The checked-path set is shared across batches."""
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"fake video bytes")
+        tmp = f.name
+    try:
+        collator = make_auto_validating_collator()
+        collator(_batch(tmp))
+        collator(_batch(tmp, tmp))
+        assert tmp in collator._checked_video_paths
+    finally:
+        os.unlink(tmp)
+
+
+def test_conversations_column_missing_detected(check_dataset_for_missing_videos):
+    """'conversations' column is scanned."""
+    ds = [
+        {
+            "conversations": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": "/nonexistent/conv.mp4"}],
+                }
+            ]
+        },
+    ]
+    with pytest.raises(FileNotFoundError):
+        check_dataset_for_missing_videos(ds)
+
+
+def test_prompt_completion_column_missing_detected(check_dataset_for_missing_videos):
+    """'prompt'/'completion' columns are scanned."""
+    ds = [
+        {
+            "prompt": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": "/nonexistent/p.mp4"}],
+                }
+            ],
+            "completion": [{"role": "assistant", "content": [{"type": "text", "text": "hi"}]}],
+        },
+    ]
+    with pytest.raises(FileNotFoundError) as exc_info:
+        check_dataset_for_missing_videos(ds)
+    assert "/nonexistent/p.mp4" in str(exc_info.value)
+
+
+def test_raw_message_list_example_missing_detected(check_dataset_for_missing_videos):
+    """Rows that are themselves message lists (no outer dict) are scanned."""
+    ds = [
+        [
+            {
+                "role": "user",
+                "content": [{"type": "video", "video": "/nonexistent/raw.mp4"}],
+            }
+        ],
+    ]
+    with pytest.raises(FileNotFoundError):
+        check_dataset_for_missing_videos(ds)
+
+
+def test_non_dict_message_entry_does_not_crash(check_dataset_for_missing_videos):
+    """Non-dict message entries are skipped."""
+    ds = [{"messages": ["not a dict", {"role": "user", "content": []}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_file_uri_percent_encoded(check_dataset_for_missing_videos, tmp_path):
+    """Percent-encoded file:// URIs decode to the real path."""
+    target = tmp_path / "my video.mp4"
+    target.write_bytes(b"x")
+    uri = "file://" + str(target).replace(" ", "%20")
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": uri}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_file_uri_localhost_host(check_dataset_for_missing_videos, tmp_path):
+    """file://localhost/<abs path> is the local machine (RFC 8089)."""
+    target = tmp_path / "clip.mp4"
+    target.write_bytes(b"x")
+    uri = f"file://localhost{target}"
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": uri}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_checked_set_reused_across_calls(check_dataset_for_missing_videos, tmp_path):
+    """A supplied checked set is populated and deduped across calls."""
+    target = tmp_path / "clip.mp4"
+    target.write_bytes(b"x")
+    shared = set()
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": str(target)}]}]}]
+    check_dataset_for_missing_videos(ds, checked = shared)
+    assert str(target) in shared
+    check_dataset_for_missing_videos(ds, checked = shared)
+    assert len(shared) == 1
+
+
+@pytest.mark.parametrize(
+    "uri",
+    [
+        "s3://bucket/clip.mp4",
+        "gs://bucket/clip.mp4",
+        "hf://datasets/u/r/clip.mp4",
+        "ftp://host/clip.mp4",
+        "az://container/clip.mp4",
+    ],
+)
+def test_non_file_remote_scheme_skipped(check_dataset_for_missing_videos, uri):
+    """Non-file URI schemes are treated as remote and skipped."""
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": uri}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_file_uri_non_localhost_host_skipped(check_dataset_for_missing_videos):
+    """file://<non-localhost>/path is remote (RFC 8089): skip local checks."""
+    ds = [
+        {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": "file://nas-server/share/clip.mp4"}],
+                }
+            ]
+        }
+    ]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+@pytest.mark.parametrize("uri", ["file://", "file://hostname"])
+def test_degenerate_file_uri_skipped(check_dataset_for_missing_videos, uri):
+    """No path component must not produce a blank missing entry."""
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": uri}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_file_uri_double_encoded_percent(check_dataset_for_missing_videos, tmp_path):
+    """%2520 must single-unquote to 'clip%20.mp4', not 'clip .mp4'."""
+    target = tmp_path / "clip%20.mp4"
+    target.write_bytes(b"x")
+    uri = "file://" + str(target).replace("%", "%25")
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": uri}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_windows_style_absolute_path_not_mistaken_for_scheme(
+    check_dataset_for_missing_videos, tmp_path
+):
+    """'C:/...' has no '://' so it is a plain path, even where urlparse
+    would yield scheme='c'."""
+    target = tmp_path / "clip.mp4"
+    target.write_bytes(b"x")
+    path = str(target)
+    if os.name != "nt":
+        # '://'-free values must round-trip unchanged; keep the real path
+        path = str(target)
+    ds = [{"messages": [{"role": "user", "content": [{"type": "video", "video": path}]}]}]
+    assert check_dataset_for_missing_videos(ds) == []
+    ds_missing = [
+        {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": "C:/definitely/missing.mp4"}],
+                }
+            ]
+        }
+    ]
+    with pytest.raises(FileNotFoundError) as exc:
+        check_dataset_for_missing_videos(ds_missing)
+    assert "C:/definitely/missing.mp4" in str(exc.value)
+
+
+def test_iterable_dataset_warns_and_skips(check_dataset_for_missing_videos):
+    """Streaming IterableDataset: warn, return [], do not exhaust it."""
+    datasets_mod = pytest.importorskip("datasets", reason = "real datasets package required")
+    if not hasattr(datasets_mod, "IterableDataset"):
+        pytest.skip("datasets.IterableDataset not available in this environment")
+    IterableDataset = datasets_mod.IterableDataset
+
+    def gen():
+        for p in ("/nonexistent/a.mp4", "/nonexistent/b.mp4"):
+            yield {"messages": [{"role": "user", "content": [{"type": "video", "video": p}]}]}
+
+    ds = IterableDataset.from_generator(gen)
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        result = check_dataset_for_missing_videos(ds)
+    assert result == []
+    assert any("IterableDataset" in str(w.message) for w in caught)
+    # generator must not have been exhausted
+    consumed = list(ds)
+    assert len(consumed) == 2
+
+
+def test_collator_applies_formatting_func_before_validation(make_auto_validating_collator):
+    """formatting_func runs before validation; super gets formatted examples
+    and must not re-apply it."""
+
+    def fmt(example):
+        return {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": example["video_id"]}],
+                }
+            ]
+        }
+
+    raise_collator = make_auto_validating_collator(formatting_func = fmt)
+    with pytest.raises(FileNotFoundError):
+        raise_collator([{"video_id": "/nonexistent/formatted.mp4"}])
+
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"x")
+        tmp = f.name
+    try:
+        ok_collator = make_auto_validating_collator(formatting_func = fmt)
+        before = ok_collator.formatting_func
+        result = ok_collator([{"video_id": tmp}])
+        assert result["ok"] is True
+        assert ok_collator.formatting_func is before
+        passed = result["examples"]
+        assert passed[0]["messages"][0]["content"][0]["video"] == tmp
+    finally:
+        os.unlink(tmp)
+
+
+def test_data_uri_skipped(check_dataset_for_missing_videos):
+    """Inline data: URIs are not flagged missing."""
+    ds = _make_video_dataset("data:video/mp4;base64,AAAABBBBCCCC")
+    assert check_dataset_for_missing_videos(ds) == []
+
+
+def test_tuple_content_entries_checked(check_dataset_for_missing_videos):
+    """Tuple message content is validated like a list."""
+    ds = [
+        {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": ({"type": "video", "video": "/nonexistent/tuple.mp4"},),
+                }
+            ]
+        }
+    ]
+    with pytest.raises(FileNotFoundError):
+        check_dataset_for_missing_videos(ds)
+
+
+def test_duplicate_missing_deduped_in_warn_mode(check_dataset_for_missing_videos):
+    """Warn mode returns each missing path once."""
+    ds = _make_video_dataset("/nonexistent/dup.mp4", "/nonexistent/dup.mp4")
+    with warnings.catch_warnings(record = True):
+        warnings.simplefilter("always")
+        missing = check_dataset_for_missing_videos(ds, raise_error = False)
+    assert missing == ["/nonexistent/dup.mp4"]
+
+
+# ── Tests: real unsloth_zoo collator integration ─────────────────────────────
+# Exercise the real trainer.py subclass against the real zoo base (the fakes
+# above don't cover super()/formatting_func); skip when unsloth can't import.
+
+
+@pytest.fixture(scope = "session")
+def real_collator_classes():
+    try:
+        from unsloth.trainer import UnslothVisionDataCollator
+        from unsloth_zoo.vision_utils import (
+            UnslothVisionDataCollator as ZooBase,
+        )
+    except Exception as exc:  # noqa: BLE001 - skip on any import failure
+        pytest.skip(f"full unsloth import unavailable: {exc!r}")
+    return UnslothVisionDataCollator, ZooBase
+
+
+def _make_real_collator(real_collator_classes, formatting_func = None):
+    """Build the real subclass without its heavy __init__ (needs a processor)."""
+    subclass, _ = real_collator_classes
+    collator = subclass.__new__(subclass)
+    collator.formatting_func = formatting_func
+    collator._checked_video_paths = set()
+    return collator
+
+
+def test_real_collator_blocks_super_on_missing_video(real_collator_classes, monkeypatch):
+    """Missing path raises before the base __call__ runs."""
+    _, zoo_base = real_collator_classes
+    calls = []
+    monkeypatch.setattr(zoo_base, "__call__", lambda self, examples: calls.append(examples))
+    collator = _make_real_collator(real_collator_classes)
+    with pytest.raises(FileNotFoundError):
+        collator(_batch("/nonexistent/real.mp4"))
+    assert calls == []  # base collator was never reached
+
+
+def test_real_collator_calls_super_with_formatting_disabled(real_collator_classes, monkeypatch):
+    """Base must see formatting_func=None and already-formatted examples;
+    the original formatting_func is restored afterwards."""
+    seen = {}
+
+    def spy(self, examples):
+        seen["formatting_func"] = self.formatting_func
+        seen["examples"] = examples
+        return {"ok": True}
+
+    _, zoo_base = real_collator_classes
+    monkeypatch.setattr(zoo_base, "__call__", spy)
+
+    def fmt(example):
+        return {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [{"type": "video", "video": example["video_id"]}],
+                }
+            ]
+        }
+
+    with tempfile.NamedTemporaryFile(suffix = ".mp4", delete = False) as f:
+        f.write(b"x")
+        tmp = f.name
+    try:
+        collator = _make_real_collator(real_collator_classes, formatting_func = fmt)
+        result = collator([{"video_id": tmp}])
+        assert result == {"ok": True}
+        assert seen["formatting_func"] is None
+        assert seen["examples"][0]["messages"][0]["content"][0]["video"] == tmp
+        assert collator.formatting_func is fmt
+        assert tmp in collator._checked_video_paths
+    finally:
+        os.unlink(tmp)
+
+
+def test_real_collator_restores_formatting_func_when_super_raises(
+    real_collator_classes, monkeypatch
+):
+    """formatting_func is restored even when the base raises."""
+
+    def boom(self, examples):
+        raise RuntimeError("base collator failed")
+
+    _, zoo_base = real_collator_classes
+    monkeypatch.setattr(zoo_base, "__call__", boom)
+
+    def fmt(example):
+        return {"messages": [{"role": "user", "content": [{"type": "text", "text": "hi"}]}]}
+
+    collator = _make_real_collator(real_collator_classes, formatting_func = fmt)
+    with pytest.raises(RuntimeError):
+        collator([{"anything": 1}])
+    assert collator.formatting_func is fmt
diff --git a/tests/test_windows_rocm_bnb_version.py b/tests/test_windows_rocm_bnb_version.py
new file mode 100644
index 0000000000..c8439b1fcd
--- /dev/null
+++ b/tests/test_windows_rocm_bnb_version.py
@@ -0,0 +1,251 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""Tests for ``maybe_set_windows_rocm_bnb_version`` (unsloth/import_fixes.py).
+
+The module is loaded in isolation (stdlib + packaging only), so no torch /
+GPU is required and unsloth's GPU init never runs."""
+
+from __future__ import annotations
+
+import importlib.util
+import os
+import types
+from pathlib import Path
+
+import pytest
+
+_IMPORT_FIXES_PATH = Path(__file__).resolve().parent.parent / "unsloth" / "import_fixes.py"
+
+
+def _load_import_fixes():
+    spec = importlib.util.spec_from_file_location(
+        "unsloth_import_fixes_under_test", _IMPORT_FIXES_PATH
+    )
+    module = importlib.util.module_from_spec(spec)
+    assert spec is not None and spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture()
+def import_fixes():
+    return _load_import_fixes()
+
+
+@pytest.fixture()
+def clean_env(monkeypatch):
+    """Unset the env vars and remove them afterwards (the function writes
+    os.environ directly, which monkeypatch does not auto-revert)."""
+    for var in (
+        "BNB_ROCM_VERSION",
+        "UNSLOTH_SKIP_BNB_ROCM_VERSION",
+        "UNSLOTH_BNB_ROCM_VERSION_SOURCE",
+    ):
+        monkeypatch.delenv(var, raising = False)
+    yield monkeypatch
+    for var in (
+        "BNB_ROCM_VERSION",
+        "UNSLOTH_SKIP_BNB_ROCM_VERSION",
+        "UNSLOTH_BNB_ROCM_VERSION_SOURCE",
+    ):
+        os.environ.pop(var, None)
+
+
+def _force(import_fixes, monkeypatch, *, win, rocm, detected):
+    monkeypatch.setattr(import_fixes.sys, "platform", "win32" if win else "linux")
+    monkeypatch.setattr(import_fixes, "_is_hip_torch_build", lambda: rocm)
+    monkeypatch.setattr(import_fixes, "_detect_installed_bnb_rocm_version", lambda: detected)
+
+
+# ---------------------------------------------------------------------------
+# _detect_installed_bnb_rocm_version
+# ---------------------------------------------------------------------------
+
+
+def test_detect_picks_highest_rocm_suffix(import_fixes, tmp_path, monkeypatch):
+    pkg = tmp_path / "bitsandbytes"
+    pkg.mkdir()
+    for name in (
+        "libbitsandbytes_rocm72.dll",
+        "libbitsandbytes_rocm713.dll",  # numerically highest -> should win
+        "libbitsandbytes_cpu.dll",
+        "__init__.py",
+    ):
+        (pkg / name).write_text("")
+    fake_spec = types.SimpleNamespace(submodule_search_locations = [str(pkg)])
+    monkeypatch.setattr(importlib.util, "find_spec", lambda name: fake_spec)
+    assert import_fixes._detect_installed_bnb_rocm_version() == "713"
+
+
+def test_detect_none_when_only_non_rocm_dlls(import_fixes, tmp_path, monkeypatch):
+    pkg = tmp_path / "bitsandbytes"
+    pkg.mkdir()
+    (pkg / "libbitsandbytes_cpu.dll").write_text("")
+    (pkg / "libbitsandbytes_cuda124.dll").write_text("")
+    fake_spec = types.SimpleNamespace(submodule_search_locations = [str(pkg)])
+    monkeypatch.setattr(importlib.util, "find_spec", lambda name: fake_spec)
+    assert import_fixes._detect_installed_bnb_rocm_version() is None
+
+
+def test_detect_none_when_bnb_absent(import_fixes, monkeypatch):
+    monkeypatch.setattr(importlib.util, "find_spec", lambda name: None)
+    assert import_fixes._detect_installed_bnb_rocm_version() is None
+
+
+# ---------------------------------------------------------------------------
+# maybe_set_windows_rocm_bnb_version
+# ---------------------------------------------------------------------------
+
+
+def test_sets_bnb_version_on_windows_rocm(import_fixes, clean_env):
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() == "72"
+    assert os.environ["BNB_ROCM_VERSION"] == "72"
+    assert os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] == "detected"
+
+
+def test_noop_off_windows(import_fixes, clean_env):
+    # Linux ROCm resolves its backend correctly from torch.version.hip.
+    _force(import_fixes, clean_env, win = False, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert "BNB_ROCM_VERSION" not in os.environ
+
+
+def test_noop_when_not_rocm_torch(import_fixes, clean_env):
+    _force(import_fixes, clean_env, win = True, rocm = False, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert "BNB_ROCM_VERSION" not in os.environ
+
+
+def test_noop_when_no_rocm_dll_installed(import_fixes, clean_env):
+    # Never force a ROCm backend name when no ROCm DLL ships (avoid breaking a
+    # non-ROCm bitsandbytes that happens to sit next to a ROCm torch build).
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = None)
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert "BNB_ROCM_VERSION" not in os.environ
+
+
+def test_respects_user_provided_value(import_fixes, clean_env):
+    clean_env.setenv("BNB_ROCM_VERSION", "999")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert os.environ["BNB_ROCM_VERSION"] == "999"
+
+
+def test_explicit_opt_out(import_fixes, clean_env):
+    clean_env.setenv("UNSLOTH_SKIP_BNB_ROCM_VERSION", "1")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert "BNB_ROCM_VERSION" not in os.environ
+
+
+def test_redetects_sitecustomize_seeded_default(import_fixes, clean_env):
+    # Studio's installer persists a default via the venv sitecustomize.py; the
+    # wheel may have changed since, so the seeded value must be redetected.
+    clean_env.setenv("BNB_ROCM_VERSION", "72")
+    clean_env.setenv("UNSLOTH_BNB_ROCM_VERSION_SOURCE", "sitecustomize")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "713")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() == "713"
+    assert os.environ["BNB_ROCM_VERSION"] == "713"
+    assert os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] == "detected"
+
+
+def test_sitecustomize_default_kept_when_no_dll_found(import_fixes, clean_env):
+    # A failed redetect must not discard the seeded value.
+    clean_env.setenv("BNB_ROCM_VERSION", "72")
+    clean_env.setenv("UNSLOTH_BNB_ROCM_VERSION_SOURCE", "sitecustomize")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = None)
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert os.environ["BNB_ROCM_VERSION"] == "72"
+    assert os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] == "sitecustomize"
+
+
+def test_user_value_with_non_sitecustomize_marker_untouched(import_fixes, clean_env):
+    # Only the sitecustomize marker makes a value redetectable.
+    clean_env.setenv("BNB_ROCM_VERSION", "999")
+    clean_env.setenv("UNSLOTH_BNB_ROCM_VERSION_SOURCE", "detected")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert os.environ["BNB_ROCM_VERSION"] == "999"
+
+
+def test_opt_out_unseats_sitecustomize_seeded_value(import_fixes, clean_env):
+    # The opt-out must also drop a default our own sitecustomize block seeded,
+    # so bitsandbytes never sees the override the user disabled.
+    clean_env.setenv("BNB_ROCM_VERSION", "72")
+    clean_env.setenv("UNSLOTH_BNB_ROCM_VERSION_SOURCE", "sitecustomize")
+    clean_env.setenv("UNSLOTH_SKIP_BNB_ROCM_VERSION", "1")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "713")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert "BNB_ROCM_VERSION" not in os.environ
+    assert "UNSLOTH_BNB_ROCM_VERSION_SOURCE" not in os.environ
+
+
+def test_opt_out_keeps_explicit_user_value(import_fixes, clean_env):
+    # Opt-out must never remove a value the user set themselves (no marker).
+    clean_env.setenv("BNB_ROCM_VERSION", "999")
+    clean_env.setenv("UNSLOTH_SKIP_BNB_ROCM_VERSION", "1")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert os.environ["BNB_ROCM_VERSION"] == "999"
+
+
+def test_empty_string_value_without_marker_is_respected(import_fixes, clean_env):
+    # "" counts as present: without the sitecustomize marker it is not ours
+    # to overwrite.
+    clean_env.setenv("BNB_ROCM_VERSION", "")
+    _force(import_fixes, clean_env, win = True, rocm = True, detected = "72")
+    assert import_fixes.maybe_set_windows_rocm_bnb_version() is None
+    assert os.environ["BNB_ROCM_VERSION"] == ""
+
+
+# ---------------------------------------------------------------------------
+# _is_hip_torch_build (the strict gate -- regression for the HIP-SDK-on-a-
+# CUDA-box false positive: env hints like HIP_PATH must NOT count)
+# ---------------------------------------------------------------------------
+
+
+def _fake_torch(hip):
+    return types.SimpleNamespace(version = types.SimpleNamespace(hip = hip))
+
+
+def test_hip_build_true_from_wheel_tag(import_fixes, monkeypatch):
+    monkeypatch.setattr(import_fixes, "importlib_version", lambda name: "2.11.0+rocm7.13.0")
+    assert import_fixes._is_hip_torch_build() is True
+
+
+def test_hip_build_true_from_torch_version_hip(import_fixes, monkeypatch):
+    # Custom/source HIP build without the +rocm tag.
+    monkeypatch.setattr(import_fixes, "importlib_version", lambda name: "2.11.0")
+    monkeypatch.setitem(__import__("sys").modules, "torch", _fake_torch("7.2.0"))
+    assert import_fixes._is_hip_torch_build() is True
+
+
+def test_hip_build_false_for_cuda_torch_despite_rocm_env_hints(import_fixes, monkeypatch):
+    """HIP SDK env vars set but CUDA torch: the strict gate must say False,
+    otherwise BNB_ROCM_VERSION gets set and CUDA bitsandbytes raises."""
+    monkeypatch.setenv("HIP_PATH", r"C:\Program Files\AMD\ROCm\6.2")
+    monkeypatch.setenv("ROCM_PATH", r"C:\Program Files\AMD\ROCm\6.2")
+    monkeypatch.setattr(import_fixes, "importlib_version", lambda name: "2.9.0+cu126")
+    monkeypatch.setitem(__import__("sys").modules, "torch", _fake_torch(None))
+    assert import_fixes._is_hip_torch_build() is False
+
+
+def test_hip_build_false_when_torch_absent(import_fixes, monkeypatch):
+    def _raise(name):
+        raise Exception("no torch dist")
+
+    monkeypatch.setattr(import_fixes, "importlib_version", _raise)
+    monkeypatch.setitem(__import__("sys").modules, "torch", None)
+    assert import_fixes._is_hip_torch_build() is False
diff --git a/tests/utils/test_batched_leftpad_generation_gpu.py b/tests/utils/test_batched_leftpad_generation_gpu.py
new file mode 100644
index 0000000000..c2610d4673
--- /dev/null
+++ b/tests/utils/test_batched_leftpad_generation_gpu.py
@@ -0,0 +1,107 @@
+"""End-to-end GPU guard for batched left-padded generation (issues #1066, #3699).
+
+For each prompt, greedy generation inside a left-padded batch must match
+generating that prompt alone at batch size 1 for the first PREFIX_TOKENS
+tokens, and the full output must not be gibberish. The bug class (#1066,
+#3699) makes padded rows diverge immediately into garbage; in contrast,
+benign batch-size-dependent kernel numerics can flip a greedy near-tie deep
+into the sequence, so an exact full-length match would be flaky. Uses a small
+instruct model (chat-templated prompts have high-margin argmaxes).
+
+Skipped automatically when CUDA is unavailable, so CPU CI is unaffected.
+Run manually on any GPU box:
+
+    python -m pytest tests/utils/test_batched_leftpad_generation_gpu.py -v
+"""
+
+import pytest
+import torch
+
+cuda_available = torch.cuda.is_available()
+
+pytestmark = pytest.mark.skipif(not cuda_available, reason = "requires a CUDA GPU")
+
+MODEL_NAME = "unsloth/Qwen2.5-0.5B-Instruct"
+MAX_NEW_TOKENS = 32
+PREFIX_TOKENS = 16
+
+PROMPTS = [
+    "Give me a short introduction to large language model.",
+    "Here is an experiment log: "
+    + " ".join(f"run {i} completed with stable throughput and no anomalies;" for i in range(1, 41))
+    + " In one sentence, what is the overall conclusion?",
+]
+
+
+@pytest.fixture(scope = "module")
+def model_and_tokenizer():
+    from unsloth import FastLanguageModel
+
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name = MODEL_NAME,
+        max_seq_length = 2048,
+        load_in_4bit = True,
+    )
+    FastLanguageModel.for_inference(model)
+    tokenizer.padding_side = "left"
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token_id = tokenizer.eos_token_id
+    return model, tokenizer
+
+
+def _chat(tokenizer, prompt):
+    return tokenizer.apply_chat_template(
+        [{"role": "user", "content": prompt}],
+        tokenize = False,
+        add_generation_prompt = True,
+    )
+
+
+def _generate(model, tokenizer, texts):
+    inputs = tokenizer(texts, return_tensors = "pt", padding = True, add_special_tokens = False).to(
+        "cuda"
+    )
+    with torch.inference_mode():
+        out = model.generate(
+            **inputs,
+            max_new_tokens = MAX_NEW_TOKENS,
+            do_sample = False,
+            temperature = None,
+            top_p = None,
+            top_k = None,
+            use_cache = True,
+            pad_token_id = tokenizer.pad_token_id,
+        )
+    suffixes = out[:, inputs["input_ids"].shape[1] :]
+    return [row.tolist() for row in suffixes]
+
+
+def _looks_gibberish(text):
+    if not text.strip():
+        return True
+    exclam = text.count("!") / max(len(text), 1)
+    nonascii = sum(1 for c in text if ord(c) > 0x2FFF) / max(len(text), 1)
+    return exclam > 0.3 or nonascii > 0.5
+
+
+def test_batched_leftpad_matches_solo_generation(model_and_tokenizer):
+    model, tokenizer = model_and_tokenizer
+
+    texts = [_chat(tokenizer, p) for p in PROMPTS]
+    solo = [_generate(model, tokenizer, [t])[0] for t in texts]
+    batched = _generate(model, tokenizer, texts)
+
+    for i, prompt in enumerate(PROMPTS):
+        solo_text = tokenizer.decode(solo[i], skip_special_tokens = True)
+        batch_text = tokenizer.decode(batched[i], skip_special_tokens = True)
+        assert batched[i][:PREFIX_TOKENS] == solo[i][:PREFIX_TOKENS], (
+            f"prompt {i} ({prompt[:30]!r}...) diverged from solo generation "
+            f"within the first {PREFIX_TOKENS} tokens inside a left-padded "
+            "batch; batched left-padded generation is broken again "
+            f"(issues #1066, #3699).\n"
+            f"solo   : {solo_text!r}\nbatched: {batch_text!r}"
+        )
+        assert not _looks_gibberish(batch_text), (
+            f"prompt {i} produced gibberish in a left-padded batch "
+            f"(issues #1066, #3699): {batch_text!r}"
+        )
diff --git a/tests/utils/test_prepare_inputs_leftpad.py b/tests/utils/test_prepare_inputs_leftpad.py
new file mode 100644
index 0000000000..b773b088b5
--- /dev/null
+++ b/tests/utils/test_prepare_inputs_leftpad.py
@@ -0,0 +1,409 @@
+"""Regression guard for batched left-padded generation (issues #1066, #3699).
+
+`_fast_prepare_inputs_for_generation` in unsloth/models/llama.py is shared by
+every decoder family wired through `fix_prepare_inputs_for_generation` (llama,
+qwen2/3, qwen3_moe, mistral, gemma/2, cohere, granite). Two historical bugs
+lived in it:
+
+  (a) the 2D attention mask was truncated to its last column during cached
+      decode, losing padding information (introduced cc4c5d77, fixed by #2216);
+  (b) position_ids were taken directly from cache_position, a global counter
+      that includes left-pad tokens, so padded rows generated garbage
+      (introduced cc4c5d77, reported in #1066/#3699, fixed by #4100).
+
+Two layers in this file, both CPU-only and deterministic:
+
+  1. AST structural checks (TestAstGuard section): parse llama.py with the
+     stdlib `ast` module only, no unsloth import, so they keep working even
+     when the package itself fails to import.
+  2. Behavioral checks: call the real function with synthetic left-padded
+     masks and fake caches; unsloth is imported lazily inside each test.
+
+Both layers fail on the historical bug patterns (validated against the code
+states immediately before #2216 and before #4100). Staging proof on GPU-less
+hosted runners: danielhanchen/unsloth-staging-2 PRs 170 (green) and 171 (red).
+
+Companion manual GPU end-to-end check (skipped without CUDA):
+tests/utils/test_batched_leftpad_generation_gpu.py
+"""
+
+import ast
+from pathlib import Path
+
+import pytest
+import torch
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+LLAMA_PY = REPO_ROOT / "unsloth" / "models" / "llama.py"
+
+FUNC_NAME = "_fast_prepare_inputs_for_generation"
+
+
+# --------------------------------------------------------------------------
+# Layer 1: AST structural guard (stdlib only, no unsloth import)
+# --------------------------------------------------------------------------
+
+# Model files that call fix_prepare_inputs_for_generation(...) and therefore
+# share the guarded function. glm4_moe and falcon_h1 are intentionally absent:
+# GLM4 MoE does not patch the Llama-compatible generation path (MLA attention)
+# and falcon_h1 ships its own _fast_prepare_inputs_for_generation variant.
+WIRED_MODEL_FILES = [
+    "mistral.py",
+    "gemma.py",
+    "gemma2.py",
+    "qwen2.py",
+    "qwen3.py",
+    "qwen3_moe.py",
+    "cohere.py",
+    "granite.py",
+]
+
+
+def _load_function():
+    tree = ast.parse(LLAMA_PY.read_text())
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == FUNC_NAME:
+            return node
+    raise AssertionError(
+        f"{FUNC_NAME} not found in {LLAMA_PY}; if it was renamed or moved, "
+        "update this guard so batched left-padded generation stays protected"
+    )
+
+
+def _names_in(node):
+    """All Name ids, attribute names and string constants in a subtree."""
+    found = set()
+    for sub in ast.walk(node):
+        if isinstance(sub, ast.Name):
+            found.add(sub.id)
+        elif isinstance(sub, ast.Attribute):
+            found.add(sub.attr)
+        elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
+            found.add(sub.value)
+    return found
+
+
+def _mentions_attention_mask(node):
+    return any("attention_mask" in name for name in _names_in(node))
+
+
+def _is_kwargs_position_ids_target(target):
+    return (
+        isinstance(target, ast.Subscript)
+        and isinstance(target.value, ast.Name)
+        and target.value.id == "kwargs"
+        and isinstance(target.slice, ast.Constant)
+        and target.slice.value == "position_ids"
+    )
+
+
+def _walk_with_paths(node, path = ()):
+    yield node, path
+    for child in ast.iter_child_nodes(node):
+        yield from _walk_with_paths(child, path + (node,))
+
+
+def _find_mask_branch(func):
+    """The If whose test checks the 2D attention mask (dim() == 2)."""
+    for node in ast.walk(func):
+        if not isinstance(node, ast.If):
+            continue
+        test_names = _names_in(node.test)
+        if "dim" in test_names and any("attention_mask" in n for n in test_names):
+            return node
+    return None
+
+
+def test_mask_derived_position_ids_branch_exists():
+    func = _load_function()
+    branch = _find_mask_branch(func)
+    assert branch is not None, (
+        f"{FUNC_NAME} no longer has a branch testing the 2D attention mask "
+        "(dim() == 2); position_ids must be derived per row from the mask for "
+        "left-padded batches (see PR #4100 / issues #1066, #3699)"
+    )
+
+    body_names = set()
+    for stmt in branch.body:
+        body_names |= _names_in(stmt)
+    assert "cumsum" in body_names and _mentions_attention_mask(
+        ast.Module(body = branch.body, type_ignores = [])
+    ), (
+        "the attention-mask branch must compute position_ids via "
+        "attention_mask.cumsum(...); reintroducing cache_position-based "
+        "positions breaks left-padded batched generation (issue #3699)"
+    )
+    assert (
+        "masked_fill_" in body_names or "masked_fill" in body_names
+    ), "the attention-mask branch must mask pad positions (masked_fill on mask == 0)"
+    assigns_kwargs = any(
+        isinstance(stmt, ast.Assign)
+        and any(_is_kwargs_position_ids_target(t) for t in stmt.targets)
+        for stmt in ast.walk(ast.Module(body = branch.body, type_ignores = []))
+    )
+    assert (
+        assigns_kwargs
+    ), 'the attention-mask branch must store the derived positions into kwargs["position_ids"]'
+
+
+def test_cache_position_only_used_as_fallback_for_position_ids():
+    func = _load_function()
+    branch = _find_mask_branch(func)
+    assert branch is not None
+
+    orelse_nodes = set()
+    for stmt in branch.orelse:
+        for sub in ast.walk(stmt):
+            orelse_nodes.add(id(sub))
+
+    offenders = []
+    for node, path in _walk_with_paths(func):
+        if not isinstance(node, ast.Assign):
+            continue
+        if not any(_is_kwargs_position_ids_target(t) for t in node.targets):
+            continue
+        value_names = _names_in(node.value)
+        # Direct use of cache_position, or the local alias `cp` the current
+        # implementation builds from it inside the fallback branch.
+        derives_from_cache_position = any("cache_position" in n for n in value_names) or bool(
+            value_names & {"cp"}
+        )
+        if derives_from_cache_position and id(node) not in orelse_nodes:
+            offenders.append(ast.unparse(node))
+
+    assert not offenders, (
+        'kwargs["position_ids"] must never be assigned from cache_position '
+        "outside the fallback (orelse) of the 2D attention-mask branch; "
+        "cache_position counts left-pad tokens, so padded rows generate "
+        f"garbage (issues #1066, #3699). Offending assignments: {offenders}"
+    )
+
+
+def test_attention_mask_never_truncated_to_last_column():
+    func = _load_function()
+    offenders = []
+    for node in ast.walk(func):
+        if not isinstance(node, ast.Assign):
+            continue
+        value = node.value
+        if not isinstance(value, ast.Subscript):
+            continue
+        if not _mentions_attention_mask(value.value):
+            continue
+        # Match a trailing [-1]-style column selection: [:, [-1]] or [:, -1:]
+        sl = value.slice
+        if isinstance(sl, ast.Tuple) and len(sl.elts) == 2:
+            col = sl.elts[1]
+            is_last_col_list = (
+                isinstance(col, ast.List)
+                and len(col.elts) == 1
+                and isinstance(col.elts[0], ast.UnaryOp)
+            )
+            is_last_col_slice = (
+                isinstance(col, ast.Slice)
+                and col.lower is not None
+                and isinstance(col.lower, ast.UnaryOp)
+                and getattr(getattr(col.lower, "operand", None), "value", None) == 1
+                and col.upper is None
+            )
+            if is_last_col_list or is_last_col_slice:
+                offenders.append(ast.unparse(node))
+    assert not offenders, (
+        "the 2D attention mask must not be truncated to its last column; this "
+        "was the pre-#2216 bug that drops padding information in cached decode "
+        f"(issue #1066). Offending assignments: {offenders}"
+    )
+
+
+def test_model_families_stay_wired_to_shared_prepare_inputs():
+    missing = []
+    for fname in WIRED_MODEL_FILES:
+        path = REPO_ROOT / "unsloth" / "models" / fname
+        if not path.exists():
+            continue
+        if "fix_prepare_inputs_for_generation(" not in path.read_text():
+            missing.append(fname)
+    assert not missing, (
+        "these model files no longer call fix_prepare_inputs_for_generation, "
+        "so they lose the guarded left-padding-safe prepare_inputs path: "
+        f"{missing}"
+    )
+
+
+# --------------------------------------------------------------------------
+# Layer 2: behavioral guard (calls the real function, lazy unsloth import)
+# --------------------------------------------------------------------------
+
+PAST_LEN = 4
+
+# Three rows with different amounts of left padding (0 = pad).
+MASK = torch.tensor(
+    [
+        [0, 0, 1, 1, 1],
+        [1, 1, 1, 1, 1],
+        [0, 1, 1, 1, 1],
+    ],
+    dtype = torch.long,
+)
+BS, SEQ = MASK.shape
+
+# Per-row positions: cumsum(-1) - 1 with pad slots filled with 1.
+EXPECTED_PREFILL_POSITIONS = torch.tensor(
+    [
+        [1, 1, 0, 1, 2],
+        [0, 1, 2, 3, 4],
+        [1, 0, 1, 2, 3],
+    ],
+    dtype = torch.long,
+)
+
+
+class FakeDynamicCache:
+    """Minimal stand-in for transformers DynamicCache with a non-empty cache."""
+
+    def __init__(self, seq_length):
+        self._seq_length = seq_length
+
+    def __len__(self):
+        return 1
+
+    def get_seq_length(self):
+        return self._seq_length
+
+
+class FakeModel:
+    """Bare-minimum `self` for _fast_prepare_inputs_for_generation."""
+
+    dtype = torch.float32
+    config = None
+
+
+class FakeModelWith4DMask(FakeModel):
+    """Variant exposing the HF 4D mask builder; records what it receives."""
+
+    def __init__(self):
+        self.mask_calls = []
+
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        self,
+        attention_mask,
+        sequence_length,
+        target_length,
+        dtype,
+        device,
+        cache_position,
+        batch_size,
+        config = None,
+        past_key_values = None,
+    ):
+        self.mask_calls.append(
+            {
+                "mask_shape": tuple(attention_mask.shape) if attention_mask is not None else None,
+                "sequence_length": sequence_length,
+                "target_length": target_length,
+                "batch_size": batch_size,
+            }
+        )
+        return torch.zeros((batch_size, 1, sequence_length, target_length), dtype = dtype)
+
+
+def _prepare(model, input_ids, attention_mask, **kwargs):
+    from unsloth.models import llama as llama_mod
+    return llama_mod._fast_prepare_inputs_for_generation(
+        model, input_ids, attention_mask = attention_mask, **kwargs
+    )
+
+
+def test_prefill_position_ids_derived_from_left_padded_mask():
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    result = _prepare(FakeModel(), input_ids, MASK)
+
+    position_ids = result.get("position_ids", None)
+    assert (
+        position_ids is not None
+    ), "prefill with a left-padded 2D attention mask must populate position_ids"
+    assert torch.equal(position_ids.long().cpu(), EXPECTED_PREFILL_POSITIONS), (
+        "prefill position_ids must be derived per row from the attention mask "
+        "(cumsum - 1, pads masked), so each row starts counting at its first "
+        f"real token; got {position_ids.tolist()}"
+    )
+    assert result["input_ids"].shape == (BS, SEQ)
+
+
+@pytest.mark.parametrize("pass_cache_position", [True, False])
+def test_cached_decode_position_ids_ignore_left_padding(pass_cache_position):
+    # Decode step: PAST_LEN tokens cached, current token is the mask's last
+    # column. Row 0 has 2 pads, so its current token sits at logical position 2,
+    # NOT at cache_position == PAST_LEN. This is exactly issue #3699.
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    kwargs = {"past_key_values": FakeDynamicCache(PAST_LEN)}
+    if pass_cache_position:
+        kwargs["cache_position"] = torch.arange(PAST_LEN, PAST_LEN + 1)
+
+    result = _prepare(FakeModel(), input_ids, MASK, **kwargs)
+
+    assert result["input_ids"].shape == (
+        BS,
+        1,
+    ), "cached decode must slice input_ids to the last token only"
+    position_ids = result.get("position_ids", None)
+    assert position_ids is not None
+    expected = torch.tensor([[2], [4], [3]], dtype = torch.long)
+    assert torch.equal(position_ids.long().cpu().reshape(BS, 1), expected), (
+        "left-padded cached decode must derive per-row position_ids from the "
+        "attention mask, not from cache_position which counts pad tokens; got "
+        f"{position_ids.tolist()}, expected {expected.tolist()} "
+        "(row 0 has 2 pads: its position must be 2, not 4)"
+    )
+
+
+def test_cached_decode_does_not_truncate_2d_attention_mask():
+    # Without a 4D mask builder the original 2D mask must survive untouched.
+    # The historical bug replaced it with attention_mask[:, [-1]].
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    result = _prepare(FakeModel(), input_ids, MASK, past_key_values = FakeDynamicCache(PAST_LEN))
+    mask_out = result["attention_mask"]
+    assert mask_out is not None
+    assert mask_out.dim() != 2 or mask_out.shape[-1] == SEQ, (
+        "the 2D attention mask must not be truncated to its last column during "
+        f"cached decode (got shape {tuple(mask_out.shape)}); padding rows lose "
+        "their pad information otherwise"
+    )
+
+
+def test_cached_decode_4d_mask_builder_receives_full_target_length():
+    model = FakeModelWith4DMask()
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    result = _prepare(model, input_ids, MASK, past_key_values = FakeDynamicCache(PAST_LEN))
+    assert len(model.mask_calls) == 1
+    call = model.mask_calls[0]
+    assert call["mask_shape"] == (BS, SEQ), (
+        "the 4D mask builder must receive the full 2D padding mask, not a "
+        f"truncated one (got {call['mask_shape']})"
+    )
+    assert call["sequence_length"] == 1
+    assert call["target_length"] == SEQ, (
+        "target_length must cover the whole mask so padded positions stay "
+        f"masked (got {call['target_length']})"
+    )
+    assert result["attention_mask"].dim() == 4
+
+
+def test_caller_supplied_position_ids_are_passed_through():
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    custom = torch.full((BS, SEQ), 7, dtype = torch.long)
+    result = _prepare(FakeModel(), input_ids, MASK, position_ids = custom)
+    assert torch.equal(
+        result["position_ids"], custom
+    ), "caller-supplied position_ids must not be overwritten"
+
+
+def test_legacy_tuple_cache_still_takes_cached_decode_path():
+    # Legacy cache format: tuple of (K, V) per layer; past length from K.shape[-2].
+    k = torch.zeros((BS, 1, PAST_LEN, 8))
+    legacy_cache = ((k, k.clone()),)
+    input_ids = torch.arange(BS * SEQ).reshape(BS, SEQ)
+    result = _prepare(FakeModel(), input_ids, MASK, past_key_values = legacy_cache)
+    assert result["input_ids"].shape == (BS, 1)
+    expected = torch.tensor([[2], [4], [3]], dtype = torch.long)
+    assert torch.equal(result["position_ids"].long().cpu().reshape(BS, 1), expected)
diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
new file mode 100644
index 0000000000..fae5a7d8a8
--- /dev/null
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -0,0 +1,312 @@
+"""Guard for config.rope_scaling being silently dropped (issue #2405).
+
+Unsloth's replacement rotary classes ignored rope_scaling when constructed
+from a config (the modern-transformers path), so Llama-3.1 ran with unscaled
+RoPE and collapsed into gibberish past ~32K tokens.
+
+Layers: (1) AST tripwire, stdlib only; (2) CPU checks of the pure helper
+_compute_config_rope_inv_freq against transformers' ROPE_INIT_FUNCTIONS;
+(3) CUDA checks instantiating the real class (skipped without a real device,
+probed by allocating a tensor so import-time CUDA spoofs cannot fool the gate).
+Layers 2 and 3 fail on the unfixed code.
+"""
+
+import ast
+import math
+from pathlib import Path
+
+import pytest
+import torch
+
+
+def _has_real_cuda():
+    try:
+        torch.zeros(1).to("cuda")
+        return True
+    except Exception:
+        return False
+
+
+HAS_REAL_CUDA = _has_real_cuda()
+requires_cuda = pytest.mark.skipif(
+    not HAS_REAL_CUDA,
+    reason = "LlamaRotaryEmbedding builds per-device CUDA caches in __init__",
+)
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+LLAMA_PY = REPO_ROOT / "unsloth" / "models" / "llama.py"
+
+CLASS_NAME = "LlamaRotaryEmbedding"
+
+# Llama-3.1-style rope_scaling.
+LLAMA3_ROPE_SCALING = {
+    "rope_type": "llama3",
+    "factor": 8.0,
+    "low_freq_factor": 1.0,
+    "high_freq_factor": 4.0,
+    "original_max_position_embeddings": 8192,
+}
+ROPE_THETA = 500000.0
+HEAD_DIM = 128
+MAX_POS = 131072
+
+
+# --- Layer 1: AST structural tripwire (stdlib only, no unsloth import) ---
+
+
+def _load_class_init():
+    tree = ast.parse(LLAMA_PY.read_text())
+    for node in ast.walk(tree):
+        if isinstance(node, ast.ClassDef) and node.name == CLASS_NAME:
+            for sub in node.body:
+                if isinstance(sub, ast.FunctionDef) and sub.name == "__init__":
+                    return sub
+    raise AssertionError(
+        f"{CLASS_NAME}.__init__ not found in {LLAMA_PY}; if it was renamed or "
+        "moved, update this guard so RoPE scaling stays protected (issue #2405)"
+    )
+
+
+def _config_branch(init_fn):
+    """The `if config is not None:` block at the top of __init__."""
+    for node in init_fn.body:
+        if isinstance(node, ast.If):
+            test = node.test
+            is_config_test = (
+                isinstance(test, ast.Compare)
+                and isinstance(test.left, ast.Name)
+                and test.left.id == "config"
+            )
+            if is_config_test:
+                return node
+    return None
+
+
+def test_config_path_inspects_rope_scaling():
+    init_fn = _load_class_init()
+    branch = _config_branch(init_fn)
+    assert branch is not None, (
+        f"{CLASS_NAME}.__init__ no longer has an `if config is not None:` "
+        "branch; the config constructor path must read config.rope_scaling so "
+        "scaled models (llama3/linear/longrope) are not silently unscaled "
+        "(issue #2405)"
+    )
+
+    names = set()
+    for stmt in branch.body:
+        for sub in ast.walk(stmt):
+            if isinstance(sub, ast.Attribute):
+                names.add(sub.attr)
+            elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
+                names.add(sub.value)
+    assert "rope_scaling" in names, (
+        f"{CLASS_NAME}.__init__ config path does not reference `rope_scaling`. "
+        "When a rotary class is built straight from a config (the path modern "
+        "transformers takes, since rotary moved to LlamaModel), the llama3 / "
+        "linear / longrope scaling must still be applied; otherwise long inputs "
+        "produce repeated-pattern gibberish (issue #2405)."
+    )
+
+    called = {
+        sub.func.id
+        for stmt in branch.body
+        for sub in ast.walk(stmt)
+        if isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)
+    }
+    assert "_compute_config_rope_inv_freq" in called, (
+        f"{CLASS_NAME}.__init__ config path no longer calls "
+        "_compute_config_rope_inv_freq; the CPU behavioral tests below cover "
+        "that helper directly, so the constructor must stay wired to it or "
+        "scaled configs silently lose RoPE scaling again (issue #2405)."
+    )
+
+
+# --- Layer 2: CPU behavioral guard (pure helper, no instantiation) ---
+
+
+def _make_config(rope_scaling):
+    from transformers import LlamaConfig
+    return LlamaConfig(
+        hidden_size = 256,
+        num_attention_heads = 2,
+        num_key_value_heads = 2,
+        head_dim = HEAD_DIM,
+        rope_theta = ROPE_THETA,
+        max_position_embeddings = MAX_POS,
+        rope_scaling = rope_scaling,
+    )
+
+
+def _unsloth_rotary(config):
+    from unsloth.models import llama as llama_mod
+    return llama_mod.LlamaRotaryEmbedding(config = config)
+
+
+def _reference_inv_freq(config, rope_type):
+    from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+    inv_freq, _attention_factor = ROPE_INIT_FUNCTIONS[rope_type](config, "cpu")
+    return inv_freq.float().cpu()
+
+
+def _vanilla_inv_freq():
+    return 1.0 / (
+        ROPE_THETA ** (torch.arange(0, HEAD_DIM, 2, dtype = torch.int64).float() / HEAD_DIM)
+    )
+
+
+def _compute_helper(config, rope_scaling):
+    from unsloth.models.llama import _compute_config_rope_inv_freq
+    return _compute_config_rope_inv_freq(config, rope_scaling)
+
+
+def test_llama3_scaling_applied_to_inv_freq():
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    got, attention_scaling = _compute_helper(config, config.rope_scaling)
+    expected = _reference_inv_freq(config, "llama3")
+    vanilla = _vanilla_inv_freq()
+
+    # Guard against a vacuous test.
+    assert not torch.allclose(
+        expected, vanilla, rtol = 1e-4
+    ), "test setup error: llama3-scaled inv_freq should differ from vanilla"
+    assert got is not None, (
+        "_compute_config_rope_inv_freq returned None for a llama3 config; the "
+        "config path is dropping config.rope_scaling, so long-context inference "
+        "degrades into repeated-pattern gibberish (issue #2405)."
+    )
+    got = got.float().cpu()
+    assert torch.allclose(got, expected, rtol = 1e-4, atol = 1e-6), (
+        "inv_freq for a llama3 config does not match transformers' llama3 RoPE "
+        "scaling (issue #2405).\n"
+        f"got[:6]={got[:6].tolist()}\nexpected[:6]={expected[:6].tolist()}"
+    )
+
+
+def test_default_rope_type_matches_vanilla_inv_freq():
+    config = _make_config(None)
+    got, attention_scaling = _compute_helper(config, {"rope_type": "default"})
+    assert got is not None
+    vanilla = _vanilla_inv_freq()
+    assert torch.allclose(got.float().cpu(), vanilla, rtol = 1e-4, atol = 1e-6), (
+        "default rope_type must equal the vanilla inv_freq; "
+        f"got[:6]={got[:6].tolist()} vanilla[:6]={vanilla[:6].tolist()}"
+    )
+
+
+def _cos_at_position(rot, position):
+    """cos row at one position, built like _set_cos_sin_cache but CPU-only."""
+    inv_freq = rot.inv_freq.float().cpu()
+    t = torch.tensor([position], dtype = torch.float32)
+    t = rot._apply_time_scaling(t.clone()) if hasattr(rot, "_apply_time_scaling") else t
+    freqs = torch.outer(t, inv_freq)
+    emb = torch.cat((freqs, freqs), dim = -1)
+    return emb.cos().squeeze(0)
+
+
+# --- Layer 3: CUDA behavioral guard (real instantiation needs a device) ---
+
+
+@requires_cuda
+def test_constructor_applies_llama3_scaling():
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    rot = _unsloth_rotary(config)
+    got = rot.inv_freq.float().cpu()
+    expected = _reference_inv_freq(config, "llama3")
+    assert torch.allclose(
+        got, expected, rtol = 1e-4, atol = 1e-6
+    ), "LlamaRotaryEmbedding built from a llama3 config produced unscaled inv_freq (issue #2405)."
+
+
+@requires_cuda
+def test_constructor_unscaled_config_uses_vanilla_inv_freq():
+    rot = _unsloth_rotary(_make_config(None))
+    got = rot.inv_freq.float().cpu()
+    vanilla = _vanilla_inv_freq()
+    assert torch.allclose(
+        got, vanilla, rtol = 1e-4, atol = 1e-6
+    ), "LlamaRotaryEmbedding with no rope_scaling must use the vanilla inv_freq"
+
+
+@requires_cuda
+def test_cos_cache_differs_between_scaled_and_unscaled_at_long_position():
+    scaled = _unsloth_rotary(_make_config(LLAMA3_ROPE_SCALING))
+    unscaled = _unsloth_rotary(_make_config(None))
+
+    pos = 10000
+    cos_scaled = _cos_at_position(scaled, pos)
+    cos_unscaled = _cos_at_position(unscaled, pos)
+    assert not torch.allclose(cos_scaled, cos_unscaled, rtol = 1e-4, atol = 1e-5), (
+        f"cos values at position {pos} are identical for a llama3-scaled and an "
+        "unscaled rotary embedding, which means scaling was dropped (issue "
+        "#2405). With correct llama3 scaling the low-frequency bands shrink by "
+        "up to 8x and must change the angles at long positions."
+    )
+
+
+@requires_cuda
+def test_extended_cache_keeps_scaling_after_growth():
+    scaled = _unsloth_rotary(_make_config(LLAMA3_ROPE_SCALING))
+    # Grow past the initial cache size (mirrors long-context decode).
+    dummy = torch.zeros(1, dtype = torch.float32)
+    scaled.extend_rope_embedding(dummy, seq_len = 40960)
+
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    expected = _reference_inv_freq(config, "llama3")
+    got = scaled.inv_freq.float().cpu()
+    assert torch.allclose(got, expected, rtol = 1e-4, atol = 1e-6), (
+        "growing the RoPE cache (extend_rope_embedding) must preserve llama3 "
+        "scaling of inv_freq; long-context decode loses scaling otherwise "
+        "(issue #2405)."
+    )
+
+
+def test_object_style_rope_scaling_does_not_crash():
+    # Object-style rope_scaling must be normalized, not .get()'d directly.
+    from dataclasses import dataclass
+
+    from unsloth.models.llama import _compute_config_rope_inv_freq
+
+    @dataclass
+    class FakeRopeScalingConfig:
+        rope_type: str = "llama3"
+        factor: float = 8.0
+        low_freq_factor: float = 1.0
+        high_freq_factor: float = 4.0
+        original_max_position_embeddings: int = 8192
+
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    inv_freq, attention_scaling = _compute_config_rope_inv_freq(config, FakeRopeScalingConfig())
+    assert inv_freq is not None, (
+        "object-style (non-dict) config.rope_scaling must be normalized, not "
+        "dropped; otherwise scaled models silently lose RoPE scaling again "
+        "(issue #2405)."
+    )
+    expected = _reference_inv_freq(config, "llama3")
+    assert torch.allclose(inv_freq.float().cpu(), expected, rtol = 1e-4, atol = 1e-6)
+
+
+def test_object_style_rope_scaling_on_config_delegates_correctly():
+    # 'linear' has no inline fallback; only the normalized-config retry passes this.
+    from dataclasses import dataclass
+
+    from unsloth.models.llama import _compute_config_rope_inv_freq
+
+    @dataclass
+    class FakeLinearRopeScalingConfig:
+        rope_type: str = "linear"
+        factor: float = 4.0
+
+    dict_config = _make_config({"rope_type": "linear", "factor": 4.0})
+    expected = _reference_inv_freq(dict_config, "linear")
+
+    object_config = _make_config({"rope_type": "linear", "factor": 4.0})
+    object_config.rope_scaling = FakeLinearRopeScalingConfig()
+    inv_freq, attention_scaling = _compute_config_rope_inv_freq(
+        object_config, object_config.rope_scaling
+    )
+    assert inv_freq is not None, (
+        "linear rope_scaling exposed as a config object was silently dropped; "
+        "delegation must retry with a config copy carrying the normalized dict "
+        "(issue #2405)."
+    )
+    assert torch.allclose(inv_freq.float().cpu(), expected, rtol = 1e-4, atol = 1e-6)
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index 172c69c5d8..5da1e27a9d 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -30,13 +30,46 @@ from .import_fixes import (
     disable_broken_causal_conv1d,
     disable_broken_vllm,
     configure_amdgpu_asic_id_table_path,
+    fix_bitsandbytes_rocm_arch_detection,
     torchvision_compatibility_check,
     fix_diffusers_warnings,
     fix_huggingface_hub,
 )
 
+# Redirect a read-only Hugging Face cache before anything below can import
+# huggingface_hub / transformers / vllm (disable_broken_vllm probes
+# `import vllm`, check_fbgemm_gpu_version imports transformers, and
+# fix_huggingface_hub imports huggingface_hub itself), all of which can
+# freeze Hub's cache constants with the un-redirected paths. unsloth_zoo
+# runs the same redirect at import, but that happens after these probes.
+# hf_cache.py is stdlib-only, so load it straight from its file without
+# triggering the full unsloth_zoo package init this early; the zoo's own
+# call later is an idempotent no-op. Older unsloth_zoo without hf_cache.py
+# is skipped silently.
+try:
+    import importlib.util as _importlib_util
+    from pathlib import Path as _Path
+
+    _zoo_spec = _importlib_util.find_spec("unsloth_zoo")
+    if _zoo_spec is not None and _zoo_spec.origin:
+        _hf_cache_file = _Path(_zoo_spec.origin).with_name("hf_cache.py")
+        if _hf_cache_file.is_file():
+            _hf_cache_spec = _importlib_util.spec_from_file_location(
+                "unsloth_zoo._early_hf_cache", _hf_cache_file
+            )
+            _hf_cache = _importlib_util.module_from_spec(_hf_cache_spec)
+            _hf_cache_spec.loader.exec_module(_hf_cache)
+            _hf_cache.redirect_hf_cache_if_readonly()
+            del _hf_cache, _hf_cache_spec
+        del _hf_cache_file
+    del _zoo_spec, _importlib_util, _Path
+except Exception:
+    pass
+
 # Configure libdrm ids table path early so ROCm can resolve AMD GPU names.
 configure_amdgpu_asic_id_table_path()
+# Must precede `import unsloth_zoo` below, which imports bnb on ROCm.
+fix_bitsandbytes_rocm_arch_detection()
 disable_broken_causal_conv1d()
 disable_broken_vllm()
 fix_message_factory_issue()
@@ -45,6 +78,7 @@ torchvision_compatibility_check()
 fix_diffusers_warnings()
 fix_huggingface_hub()
 del configure_amdgpu_asic_id_table_path
+del fix_bitsandbytes_rocm_arch_detection
 del disable_broken_causal_conv1d
 del disable_broken_vllm
 del fix_message_factory_issue
@@ -66,6 +100,13 @@ if already_imported:
     )
 del already_imported, critical_modules
 
+# Pin BNB_ROCM_VERSION before bitsandbytes is first imported (`import
+# unsloth_zoo` below pulls it in on ROCm hosts).
+from .import_fixes import maybe_set_windows_rocm_bnb_version
+
+maybe_set_windows_rocm_bnb_version()
+del maybe_set_windows_rocm_bnb_version
+
 # Multi-GPU is not yet supported (beta available on request).
 
 # Fixes https://github.com/unslothai/unsloth/issues/1266
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index b96f8cb7a0..fe078c5025 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2341,13 +2341,15 @@ extra_eos_tokens = None,
 
     You must use {INPUT}, {OUTPUT} twice, and {SYSTEM} is optional.
     """
-    # Strip only the left
+    # Strip only the left: trailing whitespace can be part of the repeated example
+    # (e.g. "{OUTPUT}\n"). Accidental trailing whitespace (#992) is retried on failure.
     chat_template = chat_template.lstrip()
 
     assert(tokenizer is not None)
 
     if extra_eos_tokens is None: extra_eos_tokens = []
     elif type(extra_eos_tokens) is str: extra_eos_tokens = [extra_eos_tokens,]
+    original_extra_eos_tokens = list(extra_eos_tokens)
 
     vocab = tokenizer.get_vocab()
     for extra_eos in extra_eos_tokens:
@@ -2454,6 +2456,20 @@ extra_eos_tokens = None,
                 f"{left_changed}"
             )
     except:
+        # Accidental trailing whitespace (#992) desyncs the two-example detection,
+        # so retry once without it. Templates that parse as-is are never altered.
+        rstripped_chat_template = chat_template.rstrip()
+        if rstripped_chat_template != chat_template:
+            try:
+                return construct_chat_template(
+                    tokenizer = tokenizer,
+                    chat_template = rstripped_chat_template,
+                    default_system_message = default_system_message,
+                    extra_eos_tokens = original_extra_eos_tokens,
+                )
+            except Exception:
+                pass
+
         output_pos = chat_template.find("{OUTPUT}")
         input_pos  = chat_template.find("{INPUT}")
         if output_pos == -1 or input_pos == -1:
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 25f5fdb9d2..695a6a577a 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -1823,6 +1823,340 @@ def configure_amdgpu_asic_id_table_path():
     return None
 
 
+# ---------------------------------------------------------------------------
+# bitsandbytes Windows ROCm fix: cextension.py runs get_rocm_gpu_arch()
+# (bnb >= 0.47) and get_rocm_warpsize() (0.49.x) at import, shelling out to
+# rocminfo / hipinfo.exe via PATH. Neither is on PATH on Windows (AMD torch
+# wheels put hipInfo.exe in venv Scripts), so every import logs ERROR +
+# WARNING, ROCM_GPU_ARCH becomes "unknown", and warp size defaults to 64:
+# wrong on RDNA (wave 32), breaking 4-bit blocksizes and
+# ALLOW_PREQUANTIZED_MODELS. Upstream fix unmerged (bitsandbytes#1969), so a
+# MetaPathFinder swaps both helpers for torch-device-props-first versions
+# right after bitsandbytes.cuda_specs executes, before cextension reads
+# them. Must run before `import unsloth_zoo` (imports bnb on ROCm).
+# ---------------------------------------------------------------------------
+
+_BNB_CUDA_SPECS_MODULE = "bitsandbytes.cuda_specs"
+_BNB_ROCM_FIX_FINDER_SENTINEL = "_unsloth_bnb_rocm_fix_finder"
+_BNB_ROCM_FIX_FUNCTION_FLAG = "__unsloth_bnb_rocm_fix__"
+
+
+def _torch_rocm_device_props():
+    """Device-0 props on a ROCm torch build with a visible GPU, else None.
+    Never raises; bnb's own import initializes the device context anyway."""
+    try:
+        import torch
+
+        if not getattr(getattr(torch, "version", None), "hip", None):
+            return None
+        if not torch.cuda.is_available():
+            return None
+        return torch.cuda.get_device_properties(0)
+    except Exception:
+        return None
+
+
+def _iter_hipinfo_paths():
+    """Yield existing hipInfo.exe paths: PATH, interpreter scripts dir (venv
+    and conda layouts), then HIP SDK / AMD installer locations."""
+    import shutil
+    import sysconfig
+
+    candidates = []
+    try:
+        resolved = shutil.which("hipinfo.exe")
+        if resolved:
+            candidates.append(resolved)
+    except Exception:
+        pass
+    try:
+        scripts_dir = sysconfig.get_path("scripts")
+        if scripts_dir:
+            candidates.append(os.path.join(scripts_dir, "hipInfo.exe"))
+    except Exception:
+        pass
+    executable_dir = os.path.dirname(sys.executable or "")
+    if executable_dir:
+        candidates.append(os.path.join(executable_dir, "hipInfo.exe"))
+        candidates.append(os.path.join(executable_dir, "Scripts", "hipInfo.exe"))
+    for env_key in ("HIP_PATH", "ROCM_PATH"):
+        root = os.environ.get(env_key, "").strip()
+        if root:
+            candidates.append(os.path.join(root, "bin", "hipInfo.exe"))
+    rocm_root = os.path.join(os.environ.get("ProgramFiles", r"C:\Program Files"), "AMD", "ROCm")
+    try:
+        if os.path.isdir(rocm_root):
+            for version_dir in sorted(os.listdir(rocm_root), reverse = True):
+                candidates.append(os.path.join(rocm_root, version_dir, "bin", "hipInfo.exe"))
+    except Exception:
+        pass
+
+    seen = set()
+    for candidate in candidates:
+        try:
+            key = os.path.normcase(os.path.normpath(candidate))
+            if key in seen:
+                continue
+            seen.add(key)
+            if os.path.isfile(candidate):
+                yield candidate
+        except Exception:
+            continue
+
+
+def _run_hipinfo(hipinfo_path):
+    """Run hipInfo.exe and return its stdout, or "" on any failure."""
+    import subprocess
+    try:
+        result = subprocess.run(
+            [hipinfo_path],
+            capture_output = True,
+            text = True,
+            timeout = 15,
+            creationflags = getattr(subprocess, "CREATE_NO_WINDOW", 0),
+        )
+        return result.stdout or ""
+    except Exception as e:
+        _log_rocm_detection(f"Unsloth: `{hipinfo_path}` failed: {e}")
+        return ""
+
+
+def _unsloth_get_rocm_gpu_arch():
+    """Replaces bnb's get_rocm_gpu_arch: torch device props first (no
+    subprocess), then hipInfo.exe by absolute path, then a quiet "unknown"."""
+    try:
+        import torch
+        if not getattr(getattr(torch, "version", None), "hip", None):
+            return "unknown"
+    except Exception:
+        return "unknown"
+    props = _torch_rocm_device_props()
+    if props is not None:
+        try:
+            # gcnArchName may carry feature flags, e.g. "gfx90a:sramecc+:xnack-"
+            arch = str(props.gcnArchName).split(":")[0].strip()
+            if arch.startswith("gfx"):
+                return arch
+        except Exception:
+            pass
+    for hipinfo_path in _iter_hipinfo_paths():
+        match = re.search(r"gcnArchName:\s+gfx([a-zA-Z\d]+)", _run_hipinfo(hipinfo_path))
+        if match:
+            return "gfx" + match.group(1)
+    _log_rocm_detection(
+        "Unsloth: Could not detect the ROCm GPU architecture - bitsandbytes will see `unknown`."
+    )
+    return "unknown"
+
+
+def _unsloth_get_rocm_warpsize():
+    """Replaces bnb 0.49.x get_rocm_warpsize: upstream defaults to 64 when
+    rocminfo is missing, wrong on RDNA (wave 32)."""
+    try:
+        import torch
+        if not getattr(getattr(torch, "version", None), "hip", None):
+            return 32  # upstream behavior: NVIDIA warp size is always 32
+    except Exception:
+        return 64  # upstream behavior: default to 64 on failure
+    props = _torch_rocm_device_props()
+    if props is not None:
+        # torch 2.11 ROCm exposes warp_size; some builds used warpSize.
+        for attribute_name in ("warp_size", "warpSize"):
+            warp_size = getattr(props, attribute_name, None)
+            if isinstance(warp_size, int) and warp_size in (32, 64):
+                return warp_size
+    for hipinfo_path in _iter_hipinfo_paths():
+        match = re.search(r"^\s*warpSize:\s+(\d+)", _run_hipinfo(hipinfo_path), re.MULTILINE)
+        if match and int(match.group(1)) in (32, 64):
+            return int(match.group(1))
+    _log_rocm_detection(
+        "Unsloth: Could not detect the ROCm warp size - defaulting to 64 "
+        "(bitsandbytes' own default)."
+    )
+    return 64
+
+
+setattr(_unsloth_get_rocm_gpu_arch, _BNB_ROCM_FIX_FUNCTION_FLAG, True)
+setattr(_unsloth_get_rocm_warpsize, _BNB_ROCM_FIX_FUNCTION_FLAG, True)
+
+
+def _bnb_rocm_helper_is_broken(function):
+    """True only for upstream's subprocess-only detectors; co_names works
+    where getsource fails. Versions consulting torch props are untouched."""
+    if function is None or not callable(function):
+        return False
+    if getattr(function, _BNB_ROCM_FIX_FUNCTION_FLAG, False):
+        return False  # Already ours.
+    try:
+        function = inspect.unwrap(function)
+    except Exception:
+        pass
+    code = getattr(function, "__code__", None)
+    co_names = getattr(code, "co_names", ()) if code is not None else ()
+    if not co_names:
+        return False  # C function or opaque wrapper -- do not touch.
+    if "get_device_properties" in co_names or "gcnArchName" in co_names:
+        return False  # Fixed upstream -- no-op.
+    return "subprocess" in co_names
+
+
+def _patch_bnb_cuda_specs_module(module):
+    """Swap broken ROCm detection helpers on an executed cuda_specs module.
+    Returns True when the module ends up patched (now or previously)."""
+    patched = False
+    for attribute_name, replacement in (
+        ("get_rocm_gpu_arch", _unsloth_get_rocm_gpu_arch),
+        ("get_rocm_warpsize", _unsloth_get_rocm_warpsize),
+    ):
+        original = getattr(module, attribute_name, None)
+        if getattr(original, _BNB_ROCM_FIX_FUNCTION_FLAG, False):
+            patched = True  # Already ours.
+            continue
+        if not _bnb_rocm_helper_is_broken(original):
+            continue
+        setattr(module, attribute_name, replacement)
+        patched = True
+        logger.info(
+            f"Unsloth: Patched bitsandbytes.cuda_specs.{attribute_name} - "
+            f"avoids PATH-dependent subprocess GPU detection on Windows ROCm."
+        )
+    return patched
+
+
+class _BnbCudaSpecsPatchLoader(importlib.abc.Loader):
+    __slots__ = ("_loader",)
+
+    def __init__(self, loader):
+        self._loader = loader
+
+    def create_module(self, spec):
+        create_module = getattr(self._loader, "create_module", None)
+        if create_module is None:
+            return None
+        return create_module(spec)
+
+    def exec_module(self, module):
+        self._loader.exec_module(module)
+        # Patch after the module body ran, before cextension calls it. The
+        # finder stays on sys.meta_path (same lifecycle as the blockers
+        # above) so importlib.reload(bitsandbytes.cuda_specs) re-patches.
+        try:
+            _patch_bnb_cuda_specs_module(module)
+        except Exception as e:
+            _log_rocm_detection(f"Unsloth: bitsandbytes ROCm detection patch failed: {e}")
+
+    def __getattr__(self, name):
+        # Delegate get_source / get_filename etc. so introspection works.
+        return getattr(self._loader, name)
+
+
+class _BnbCudaSpecsPatchFinder(importlib.abc.MetaPathFinder):
+    __slots__ = (_BNB_ROCM_FIX_FINDER_SENTINEL,)
+
+    def __init__(self):
+        setattr(self, _BNB_ROCM_FIX_FINDER_SENTINEL, True)
+
+    def find_spec(
+        self,
+        fullname,
+        path = None,
+        target = None,
+    ):
+        if fullname != _BNB_CUDA_SPECS_MODULE:
+            return None
+        # Delegate to remaining finders (editable installs, frozen apps)
+        # and wrap the loader that would actually be used.
+        spec = None
+        for finder in sys.meta_path:
+            if finder is self or getattr(finder, _BNB_ROCM_FIX_FINDER_SENTINEL, False):
+                continue
+            finder_find_spec = getattr(finder, "find_spec", None)
+            if finder_find_spec is None:
+                continue
+            try:
+                spec = finder_find_spec(fullname, path, target)
+            except Exception:
+                spec = None
+            if spec is not None:
+                break
+        if spec is None or spec.loader is None:
+            return None
+        if not hasattr(spec.loader, "exec_module"):
+            return None  # Legacy loader -- let the stock machinery handle it.
+        spec.loader = _BnbCudaSpecsPatchLoader(spec.loader)
+        return spec
+
+
+def _repair_imported_bitsandbytes_rocm_constants():
+    """bnb imported before unsloth: noise already fired, but fix detectors
+    and cached constants, incl. by-value ROCM_WARP_SIZE_64 copies."""
+    cuda_specs = sys.modules.get(_BNB_CUDA_SPECS_MODULE)
+    if cuda_specs is None:
+        return
+    if not _patch_bnb_cuda_specs_module(cuda_specs):
+        return
+
+    try:
+        arch = cuda_specs.get_rocm_gpu_arch()
+    except Exception:
+        arch = "unknown"
+    warp_size_64 = None
+    get_rocm_warpsize = getattr(cuda_specs, "get_rocm_warpsize", None)
+    if callable(get_rocm_warpsize):
+        try:
+            warp_size_64 = get_rocm_warpsize() == 64
+        except Exception:
+            warp_size_64 = None
+
+    for module_name, module in list(sys.modules.items()):
+        if module is None or module is cuda_specs:
+            continue
+        if module_name != "bitsandbytes" and not module_name.startswith("bitsandbytes."):
+            continue
+        try:
+            if arch != "unknown" and getattr(module, "ROCM_GPU_ARCH", None) == "unknown":
+                module.ROCM_GPU_ARCH = arch
+            if warp_size_64 is not None and isinstance(
+                getattr(module, "ROCM_WARP_SIZE_64", None), bool
+            ):
+                module.ROCM_WARP_SIZE_64 = warp_size_64
+        except Exception:
+            continue
+    logger.info("Unsloth: Repaired bitsandbytes ROCm arch / warp-size constants in place.")
+
+
+def fix_bitsandbytes_rocm_arch_detection():
+    """Fix bnb's import-time ROCm arch / warp-size detection on Windows
+    (see header above). No-op on non-Windows, non-ROCm, missing or
+    upstream-fixed bnb. Idempotent. Opt out: UNSLOTH_DISABLE_BNB_ROCM_FIX=1."""
+    if os.environ.get("UNSLOTH_DISABLE_BNB_ROCM_FIX", "0") == "1":
+        return
+    if sys.platform != "win32":
+        return
+    if not _is_rocm_torch_build():
+        return
+
+    # Already imported: prevention impossible, repair in place instead.
+    if _BNB_CUDA_SPECS_MODULE in sys.modules:
+        try:
+            _repair_imported_bitsandbytes_rocm_constants()
+        except Exception:
+            pass
+        return
+
+    try:
+        if importlib.util.find_spec("bitsandbytes") is None:
+            return
+    except Exception:
+        return
+
+    for finder in sys.meta_path:
+        if getattr(finder, _BNB_ROCM_FIX_FINDER_SENTINEL, False):
+            return  # Already installed -- idempotent.
+    sys.meta_path.insert(0, _BnbCudaSpecsPatchFinder())
+    _log_rocm_detection("Unsloth: Installed the bitsandbytes ROCm arch detection patch hook.")
+
+
 def _is_causal_conv1d_name(module_name: str) -> bool:
     return module_name == _CAUSAL_CONV1D_PREFIX or module_name.startswith(
         _CAUSAL_CONV1D_PREFIX + "."
@@ -2162,3 +2496,91 @@ def disable_broken_causal_conv1d():
         "Unsloth: Detected broken causal_conv1d binary; "
         "disabling causal_conv1d fast path and continuing import."
     )
+
+
+_BNB_ROCM_DLL_RE = re.compile(r"libbitsandbytes_rocm(\d+)\.dll", re.IGNORECASE)
+
+
+def _is_hip_torch_build():
+    """True only when torch itself is a HIP/ROCm build. Env hints (HIP_PATH
+    etc.) do not count: CUDA bitsandbytes raises at import when the ROCm
+    override is set. Wheel tag first (no torch import); torch.version.hip
+    fallback for source builds."""
+    try:
+        if "rocm" in str(importlib_version("torch")).lower():
+            return True
+    except Exception:
+        pass
+    try:
+        import torch
+        return bool(getattr(torch.version, "hip", None))
+    except Exception:
+        return False
+
+
+def _detect_installed_bnb_rocm_version():
+    """Highest installed ``libbitsandbytes_rocm<NN>.dll`` suffix ("72", "713")
+    or ``None``. Listing order is unordered, so take the numeric max."""
+    try:
+        spec = importlib.util.find_spec("bitsandbytes")
+    except Exception:
+        return None
+    if spec is None or not spec.submodule_search_locations:
+        return None
+
+    suffixes = []
+    for pkg_dir in spec.submodule_search_locations:
+        try:
+            entries = os.listdir(pkg_dir)
+        except Exception:
+            continue
+        for entry in entries:
+            match = _BNB_ROCM_DLL_RE.fullmatch(entry)
+            if match is not None:
+                suffixes.append(match.group(1))
+    if not suffixes:
+        return None
+    return max(suffixes, key = lambda value: int(value))
+
+
+def maybe_set_windows_rocm_bnb_version():
+    """Pin ``BNB_ROCM_VERSION`` from the installed wheel on Windows + ROCm torch.
+
+    AMD's Windows wheel ships one ``libbitsandbytes_rocm<NN>.dll`` whose
+    suffix can disagree with ``torch.version.hip`` (HIP 7.13 vs rocm72.dll),
+    breaking the native 4-bit/8-bit paths. Pin the installed suffix before
+    bitsandbytes is first imported.
+
+    No-op unless ALL of: Windows, a real HIP torch build (env hints like
+    HIP_PATH do not count), a ROCm DLL installed, and no explicit user value.
+    Linux is untouched. Values seeded by Studio's venv sitecustomize.py
+    (marked ``UNSLOTH_BNB_ROCM_VERSION_SOURCE=sitecustomize``) are
+    redetectable defaults, not overrides; ``UNSLOTH_SKIP_BNB_ROCM_VERSION=1``
+    opts out and drops a seeded default. Returns the value set, else None.
+    """
+    if sys.platform != "win32":
+        return None
+    if os.environ.get("UNSLOTH_SKIP_BNB_ROCM_VERSION") == "1":
+        # Real opt-out: drop our seeded default (marker present); explicit
+        # user values carry no marker and are kept.
+        if os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") == "sitecustomize":
+            os.environ.pop("BNB_ROCM_VERSION", None)
+            os.environ.pop("UNSLOTH_BNB_ROCM_VERSION_SOURCE", None)
+        return None
+    if "BNB_ROCM_VERSION" in os.environ and (
+        os.environ.get("UNSLOTH_BNB_ROCM_VERSION_SOURCE") != "sitecustomize"
+    ):
+        return None
+    if not _is_hip_torch_build():
+        return None
+    version = _detect_installed_bnb_rocm_version()
+    if version is None:
+        return None
+    os.environ["BNB_ROCM_VERSION"] = version
+    os.environ["UNSLOTH_BNB_ROCM_VERSION_SOURCE"] = "detected"
+    if UNSLOTH_ENABLE_LOGGING:
+        logger.info(
+            f"Unsloth: set BNB_ROCM_VERSION={version} "
+            "(detected from the installed bitsandbytes ROCm wheel on Windows)."
+        )
+    return version
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index c8dfc6cf67..ab7f19133c 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.1"
+__version__ = "2026.6.3"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
@@ -1432,6 +1432,18 @@ if is_openai_available():
 from transformers import AutoTokenizer
 from transformers.utils.import_utils import _is_package_available
 
+
+def _package_available(pkg_name: str) -> bool:
+    # transformers >= 5.x makes `_is_package_available` always return a
+    # `(exists, version)` tuple, which is truthy even when the package is
+    # absent; older versions returned a plain bool. Normalise to a bool so
+    # callers don't take "package present" branches for missing packages.
+    result = _is_package_available(pkg_name)
+    if isinstance(result, tuple):
+        return bool(result[0])
+    return bool(result)
+
+
 SUPPORTS_BFLOAT16 = False
 HAS_FLASH_ATTENTION = False
 HAS_FLASH_ATTENTION_SOFTCAPPING = False
@@ -1442,7 +1454,7 @@ if DEVICE_TYPE == "cuda":
 
     if major_version >= 8:
         SUPPORTS_BFLOAT16 = True
-        if _is_package_available("flash_attn"):
+        if _package_available("flash_attn"):
             # Check for CUDA linking errors "undefined symbol: _ZNK3c106SymIntltEl"
             try:
                 try:
@@ -1487,7 +1499,7 @@ if DEVICE_TYPE == "cuda":
         HAS_FLASH_ATTENTION = False
 elif DEVICE_TYPE == "hip":
     SUPPORTS_BFLOAT16 = True
-    if _is_package_available("flash_attn"):
+    if _package_available("flash_attn"):
         # Check for CUDA linking errors "undefined symbol: _ZNK3c106SymIntltEl"
         try:
             try:
@@ -2152,7 +2164,7 @@ def is_bfloat16_supported():
 
 
 def is_vLLM_available():
-    return _is_package_available("vllm")
+    return _package_available("vllm")
 
 
 # Patches models to add RoPE Scaling
diff --git a/unsloth/models/diffusion.py b/unsloth/models/diffusion.py
new file mode 100644
index 0000000000..12596b432e
--- /dev/null
+++ b/unsloth/models/diffusion.py
@@ -0,0 +1,301 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+FastDiffusionModel: a transformers-only slow path for text-diffusion models (e.g. DiffusionGemma).
+
+These models use a block-diffusion sampling loop (custom generate) and a novel backbone, so we skip
+Unsloth's autoregressive kernel/compile patching and load the unmodified HF model (outputs stay
+bit-identical to transformers), keeping only the safe conveniences: 4bit/8bit loading, PEFT LoRA, the
+(model, tokenizer) API, and for_inference/for_training. Extend DIFFUSION_MODEL_TYPES as more land.
+"""
+
+import os
+import torch
+from transformers import AutoConfig, AutoProcessor, AutoTokenizer
+
+from ._utils import is_bfloat16_supported
+from .llama import logger
+
+__all__ = ["FastDiffusionModel", "DIFFUSION_MODEL_TYPES", "is_diffusion_model_type"]
+
+# transformers model_type strings routed to this slow path
+DIFFUSION_MODEL_TYPES = ("diffusion_gemma", "diffusion_gemma4")
+
+# Default LoRA targets: standard nn.Linear modules in the shared Gemma-4 backbone. The 128 MoE experts
+# are fused 3D Parameters (gate_up_proj/down_proj), not nn.Linear, so PEFT LoRA cannot target them.
+DIFFUSION_LORA_TARGETS = [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",  # attention
+    "gate_proj",
+    "up_proj",
+    "down_proj",  # dense (non-expert) MLP
+]
+
+# Vision tower uses a custom Linear with the same suffix names; exclude it so only the text path is wrapped.
+DIFFUSION_LORA_EXCLUDE = r".*(vision_tower|embed_vision).*"
+
+
+def is_diffusion_model_type(model_types):
+    """model_types: str or iterable -> True if any is a known diffusion model_type."""
+    if isinstance(model_types, str):
+        model_types = (model_types,)
+    return any(mt in DIFFUSION_MODEL_TYPES for mt in model_types)
+
+
+def _resolve_diffusion_model_class(config):
+    """Resolve the HF model class for a diffusion checkpoint from config.architectures."""
+    import transformers
+
+    archs = getattr(config, "architectures", None) or []
+    for arch in archs:
+        cls = getattr(transformers, arch, None)
+        if cls is not None:
+            return cls
+    # Fallbacks across naming revisions.
+    for name in (
+        "DiffusionGemmaForBlockDiffusion",
+        "DiffusionGemma4ModelForBlockDiffusion",
+        "DiffusionGemma4ForBlockDiffusion",
+    ):
+        cls = getattr(transformers, name, None)
+        if cls is not None:
+            return cls
+    raise RuntimeError(
+        f"Unsloth: could not resolve a diffusion model class from architectures={archs}. "
+        "Ensure you have the transformers build that ships the DiffusionGemma implementation."
+    )
+
+
+def _load_diffusion_config(model_name, token, trust_remote_code, revision, local_files_only):
+    """Load the config, aliasing the legacy ``diffusion_gemma`` model_type to the ``diffusion_gemma4``
+    classes current transformers ships. AutoConfig raises on the legacy type; catch that, rewrite the
+    type/arch names in-memory, and rebuild."""
+    try:
+        return AutoConfig.from_pretrained(
+            model_name,
+            token = token,
+            trust_remote_code = trust_remote_code,
+            revision = revision,
+            local_files_only = local_files_only,
+        )
+    except ValueError as e:
+        if "diffusion_gemma" not in str(e):
+            raise
+        import json
+        from transformers.utils import cached_file
+
+        cfg_path = cached_file(
+            model_name,
+            "config.json",
+            token = token,
+            revision = revision,
+            local_files_only = local_files_only,
+        )
+        with open(cfg_path, encoding = "utf-8") as f:
+            cd = json.load(f)
+        cd["model_type"] = "diffusion_gemma4"
+        cd.setdefault("architectures", ["DiffusionGemma4ModelForBlockDiffusion"])
+        if isinstance(cd.get("text_config"), dict):
+            cd["text_config"]["model_type"] = "diffusion_gemma4_text"
+        if isinstance(cd.get("vision_config"), dict):
+            cd["vision_config"]["model_type"] = "diffusion_gemma4_vision"
+        from transformers import DiffusionGemma4Config
+
+        return DiffusionGemma4Config.from_dict(cd)
+
+
+class FastDiffusionModel:
+    """transformers-only slow path for text-diffusion models."""
+
+    @staticmethod
+    def from_pretrained(
+        model_name = "google/diffusiongemma-26B-A4B-it",
+        max_seq_length = None,  # API-compat; diffusion uses canvas_length
+        dtype = None,
+        load_in_4bit = False,
+        load_in_8bit = False,
+        load_in_16bit = False,
+        full_finetuning = False,
+        token = None,
+        device_map = "auto",
+        trust_remote_code = False,
+        attn_implementation = "eager",  # exact match with the reference golden logits
+        revision = None,
+        return_tokenizer = True,
+        **kwargs,
+    ):
+        SUPPORTS_BFLOAT16 = is_bfloat16_supported()
+        if dtype is None:
+            dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
+        elif dtype == torch.bfloat16 and not SUPPORTS_BFLOAT16:
+            logger.warning_once("Device does not support bfloat16. Will change to float16.")
+            dtype = torch.float16
+        assert dtype in (torch.float16, torch.bfloat16, torch.float32)
+
+        # Honor an explicit local_files_only; else fall back to the offline env vars.
+        local_files_only = kwargs.pop("local_files_only", None)
+        if local_files_only is None:
+            local_files_only = (
+                os.environ.get("HF_HUB_OFFLINE", "0") == "1"
+                or os.environ.get("TRANSFORMERS_OFFLINE", "0") == "1"
+            )
+        config = _load_diffusion_config(
+            model_name,
+            token,
+            trust_remote_code,
+            revision,
+            local_files_only,
+        )
+        model_type = getattr(config, "model_type", None)
+        if not is_diffusion_model_type(model_type):
+            raise RuntimeError(
+                f"Unsloth: FastDiffusionModel only supports diffusion model_types {DIFFUSION_MODEL_TYPES}, "
+                f"got '{model_type}'. Use FastModel/FastLanguageModel for autoregressive models."
+            )
+
+        model_cls = _resolve_diffusion_model_class(config)
+
+        load_kwargs = dict(
+            dtype = dtype,
+            device_map = device_map,
+            token = token,
+            trust_remote_code = trust_remote_code,
+            attn_implementation = attn_implementation,
+            revision = revision,
+            local_files_only = local_files_only,
+        )
+
+        # Optional bitsandbytes quant. The MoE experts (3D Parameters) are not nn.Linear so bnb skips
+        # them; only attention + dense MLP Linears quantize, lm_head/embeddings stay full precision.
+        if load_in_4bit or load_in_8bit:
+            from transformers import BitsAndBytesConfig
+            if load_in_4bit:
+                qcfg = BitsAndBytesConfig(
+                    load_in_4bit = True,
+                    bnb_4bit_use_double_quant = True,
+                    bnb_4bit_quant_type = "nf4",
+                    bnb_4bit_compute_dtype = dtype,
+                    llm_int8_skip_modules = [
+                        "lm_head",
+                        "embed_tokens",
+                        "experts",
+                        "self_conditioning",
+                        "router",
+                    ],
+                )
+            else:
+                qcfg = BitsAndBytesConfig(load_in_8bit = True)
+            load_kwargs["quantization_config"] = qcfg
+
+        print(f"==((  Unsloth: FastDiffusionModel (slow / transformers-only path)  ))==")
+        print(f"   Model: {model_name}  | class: {model_cls.__name__}  | model_type: {model_type}")
+        print(
+            f"   dtype: {dtype} | 4bit: {load_in_4bit} | 8bit: {load_in_8bit} | attn: {attn_implementation}"
+        )
+
+        model = model_cls.from_pretrained(model_name, **load_kwargs).eval()
+        # Mark before any early return so get_peft_model/for_* route to the slow path.
+        model._unsloth_slow_diffusion = True
+
+        if not return_tokenizer:
+            return model, None
+
+        # Prefer the processor (chat template + tokenizer); fall back to a bare tokenizer. Returned as
+        # "tokenizer" to match the Unsloth (model, tokenizer) contract.
+        try:
+            tokenizer = AutoProcessor.from_pretrained(
+                model_name,
+                token = token,
+                trust_remote_code = trust_remote_code,
+                revision = revision,
+                local_files_only = local_files_only,
+            )
+        except Exception:
+            tokenizer = AutoTokenizer.from_pretrained(
+                model_name,
+                token = token,
+                trust_remote_code = trust_remote_code,
+                revision = revision,
+                local_files_only = local_files_only,
+            )
+
+        return model, tokenizer
+
+    @staticmethod
+    def get_peft_model(
+        model,
+        r = 16,
+        target_modules = None,
+        lora_alpha = 16,
+        lora_dropout = 0.0,
+        bias = "none",
+        use_gradient_checkpointing = True,
+        random_state = 3407,
+        task_type = None,
+        **kwargs,
+    ):
+        """Attach a PEFT LoRA to the diffusion backbone (attention + dense MLP). No fused kernels."""
+        from peft import LoraConfig, get_peft_model as peft_get_peft_model
+
+        if target_modules is None:
+            target_modules = DIFFUSION_LORA_TARGETS
+
+        lora_kwargs = dict(
+            r = r,
+            lora_alpha = lora_alpha,
+            lora_dropout = lora_dropout,
+            bias = bias,
+            target_modules = target_modules,
+            task_type = task_type,  # None: diffusion has no standard CAUSAL_LM head
+            **{k: v for k, v in kwargs.items() if k in ("modules_to_save", "init_lora_weights")},
+        )
+        # Exclude the vision tower's custom (non-Linear) modules that share suffix names.
+        exclude = kwargs.get("exclude_modules", DIFFUSION_LORA_EXCLUDE)
+        try:
+            lora_config = LoraConfig(exclude_modules = exclude, **lora_kwargs)
+        except TypeError:
+            # Older PEFT without exclude_modules: scope the target to the text decoder by regex.
+            lora_kwargs["target_modules"] = (
+                r".*model\.decoder\.layers\.\d+\.(self_attn\.[qkvo]_proj|mlp\.(gate|up|down)_proj)"
+            )
+            lora_config = LoraConfig(**lora_kwargs)
+        if use_gradient_checkpointing:
+            model.gradient_checkpointing_enable()
+            if hasattr(model, "enable_input_require_grads"):
+                model.enable_input_require_grads()
+
+        model = peft_get_peft_model(model, lora_config)
+        model._unsloth_slow_diffusion = True
+        try:
+            model.print_trainable_parameters()
+        except Exception:
+            pass
+        return model
+
+    @staticmethod
+    def for_inference(model):
+        model.eval()
+        for _, m in model.named_modules():
+            if hasattr(m, "gradient_checkpointing"):
+                m.gradient_checkpointing = False
+        return model
+
+    @staticmethod
+    def for_training(model, use_gradient_checkpointing = True):
+        model.train()
+        if use_gradient_checkpointing and hasattr(model, "gradient_checkpointing_enable"):
+            model.gradient_checkpointing_enable()
+        return model
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 6018a4ccb4..2d31f71ab1 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1622,6 +1622,93 @@ def _get_rope_theta(config, default = 10000.0):
     return default
 
 
+def _rope_scaling_as_dict(rope_scaling):
+    """Normalize config.rope_scaling (dict or config object) to a dict; {} on failure."""
+    if isinstance(rope_scaling, dict):
+        return rope_scaling
+    for converter in ("to_dict", "dict"):
+        fn = getattr(rope_scaling, converter, None)
+        if callable(fn):
+            try:
+                d = fn()
+                if isinstance(d, dict):
+                    return d
+            except Exception:
+                pass
+    try:
+        return {k: v for k, v in vars(rope_scaling).items() if not k.startswith("_")}
+    except TypeError:
+        return {}
+
+
+def _llama3_inv_freq_from_config(
+    config,
+    rope_scaling,
+    device = "cpu",
+):
+    """llama3 inv_freq with factors from config; fallback when modeling_rope_utils is missing."""
+    base = _get_rope_theta(config, default = 10000.0)
+    dim = getattr(config, "head_dim", None)
+    if dim is None:
+        dim = int(config.hidden_size // config.num_attention_heads)
+    inv_freq = 1.0 / (
+        base ** (torch.arange(0, dim, 2, dtype = torch.int64, device = device).float() / dim)
+    )
+
+    scale_factor = rope_scaling.get("factor", 8.0)
+    low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+    high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+    old_context_len = rope_scaling.get("original_max_position_embeddings", 8192)
+
+    low_freq_wavelen = old_context_len / low_freq_factor
+    high_freq_wavelen = old_context_len / high_freq_factor
+    assert low_freq_wavelen != high_freq_wavelen
+
+    # Vectorized meta-llama bands: high freqs kept, low divided by factor, medium blended.
+    wavelen = 2 * math.pi / inv_freq
+    scaled = torch.where(wavelen > low_freq_wavelen, inv_freq / scale_factor, inv_freq)
+    smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+    smoothed = (1 - smooth) * inv_freq / scale_factor + smooth * inv_freq
+    is_medium = (wavelen >= high_freq_wavelen) & (wavelen <= low_freq_wavelen)
+    return torch.where(is_medium, smoothed, scaled)
+
+
+def _compute_config_rope_inv_freq(config, rope_scaling):
+    """(inv_freq, attention_scaling) per config.rope_scaling via transformers'
+    ROPE_INIT_FUNCTIONS, with an inline llama3 fallback; (None, 1.0) on failure."""
+    original_rope_scaling = rope_scaling
+    rope_scaling = _rope_scaling_as_dict(rope_scaling)
+    rope_type = rope_scaling.get("rope_type", None) or rope_scaling.get("type", None)
+    try:
+        from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+
+        rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type]
+        try:
+            inv_freq, attention_scaling = rope_init_fn(config, torch.device("cpu"))
+        except Exception:
+            # Object-style rope_scaling: retry with a config copy carrying the plain dict.
+            if isinstance(original_rope_scaling, dict):
+                raise
+            import copy as _copy
+
+            config_copy = _copy.copy(config)
+            config_copy.rope_scaling = rope_scaling
+            inv_freq, attention_scaling = rope_init_fn(config_copy, torch.device("cpu"))
+        return inv_freq.to(dtype = torch.float32, device = "cpu"), float(attention_scaling)
+    except Exception as exception:
+        if rope_type == "llama3":
+            try:
+                return _llama3_inv_freq_from_config(config, rope_scaling), 1.0
+            except Exception:
+                pass
+        logger.warning_once(
+            f"Unsloth: Could not apply RoPE scaling '{rope_type}' from config "
+            f"({type(exception).__name__}: {exception}); falling back to unscaled RoPE. "
+            "Long-context generation may degrade."
+        )
+        return None, 1.0
+
+
 # Solves https://github.com/unslothai/unsloth/issues/168
 # Static KV Cache was introduced in 4.38.0, causing training to be much slower.
 # Inference can now be CUDAGraphed, but we shall retain the old rotary embeddings.
@@ -1640,6 +1727,12 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         config = None,  # [TODO] Hack to pass in config - need to remove later
     ):
         super().__init__()
+        # cos/sin multiplier (1.0 except yarn / longrope); set before any cache build.
+        self.attention_scaling = 1.0
+        # Base-class-from-config path (modern transformers): derive inv_freq like
+        # transformers so config.rope_scaling is not dropped (#2405). Scaled
+        # subclasses are excluded to avoid double-scaling.
+        config_inv_freq = None
         if config is not None:
             # [TODO] Hack to pass in config - need to remove later
             base = _get_rope_theta(config, default = base)
@@ -1652,6 +1745,13 @@ class LlamaRotaryEmbedding(torch.nn.Module):
             device = DEVICE_TYPE_TORCH
             max_position_embeddings = config.max_position_embeddings
 
+            rope_scaling = getattr(config, "rope_scaling", None)
+            if rope_scaling is not None and type(self) is LlamaRotaryEmbedding:
+                config_inv_freq, self.attention_scaling = _compute_config_rope_inv_freq(
+                    config,
+                    rope_scaling,
+                )
+
         self.dim = dim
         self.max_position_embeddings = max_position_embeddings
         self.base = base
@@ -1660,12 +1760,17 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         self.multi_gpu_cos_cached = [None] * DEVICE_COUNT
         self.multi_gpu_sin_cached = [None] * DEVICE_COUNT
 
-        # Normal Llama-3 RoPE
-        inv_freq = 1.0 / (
-            self.base
-            ** (torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim)
-        )
-        inv_freq = self._apply_inv_freq_scaling(inv_freq)
+        if config_inv_freq is not None:
+            inv_freq = config_inv_freq  # already scaled; skip subclass scaling
+        else:
+            # Normal Llama-3 RoPE
+            inv_freq = 1.0 / (
+                self.base
+                ** (
+                    torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim
+                )
+            )
+            inv_freq = self._apply_inv_freq_scaling(inv_freq)
         self.register_buffer("inv_freq", inv_freq, persistent = False)
 
         # Build here to make `torch.jit.trace` work.
@@ -1704,8 +1809,10 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         freqs = torch.outer(t, self.inv_freq)
         # Different from paper, but it uses a different permutation in order to obtain the same calculation
         emb = torch.cat((freqs, freqs), dim = -1)
-        cos = emb.cos().to(dtype = dtype, device = device, non_blocking = True)
-        sin = emb.sin().to(dtype = dtype, device = device, non_blocking = True)
+        # Applied here so attention_scaling survives extend_rope_embedding rebuilds;
+        # default 1.0 keeps unscaled paths bit-identical.
+        cos = (emb.cos() * self.attention_scaling).to(dtype = dtype, device = device, non_blocking = True)
+        sin = (emb.sin() * self.attention_scaling).to(dtype = dtype, device = device, non_blocking = True)
         self.multi_gpu_cos_cached[device.index] = cos
         self.multi_gpu_sin_cached[device.index] = sin
         return cos, sin
@@ -3467,6 +3574,14 @@ class FastLlamaModel:
             embeddings = model.get_output_embeddings()
             if hasattr(embeddings, "training"):
                 embeddings.training = False
+
+        # Restore use_cache values that prepare_model_for_training disabled
+        # for gradient checkpointing (older unsloth_zoo has no restore helper)
+        try:
+            from unsloth_zoo.training_utils import restore_use_cache
+            restore_use_cache(model)
+        except ImportError:
+            pass
         return model
 
     @staticmethod
@@ -3514,6 +3629,18 @@ class FastLlamaModel:
             embeddings = model.get_output_embeddings()
             if hasattr(embeddings, "training"):
                 embeddings.training = True
+
+        # Re-disable use_cache if prepare_model_for_training had disabled it
+        # and for_inference restored it (record only exists after a disable)
+        if (
+            use_gradient_checkpointing
+            and getattr(model, "_unsloth_use_cache_originals", None) is not None
+        ):
+            try:
+                from unsloth_zoo.training_utils import disable_use_cache
+                disable_use_cache(model)
+            except ImportError:
+                pass
         return model
 
 
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 6049fe343f..d9f092dee1 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -723,6 +723,14 @@ class FastLanguageModel(FastLlamaModel):
             load_in_4bit_kwargs = False
             load_in_8bit_kwargs = False
 
+        # Mirror FastModel: bitsandbytes < 0.46.0 needs dynamo disabled.
+        # Best effort: never crash the load (old unsloth_zoo without the
+        # zoo #710 fix raises NameError here on Python 3.13).
+        try:
+            patch_compiling_bitsandbytes()
+        except Exception as e:
+            print(f"Unsloth: Could not patch bitsandbytes for torch.compile - {e}")
+
         model, tokenizer = dispatch_model.from_pretrained(
             model_name = model_name,
             max_seq_length = max_seq_length,
@@ -822,6 +830,7 @@ from ..kernels import (
     post_patch_loss_function,
 )
 from .vision import FastBaseModel
+from .diffusion import FastDiffusionModel, is_diffusion_model_type
 from transformers import (
     AutoModelForCausalLM,
 )
@@ -839,6 +848,25 @@ class FastModel(FastBaseModel):
         model = _prepare_model_for_qat(model, qat_scheme)
         return model
 
+    @staticmethod
+    def get_peft_model(model, *args, **kwargs):
+        # Route text-diffusion models (slow path) to the transformers-only PEFT helper.
+        if getattr(model, "_unsloth_slow_diffusion", False):
+            return FastDiffusionModel.get_peft_model(model, *args, **kwargs)
+        return FastBaseModel.get_peft_model(model, *args, **kwargs)
+
+    @staticmethod
+    def for_inference(model):
+        if getattr(model, "_unsloth_slow_diffusion", False):
+            return FastDiffusionModel.for_inference(model)
+        return FastBaseModel.for_inference(model)
+
+    @staticmethod
+    def for_training(model, use_gradient_checkpointing = True):
+        if getattr(model, "_unsloth_slow_diffusion", False):
+            return FastDiffusionModel.for_training(model, use_gradient_checkpointing)
+        return FastBaseModel.for_training(model, use_gradient_checkpointing)
+
     @staticmethod
     def from_pretrained(
         model_name = "unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit",
@@ -1058,6 +1086,24 @@ class FastModel(FastBaseModel):
                 local_files_only = True
                 kwargs["local_files_only"] = True
 
+        # Text-diffusion slow-path dispatch, factored so both the normal route (below) and the
+        # legacy-config fallback (in the AutoConfig except handler) share one call site.
+        def _dispatch_diffusion():
+            return FastDiffusionModel.from_pretrained(
+                model_name = model_name,
+                max_seq_length = max_seq_length,
+                dtype = dtype,
+                load_in_4bit = load_in_4bit,
+                load_in_8bit = load_in_8bit,
+                load_in_16bit = load_in_16bit,
+                full_finetuning = full_finetuning,
+                token = token,
+                device_map = device_map,
+                trust_remote_code = trust_remote_code,
+                revision = revision,
+                **kwargs,
+            )
+
         try:
             model_config = AutoConfig.from_pretrained(
                 model_name,
@@ -1071,6 +1117,12 @@ class FastModel(FastBaseModel):
             raise
         except Exception as error:
             autoconfig_error = str(error)
+            # Legacy text-diffusion configs use model_type "diffusion_gemma", which current
+            # transformers does not register by name (it ships "diffusion_gemma4"). AutoConfig
+            # raises before we can dispatch; route straight to the diffusion slow path, whose
+            # loader aliases the legacy type to the gemma4 classes.
+            if "diffusion_gemma" in autoconfig_error and is_diffusion_model_type("diffusion_gemma"):
+                return _dispatch_diffusion()
             if "architecture" in autoconfig_error:
                 if "qwen3_5" in autoconfig_error:
                     raise ImportError(
@@ -1119,6 +1171,13 @@ class FastModel(FastBaseModel):
         )
         model_types_all = ",".join(model_types) + ","
 
+        # ---- Text-diffusion models (e.g. DiffusionGemma) take a transformers-only slow path. ----
+        # These use a custom block-diffusion `generate` and a novel backbone, so we skip Unsloth's
+        # autoregressive kernel/compile patching and load the unmodified HF model (bit-identical to
+        # naive transformers), keeping only 4bit/8bit + PEFT LoRA conveniences.
+        if is_diffusion_model_type(model_types):
+            return _dispatch_diffusion()
+
         # Save model types and loading method
         lowered_model_name = model_name.lower()
         string = os.environ.get("UNSLOTH_MODEL_NAME", "") + model_types_all
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index daa4a4835c..a52ab359bd 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -283,6 +283,9 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
         input_ids = kwargs["input"]
     elif "input_features" in kwargs:
         input_ids = kwargs["input_features"]
+    elif "inputs_embeds" in kwargs:
+        # canonical HF name for embedding inputs (e.g. multimodal generate)
+        input_ids = kwargs["inputs_embeds"]
     elif "input_embeds" in kwargs:
         input_ids = kwargs["input_embeds"]
     elif "inputs" in kwargs:
@@ -1280,6 +1283,15 @@ class FastBaseModel:
         tokenizer.padding_side = "left"  # Force inference
         if hasattr(tokenizer, "tokenizer"):
             tokenizer.tokenizer.padding_side = "left"  # Force inference
+        # Audio feature extractors must stay right padded: left (a text setting,
+        # forwarded by from_pretrained) shifts Whisper mels and desyncs Gemma 4
+        # audio token counts (crash on transformers < 5.10).
+        feature_extractor = getattr(tokenizer, "feature_extractor", None)
+        if (
+            feature_extractor is not None
+            and getattr(feature_extractor, "padding_side", None) == "left"
+        ):
+            feature_extractor.padding_side = "right"
         m = model
         while hasattr(m, "model"):
             m.max_seq_length = max_seq_length
@@ -1706,6 +1718,14 @@ class FastBaseModel:
             embeddings = model.get_output_embeddings()
             if hasattr(embeddings, "training"):
                 embeddings.training = False
+        # Restore use_cache values that prepare_model_for_training disabled
+        # for gradient checkpointing (older unsloth_zoo has no restore helper)
+        try:
+            from unsloth_zoo.training_utils import restore_use_cache
+            restore_use_cache(model)
+        except ImportError:
+            pass
+
         # Must disable returning hidden states in the case for GRPO
         os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
         # Must enable returning logits
@@ -1764,9 +1784,143 @@ class FastBaseModel:
             embeddings = model.get_output_embeddings()
             if hasattr(embeddings, "training"):
                 embeddings.training = True
+        # Re-disable use_cache if prepare_model_for_training had disabled it
+        # and for_inference restored it (record only exists after a disable)
+        if (
+            use_gradient_checkpointing
+            and getattr(model, "_unsloth_use_cache_originals", None) is not None
+        ):
+            try:
+                from unsloth_zoo.training_utils import disable_use_cache
+                disable_use_cache(model)
+            except ImportError:
+                pass
+
         # Can re-enable not returning logits
         os.environ["UNSLOTH_RETURN_LOGITS"] = "0"
         # Turn off skip guards and set stance to default
         if torch_compiler_set_stance is not None:
             torch_compiler_set_stance(stance = "default", skip_guard_eval_unsafe = False)
         return model
+
+
+def _looks_like_message_list(value):
+    return isinstance(value, list) and (len(value) == 0 or isinstance(value[0], dict))
+
+
+def _iter_message_lists(example, column):
+    if _looks_like_message_list(example):
+        yield example
+        return
+    if not isinstance(example, dict):
+        return
+    seen_keys = set()
+    for key in (column, "messages", "conversations", "prompt", "completion"):
+        if key in seen_keys:
+            continue
+        seen_keys.add(key)
+        value = example.get(key)
+        if _looks_like_message_list(value):
+            yield value
+
+
+def _local_path_from_video_value(video_path):
+    # data: URIs are inline payloads, not files, and contain no "://"
+    if video_path.startswith("data:"):
+        return None
+    if "://" not in video_path:
+        return video_path
+    if not video_path.startswith("file://"):
+        return None
+    from urllib.parse import urlparse
+    from urllib.request import url2pathname
+
+    parsed = urlparse(video_path)
+    # RFC 8089: only an empty authority or "localhost" is the local machine
+    if parsed.netloc and parsed.netloc != "localhost":
+        return None
+    path = url2pathname(parsed.path)
+    return path or None
+
+
+def check_dataset_for_missing_videos(
+    dataset,
+    column = "messages",
+    raise_error = True,
+    checked = None,
+):
+    """
+    Validate that local video paths referenced in a dataset exist, catching
+    missing files before training (torchvision otherwise returns an empty
+    tensor and the model silently receives no video signal).
+
+    Args:
+        dataset:     Map-style Dataset, list of dicts, or iterable of examples
+                     (not a streaming IterableDataset - iterating consumes it).
+        column:      Chat-messages column, default "messages"; "conversations",
+                     "prompt" and "completion" are also scanned.
+        raise_error: True (default) raises FileNotFoundError listing missing
+                     files; False warns and returns them.
+        checked:     Optional set of known-good paths for cross-call dedup.
+
+    Returns:
+        List[str]: Missing file paths (empty when all exist).
+    """
+    try:
+        from datasets import IterableDataset as _IterableDataset
+        if isinstance(dataset, _IterableDataset):
+            warnings.warn(
+                "Unsloth: check_dataset_for_missing_videos received a streaming "
+                "IterableDataset; iterating would exhaust it and training would "
+                "see zero samples. Skipping validation - pass a map-style Dataset "
+                "or rely on the UnslothVisionDataCollator's per-batch check.",
+                stacklevel = 2,
+            )
+            return []
+    except ImportError:
+        pass
+
+    missing = []
+    # Report each missing path once; only confirmed-existing paths enter
+    # `checked`, so retries after an error re-check previously missing files.
+    seen_missing = set()
+    if checked is None:
+        checked = set()
+
+    for example in dataset:
+        for messages in _iter_message_lists(example, column):
+            for msg in messages:
+                if not isinstance(msg, dict):
+                    continue
+                content = msg.get("content", [])
+                if not isinstance(content, (list, tuple)):
+                    continue
+                for item in content:
+                    if not isinstance(item, dict) or item.get("type") != "video":
+                        continue
+                    video_path = item.get("video", "")
+                    if not isinstance(video_path, str) or not video_path:
+                        continue
+                    path = _local_path_from_video_value(video_path)
+                    if path is None or path in checked or path in seen_missing:
+                        continue
+                    if not os.path.isfile(path):
+                        seen_missing.add(path)
+                        missing.append(path)
+                    else:
+                        checked.add(path)
+
+    if missing:
+        missing_list = "\n".join(f"  - {p}" for p in missing)
+        error_msg = (
+            f"Unsloth: {len(missing)} video file(s) referenced in your dataset could not be found.\n"
+            "Training would silently continue with empty video tensors - the model would receive\n"
+            "no actual video signal while loss still appears to decrease.\n\n"
+            f"Missing files:\n{missing_list}\n\n"
+            "Fix: verify the video file paths in your dataset before calling the trainer."
+        )
+        if raise_error:
+            raise FileNotFoundError(error_msg)
+        warnings.warn(error_msg, stacklevel = 2)
+
+    return missing
diff --git a/unsloth/save.py b/unsloth/save.py
index 4ff56cc477..629cbb9548 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -1092,7 +1092,6 @@ def unsloth_save_model(
     gc.collect()
 
     # Remove temporary location
-    import shutil
 
     shutil.rmtree(temporary_location, ignore_errors = True)
 
@@ -1224,7 +1223,6 @@ def install_llama_cpp_old(version = -10):
         for i in range(30):
             print(f"**[WARNING]** Deleting llama.cpp directory... {30-i} seconds left.")
             time.sleep(1)
-        import shutil
 
         shutil.rmtree("llama.cpp", ignore_errors = True)
 
@@ -2494,7 +2492,6 @@ def unsloth_push_to_hub_gguf(
 
     except Exception as e:
         if cleanup_temp:
-            import shutil
             for d in [save_directory, f"{save_directory}_gguf"]:
                 try:
                     shutil.rmtree(d)
@@ -2681,7 +2678,6 @@ This model was finetuned and converted to GGUF format using [Unsloth](https://gi
         # Clean up temporary directory
         if cleanup_temp:
             print("Unsloth: Cleaning up temporary files...")
-            import shutil
             for d in [save_directory, f"{save_directory}_gguf"]:
                 if os.path.exists(d):
                     try:
diff --git a/unsloth/trainer.py b/unsloth/trainer.py
index 9b80b10705..8790099fe3 100644
--- a/unsloth/trainer.py
+++ b/unsloth/trainer.py
@@ -36,8 +36,9 @@ from unsloth_zoo.training_utils import (
     unsloth_train as _unsloth_train,
 )
 from unsloth_zoo.vision_utils import (
-    UnslothVisionDataCollator,
+    UnslothVisionDataCollator as _UnslothVisionDataCollatorBase,
 )
+from unsloth.models.vision import check_dataset_for_missing_videos
 from unsloth_zoo.hf_utils import get_transformers_model_type
 from unsloth_zoo.utils import Version
 import dataclasses
@@ -49,10 +50,48 @@ __all__ = [
     "_patch_trl_trainer",
     "UnslothVisionDataCollator",
     "QGaloreConfig",
+    "check_dataset_for_missing_videos",
 ]
 
 logger = logging.getLogger(__name__)
 
+
+class UnslothVisionDataCollator(_UnslothVisionDataCollatorBase):
+    """
+    Drop-in zoo collator that validates local video paths on every batch
+    (deduped across batches), applying formatting_func first so formatter-made
+    paths are checked too. Raises FileNotFoundError on missing files instead
+    of silently training on empty video tensors (issue #5085).
+    """
+
+    __slots__ = ("_checked_video_paths",)
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._checked_video_paths = set()
+
+    def __call__(self, examples):
+        formatting_func = self.formatting_func
+        if formatting_func is not None:
+            examples = [formatting_func(example) for example in examples]
+
+        check_dataset_for_missing_videos(
+            examples,
+            raise_error = True,
+            checked = self._checked_video_paths,
+        )
+
+        if formatting_func is None:
+            return super().__call__(examples)
+
+        # why: base __call__ would reapply formatting_func; applied above.
+        self.formatting_func = None
+        try:
+            return super().__call__(examples)
+        finally:
+            self.formatting_func = formatting_func
+
+
 _AUTO_PADDING_FREE_ENV_DISABLED = os.environ.get(
     "UNSLOTH_DISABLE_AUTO_PADDING_FREE", ""
 ).strip().lower() in {"1", "true", "yes", "on"}
diff --git a/unsloth_cli/__init__.py b/unsloth_cli/__init__.py
index c8ec4c66c0..08d7e17e8c 100644
--- a/unsloth_cli/__init__.py
+++ b/unsloth_cli/__init__.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import os.path as _osp
+import os as _os
 import sys as _sys
 
 import typer
@@ -10,6 +10,7 @@ from importlib.metadata import version as package_version, PackageNotFoundError
 
 from unsloth_cli.commands.train import train
 from unsloth_cli.commands.inference import inference
+from unsloth_cli.commands.chat import chat
 from unsloth_cli.commands.export import export, list_checkpoints
 from unsloth_cli.commands.studio import (
     run as studio_run,
@@ -20,7 +21,7 @@ from unsloth_cli.commands.studio import (
 
 # Canonicalise `-np<N>` only under the `unsloth` console-script;
 # third-party scripts that import unsloth_cli keep their argv intact.
-_entry_base = _osp.basename(_sys.argv[0]).lower() if _sys.argv else ""
+_entry_base = _os.path.basename(_sys.argv[0]).lower() if _sys.argv else ""
 if _entry_base in {"unsloth", "unsloth.exe"}:
     _expand_attached_np_short()
 del _entry_base
@@ -53,11 +54,26 @@ def main(
         help = "Show version and exit.",
     ),
 ):
-    pass
+    if (
+        _sys.platform == "win32"
+    ):  # this block catches unsloth running inside of System32 or any subdirs, this WILL cause errors if not prevented.
+        _cwd = _os.path.normcase(_os.path.normpath(_os.getcwd()))
+        _system32 = _os.path.normcase(
+            _os.path.normpath(_os.path.join(_os.environ.get("WINDIR", r"C:\Windows"), "System32"))
+        )
+        if _cwd == _system32 or _cwd.startswith(_system32 + _os.sep):
+            typer.secho(
+                "Refusing to run Unsloth inside System32 as it will lead to Errors.\n"
+                "cd to a normal working directory and try again.",
+                fg = "red",
+                err = True,
+            )
+            raise typer.Exit(code = 1)
 
 
 app.command()(train)
 app.command()(inference)
+app.command()(chat)
 app.command()(export)
 app.command("list-checkpoints")(list_checkpoints)
 app.add_typer(studio_app, name = "studio", help = "Unsloth Studio commands.")
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
new file mode 100644
index 0000000000..5e734ee2a1
--- /dev/null
+++ b/unsloth_cli/_inference.py
@@ -0,0 +1,417 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Model loading and streaming shared by `inference` and `chat`."""
+
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+
+import typer
+
+_THINK_OPEN = "<think>"
+_THINK_BLOCK = re.compile(rf"{re.escape(_THINK_OPEN)}.*?</think>", re.DOTALL)
+
+
+def ensure_studio_backend_path() -> None:
+    backend_dir = str(Path(__file__).resolve().parents[1] / "studio" / "backend")
+    if backend_dir not in sys.path:
+        sys.path.insert(0, backend_dir)
+
+
+def configure_quiet_logging() -> None:
+    import logging
+
+    import structlog
+
+    # The CLI never configures structlog, so without this every backend INFO
+    # line prints. LOG_LEVEL is exported so the worker subprocess inherits it.
+    level_name = os.environ.setdefault("LOG_LEVEL", "WARNING").upper()
+    level = getattr(logging, level_name, logging.WARNING)
+    structlog.configure(wrapper_class = structlog.make_filtering_bound_logger(level))
+    os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
+
+
+def visible_text(text: str, show_thinking: bool) -> str:
+    if show_thinking:
+        return text
+    text = _THINK_BLOCK.sub("", text)
+    # Hold back an unclosed trailing <think> so reasoning never leaks mid-stream.
+    open_idx = text.find(_THINK_OPEN)
+    if open_idx != -1:
+        text = text[:open_idx]
+    max_prefix = min(len(text), len(_THINK_OPEN) - 1)
+    for size in range(max_prefix, 0, -1):
+        if _THINK_OPEN.startswith(text[-size:]):
+            return text[:-size]
+    return text
+
+
+def stream_to_stdout(stream, show_thinking: bool) -> str:
+    # Backends yield the full text-so-far on each step (llama.cpp ends with a
+    # metadata dict, skipped); print the growing tail, return the raw text.
+    raw = ""
+    shown = ""
+    for chunk in stream:
+        if not isinstance(chunk, str):
+            continue
+        raw = chunk
+        rendered = visible_text(chunk, show_thinking)
+        delta = rendered[len(shown) :]
+        if delta:
+            sys.stdout.write(delta)
+            sys.stdout.flush()
+        shown = rendered
+    sys.stdout.write("\n")
+    sys.stdout.flush()
+    return raw
+
+
+def stream_markdown(stream, show_thinking: bool, *, console) -> str:
+    from rich.live import Live
+    from rich.markdown import Markdown
+    from rich.text import Text
+
+    raw = ""
+    with Live(console = console, refresh_per_second = 12, vertical_overflow = "visible") as live:
+        for chunk in stream:
+            if not isinstance(chunk, str):
+                continue
+            raw = chunk
+            visible = visible_text(chunk, show_thinking)
+            live.update(Markdown(visible) if visible.strip() else Text(""))
+    return raw
+
+
+def collect_stream(stream, show_thinking: bool) -> str:
+    raw = ""
+    for chunk in stream:
+        if isinstance(chunk, str):
+            raw = chunk
+    return visible_text(raw, show_thinking)
+
+
+def render_columns(
+    left_label: str,
+    left_text: str,
+    right_label: str,
+    right_text: str,
+    *,
+    console = None,
+) -> None:
+    from rich import box
+    from rich.console import Console
+    from rich.table import Table
+
+    table = Table(box = box.MINIMAL, expand = True, padding = (0, 1), pad_edge = False)
+    table.add_column(left_label, header_style = "bold yellow", ratio = 1, overflow = "fold")
+    table.add_column(right_label, header_style = "bold magenta", ratio = 1, overflow = "fold")
+    table.add_row(left_text or "", right_text or "")
+    (console or Console()).print(table)
+
+
+class ChatBackend:
+    """Uniform stream()/close() over the llama-server and Unsloth backends."""
+
+    def __init__(self, kind: str, backend) -> None:
+        self._kind = kind  # "gguf" | "unsloth"
+        self._backend = backend
+
+    def stream(
+        self,
+        messages: list,
+        *,
+        system_prompt: str,
+        temperature: float,
+        top_p: float,
+        top_k: int,
+        max_new_tokens: int,
+        repetition_penalty: float,
+        enable_thinking: bool,
+        use_adapter: Optional[bool] = None,
+    ):
+        if self._kind == "gguf":
+            # llama-server takes the system prompt as the first message.
+            msgs = list(messages)
+            if system_prompt:
+                msgs = [{"role": "system", "content": system_prompt}, *msgs]
+            return self._backend.generate_chat_completion(
+                messages = msgs,
+                temperature = temperature,
+                top_p = top_p,
+                top_k = top_k,
+                max_tokens = max_new_tokens,
+                repetition_penalty = repetition_penalty,
+                enable_thinking = enable_thinking,
+            )
+        gen_kwargs = dict(
+            messages = messages,
+            system_prompt = system_prompt,
+            temperature = temperature,
+            top_p = top_p,
+            top_k = top_k,
+            max_new_tokens = max_new_tokens,
+            repetition_penalty = repetition_penalty,
+            enable_thinking = enable_thinking,
+        )
+        if use_adapter is not None:
+            return self._backend.generate_with_adapter_control(
+                use_adapter = use_adapter, **gen_kwargs
+            )
+        return self._backend.generate_chat_response(**gen_kwargs)
+
+    def close(self) -> None:
+        # Shut the worker down directly: the graceful unload_model waits for
+        # an ack that compare mode can swallow, hanging exit for minutes.
+        try:
+            if self._kind == "gguf":
+                self._backend.unload_model()
+            else:
+                self._backend._shutdown_subprocess(timeout = 2.0)
+        except Exception:
+            pass
+
+
+def resolve_model_config(model: str, *, hf_token: Optional[str]):
+    ensure_studio_backend_path()
+    from utils.models import ModelConfig
+
+    model_config = ModelConfig.from_identifier(model_id = model, hf_token = hf_token)
+    if not model_config:
+        typer.echo("Could not resolve model config", err = True)
+        raise typer.Exit(code = 1)
+    return model_config
+
+
+def _load_gguf_backend(model_config, *, hf_token, max_seq_length):
+    ensure_studio_backend_path()
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    llama_backend = LlamaCppBackend()
+    common = dict(
+        hf_variant = model_config.gguf_variant,
+        model_identifier = model_config.identifier,
+        is_vision = model_config.is_vision,
+        n_ctx = max_seq_length,
+    )
+    if model_config.gguf_hf_repo:
+        loaded = llama_backend.load_model(
+            hf_repo = model_config.gguf_hf_repo, hf_token = hf_token, **common
+        )
+    else:
+        loaded = llama_backend.load_model(
+            gguf_path = model_config.gguf_file,
+            mmproj_path = model_config.gguf_mmproj_file,
+            mtp_draft_path = model_config.gguf_mtp_file,
+            **common,
+        )
+    if not loaded:
+        typer.echo("Model load failed", err = True)
+        raise typer.Exit(code = 1)
+    return ChatBackend("gguf", llama_backend)
+
+
+def load_chat_backend(
+    model: str,
+    *,
+    hf_token: Optional[str],
+    max_seq_length: int,
+    load_in_4bit: bool,
+    model_config = None,
+    fresh_backend: bool = False,
+):
+    """Load `model` in-process: GGUF via llama-server, else the orchestrator.
+
+    fresh_backend uses a private orchestrator so a second model (compare's
+    base column) can run alongside the main one.
+    """
+    if model_config is None:
+        model_config = resolve_model_config(model, hf_token = hf_token)
+
+    typer.echo(f"Loading {model}", err = True)
+
+    if model_config.is_gguf:
+        return _load_gguf_backend(model_config, hf_token = hf_token, max_seq_length = max_seq_length)
+
+    if fresh_backend:
+        ensure_studio_backend_path()
+        from core.inference import InferenceOrchestrator
+        backend = InferenceOrchestrator()
+    else:
+        ensure_studio_backend_path()
+        from core.inference import get_inference_backend
+        backend = get_inference_backend()
+    if not backend.load_model(
+        config = model_config,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        hf_token = hf_token,
+    ):
+        typer.echo("Model load failed", err = True)
+        raise typer.Exit(code = 1)
+    return ChatBackend("unsloth", backend)
+
+
+def find_studio_server(timeout: float = 0.4) -> Optional[str]:
+    import urllib.request
+    base = os.environ.get("UNSLOTH_STUDIO_URL", "http://127.0.0.1:8888").rstrip("/")
+    try:
+        with urllib.request.urlopen(f"{base}/api/health", timeout = timeout):
+            return base
+    except Exception:
+        return None
+
+
+def _studio_token() -> Optional[str]:
+    """Self-issue a JWT: the CLI runs as the same OS user as the server, so it
+    signs with the same stored secret the server validates against."""
+    try:
+        import studio.backend.core  # noqa: F401  puts studio/backend on sys.path
+
+        from studio.backend.auth import storage
+        from studio.backend.auth.authentication import create_access_token
+
+        row = storage.get_connection().execute("SELECT username FROM auth_user LIMIT 1").fetchone()
+        return create_access_token(row[0], desktop = True) if row else None
+    except Exception:
+        return None
+
+
+class HttpChatBackend:
+    """Chat against a running Studio server over its OpenAI-compatible API.
+
+    close() leaves the model loaded on purpose — the next session (or the
+    UI) starts instantly.
+    """
+
+    def __init__(self, base_url: str, token: str) -> None:
+        self._base = base_url
+        self._token = token
+
+    def _request(
+        self,
+        method: str,
+        path: str,
+        payload = None,
+        timeout = None,
+    ):
+        import json
+        import urllib.request
+
+        request = urllib.request.Request(
+            self._base + path,
+            data = None if payload is None else json.dumps(payload).encode(),
+            headers = {
+                "Authorization": f"Bearer {self._token}",
+                "Content-Type": "application/json",
+            },
+            method = method,
+        )
+        return urllib.request.urlopen(request, timeout = timeout)
+
+    def ensure_loaded(self, model: str, *, hf_token, max_seq_length, load_in_4bit) -> None:
+        typer.echo(f"Loading {model} on the Studio server", err = True)
+        try:
+            self._request(
+                "POST",
+                "/api/inference/load",
+                {
+                    "model_path": model,
+                    "hf_token": hf_token,
+                    "max_seq_length": max_seq_length,
+                    "load_in_4bit": load_in_4bit,
+                },
+            ).close()
+        except Exception as exc:
+            typer.echo(f"Model load failed: {exc}", err = True)
+            raise typer.Exit(code = 1)
+
+    def stream(
+        self,
+        messages: list,
+        *,
+        system_prompt: str,
+        temperature: float,
+        top_p: float,
+        top_k: int,
+        max_new_tokens: int,
+        repetition_penalty: float,
+        enable_thinking: bool,
+        use_adapter: Optional[bool] = None,
+    ):
+        import json
+
+        msgs = list(messages)
+        if system_prompt:
+            msgs = [{"role": "system", "content": system_prompt}, *msgs]
+        resp = self._request(
+            "POST",
+            "/v1/chat/completions",
+            {
+                "model": "default",
+                "messages": msgs,
+                "stream": True,
+                "temperature": temperature,
+                "top_p": top_p,
+                "top_k": top_k,
+                "max_tokens": max_new_tokens,
+                "repetition_penalty": repetition_penalty,
+                "enable_thinking": enable_thinking,
+            },
+        )
+
+        def cumulative():
+            # Accumulate SSE deltas into the full-text-so-far convention the
+            # stream helpers expect.
+            text = ""
+            with resp:
+                for raw_line in resp:
+                    line = raw_line.decode("utf-8", "replace").strip()
+                    if not line.startswith("data:"):
+                        continue
+                    data = line[len("data:") :].strip()
+                    if data == "[DONE]":
+                        break
+                    try:
+                        parsed = json.loads(data)
+                    except ValueError:
+                        continue
+                    if "error" in parsed:
+                        raise RuntimeError(
+                            f"Server error: {parsed['error'].get('message', 'Unknown server error')}"
+                        )
+                    try:
+                        delta = parsed["choices"][0]["delta"].get("content")
+                    except (KeyError, IndexError):
+                        continue
+                    if not delta:
+                        continue
+                    text += delta
+                    # An emoji can arrive split across two deltas as lone
+                    # surrogate halves: hold back a trailing half, merge pairs.
+                    visible = text
+                    if "\ud800" <= visible[-1] <= "\udbff":
+                        visible = visible[:-1]
+                    yield visible.encode("utf-16", "surrogatepass").decode("utf-16", "replace")
+
+        return cumulative()
+
+    def close(self) -> None:
+        pass
+
+
+def connect_studio_server(model: str, *, hf_token, max_seq_length, load_in_4bit):
+    """Backend on a running Studio server, or None (caller loads locally)."""
+    base_url = find_studio_server()
+    if not base_url:
+        return None
+    token = _studio_token()
+    if not token:
+        return None
+    backend = HttpChatBackend(base_url, token)
+    backend.ensure_loaded(
+        model, hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit
+    )
+    return backend
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
new file mode 100644
index 0000000000..c62916bc75
--- /dev/null
+++ b/unsloth_cli/commands/chat.py
@@ -0,0 +1,340 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from typing import Optional
+
+import typer
+from rich.console import Console
+
+from unsloth_cli._inference import (
+    collect_stream,
+    configure_quiet_logging,
+    connect_studio_server,
+    ensure_studio_backend_path,
+    load_chat_backend,
+    render_columns,
+    resolve_model_config,
+    stream_markdown,
+    visible_text,
+)
+
+_HELP = (
+    "Commands: /exit (quit), /reset (clear history), "
+    "/think (toggle reasoning), /compare (base vs tuned), /help"
+)
+
+
+def _you_prompt(colors: bool) -> str:
+    # The prompt must go through input(), not a separate print — readline
+    # redraws erase anything they didn't draw, eating the label. GNU readline
+    # wants colors wrapped in \001/\002; libedit (macOS) prints those
+    # literally, so it gets raw ANSI.
+    try:
+        import readline
+    except ImportError:
+        return "\n\x1b[1;36mYou: \x1b[0m" if colors else "\nYou: "
+    libedit = (
+        "libedit" in (readline.__doc__ or "") or getattr(readline, "backend", "") == "editline"
+    )
+    if not colors:
+        return "\nYou: "
+    if libedit:
+        return "\n\x1b[1;36mYou: \x1b[0m"
+    return "\n\001\x1b[1;36m\002You: \001\x1b[0m\002"
+
+
+def _compare_blocked_reason(model_config) -> Optional[str]:
+    if model_config.is_gguf:
+        return (
+            "GGUF models can't toggle adapters — load a LoRA fine-tune "
+            "(transformers backend) to compare base vs tuned."
+        )
+    if not model_config.is_lora:
+        return (
+            "this isn't a LoRA adapter — compare turns the adapter off for the "
+            "'base' column, so there's nothing to compare against."
+        )
+    return None
+
+
+def _get_base_load_in_4bit(model_config) -> bool:
+    """Determine load_in_4bit for base model based on tuned adapter precision."""
+    if not model_config.is_lora or not model_config.path:
+        # Fallback to default if not a LoRA or no path
+        return True
+
+    try:
+        import json
+        from pathlib import Path
+
+        adapter_cfg_path = Path(model_config.path) / "adapter_config.json"
+        if not adapter_cfg_path.exists():
+            return True
+
+        with open(adapter_cfg_path) as f:
+            adapter_cfg = json.load(f)
+
+        training_method = adapter_cfg.get("unsloth_training_method")
+        if training_method == "lora":
+            return False
+        elif training_method == "qlora":
+            return True
+        elif not training_method:
+            # Fallback: check base model name for -bnb-4bit suffix
+            if model_config.base_model and "-bnb-4bit" not in model_config.base_model.lower():
+                return False
+            return True
+        return True
+    except Exception:
+        return True
+
+
+def _compare_needs_second_model() -> bool:
+    # MLX can't toggle the adapter off, so compare loads the base separately.
+    # detect_hardware() would print into the chat (and import torch), so
+    # probe its MLX condition quietly: Apple Silicon with mlx installed.
+    try:
+        from studio.backend.utils.hardware import hardware as hw
+
+        if hw.DEVICE is not None:
+            return hw.DEVICE == hw.DeviceType.MLX
+        if not hw.is_apple_silicon():
+            return False
+        import mlx.core  # noqa: F401
+
+        return True
+    except Exception:
+        return False
+
+
+def _pick_trained_model(console) -> str:
+    ensure_studio_backend_path()
+    from utils.models import scan_trained_models
+
+    trained = scan_trained_models()
+    if not trained:
+        typer.echo(
+            "No trained models found in your outputs folder. "
+            "Pass a model id or path: `unsloth chat <model>`.",
+            err = True,
+        )
+        raise typer.Exit(code = 1)
+
+    console.print("Your trained models (newest first):", style = "bold")
+    for i, (display_name, _, model_type) in enumerate(trained, 1):
+        console.print(f"  {i}. {display_name}  ({model_type})", markup = False)
+
+    while True:
+        try:
+            raw = input(f"Chat with [1-{len(trained)}, Enter = 1]: ").strip()
+        except (EOFError, KeyboardInterrupt):
+            raise typer.Exit(code = 1)
+        if not raw:
+            return trained[0][1]
+        if raw.isdigit() and 1 <= int(raw) <= len(trained):
+            return trained[int(raw) - 1][1]
+        console.print(f"Pick a number between 1 and {len(trained)}.", style = "yellow")
+
+
+def chat(
+    model: Optional[str] = typer.Argument(
+        None, help = "HF model id or local path. Omit to pick one of your trained models."
+    ),
+    hf_token: Optional[str] = typer.Option(
+        None, "--hf-token", envvar = "HF_TOKEN", help = "Hugging Face token if needed."
+    ),
+    temperature: float = typer.Option(0.7, "--temperature"),
+    top_p: float = typer.Option(0.9, "--top-p"),
+    top_k: int = typer.Option(40, "--top-k"),
+    max_new_tokens: int = typer.Option(512, "--max-new-tokens"),
+    repetition_penalty: float = typer.Option(1.1, "--repetition-penalty"),
+    system_prompt: str = typer.Option(
+        "", "--system-prompt", help = "Optional system prompt for the conversation."
+    ),
+    max_seq_length: int = typer.Option(4096, "--max-seq-length"),
+    load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
+    think: bool = typer.Option(
+        False,
+        "--think/--no-think",
+        help = "Start with the model's <think> reasoning shown. Toggle live with /think.",
+    ),
+    compare: bool = typer.Option(
+        False,
+        "--compare/--no-compare",
+        help = "Answer each prompt twice — base vs fine-tuned — side by side. "
+        "Needs a LoRA adapter. Toggle live with /compare.",
+    ),
+    verbose: bool = typer.Option(
+        False, "--verbose", "-v", help = "Show backend and llama-server logs."
+    ),
+    no_server: bool = typer.Option(
+        False,
+        "--no-server",
+        help = "Load the model in-process even if a Studio server is running.",
+    ),
+):
+    """Start an interactive chat with a model (loads once, stays warm)."""
+    if not verbose:
+        configure_quiet_logging()
+
+    console = Console()
+    err = Console(stderr = True)
+
+    if model is None:
+        model = _pick_trained_model(console)
+
+    # Resolve first so --compare can be rejected before the slow load.
+    model_config = resolve_model_config(model, hf_token = hf_token)
+    compare_blocked = _compare_blocked_reason(model_config)
+    if compare and compare_blocked:
+        err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
+        raise typer.Exit(code = 1)
+
+    load_opts = dict(hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit)
+
+    # Prefer a running Studio server: instant starts, model shared with the UI.
+    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    server_mode = chat_backend is not None
+    if server_mode:
+        console.print(
+            "(Studio server connected — model stays warm after /exit)",
+            style = "bright_black",
+        )
+    else:
+        chat_backend = load_chat_backend(model, model_config = model_config, **load_opts)
+
+    name = model_config.display_name or model
+    show_thinking = think
+    compare_mode = compare
+    messages = []
+
+    # Compare's base column: server mode keeps the tuned model remote and
+    # loads the base locally; local MLX (no adapter toggle) does the same;
+    # local CUDA just toggles the adapter on the one loaded model.
+    dual_compare = compare_blocked is None and (server_mode or _compare_needs_second_model())
+    base_backend = None
+
+    def load_base_for_compare():
+        nonlocal base_backend
+        if base_backend is not None:
+            return True
+        base_id = model_config.base_model
+        if not base_id:
+            console.print(
+                "(compare unavailable: this adapter doesn't record its base model)",
+                style = "yellow",
+            )
+            return False
+        console.print(
+            f"(loading base model {base_id} for compare — keeps two models in memory)",
+            style = "bright_black",
+            markup = False,
+        )
+        try:
+            # Use the same precision as the tuned model for fair comparison
+            base_load_opts = dict(load_opts)  # Copy original options
+            base_load_opts["load_in_4bit"] = _get_base_load_in_4bit(model_config)
+            base_backend = load_chat_backend(base_id, fresh_backend = True, **base_load_opts)
+        except Exception as exc:
+            err.print(f"(base model load failed: {exc})", style = "red", markup = False)
+            return False
+        return True
+
+    if compare and dual_compare and not load_base_for_compare():
+        raise typer.Exit(code = 1)
+
+    def generate(backend = None, use_adapter = None):
+        # Reads messages and show_thinking live, so /reset and /think apply.
+        return (backend or chat_backend).stream(
+            messages,
+            system_prompt = system_prompt,
+            temperature = temperature,
+            top_p = top_p,
+            top_k = top_k,
+            max_new_tokens = max_new_tokens,
+            repetition_penalty = repetition_penalty,
+            enable_thinking = show_thinking,
+            use_adapter = use_adapter,
+        )
+
+    console.print()
+    console.print(f"Chatting with {name}", style = "bold green", markup = False)
+    console.print(_HELP, style = "bright_black")
+
+    # legacy_windows: pre-VT consoles print raw ANSI as ←[1;36m garbage.
+    you_prompt = _you_prompt(console.is_terminal and not console.legacy_windows)
+    assistant_label = "[bold magenta]Assistant:[/bold magenta]"
+
+    try:
+        while True:
+            try:
+                user = input(you_prompt).strip()
+            except (EOFError, KeyboardInterrupt):
+                console.print()
+                break
+
+            if not user:
+                continue
+            if user in ("/exit", "/quit"):
+                break
+            if user == "/reset":
+                messages = []
+                console.print("(history cleared)", style = "bright_black")
+                continue
+            if user == "/think":
+                show_thinking = not show_thinking
+                state = "on" if show_thinking else "off"
+                console.print(f"(thinking {state})", style = "bright_black")
+                continue
+            if user == "/compare":
+                if compare_blocked:
+                    console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
+                    continue
+                if not compare_mode and dual_compare and not load_base_for_compare():
+                    continue
+                compare_mode = not compare_mode
+                state = "on" if compare_mode else "off"
+                console.print(f"(compare {state})", style = "bright_black")
+                continue
+            if user in ("/help", "/?"):
+                console.print(_HELP, style = "bright_black")
+                continue
+
+            messages.append({"role": "user", "content": user})
+
+            try:
+                if compare_mode:
+                    console.print("(comparing base vs tuned…)", style = "bright_black")
+                    if dual_compare:
+                        base_text = collect_stream(generate(backend = base_backend), show_thinking)
+                        tuned_text = collect_stream(generate(), show_thinking)
+                    else:
+                        base_text = collect_stream(generate(use_adapter = False), show_thinking)
+                        tuned_text = collect_stream(generate(use_adapter = True), show_thinking)
+                    console.print()
+                    render_columns(
+                        "base", base_text, f"{name} (tuned)", tuned_text, console = console
+                    )
+                    # History continues as the tuned model; base is just the reference.
+                    answer = tuned_text
+                else:
+                    console.print(assistant_label)
+                    answer = stream_markdown(generate(), show_thinking, console = console)
+            except KeyboardInterrupt:
+                # Ctrl-C aborts this answer only; drop the unanswered turn.
+                console.print("\n(interrupted)", style = "bright_black")
+                messages.pop()
+                continue
+            except Exception as exc:
+                err.print(f"\n(error: {exc})", style = "red", markup = False)
+                messages.pop()
+                continue
+
+            messages.append(
+                {"role": "assistant", "content": visible_text(answer, show_thinking = False)}
+            )
+    finally:
+        chat_backend.close()
+        if base_backend is not None:
+            base_backend.close()
+        err.print("\nBye.", style = "bright_black")
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 18de68c4d9..5dbc32c7d2 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.py
@@ -1,11 +1,17 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import sys
 from typing import Optional
 
 import typer
 
+from unsloth_cli._inference import (
+    configure_quiet_logging,
+    connect_studio_server,
+    load_chat_backend,
+    stream_to_stdout,
+)
+
 
 def inference(
     model: str = typer.Argument(..., help = "HF model id or local path."),
@@ -25,45 +31,46 @@ def inference(
     ),
     max_seq_length: int = typer.Option(2048, "--max-seq-length"),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
+    think: bool = typer.Option(
+        False,
+        "--think/--no-think",
+        help = "Show the model's <think> reasoning. Off by default so reasoning "
+        "models answer directly instead of spending the token budget thinking.",
+    ),
+    verbose: bool = typer.Option(
+        False,
+        "--verbose",
+        "-v",
+        help = "Show backend and llama-server logs (otherwise only the answer).",
+    ),
+    no_server: bool = typer.Option(
+        False,
+        "--no-server",
+        help = "Load the model in-process even if a Studio server is running.",
+    ),
 ):
     """Run a single inference using the specified model."""
-    from studio.backend.core import ModelConfig, get_inference_backend
+    if not verbose:
+        configure_quiet_logging()
 
-    inference_backend = get_inference_backend()
-    model_config = ModelConfig.from_ui_selection(
-        dropdown_value = model, search_value = None, hf_token = hf_token, is_lora = False
-    )
-    if not model_config:
-        typer.echo("Could not resolve model config", err = True)
-        raise typer.Exit(code = 1)
-
-    if not inference_backend.load_model(
-        config = model_config,
-        max_seq_length = max_seq_length,
-        load_in_4bit = load_in_4bit,
-        hf_token = hf_token,
-    ):
-        typer.echo("Model load failed", err = True)
-        raise typer.Exit(code = 1)
-
-    messages = [{"role": "user", "content": prompt}]
-    stream = inference_backend.generate_chat_response(
-        messages = messages,
-        system_prompt = system_prompt,
-        temperature = temperature,
-        top_p = top_p,
-        top_k = top_k,
-        max_new_tokens = max_new_tokens,
-        repetition_penalty = repetition_penalty,
-    )
-
-    typer.echo("Assistant:", nl = True)
-    previous = ""
-    for chunk in stream:
-        delta = chunk[len(previous) :]
-        if delta:
-            sys.stdout.write(delta)
-            sys.stdout.flush()
-        previous = chunk
-    sys.stdout.write("\n")
-    sys.stdout.flush()
+    # A running Studio server keeps the model warm between runs, which is
+    # exactly what a one-shot command wants.
+    load_opts = dict(hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit)
+    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    if chat_backend is None:
+        chat_backend = load_chat_backend(model, **load_opts)
+    try:
+        stream = chat_backend.stream(
+            [{"role": "user", "content": prompt}],
+            system_prompt = system_prompt,
+            temperature = temperature,
+            top_p = top_p,
+            top_k = top_k,
+            max_new_tokens = max_new_tokens,
+            repetition_penalty = repetition_penalty,
+            enable_thinking = think,
+        )
+        typer.echo("Assistant:")
+        stream_to_stdout(stream, show_thinking = think)
+    finally:
+        chat_backend.close()
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index a458f195b4..fb63fbf208 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -570,6 +570,19 @@ def _load_model_via_http(
         raise RuntimeError(f"Model load failed (HTTP {exc.code}): {body}") from exc
 
 
+def _format_context_length_line(load_result: dict) -> Optional[str]:
+    value = load_result.get("context_length")
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    if value_int <= 0:
+        return None
+    return f"  Context length: {value_int} tokens"
+
+
 # ── unsloth studio (server) ──────────────────────────────────────────
 
 
@@ -597,6 +610,11 @@ def studio_default(
             f"defaults to {_PARALLEL_DEFAULT_RUN}."
         ),
     ),
+    cloudflare: bool = typer.Option(
+        True,
+        "--cloudflare/--no-cloudflare",
+        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 (default on).",
+    ),
 ):
     """Launch the Unsloth Studio server."""
     # Runs before every subcommand (run/setup/update/...).
@@ -614,6 +632,16 @@ def studio_default(
                 err = True,
             )
             raise typer.Exit(2)
+        # Same for --no-cloudflare: it would not reach the subcommand.
+        if not cloudflare:
+            typer.echo(
+                f"Error: --no-cloudflare on `unsloth studio` applies to the "
+                f"plain-server path only. For `unsloth studio "
+                f"{ctx.invoked_subcommand}`, put it after the subcommand: "
+                f"`unsloth studio {ctx.invoked_subcommand} --no-cloudflare ...`",
+                err = True,
+            )
+            raise typer.Exit(2)
         return
 
     # Use the studio venv if it exists and we aren't already in it.
@@ -648,6 +676,8 @@ def studio_default(
                 args.append("--silent")
             if api_only:
                 args.append("--api-only")
+            # Forward the explicit polarity (matches run.py's BooleanOptionalAction).
+            args.append("--cloudflare" if cloudflare else "--no-cloudflare")
             # On Windows os.execvp keeps the parent alive, so Ctrl+C
             # would orphan the child; use Popen+wait instead.
             if sys.platform == "win32":
@@ -689,6 +719,7 @@ def studio_default(
         silent = silent,
         api_only = api_only,
         llama_parallel_slots = parallel,
+        cloudflare = cloudflare,
     )
     if frontend is not None:
         run_kwargs["frontend_path"] = frontend
@@ -823,7 +854,10 @@ def run(
         None, "--gguf-variant", help = "GGUF quant variant (e.g. UD-Q4_K_XL)"
     ),
     max_seq_length: int = typer.Option(
-        0, "--max-seq-length", help = "Max sequence length (0 = model default)"
+        0,
+        "--max-seq-length",
+        "--context-length",
+        help = "Runtime context length in tokens (0 = model default for GGUF; 2048 for hub models)",
     ),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
     api_key_name: str = typer.Option(
@@ -861,6 +895,11 @@ def run(
             f"{_PARALLEL_DEFAULT_RUN} (pre-PR hardcoded value)."
         ),
     ),
+    cloudflare: bool = typer.Option(
+        True,
+        "--cloudflare/--no-cloudflare",
+        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 (default on).",
+    ),
 ):
     """Start Studio, load a model, print an API key -- one-liner server.
 
@@ -980,6 +1019,8 @@ def run(
         # Typer claims --parallel outside ctx.args; without this the
         # child reverts to its default and silently drops the value.
         args.extend(["--parallel", str(parallel)])
+        # Forward the explicit polarity (same rationale as --load-in-4bit above).
+        args.append("--cloudflare" if cloudflare else "--no-cloudflare")
         # llama-server pass-through extras → child ctx.args → load payload.
         if extra_llama_args:
             args.extend(extra_llama_args)
@@ -997,7 +1038,13 @@ def run(
     # ── 2. Start server (always suppress built-in banner) ─────────────
     from studio.backend.run import run_server, _resolve_external_ip
 
-    run_kwargs = dict(host = host, port = port, silent = True, llama_parallel_slots = parallel)
+    run_kwargs = dict(
+        host = host,
+        port = port,
+        silent = True,
+        llama_parallel_slots = parallel,
+        cloudflare = cloudflare,
+    )
     if frontend is not None:
         run_kwargs["frontend_path"] = frontend
     app = run_server(**run_kwargs)
@@ -1011,40 +1058,52 @@ def run(
 
     set_tool_policy(enable_tools)
 
-    # 3. Wait for server health.
-    if not silent:
-        typer.echo("Starting Unsloth Studio...")
-    if not _wait_for_server(actual_port):
-        typer.echo("Error: server did not become healthy within 30 seconds.", err = True)
-        raise typer.Exit(1)
+    # Steps 3-5 can abort (health timeout, model-load error, or Ctrl+C during the
+    # slow load); tear the server and its children (llama-server, cloudflared) down
+    # on any abort so they never orphan.
+    from studio.backend.run import _graceful_shutdown, _server
 
-    # 4. Create API key in-process.
-    api_key = _create_api_key_inprocess(api_key_name)
-
-    # 5. Load model via HTTP.
-    if not silent:
-        typer.echo(f"Loading model: {model}...")
     try:
-        result = _load_model_via_http(
-            port = actual_port,
-            api_key = api_key,
-            model = model,
-            gguf_variant = gguf_variant,
-            max_seq_length = max_seq_length,
-            load_in_4bit = load_in_4bit,
-            llama_extra_args = extra_llama_args,
-        )
-    except RuntimeError as exc:
-        typer.echo(f"Error: {exc}", err = True)
-        raise typer.Exit(1)
+        # 3. Wait for server health.
+        if not silent:
+            typer.echo("Starting Unsloth Studio...")
+        if not _wait_for_server(actual_port):
+            typer.echo("Error: server did not become healthy within 30 seconds.", err = True)
+            raise typer.Exit(1)
+
+        # 4. Create API key in-process.
+        api_key = _create_api_key_inprocess(api_key_name)
+
+        # 5. Load model via HTTP.
+        if not silent:
+            typer.echo(f"Loading model: {model}...")
+        try:
+            result = _load_model_via_http(
+                port = actual_port,
+                api_key = api_key,
+                model = model,
+                gguf_variant = gguf_variant,
+                max_seq_length = max_seq_length,
+                load_in_4bit = load_in_4bit,
+                llama_extra_args = extra_llama_args,
+            )
+        except RuntimeError as exc:
+            typer.echo(f"Error: {exc}", err = True)
+            raise typer.Exit(1)
+    except BaseException:
+        _graceful_shutdown(_server)
+        raise
 
     loaded_model = result.get("model", model)
     display_variant = f" ({gguf_variant})" if gguf_variant else ""
+    context_length_line = _format_context_length_line(result)
 
     # 6. Print banner.
     display_host = _resolve_external_ip() if host == "0.0.0.0" else host
     base_url = f"http://{display_host}:{actual_port}"
     sdk_base_url = f"{base_url}/v1"
+    # run_server started the tunnel during the silent run above (0.0.0.0 only).
+    _cf_url = getattr(app.state, "cloudflare_url", None)
 
     # Orange so the tool-policy notice stands out; printed under
     # --silent / --yes too so the policy is never invisible.
@@ -1074,7 +1133,11 @@ def run(
         typer.echo("")
         typer.echo("=" * 56)
         typer.echo(f"  Unsloth Studio running at {base_url}")
+        if _cf_url:
+            typer.echo(f"  Secure link access via Cloudflare: {_cf_url}")
         typer.echo(f"  Model loaded: {loaded_model}{display_variant}")
+        if context_length_line:
+            typer.echo(context_length_line)
         typer.echo(f"  API Key:      {api_key}")
         typer.echo("")
         typer.echo("  OpenAI / Anthropic SDK base URL:")
@@ -1107,6 +1170,10 @@ def run(
     else:
         # Silent still prints URL + API key + tool-status policy.
         typer.echo(f"URL:     {base_url}")
+        if _cf_url:
+            typer.echo(f"Secure link access via Cloudflare: {_cf_url}")
+        if context_length_line:
+            typer.echo(context_length_line.strip())
         typer.echo(f"API Key: {api_key}")
         typer.secho(_tool_notice, fg = _tool_notice_fg, bold = True)
 
@@ -1130,6 +1197,33 @@ def run(
 _PID_FILE = STUDIO_HOME / "studio.pid"
 
 
+def _pid_alive(pid: int) -> bool:
+    """Return True if a process with ``pid`` exists.
+
+    ``os.kill(pid, 0)`` raises OSError (WinError 87) for every pid on Windows,
+    so use ``tasklist`` there and the signal-0 probe elsewhere.
+    """
+    if sys.platform == "win32":
+        try:
+            out = subprocess.run(
+                ["tasklist", "/FI", f"PID eq {int(pid)}", "/NH", "/FO", "CSV"],
+                capture_output = True,
+                text = True,
+                timeout = 10,
+            ).stdout
+        except Exception:
+            # Can't determine -- assume alive; taskkill no-ops if already gone.
+            return True
+        return f'"{int(pid)}"' in out
+    try:
+        os.kill(pid, 0)
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+    return True
+
+
 @studio_app.command()
 def stop():
     """Stop a running Unsloth Studio server.
@@ -1151,15 +1245,11 @@ def stop():
 
     pid = int(pid_text)
 
-    # Check if the process is still alive
-    try:
-        os.kill(pid, 0)
-    except ProcessLookupError:
+    # Check if still alive (os.kill(pid, 0) is invalid on Windows -- see _pid_alive).
+    if not _pid_alive(pid):
         typer.echo(f"Studio server (PID {pid}) is not running. Cleaning up stale PID file.")
         _PID_FILE.unlink(missing_ok = True)
         raise typer.Exit(0)
-    except PermissionError:
-        pass  # process exists but we may not own it; try to signal anyway
 
     # Send SIGTERM (graceful shutdown) or TerminateProcess on Windows
     try:
@@ -1176,17 +1266,13 @@ def stop():
         typer.echo(f"Failed to stop Studio server (PID {pid}): {e}", err = True)
         raise typer.Exit(1)
 
-    # Wait briefly for the process to exit and clean up
+    # Wait briefly for the process to exit and clean up.
     for _ in range(10):
         time.sleep(0.5)
-        try:
-            os.kill(pid, 0)
-        except ProcessLookupError:
+        if not _pid_alive(pid):
             _PID_FILE.unlink(missing_ok = True)
             typer.echo("Studio server stopped.")
             raise typer.Exit(0)
-        except PermissionError:
-            break
 
     typer.echo("Studio server is shutting down (may take a few seconds).")
 
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
new file mode 100644
index 0000000000..629cb46405
--- /dev/null
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -0,0 +1,401 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the `unsloth chat` / `unsloth inference` CLI — fakes only, no model loads."""
+
+from __future__ import annotations
+
+import inspect
+import sys
+import types
+from pathlib import Path
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+import typer
+from rich.console import Console
+from typer.testing import CliRunner
+
+import unsloth_cli.commands.chat as chatmod
+from unsloth_cli._inference import (
+    ChatBackend,
+    HttpChatBackend,
+    collect_stream,
+    render_columns,
+    visible_text,
+)
+
+
+class _FakeConfig:
+    is_gguf = False
+    is_lora = True
+    display_name = "fake-model"
+    base_model = "fake/base"
+    path = None
+
+
+def _chat_app():
+    cli = typer.Typer()
+    cli.command()(chatmod.chat)
+    return cli
+
+
+def test_visible_text_passthrough_when_shown():
+    text = "<think>reasoning</think>answer"
+    assert visible_text(text, show_thinking = True) == text
+
+
+def test_visible_text_strips_closed_think_block():
+    text = "<think>step 1\nstep 2</think>The answer is 42."
+    assert visible_text(text, show_thinking = False) == "The answer is 42."
+
+
+def test_visible_text_holds_unclosed_think():
+    # An open <think> is held back so partial reasoning never leaks mid-stream.
+    assert visible_text("<think>still thinking", show_thinking = False) == ""
+    assert visible_text("done.<think>more thinking", show_thinking = False) == "done."
+
+
+def test_visible_text_holds_partial_think_prefix():
+    # Streams are cumulative, so the opening tag can arrive as "<", "<thi",
+    # then "<think>". Hold possible tag prefixes until they are disambiguated.
+    assert visible_text("<", show_thinking = False) == ""
+    assert visible_text("<thi", show_thinking = False) == ""
+    assert visible_text("done.<thi", show_thinking = False) == "done."
+    assert visible_text("2 < 3", show_thinking = False) == "2 < 3"
+
+
+def _option(command_fn, name):
+    return inspect.signature(command_fn).parameters[name].default
+
+
+def test_inference_think_defaults_off():
+    from unsloth_cli.commands.inference import inference
+
+    opt = _option(inference, "think")
+    assert getattr(opt, "default", None) is False
+    # typer stores a flag/--no-flag pair as one combined decl.
+    assert "--think/--no-think" in (getattr(opt, "param_decls", None) or [])
+
+
+def test_chat_command_is_registered_with_options():
+    params = inspect.signature(chatmod.chat).parameters
+    assert "model" in params
+
+    think = _option(chatmod.chat, "think")
+    assert "--think/--no-think" in (getattr(think, "param_decls", None) or [])
+
+    compare = _option(chatmod.chat, "compare")
+    assert "--compare/--no-compare" in (getattr(compare, "param_decls", None) or [])
+
+    verbose = _option(chatmod.chat, "verbose")
+    assert {"--verbose", "-v"} <= set(getattr(verbose, "param_decls", None) or [])
+
+
+class _FakeBackend:
+    def __init__(self):
+        self.calls = []
+
+    def generate_chat_response(self, **kwargs):
+        self.calls.append(("plain", None, kwargs))
+        return iter(["hi"])
+
+    def generate_with_adapter_control(self, *, use_adapter, **kwargs):
+        self.calls.append(("adapter", use_adapter, kwargs))
+        return iter(["hi"])
+
+
+_STREAM_KWARGS = dict(
+    system_prompt = "",
+    temperature = 0.7,
+    top_p = 0.9,
+    top_k = 40,
+    max_new_tokens = 8,
+    repetition_penalty = 1.1,
+    enable_thinking = False,
+)
+
+
+def test_chatbackend_routes_compare_to_adapter_control():
+    fake = _FakeBackend()
+    backend = ChatBackend("unsloth", fake)
+
+    list(backend.stream([{"role": "user", "content": "x"}], use_adapter = False, **_STREAM_KWARGS))
+    list(backend.stream([{"role": "user", "content": "x"}], use_adapter = True, **_STREAM_KWARGS))
+
+    assert [(path, flag) for path, flag, _ in fake.calls] == [
+        ("adapter", False),
+        ("adapter", True),
+    ]
+
+
+def test_chatbackend_normal_path_skips_adapter_control():
+    fake = _FakeBackend()
+    backend = ChatBackend("unsloth", fake)
+
+    list(backend.stream([{"role": "user", "content": "x"}], **_STREAM_KWARGS))
+
+    assert fake.calls[0][0] == "plain"
+
+
+def test_collect_stream_returns_last_cumulative_think_stripped():
+    stream = iter(["<think>r</think>hel", "<think>r</think>hello"])
+    assert collect_stream(stream, show_thinking = False) == "hello"
+
+
+def test_render_columns_emits_both_answers_with_separator(capsys):
+    render_columns("base", "alpha", "tuned", "beta")
+    out = capsys.readouterr().out
+    assert "base" in out and "tuned" in out
+    assert "alpha" in out and "beta" in out
+    assert "│" in out
+
+
+def test_you_prompt_matches_readline_backend(monkeypatch):
+    gnu = types.ModuleType("readline")
+    gnu.__doc__ = "Importing this module enables command line editing using GNU readline."
+    monkeypatch.setitem(sys.modules, "readline", gnu)
+    prompt = chatmod._you_prompt(colors = True)
+    assert "You: " in prompt and "\001" in prompt
+
+    libedit = types.ModuleType("readline")
+    libedit.__doc__ = "Importing this module enables command line editing using libedit readline."
+    monkeypatch.setitem(sys.modules, "readline", libedit)
+    assert chatmod._you_prompt(colors = True) == "\n\x1b[1;36mYou: \x1b[0m"
+    assert chatmod._you_prompt(colors = False) == "\nYou: "
+
+    # Windows: no readline module at all; the console's own line editing
+    # handles backspace, so plain ANSI color (no markers) is safe.
+    monkeypatch.setitem(sys.modules, "readline", None)
+    assert chatmod._you_prompt(colors = True) == "\n\x1b[1;36mYou: \x1b[0m"
+    assert chatmod._you_prompt(colors = False) == "\nYou: "
+
+
+def test_chat_registered_on_app():
+    from unsloth_cli import app
+
+    # cmd.name is None until typer resolves it from the callback name.
+    names = {(cmd.name or cmd.callback.__name__) for cmd in app.registered_commands}
+    assert "chat" in names
+
+
+def test_chat_exits_cleanly_on_slash_exit(monkeypatch):
+    closed = []
+
+    class _FakeChatBackend:
+        def stream(self, *a, **k):
+            return iter(["hello"])
+
+        def close(self):
+            closed.append(True)
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: None)
+
+    runner = CliRunner()
+    for args in (["fake-model"], ["fake-model", "--compare"]):
+        closed.clear()
+        result = runner.invoke(_chat_app(), args, input = "hi\n/exit\n")
+        assert result.exit_code == 0, result.output
+        assert closed == [True]
+        assert "Bye." in result.output
+        # The prompt must go through input() (readline-safe), not a print.
+        assert "You: " in result.output
+        assert "You: You:" not in result.output
+
+
+def test_pick_trained_model_lists_and_selects(monkeypatch):
+    fake_models = types.ModuleType("utils.models")
+    fake_models.scan_trained_models = lambda: [
+        ("run-new", "outputs/run-new", "lora"),
+        ("run-old", "outputs/run-old", "merged"),
+    ]
+    monkeypatch.setitem(sys.modules, "utils.models", fake_models)
+
+    monkeypatch.setattr("builtins.input", lambda prompt = "": "2")
+    assert chatmod._pick_trained_model(Console()) == "outputs/run-old"
+
+    monkeypatch.setattr("builtins.input", lambda prompt = "": "")
+    assert chatmod._pick_trained_model(Console()) == "outputs/run-new"
+
+
+def test_chat_no_arg_chats_with_picked_trained_model(monkeypatch):
+    class _FakeChatBackend:
+        def stream(self, *a, **k):
+            return iter(["hello"])
+
+        def close(self):
+            pass
+
+    resolved = []
+    monkeypatch.setattr(chatmod, "_pick_trained_model", lambda console: "outputs/run-42")
+    monkeypatch.setattr(
+        chatmod,
+        "resolve_model_config",
+        lambda model, **k: (resolved.append(model), _FakeConfig())[1],
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: None)
+
+    result = CliRunner().invoke(_chat_app(), [], input = "/exit\n")
+    assert result.exit_code == 0, result.output
+    assert resolved == ["outputs/run-42"]
+
+
+def test_find_studio_server_none_when_not_running(monkeypatch):
+    import urllib.request
+
+    from unsloth_cli import _inference
+
+    def refuse(*a, **k):
+        raise OSError("connection refused")
+
+    monkeypatch.setattr(urllib.request, "urlopen", refuse)
+    assert _inference.find_studio_server() is None
+
+
+class _FakeSSEResponse:
+    def __init__(self, lines):
+        self._lines = lines
+
+    def __iter__(self):
+        return iter(self._lines)
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+def test_http_backend_streams_cumulative_text(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    response = _FakeSSEResponse(
+        [
+            b'data: {"choices":[{"delta":{"content":"He"}}]}\n',
+            b"\n",
+            b'data: {"choices":[{"delta":{"content":"llo"}}]}\n',
+            b"data: [DONE]\n",
+        ]
+    )
+    monkeypatch.setattr(backend, "_request", lambda *a, **k: response)
+
+    out = list(backend.stream([{"role": "user", "content": "hi"}], **_STREAM_KWARGS))
+    assert out == ["He", "Hello"]
+
+
+def test_http_backend_merges_emoji_split_across_deltas(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    response = _FakeSSEResponse(
+        [
+            b'data: {"choices":[{"delta":{"content":"hi "}}]}\n',
+            b'data: {"choices":[{"delta":{"content":"\\ud83d"}}]}\n',
+            b'data: {"choices":[{"delta":{"content":"\\ude0a"}}]}\n',
+            b"data: [DONE]\n",
+        ]
+    )
+    monkeypatch.setattr(backend, "_request", lambda *a, **k: response)
+
+    out = list(backend.stream([{"role": "user", "content": "hi"}], **_STREAM_KWARGS))
+    # The lone high surrogate is held back, then merged with its other half.
+    assert out == ["hi ", "hi ", "hi 😊"]
+
+
+def test_chat_prefers_running_studio_server(monkeypatch):
+    closed = []
+
+    class _FakeHttpBackend:
+        def stream(self, *a, **k):
+            return iter(["hello"])
+
+        def close(self):
+            closed.append("http")
+
+    local_loads = []
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: _FakeHttpBackend())
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: local_loads.append(1))
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert local_loads == []
+    assert "stays warm" in result.output
+    assert closed == ["http"]
+
+
+def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
+    streamed, closed, base_loads = [], [], []
+
+    class _FakeHttpBackend:
+        def stream(self, *a, **k):
+            streamed.append("tuned")
+            return iter(["tuned-answer"])
+
+        def close(self):
+            closed.append("http")
+
+    class _FakeBaseBackend:
+        def stream(self, *a, **k):
+            streamed.append("base")
+            return iter(["base-answer"])
+
+        def close(self):
+            closed.append("base")
+
+    def fake_local_load(model, **kwargs):
+        base_loads.append((model, kwargs.get("fresh_backend", False)))
+        return _FakeBaseBackend()
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: _FakeHttpBackend())
+    monkeypatch.setattr(chatmod, "load_chat_backend", fake_local_load)
+
+    result = CliRunner().invoke(_chat_app(), ["tuned-run"], input = "/compare\nhi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert "(compare on)" in result.output
+    # Only the base model loaded locally, on its own private backend.
+    assert base_loads == [("fake/base", True)]
+    assert streamed == ["base", "tuned"]
+    assert set(closed) == {"http", "base"}
+
+
+def test_chat_compare_on_mlx_loads_base_model_side_by_side(monkeypatch):
+    loads, streamed, closed = [], [], []
+
+    class _FakeLocalBackend:
+        def __init__(self, role):
+            self.role = role
+
+        def stream(self, *a, **k):
+            streamed.append((self.role, k.get("use_adapter")))
+            return iter([f"{self.role}-answer"])
+
+        def close(self):
+            closed.append(self.role)
+
+    def fake_load(model, **kwargs):
+        fresh = kwargs.get("fresh_backend", False)
+        loads.append((model, fresh))
+        return _FakeLocalBackend("base" if fresh else "tuned")
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(chatmod, "load_chat_backend", fake_load)
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: True)
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: None)
+
+    result = CliRunner().invoke(_chat_app(), ["tuned-run", "--compare"], input = "hi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert loads == [("tuned-run", False), ("fake/base", True)]
+    # Both models answered the turn, via plain generation (no adapter toggle).
+    assert ("base", None) in streamed and ("tuned", None) in streamed
+    assert set(closed) == {"tuned", "base"}
diff --git a/unsloth_cli/tests/test_studio_cloudflare_flag.py b/unsloth_cli/tests/test_studio_cloudflare_flag.py
new file mode 100644
index 0000000000..5c2a039547
--- /dev/null
+++ b/unsloth_cli/tests/test_studio_cloudflare_flag.py
@@ -0,0 +1,292 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the `--cloudflare/--no-cloudflare` Studio flag.
+
+Pins the typer Option (default on) on both `unsloth studio` and
+`unsloth studio run`, and that the chosen polarity reaches the re-exec'd
+child and run_server. Modeled on test_studio_run_parallel_flag.py.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+from typer.testing import CliRunner
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+def _studio():
+    from unsloth_cli.commands import studio as _studio_mod
+    return _studio_mod
+
+
+_BASE = ["--model", "unsloth/Qwen3-1.7B-GGUF"]
+
+
+# ── option registration ──────────────────────────────────────────────
+
+
+def test_run_exposes_cloudflare_option_default_on():
+    import inspect
+
+    sig = inspect.signature(_studio().run)
+    assert "cloudflare" in sig.parameters
+    opt = sig.parameters["cloudflare"].default
+    decls = set(getattr(opt, "param_decls", []) or [])
+    assert "--cloudflare/--no-cloudflare" in decls
+    assert getattr(opt, "default", None) is True
+
+
+def test_studio_default_exposes_cloudflare_option_default_on():
+    import inspect
+
+    sig = inspect.signature(_studio().studio_default)
+    assert "cloudflare" in sig.parameters
+    opt = sig.parameters["cloudflare"].default
+    assert getattr(opt, "default", None) is True
+
+
+# ── re-exec forwarding: `unsloth studio run` ─────────────────────────
+
+
+class _ExecCaptured(SystemExit):
+    def __init__(self, argv):
+        super().__init__(0)
+        self.argv = list(argv)
+
+
+def _install_run_reexec_capture(monkeypatch, *, platform = "linux"):
+    studio_mod = _studio()
+    captured = []
+
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    fake_bin = fake_venv / "bin" / "unsloth"
+    real_is_file = Path.is_file
+    monkeypatch.setattr(
+        Path,
+        "is_file",
+        lambda self: True if str(self) == str(fake_bin) else real_is_file(self),
+    )
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+    monkeypatch.setattr(sys, "platform", platform)
+
+    def fake_execvp(file, argv):
+        captured.append(list(argv))
+        raise _ExecCaptured(argv)
+
+    monkeypatch.setattr(studio_mod.os, "execvp", fake_execvp)
+    return captured
+
+
+def _invoke_run(monkeypatch, args):
+    import typer as _typer
+
+    studio_mod = _studio()
+    captured = _install_run_reexec_capture(monkeypatch)
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    CliRunner().invoke(app, args, catch_exceptions = True)
+    return captured
+
+
+@pytest.mark.parametrize(
+    "user_flag,expected,unexpected",
+    [
+        (None, "--cloudflare", "--no-cloudflare"),  # default on
+        ("--cloudflare", "--cloudflare", "--no-cloudflare"),
+        ("--no-cloudflare", "--no-cloudflare", "--cloudflare"),
+    ],
+)
+def test_run_reexec_forwards_cloudflare_polarity(monkeypatch, user_flag, expected, unexpected):
+    extras = [user_flag] if user_flag else []
+    captured = _invoke_run(monkeypatch, _BASE + extras)
+    assert len(captured) == 1, captured
+    argv = captured[0]
+    assert expected in argv, f"expected {expected} in child argv; got {argv}"
+    assert unexpected not in argv, f"unexpected {unexpected} in child argv; got {argv}"
+
+
+# ── re-exec forwarding: plain `unsloth studio` ───────────────────────
+
+
+def _invoke_studio_default(
+    monkeypatch,
+    args,
+    *,
+    platform = "linux",
+):
+    import typer as _typer
+
+    studio_mod = _studio()
+    captured = []
+
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    monkeypatch.setattr(studio_mod, "_ensure_studio_env_exported", lambda: None)
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)
+    monkeypatch.setattr(sys, "platform", platform)
+
+    def fake_execvp(file, argv):
+        captured.append(list(argv))
+        raise _ExecCaptured(argv)
+
+    monkeypatch.setattr(studio_mod.os, "execvp", fake_execvp)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    CliRunner().invoke(app, args, catch_exceptions = True)
+    return captured
+
+
+@pytest.mark.parametrize(
+    "user_flag,expected,unexpected",
+    [
+        (None, "--cloudflare", "--no-cloudflare"),
+        ("--no-cloudflare", "--no-cloudflare", "--cloudflare"),
+    ],
+)
+def test_studio_default_reexec_forwards_cloudflare(monkeypatch, user_flag, expected, unexpected):
+    extras = [user_flag] if user_flag else []
+    captured = _invoke_studio_default(monkeypatch, ["-H", "0.0.0.0"] + extras)
+    assert len(captured) == 1, captured
+    argv = captured[0]
+    assert expected in argv, f"expected {expected}; got {argv}"
+    assert unexpected not in argv, f"unexpected {unexpected}; got {argv}"
+
+
+# ── in-venv path forwards cloudflare into run_server ─────────────────
+
+
+class _RunServerCaptured(SystemExit):
+    def __init__(self, kwargs):
+        super().__init__(0)
+        self.kwargs = dict(kwargs)
+
+
+@pytest.mark.parametrize("user_flag,expected", [(None, True), ("--no-cloudflare", False)])
+def test_run_in_venv_passes_cloudflare_to_run_server(monkeypatch, user_flag, expected):
+    import types
+
+    studio_mod = _studio()
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
+
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+
+    captured: dict = {}
+
+    def fake_run_server(**kwargs):
+        captured.update(kwargs)
+        raise _RunServerCaptured(kwargs)
+
+    fake_backend_run = sys.modules.setdefault(
+        "studio.backend.run", types.ModuleType("studio.backend.run")
+    )
+    fake_backend_run.run_server = fake_run_server
+    fake_backend_run._resolve_external_ip = lambda: "127.0.0.1"
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    extras = [user_flag] if user_flag else []
+    CliRunner().invoke(app, _BASE + extras, catch_exceptions = True)
+
+    assert captured.get("cloudflare") is expected, captured
+
+
+# ── parent-level --no-cloudflare with a subcommand is rejected ───────
+
+
+def test_studio_default_rejects_no_cloudflare_with_subcommand(monkeypatch):
+    # `unsloth studio --no-cloudflare run ...` would not reach the subcommand,
+    # so it must error (mirrors --parallel) rather than silently still tunnel.
+    import typer as _typer
+
+    studio_mod = _studio()
+    app = _typer.Typer()
+    app.add_typer(studio_mod.studio_app, name = "studio")
+    result = CliRunner().invoke(app, ["studio", "--no-cloudflare", "run", "--model", "X"])
+    assert result.exit_code == 2, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "--no-cloudflare" in combined, combined
+
+
+# ── run() tears the server + tunnel down if startup aborts ───────────
+
+
+def test_run_in_venv_shuts_down_on_startup_abort(monkeypatch):
+    import types
+
+    studio_mod = _studio()
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
+
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+
+    class _App:
+        class state:
+            server_port = 8888
+
+    shutdown_calls = []
+    backend = sys.modules.setdefault("studio.backend.run", types.ModuleType("studio.backend.run"))
+    backend.run_server = lambda **k: _App()
+    backend._resolve_external_ip = lambda: "1.2.3.4"
+    backend._server = object()
+    backend._shutdown_event = None
+    backend._graceful_shutdown = lambda server: shutdown_calls.append(server)
+
+    # set_tool_policy is imported as `from state.tool_policy import set_tool_policy`.
+    state_mod = sys.modules.setdefault("state", types.ModuleType("state"))
+    tp_mod = sys.modules.setdefault("state.tool_policy", types.ModuleType("state.tool_policy"))
+    tp_mod.set_tool_policy = lambda *a, **k: None
+    state_mod.tool_policy = tp_mod
+
+    # Force the health check to fail so startup aborts after run_server().
+    monkeypatch.setattr(studio_mod, "_wait_for_server", lambda *a, **k: False)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    result = CliRunner().invoke(app, _BASE + ["-H", "0.0.0.0"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert len(shutdown_calls) == 1, "startup abort must call _graceful_shutdown"
diff --git a/unsloth_cli/tests/test_studio_run_parallel_flag.py b/unsloth_cli/tests/test_studio_run_parallel_flag.py
index abc0b3c7c4..8870fa348d 100644
--- a/unsloth_cli/tests/test_studio_run_parallel_flag.py
+++ b/unsloth_cli/tests/test_studio_run_parallel_flag.py
@@ -51,6 +51,18 @@ def test_parallel_option_is_registered():
         assert required in flags, f"flag {required!r} missing from --parallel option"
 
 
+def test_context_length_alias_is_registered():
+    """`--context-length` is an operator-facing alias for --max-seq-length."""
+    studio_mod = _load_run_command()
+    import inspect
+
+    sig = inspect.signature(studio_mod.run)
+    opt = sig.parameters["max_seq_length"].default
+    flags = set(getattr(opt, "param_decls", None) or [])
+    assert "--max-seq-length" in flags
+    assert "--context-length" in flags
+
+
 def test_parallel_default_is_four():
     """Default must stay at 4 so plain `unsloth studio run` is unchanged."""
     studio_mod = _load_run_command()
@@ -249,6 +261,15 @@ def test_reexec_np_is_first_class_alias(monkeypatch):
     assert _value_after(argv, "--port") == "8888", argv
 
 
+def test_reexec_forwards_context_length_alias(monkeypatch):
+    """Alias should normalize to the existing child --max-seq-length flag."""
+    result, captured = _invoke_run(monkeypatch, _BASE + ["--context-length", "8192"])
+    assert len(captured) == 1, result.output
+    argv = captured[0]["argv"]
+    assert _value_after(argv, "--max-seq-length") == "8192", argv
+    assert "--context-length" not in argv, argv
+
+
 def test_reexec_mixed_parallel_with_passthrough(monkeypatch):
     """--parallel + llama-server pass-through flags must all reach the child."""
     result, captured = _invoke_run(
@@ -262,6 +283,22 @@ def test_reexec_mixed_parallel_with_passthrough(monkeypatch):
     assert _value_after(argv, "--temp") == "0.7", argv
 
 
+def test_context_length_banner_line_formats_ints():
+    studio_mod = _load_run_command()
+    assert studio_mod._format_context_length_line({"context_length": 4096}) == (
+        "  Context length: 4096 tokens"
+    )
+    assert studio_mod._format_context_length_line({"context_length": "8192"}) == (
+        "  Context length: 8192 tokens"
+    )
+
+
+@pytest.mark.parametrize("value", [None, 0, -1, True, ""])
+def test_context_length_banner_line_omits_unknown_values(value):
+    studio_mod = _load_run_command()
+    assert studio_mod._format_context_length_line({"context_length": value}) is None
+
+
 @pytest.mark.parametrize(
     "user_flag,expected_in_child",
     [