diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 71c87c6859..7978a200c0 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -209,7 +209,7 @@ jobs:
             'peft>=0.18,<0.20' 'accelerate>=0.34,<2' \
             ipython
           # torchvision: unsloth_zoo.vision_utils imports it at module scope.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           # transformers + trl from the matrix combo.
           pip install "$RESOLVED_TRANSFORMERS_SPEC"
@@ -2174,7 +2174,7 @@ jobs:
           python -m pip install --upgrade pip
           # Match the matrix job's torch path so unsloth_zoo's
           # `import torch` resolves to the same CPU build.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           pip install \
             'numpy<3' protobuf sentencepiece \
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 864630f9f0..a2f716a93c 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -163,7 +163,7 @@ jobs:
             'pytest==9.0.3' \
             'pytest-asyncio==1.3.0' \
             'httpx==0.28.1'
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch==2.10.0'
           # github.com occasionally 500s on the git fetch; retry the
           # zoo install so a single upstream blip does not fail CI.
@@ -231,99 +231,6 @@ jobs:
             tests/studio/test_is_mlx_dispatch_gate.py \
             tests/studio/test_mlx_training_worker_behaviors.py
 
-      # 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:
-          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
-          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
-          # install_llama_prebuilt.py hits the GitHub releases API to
-          # resolve the asset URL. Anonymous calls share the runner-IP
-          # rate-limit bucket and 403 quickly -- pass the workflow's
-          # automatic GITHUB_TOKEN to bump us to the 5000/hr authenticated
-          # bucket.
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          set -euo pipefail
-          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
-          rm -rf "$INSTALL_DIR"
-          # Mirror studio/setup.sh on macOS (the install.sh user path):
-          # it plans against the unslothai/llama.cpp fork's latest
-          # release with no policy or tag flags.
-          python studio/install_llama_prebuilt.py \
-            --install-dir "$INSTALL_DIR" \
-            --published-repo unslothai/llama.cpp
-
-          # Studio bundles only llama-server + llama-quantize from the
-          # prebuilt (not llama-cli) -- inference goes through
-          # llama-server's HTTP /completion endpoint. Validate both:
-          # llama-quantize --help proves the dynamic libs link, then
-          # spin up llama-server and POST a /completion request on a
-          # tiny published GGUF.
-          LLAMA_SERVER="$INSTALL_DIR/build/bin/llama-server"
-          LLAMA_QUANT="$INSTALL_DIR/build/bin/llama-quantize"
-          [ -x "$LLAMA_SERVER" ] || { echo "::error::llama-server missing at $LLAMA_SERVER"; find "$INSTALL_DIR/build" -type f | head -40; exit 1; }
-          [ -x "$LLAMA_QUANT" ]  || { echo "::error::llama-quantize missing at $LLAMA_QUANT"; exit 1; }
-          echo "llama-server : $LLAMA_SERVER"
-          echo "llama-quantize: $LLAMA_QUANT"
-          "$LLAMA_QUANT" --help >/dev/null && echo "  llama-quantize loads OK"
-
-          mkdir -p /tmp/ggufs
-          bash .github/scripts/hf-download-with-retry.sh \
-            'unsloth/gemma-3-270m-it-GGUF' \
-            'gemma-3-270m-it-Q4_K_M.gguf' \
-            /tmp/ggufs
-
-          PORT=18080
-          echo "=== starting llama-server on 127.0.0.1:$PORT ==="
-          "$LLAMA_SERVER" \
-            -m /tmp/ggufs/gemma-3-270m-it-Q4_K_M.gguf \
-            --host 127.0.0.1 \
-            --port "$PORT" \
-            -c 256 \
-            -n 16 \
-            --no-warmup \
-            > /tmp/llama-server.log 2>&1 &
-          SERVER_PID=$!
-          trap 'kill "$SERVER_PID" 2>/dev/null || true' EXIT
-
-          # Wait for /health to come up
-          for i in $(seq 1 30); do
-            if curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
-              echo "  server up after ${i}s"
-              break
-            fi
-            sleep 1
-          done
-          if ! curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
-            echo "::error::llama-server never became healthy"
-            tail -40 /tmp/llama-server.log
-            exit 1
-          fi
-
-          PROMPT="Hello, my name is"
-          echo "=== POST /completion ==="
-          RESP=$(curl -sf -X POST "http://127.0.0.1:$PORT/completion" \
-            -H 'Content-Type: application/json' \
-            -d "{\"prompt\":\"$PROMPT\",\"n_predict\":16,\"temperature\":0,\"seed\":3407}")
-          echo "raw response (head): $(echo "$RESP" | head -c 600)"
-          CONTENT=$(echo "$RESP" | python -c "import json,sys; print(json.loads(sys.stdin.read()).get('content',''))")
-          echo "completion content: $CONTENT"
-
-          if [ -z "$CONTENT" ]; then
-            echo "::error::llama-server /completion returned empty content"
-            tail -40 /tmp/llama-server.log
-            exit 1
-          fi
-          echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works"
-
       # Real MLX training + inference smoke test. Trains
       # unsloth/gemma-3-270m-it for 7 deterministic LoRA steps
       # (batch_size=2, gradient_accumulation_steps=3) on a single
@@ -338,6 +245,9 @@ jobs:
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           mkdir -p mlx_workdir
+          # Authenticate llama.cpp's release-API lookup (anonymous 403s on rate-limit);
+          # read-only GITHUB_TOKEN scoped here only, never to steps that run binaries.
+          GH_TOKEN="${{ secrets.GITHUB_TOKEN }}" GITHUB_TOKEN="${{ secrets.GITHUB_TOKEN }}" \
           python tests/studio/run_real_mlx_smoke.py train \
             --workdir "$PWD/mlx_workdir"
 
@@ -406,3 +316,88 @@ jobs:
             cat "$f" 2>/dev/null || echo "(missing)"
             echo
           done
+
+      # Validates the macOS prebuilt path Studio's setup.sh uses (#5963): install the
+      # unslothai/llama.cpp fork's latest release, download a small public GGUF, and
+      # check llama-server /completion end to end. Split and placed last so the
+      # untrusted binary runs only in the final smoke step, after every HF_TOKEN step,
+      # leaving no token-bearing step or shared workspace for a tampered prebuilt to
+      # corrupt. GH_TOKEN: releases API; HF_TOKEN (withheld on PR): probe + GGUF fetch.
+      - name: Studio prebuilt llama.cpp install + GGUF download (Mac M1)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
+        run: |
+          set -euo pipefail
+          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
+          rm -rf "$INSTALL_DIR"
+          # Download only -- no llama-quantize / llama-server launch in this step.
+          python studio/install_llama_prebuilt.py \
+            --install-dir "$INSTALL_DIR" \
+            --published-repo unslothai/llama.cpp
+          mkdir -p /tmp/ggufs
+          bash .github/scripts/hf-download-with-retry.sh \
+            'unsloth/gemma-3-270m-it-GGUF' \
+            'gemma-3-270m-it-Q4_K_M.gguf' \
+            /tmp/ggufs
+
+      # Final step: runs the downloaded binaries with no secrets present, and clears
+      # the GitHub Actions command files so a tampered prebuilt cannot influence the job.
+      - name: Studio prebuilt llama.cpp GGUF inference smoke (Mac M1)
+        run: |
+          set -euo pipefail
+          unset GITHUB_ENV GITHUB_PATH GITHUB_OUTPUT GITHUB_STEP_SUMMARY
+          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
+          # Studio bundles only llama-server + llama-quantize (not llama-cli);
+          # inference goes through llama-server's HTTP /completion endpoint.
+          LLAMA_SERVER="$INSTALL_DIR/build/bin/llama-server"
+          LLAMA_QUANT="$INSTALL_DIR/build/bin/llama-quantize"
+          [ -x "$LLAMA_SERVER" ] || { echo "::error::llama-server missing at $LLAMA_SERVER"; find "$INSTALL_DIR/build" -type f | head -40; exit 1; }
+          [ -x "$LLAMA_QUANT" ]  || { echo "::error::llama-quantize missing at $LLAMA_QUANT"; exit 1; }
+          echo "llama-server : $LLAMA_SERVER"
+          echo "llama-quantize: $LLAMA_QUANT"
+          "$LLAMA_QUANT" --help >/dev/null && echo "  llama-quantize loads OK"
+
+          PORT=18080
+          echo "=== starting llama-server on 127.0.0.1:$PORT ==="
+          "$LLAMA_SERVER" \
+            -m /tmp/ggufs/gemma-3-270m-it-Q4_K_M.gguf \
+            --host 127.0.0.1 \
+            --port "$PORT" \
+            -c 256 \
+            -n 16 \
+            --no-warmup \
+            > /tmp/llama-server.log 2>&1 &
+          SERVER_PID=$!
+          trap 'kill "$SERVER_PID" 2>/dev/null || true' EXIT
+
+          # Wait for /health to come up
+          for i in $(seq 1 30); do
+            if curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
+              echo "  server up after ${i}s"
+              break
+            fi
+            sleep 1
+          done
+          if ! curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
+            echo "::error::llama-server never became healthy"
+            tail -40 /tmp/llama-server.log
+            exit 1
+          fi
+
+          PROMPT="Hello, my name is"
+          echo "=== POST /completion ==="
+          RESP=$(curl -sf -X POST "http://127.0.0.1:$PORT/completion" \
+            -H 'Content-Type: application/json' \
+            -d "{\"prompt\":\"$PROMPT\",\"n_predict\":16,\"temperature\":0,\"seed\":3407}")
+          echo "raw response (head): $(echo "$RESP" | head -c 600)"
+          CONTENT=$(echo "$RESP" | python -c "import json,sys; print(json.loads(sys.stdin.read()).get('content',''))")
+          echo "completion content: $CONTENT"
+
+          if [ -z "$CONTENT" ]; then
+            echo "::error::llama-server /completion returned empty content"
+            tail -40 /tmp/llama-server.log
+            exit 1
+          fi
+          echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works"
diff --git a/.github/workflows/notebooks-ci.yml b/.github/workflows/notebooks-ci.yml
index 2edcae8ab2..0e0b35dd4d 100644
--- a/.github/workflows/notebooks-ci.yml
+++ b/.github/workflows/notebooks-ci.yml
@@ -263,7 +263,7 @@ jobs:
           # unsloth_zoo.vision_utils imports PIL at module top, and the
           # easiest way to get a torch-compatible PIL on a CPU runner is
           # to let torchvision pull the right Pillow version.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
                       'torch>=2.8,<2.11' 'torchvision<0.26'
           # Pin to the same versions update_all_notebooks.py installs in
           # generated notebooks. Keep these in lockstep with PIN_TRL /
diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index ea60252cf6..bce355458a 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -76,7 +76,7 @@ jobs:
           # Torch CPU + transformers are required by a chunk of the backend test
           # suite (gpu_selection, kv_cache_estimation, utils). CPU-only torch
           # keeps the install ~250 MB / ~1 min on a clean runner.
-          pip install --index-url https://download.pytorch.org/whl/cpu 'torch>=2.4,<2.11'
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple 'torch>=2.4,<2.11'
           pip install 'transformers>=4.51,<5.5'
 
       - name: Backend tests
@@ -137,7 +137,7 @@ jobs:
             pyyaml jinja2 mammoth unpdf requests typer \
             'numpy<3' pytest pytest-asyncio httpx
           # torchvision: unsloth_zoo.vision_utils imports it at module scope.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           pip install 'transformers>=4.51,<5.5'
           # bitsandbytes: hard import in unsloth/models/_utils.py. Recent
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 512af54d53..20ca247b9f 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -185,13 +185,14 @@ jobs:
         # Retry up to 3 times to absorb known macos-14 free-runner
         # flakes: (1) Playwright Node 24 pipeTransport.js 'Unexpected
         # end of JSON input' crash when the Chromium browser process
-        # dies mid-test, and (2) Chromium net::ERR_NO_BUFFER_SPACE
-        # when the runner's kernel briefly runs out of socket buffers.
-        # The retry FULLY resets Studio (kill, reset-password, reboot,
-        # wait /api/health, re-export bootstrap pw) before re-running
-        # the script. A real test failure (assertion / timeout) does
-        # NOT match either pattern so it bypasses retry and surfaces
-        # immediately.
+        # dies mid-test, (2) Chromium net::ERR_NO_BUFFER_SPACE when the
+        # runner's kernel briefly runs out of socket buffers, and (3) a
+        # goto 'interrupted by another navigation' when the SPA auth
+        # guard redirects mid-navigation. The retry FULLY resets Studio
+        # (kill, reset-password, reboot, wait /api/health, re-export
+        # bootstrap pw) before re-running the script. A real test failure
+        # (assertion / timeout) does NOT match any pattern so it bypasses
+        # retry and surfaces immediately.
         run: |
           mkdir -p logs/playwright
           attempt=1
@@ -204,8 +205,9 @@ jobs:
             if [ "$rc" -eq 0 ]; then
               break
             fi
-            if { grep -q "Unexpected end of JSON input" logs/playwright_attempt_${attempt}.log \
-                 || grep -q "ERR_NO_BUFFER_SPACE"        logs/playwright_attempt_${attempt}.log; } \
+            if { grep -q "Unexpected end of JSON input"       logs/playwright_attempt_${attempt}.log \
+                 || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_attempt_${attempt}.log \
+                 || grep -q "interrupted by another navigation" logs/playwright_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
               kill "${STUDIO_PID}" 2>/dev/null || true
@@ -280,8 +282,8 @@ jobs:
           STUDIO_UI_TURN_TIMEOUT_MS: '540000'
           GGUF_REPO: ${{ env.GGUF_REPO }}
           GGUF_VARIANT: ${{ env.GGUF_VARIANT }}
-        # Same flake-retry shape as "Drive the chat UI with Playwright"
-        # -- catches pipeTransport JSON crash and ERR_NO_BUFFER_SPACE.
+        # Same flake-retry shape as "Drive the chat UI with Playwright" -- catches
+        # pipeTransport JSON crash, ERR_NO_BUFFER_SPACE, and nav interrupts.
         run: |
           mkdir -p logs/playwright_extra
           attempt=1
@@ -294,8 +296,9 @@ jobs:
             if [ "$rc" -eq 0 ]; then
               break
             fi
-            if { grep -q "Unexpected end of JSON input" logs/playwright_extra_attempt_${attempt}.log \
-                 || grep -q "ERR_NO_BUFFER_SPACE"        logs/playwright_extra_attempt_${attempt}.log; } \
+            if { grep -q "Unexpected end of JSON input"       logs/playwright_extra_attempt_${attempt}.log \
+                 || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_extra_attempt_${attempt}.log \
+                 || grep -q "interrupted by another navigation" logs/playwright_extra_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
               kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index c44c68278d..8186c07211 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1338,11 +1338,19 @@ jobs:
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # A Program Files dir can hold a transient handle (Defender / MSBuild node)
+          # so Rename-Item intermittently fails with "Access is denied"; retry to ride it out.
+          function Rename-WithRetry($Path, $NewName) {
+            for ($i = 1; $i -le 6; $i++) {
+              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
+              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+            }
+          }
           # Rename the Visual Studio install roots (incl. the Installer that holds
           # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
           foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
             if (Test-Path -LiteralPath $d) {
-              Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff')
+              Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff')
               Write-Host "Hid VS: $d"
             }
           }
@@ -1351,7 +1359,7 @@ jobs:
           $hidden = @()
           foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
             if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
-              Rename-Item -LiteralPath $c.Source -NewName ((Split-Path $c.Source -Leaf) + '.off')
+              Rename-WithRetry $c.Source ((Split-Path $c.Source -Leaf) + '.off')
               $hidden += $c.Source
               Write-Host "Hid cmake: $($c.Source)"
             }
@@ -1376,7 +1384,7 @@ jobs:
       - name: PyTorch CPU wheel installs and imports (no Visual Studio)
         run: |
           python -m pip install --upgrade pip
-          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple
           python -c "import torch; print('torch', torch.__version__, 'cuda?', torch.cuda.is_available())"
 
       - name: Install Studio (--local, --no-torch) with no build tools present
@@ -1536,8 +1544,16 @@ jobs:
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # Retry the rename: a Program Files dir can hold a transient handle that
+          # makes Rename-Item intermittently fail with "Access is denied".
+          function Rename-WithRetry($Path, $NewName) {
+            for ($i = 1; $i -le 6; $i++) {
+              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
+              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+            }
+          }
           foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) { Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
+            if (Test-Path -LiteralPath $d) { Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
           }
 
       - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index 599b53df1d..e492d21e99 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -242,7 +242,7 @@ jobs:
         run: |
           python -m pip install --upgrade pip
           # CPU torch (vllm/peft/st all depend on it).
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26' 'torchcodec<0.10'
           # torchcodec is a hard requirement on transformers 5.x:
           # transformers/audio_utils.py:55 does
diff --git a/README.md b/README.md
index 9162d29b1c..e3fd4e6980 100644
--- a/README.md
+++ b/README.md
@@ -246,6 +246,11 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_STUDIO_HOME=/abs/path sh
 $env:UNSLOTH_STUDIO_HOME='C:\path'; irm https://unsloth.ai/install.ps1 | iex
 ```
 
+On macOS, the installer defaults to the system certificate store (`UV_SYSTEM_CERTS=1`) so uv trusts the CAs in your Keychain, needed behind TLS-inspecting proxies (Cisco Umbrella, Zscaler, etc.). Opt out with:
+```bash
+curl -fsSL https://unsloth.ai/install.sh | UV_SYSTEM_CERTS=0 sh
+```
+
 Point the frontend build at a corporate npm mirror/proxy with `UNSLOTH_NPM_REGISTRY` (for the developer install behind a firewall that blocks `registry.npmjs.org`):
 ```bash
 UNSLOTH_NPM_REGISTRY=https://artifactory.example.com/api/npm/npm/ ./install.sh --local
diff --git a/install.sh b/install.sh
index 548e6f702a..7a9f0be87f 100755
--- a/install.sh
+++ b/install.sh
@@ -1636,6 +1636,21 @@ export UV_HTTP_RETRIES
 : "${UV_HTTP_TIMEOUT:=180}"
 export UV_HTTP_TIMEOUT
 
+# macOS: trust the system Keychain so uv uses SecureTransport instead of rustls.
+# Required behind TLS-inspecting proxies (Cisco Umbrella, Zscaler, etc.) which
+# present their own CA certificate. rustls (uv's default) ignores the Keychain
+# and rejects intercepted connections with "invalid peer certificate: UnknownIssuer".
+# Set both vars: UV_SYSTEM_CERTS is the modern one (uv >= 0.11), UV_NATIVE_TLS the
+# legacy one understood by uv 0.8.16-0.10.x, which the installer keeps if already
+# present (UV_MIN_VERSION) and which ignores UV_SYSTEM_CERTS. Mirror the choice onto
+# both so it works on either uv. Opt out with UV_SYSTEM_CERTS=0.
+if [ "$OS" = "macos" ]; then
+    : "${UV_SYSTEM_CERTS:=1}"
+    : "${UV_NATIVE_TLS:=$UV_SYSTEM_CERTS}"
+fi
+[ -n "${UV_SYSTEM_CERTS:-}" ] && export UV_SYSTEM_CERTS
+[ -n "${UV_NATIVE_TLS:-}" ] && export UV_NATIVE_TLS
+
 version_ge() {
     # returns 0 if $1 >= $2
     _a=$1
diff --git a/pyproject.toml b/pyproject.toml
index 13c421d8ea..844ead2454 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -255,10 +255,6 @@ cu118onlytorch270 = [
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-manylinux_2_28_x86_64.whl ; python_version=='3.10' and ('linux' in sys_platform)",
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-manylinux_2_28_x86_64.whl ; python_version=='3.11' and ('linux' in sys_platform)",
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-manylinux_2_28_x86_64.whl ; python_version=='3.12' and ('linux' in sys_platform)",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp39-cp39-win_amd64.whl ; python_version=='3.9' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-win_amd64.whl ; python_version=='3.10' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-win_amd64.whl ; python_version=='3.11' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-win_amd64.whl ; python_version=='3.12' and (sys_platform == 'win32')",
 ]
 cu126onlytorch270 = [
     "xformers @ https://download.pytorch.org/whl/cu126/xformers-0.0.30-cp39-cp39-manylinux_2_28_x86_64.whl ; python_version=='3.9' and ('linux' in sys_platform)",
@@ -282,7 +278,6 @@ cu128onlytorch270 = [
 ]
 cu118onlytorch271 = [
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.31.post1-cp39-abi3-manylinux_2_28_x86_64.whl ; ('linux' in sys_platform)",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.31.post1-cp39-abi3-win_amd64.whl ; (sys_platform == 'win32')",
 ]
 cu126onlytorch271 = [
     "xformers @ https://download.pytorch.org/whl/cu126/xformers-0.0.31.post1-cp39-abi3-manylinux_2_28_x86_64.whl ; ('linux' in sys_platform)",
@@ -879,14 +874,12 @@ flashattentiontorch240abiFALSEcu12x = [
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.10'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.11'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp312-cp312-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.12'",
-    "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp313-cp313-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.13'",
 ]
 flashattentiontorch240abiTRUEcu12x = [
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp39-cp39-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.9'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp310-cp310-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.10'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp311-cp311-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.11'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp312-cp312-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.12'",
-    "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp313-cp313-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.13'",
 ]
 intelgputorch260 = [
     "unsloth_zoo[intelgpu]",
@@ -1174,14 +1167,14 @@ intelgputorch2120 = [
     "unsloth_zoo[intelgpu]",
     "unsloth[huggingfacenotorch]",
 
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=844d981cb1b3948085e8cfa62c74de9f100259f6131959aa70be49123b88ae81 ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=a16b1d00e94ad87d62af3512e390348b8656419598004100c56028bf494f086b ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=4e46e71e077cf483404a4c17ce40d71c5f0e13a81459139d4346ca427b1dd455 ; platform_system == 'Linux' and python_version == '3.12' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=4fdaed1bafc51d3a2834656a3420a6686a74ea226508765a49bf15d58ff3a930 ; platform_system == 'Linux' and python_version == '3.13' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-win_amd64.whl#sha256=2778b46b22e9fa0916398db299a125027a1b2331c1173b3dd2b9e2cab6263a31 ; sys_platform == 'win32' and python_version == '3.10' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-win_amd64.whl#sha256=ad5b147d04ee0d40f3d4d32f85f5aa3a3beb6cd5799ca026d3d7f4afa3d9e24f ; sys_platform == 'win32' and python_version == '3.11' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-win_amd64.whl#sha256=d9482063af2a308543f23333e32edd738ea87cbb33ade68afda9ae0fd704ccd9 ; sys_platform == 'win32' and python_version == '3.12' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-win_amd64.whl#sha256=5d4d67f0deb1e851c01b293e602b8dcddad26ca2be61221cee3dc0e1aa0cdefd ; sys_platform == 'win32' and python_version == '3.13' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=81ff0eb0c4fc8e19d2510b28c3e1d9382a3c7d6fdaf6a9f9631a93a030d841cf ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=55574a68d275b85cd4d5cbf185084bae019ebf09c3f43b0bd2831b14935ec8e7 ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=a31c058c5c2e78ebe490a2e69f2f50caec6b1307ac096e944f116fdc06819d9a ; platform_system == 'Linux' and python_version == '3.12' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=e701a31efa0334775f357c98716f3821775aa944219f7888e13c2dfe2daabe2a ; platform_system == 'Linux' and python_version == '3.13' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-win_amd64.whl#sha256=0d7730651c3e52fbf3a430cc201455f0c6600dc72e681aec495f131ea44f341a ; sys_platform == 'win32' and python_version == '3.10' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-win_amd64.whl#sha256=8f4a63de73e3d632098f93c8f0bd77244958a47d7c5f728b8ff35f8a91fdb983 ; sys_platform == 'win32' and python_version == '3.11' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-win_amd64.whl#sha256=6589ece3adc2b1ab88d90ff1267afc25df5c7b868f0b633e732cac70df36cbde ; sys_platform == 'win32' and python_version == '3.12' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-win_amd64.whl#sha256=2fdf001a9b0575e8b1827127259bb9b13bf36e659882be74c2dfab46597d3e7a ; sys_platform == 'win32' and python_version == '3.13' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
 
     "torch @ https://download.pytorch.org/whl/xpu/torch-2.12.0%2Bxpu-cp310-cp310-linux_x86_64.whl#sha256=e8923cd1fe560472904b1461b745d2f1826bb9c1bc0808225d5f28a450e4d553 ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
     "torch @ https://download.pytorch.org/whl/xpu/torch-2.12.0%2Bxpu-cp311-cp311-linux_x86_64.whl#sha256=f7c082b2fc9b61def594d30ea57762dc4a8bc7111a9a9593953ed948de242e28 ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 255fc4140f..8ec11fa79f 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1271,6 +1271,9 @@ class LlamaCppBackend:
         self._cache_type_kv: Optional[str] = None
         # Whether --split-mode tensor was applied on the active load.
         self._tensor_parallel: bool = False
+        # Layer load kept multi-GPU only to honor a downgraded tensor request, so a
+        # later explicit tensor-off reloads instead of deduping to it (#6659).
+        self._layer_preserves_tensor_intent: bool = False
         self._reasoning_default: bool = True
         self._speculative_type: Optional[str] = None
         # Canonical UI-facing mode the user requested
@@ -1643,6 +1646,11 @@ class LlamaCppBackend:
         """Whether --split-mode tensor is active on the loaded server."""
         return self._tensor_parallel
 
+    @property
+    def layer_preserves_tensor_intent(self) -> bool:
+        """True when a downgraded tensor request kept this layer load multi-GPU."""
+        return self._layer_preserves_tensor_intent
+
     @property
     def speculative_type(self) -> Optional[str]:
         return self._speculative_type
@@ -2430,6 +2438,37 @@ class LlamaCppBackend:
     # aborts a --split-mode tensor load, so it's dropped for the tensor attempt.
     _TENSOR_PARALLEL_KV_TYPES = frozenset({"f16", "bf16", "f32"})
 
+    # (binary, mtime, model) that aborted on --split-mode tensor this process (#6415
+    # geometry limit, e.g. MQA n_head_kv=1). Model-keyed so one model's abort doesn't
+    # skip tensor for others; tensor is tried by default, recorded only on a real abort.
+    _tensor_split_abort_keys: set[tuple[str, int, str]] = set()
+
+    @classmethod
+    def _tensor_split_cache_key(
+        cls, binary: Optional[str], model: Optional[str]
+    ) -> Optional[tuple[str, int, str]]:
+        """(path, mtime_ns, model) key; ns mtime re-probes a same-second binary swap."""
+        if not binary or not model:
+            return None
+        try:
+            mtime = Path(binary).stat().st_mtime_ns
+        except OSError:
+            mtime = 0
+        return (binary, mtime, model)
+
+    @classmethod
+    def _tensor_split_aborts(cls, binary: Optional[str], model: Optional[str]) -> bool:
+        """True if (binary, model) aborted on --split-mode tensor this session."""
+        key = cls._tensor_split_cache_key(binary, model)
+        return key is not None and key in cls._tensor_split_abort_keys
+
+    @classmethod
+    def _record_tensor_split_abort(cls, binary: Optional[str], model: Optional[str]) -> None:
+        """Remember a (binary, model) that aborts on --split-mode tensor."""
+        key = cls._tensor_split_cache_key(binary, model)
+        if key is not None:
+            cls._tensor_split_abort_keys.add(key)
+
     @staticmethod
     def _windows_pip_nvidia_dll_dirs(prefix: str) -> list[str]:
         """Return DLL dirs from pip-installed CUDA wheels under
@@ -2569,9 +2608,13 @@ class LlamaCppBackend:
         usable_fraction: Optional[float] = None,
         total_by_idx: Optional[dict[int, int]] = None,
         per_device_overhead_bytes: int = 0,
+        min_gpus: int = 1,
     ) -> tuple[Optional[list[int]], bool]:
         """Pick GPU(s) for a model from estimated VRAM and free memory.
 
+        ``min_gpus`` (default 1, capped at ``len(gpus)``) keeps a downgraded
+        tensor/multi-GPU request spread instead of collapsing to one card.
+
         ``model_size_bytes`` should include weights and estimated KV cache.
         ``usable_fraction`` (default ``_GPU_PIN_VRAM_FRACTION``) provides
         headroom for compute buffers, CUDA context, and other runtime
@@ -2590,9 +2633,11 @@ class LlamaCppBackend:
         if not gpus:
             return None, True
 
+        min_gpus = max(1, min(min_gpus, len(gpus)))
         model_size_mib = model_size_bytes / (1024 * 1024)
         if usable_fraction is None:
             usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION
+        overhead_mib = per_device_overhead_bytes / (1024 * 1024)
 
         # Per-GPU usable budget: free - (1-frac)*total when total is known, else
         # the legacy free*frac (also covers a total-0 two-column probe).
@@ -2606,19 +2651,26 @@ class LlamaCppBackend:
         # card can have less usable room than a less-used small one.
         ranked = sorted(gpus, key = lambda g: _usable(g[0], g[1]), reverse = True)
 
-        # Try 1 GPU at the usable-VRAM threshold.
-        if _usable(ranked[0][0], ranked[0][1]) >= model_size_mib:
+        # Cap a downgraded multi-GPU request to the usable count so it doesn't pull
+        # in a near-full card to hit min_gpus. No-op for the default min_gpus == 1.
+        usable_count = sum(1 for idx, free_mib in ranked if _usable(idx, free_mib) > overhead_mib)
+        min_gpus = max(1, min(min_gpus, usable_count or 1))
+
+        # Try 1 GPU at the usable-VRAM threshold (only when one device is allowed).
+        if min_gpus <= 1 and _usable(ranked[0][0], ranked[0][1]) >= model_size_mib:
             return [ranked[0][0]], False
 
-        # Try N GPUs (accumulate usable memory from most-free). Each GPU past the
-        # first adds a fixed per-device overhead the pool must hold.
-        overhead_mib = per_device_overhead_bytes / (1024 * 1024)
+        # Try N GPUs (most-free first); each past the first adds per-device overhead.
+        # Require at least min_gpus devices before accepting a fit.
         cumulative = 0.0
         selected = []
         for idx, free_mib in ranked:
             selected.append(idx)
             cumulative += _usable(idx, free_mib)
-            if cumulative >= model_size_mib + (len(selected) - 1) * overhead_mib:
+            if (
+                len(selected) >= min_gpus
+                and cumulative >= model_size_mib + (len(selected) - 1) * overhead_mib
+            ):
                 return sorted(selected), False
 
         # Too large even for all GPUs; let --fit handle it
@@ -3868,7 +3920,7 @@ class LlamaCppBackend:
                     logger.debug(f"Could not list repo files for {label}: {e}")
                     break
                 logger.debug(
-                    f"Could not list repo files for {label} " f"(attempt {attempt + 1}/3): {e}"
+                    f"Could not list repo files for {label} (attempt {attempt + 1}/3): {e}"
                 )
                 if attempt < 2:
                     self._cancel_event.wait(2**attempt)
@@ -4332,6 +4384,17 @@ class LlamaCppBackend:
             )
         )
 
+    @staticmethod
+    def _is_tensor_split_assert(output: str) -> bool:
+        """True only for the #6415 split-axis warmup assert (GGML_BACKEND_SPLIT_AXIS_*),
+        not any ggml assert/abort, so an unrelated invariant isn't cached. stderr is
+        merged into output."""
+        text = (output or "").lower()
+        if "ggml_assert" not in text and "ggml_abort" not in text:
+            return False
+        # the split-axis enum token, unique to this assert (not the source file).
+        return "split_axis" in text
+
     @staticmethod
     def _is_signal_crash(returncode: Optional[int]) -> bool:
         """True only on a hard fault (SIGSEGV/SIGABRT/SIGILL/SIGFPE/SIGBUS or a
@@ -4344,6 +4407,20 @@ class LlamaCppBackend:
             return True
         return -returncode in (4, 6, 7, 8, 11)  # SIGILL SIGABRT SIGBUS SIGFPE SIGSEGV
 
+    @staticmethod
+    def _is_abort_exit(returncode: Optional[int]) -> bool:
+        """Windows CRT abort() exit code (3) from GGML_ASSERT on MSVC -- not a POSIX
+        signal or 0xC0000000+ NTSTATUS."""
+        return returncode == 3
+
+    @classmethod
+    def _should_record_tensor_split_abort(cls, returncode: Optional[int], output: str) -> bool:
+        """The #6415 split-axis abort: the marker plus a hard crash (POSIX signal or
+        Windows abort exit). Marker required so a generic crash isn't cached."""
+        return cls._is_tensor_split_assert(output) and (
+            cls._is_signal_crash(returncode) or cls._is_abort_exit(returncode)
+        )
+
     @staticmethod
     def _with_flash_attn_off(cmd: list[str]) -> Optional[list[str]]:
         """Return cmd with flash attention forced off, or None when its effective
@@ -4488,6 +4565,8 @@ class LlamaCppBackend:
         n_gpu_layers: Optional[int] = None,  # caller compat, unused
         n_parallel: int = 1,
         extra_args: Optional[List[str]] = None,
+        # Route-level tensor->layer fallback retry: keep the layer split multi-GPU.
+        preserve_multi_gpu_on_layer: bool = False,
     ) -> bool:
         """Start llama-server with a GGUF model.
 
@@ -4518,6 +4597,8 @@ class LlamaCppBackend:
             "n_gpu_layers": n_gpu_layers,
             "n_parallel": n_parallel,
             "extra_args": list(extra_args) if extra_args is not None else None,
+            # Replayed by _respawn_if_dead so a downgraded model stays multi-GPU.
+            "preserve_multi_gpu_on_layer": preserve_multi_gpu_on_layer,
         }
         # Serialise the whole load so concurrent /load calls never leave two
         # llama-server processes alive (#5401 / #5161). Doesn't block /unload.
@@ -4541,6 +4622,7 @@ class LlamaCppBackend:
                 chat_template_override = chat_template_override,
                 extra_args = extra_args,
                 is_vision = is_vision,
+                preserve_multi_gpu_on_layer = preserve_multi_gpu_on_layer,
             ):
                 logger.info(
                     f"load_model: backend already in target state for "
@@ -4626,6 +4708,9 @@ class LlamaCppBackend:
             # Block-diffusion GGUFs (DiffusionGemma) cannot run on llama-server;
             # serve them with the diffusion runner (same OpenAI-compat interface).
             if self._is_diffusion:
+                # Not a tensor/layer GGUF: clear any preserved-fallback flag from a
+                # prior load (this path skips the command builder that clears it).
+                self._layer_preserves_tensor_intent = False
                 with self._lock:
                     if self._cancel_event.is_set():
                         logger.info("Load cancelled before diffusion server start")
@@ -4780,6 +4865,9 @@ class LlamaCppBackend:
                         "image input will be disabled for this session"
                     )
                 model_size = None  # set in the fit try; used by the APU RAM guard
+                # Layer-fallback min GPUs; raised below on a tensor downgrade. Bound
+                # before the try so the --fit-on except path still has it (no UnboundLocal).
+                _layer_min_gpus = 1
                 try:
                     gguf_size = self._get_gguf_size_bytes(model_path)
                     # Include GPU-loaded mmproj in the fit budget (#5825).
@@ -5064,10 +5152,8 @@ class LlamaCppBackend:
                     _apple_budget_mib = self._apple_metal_memory_budget_bytes() // (1024 * 1024)
 
                     def _restore_after_tensor_downgrade():
-                        # Tensor mode dropped a quantized KV and stripped the cache
-                        # extras (it rejects quantized); layer split supports them, so
-                        # restore the original type + extras (minus --split-mode) and
-                        # clear the env flag so the layer launch re-emits them.
+                        # Restore the quantized KV + extras tensor dropped (layer
+                        # split supports them), minus --split-mode.
                         nonlocal cache_type_kv, _cache_type_from_env, extra_args
                         if _tensor_dropped_cache_type_kv is not None:
                             cache_type_kv = _tensor_dropped_cache_type_kv
@@ -5078,13 +5164,22 @@ class LlamaCppBackend:
                             else extra_args
                         )
 
-                    if tensor_parallel and effective_is_vision:
+                    # The route fallback retry is tensor-off; keep it multi-GPU.
+                    if preserve_multi_gpu_on_layer:
+                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))
+
+                    if tensor_parallel and self._tensor_split_aborts(binary, model_identifier):
+                        # Aborted on tensor for this model this session (#6415); skip
+                        # tensor upfront, layer split serves it.
                         logger.info(
-                            "Tensor parallelism skipped for vision model: "
-                            "--split-mode tensor is incompatible with --mmproj "
-                            "in the current llama.cpp build; using layer split."
+                            "Tensor parallelism skipped: this llama.cpp build aborted "
+                            "on --split-mode tensor for this model earlier this "
+                            "session; using layer split across %d GPU(s).",
+                            len(gpus),
                         )
                         tensor_parallel = False
+                        # Keep the multi-GPU request (gated on it, not the cache).
+                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                         _restore_after_tensor_downgrade()
 
                     # Tensor mode replicates a compute buffer on every GPU, so drop
@@ -5124,6 +5219,11 @@ class LlamaCppBackend:
                             len(gpus),
                         )
                         tensor_parallel = False
+                        # GPUs below tensor's compute-buffer reserve can still do layer
+                        # split, so keep multi-GPU (mirrors the budget/geometry drops);
+                        # _select_gpus caps unusable cards.
+                        if len(gpus) >= 2:
+                            _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                         # Layer split supports a quantized KV the tensor attempt
                         # dropped; restore the original cache type + extras (minus
                         # --split-mode) so the layer launch re-emits them.
@@ -5160,8 +5260,12 @@ class LlamaCppBackend:
                                 "per-device compute buffers; falling back to layer split."
                             )
                             tensor_parallel = False
-                            # Restore the dropped quantized KV + original cache extras
-                            # (minus --split-mode); layer split supports them.
+                            # Weights needed >1 card, so keep multi-GPU across the
+                            # usable tensor GPUs.
+                            if len(tp_gpus) >= 2:
+                                _layer_min_gpus = max(_layer_min_gpus, len(tp_gpus))
+                            # Restore the dropped quantized KV + cache extras (minus
+                            # --split-mode); layer split supports them.
                             _restore_after_tensor_downgrade()
 
                     if tensor_parallel and tp_gpus:
@@ -5263,6 +5367,7 @@ class LlamaCppBackend:
                                 usable_fraction = _pin_fraction,
                                 total_by_idx = total_by_idx,
                                 per_device_overhead_bytes = _pipeline_overhead_bytes,
+                                min_gpus = _layer_min_gpus,
                             )
                             # No silent shrink: effective_ctx stays == requested_ctx.
                         else:
@@ -5273,7 +5378,22 @@ class LlamaCppBackend:
                             ranked = sorted(
                                 gpus, key = lambda g: _gpu_usable(g, pin_fraction), reverse = True
                             )
-                            for n_gpus in range(1, len(ranked) + 1):
+                            # Skips _select_gpus, so apply its cap: count only cards
+                            # whose usable VRAM clears the per-device layer overhead.
+                            _pipeline_overhead_mib = _pipeline_overhead_bytes / (1024 * 1024)
+                            _auto_min_gpus = max(
+                                1,
+                                min(
+                                    _layer_min_gpus,
+                                    sum(
+                                        1
+                                        for g in ranked
+                                        if _gpu_usable(g, pin_fraction) > _pipeline_overhead_mib
+                                    )
+                                    or 1,
+                                ),
+                            )
+                            for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                 subset = ranked[:n_gpus]
                                 pool_budget = _pool_budget_mib(subset, pin_fraction)
                                 _ms = _subset_model_size(n_gpus)
@@ -5303,7 +5423,7 @@ class LlamaCppBackend:
                                 # at 131k may pin fine with a 4096 KV (#5106).
                                 effective_ctx = min(4096, effective_ctx)
                                 if effective_ctx > 0:
-                                    for n_gpus in range(1, len(ranked) + 1):
+                                    for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                         subset = ranked[:n_gpus]
                                         kv = self._estimate_kv_cache_bytes(
                                             effective_ctx,
@@ -5339,6 +5459,7 @@ class LlamaCppBackend:
                             usable_fraction = _pin_fraction,
                             total_by_idx = total_by_idx,
                             per_device_overhead_bytes = _pipeline_overhead_bytes,
+                            min_gpus = _layer_min_gpus,
                         )
                         if use_fit and not explicit_ctx:
                             # Weights don't fit on any subset; default UI to 4096
@@ -5476,6 +5597,15 @@ class LlamaCppBackend:
                     "--no-context-shift",
                 ]
 
+                # Report a clean public model id (matching GET /v1/models) rather
+                # than the raw -m path in llama-server's own /v1/models and the
+                # "model" field of its chat/completions responses.
+                from core.inference.model_ids import public_model_id
+
+                _alias = public_model_id(self._model_identifier or model_path)
+                if _alias:
+                    cmd.extend(["--alias", _alias])
+
                 fully_gpu_offloaded = False
                 if use_fit:
                     cmd.extend(["--fit", "on"])
@@ -5569,12 +5699,15 @@ class LlamaCppBackend:
                             ]
                         )
                     self._tensor_parallel = True
+                    self._layer_preserves_tensor_intent = False
                     logger.info(
                         "Tensor parallelism: --split-mode tensor, --tensor-split %s",
                         tp_tensor_split,
                     )
                 else:
                     self._tensor_parallel = False
+                    # > 1 only when a tensor request was downgraded but kept multi-GPU.
+                    self._layer_preserves_tensor_intent = _layer_min_gpus > 1
 
                 # Speculative decoding. See _build_speculative_flags for the
                 # mode resolution, benchmarks, and llama.cpp references.
@@ -5858,7 +5991,17 @@ class LlamaCppBackend:
                         _startup_crashed = (
                             self._process.poll() is not None and self._process.returncode != 0
                         )
-                        if _spawn_attempt == 0 and _fit_retry_allowed and _startup_crashed:
+                        # A split-axis abort (#6415) is fit-independent: skip the
+                        # --fit off retry and let the caller latch it.
+                        _split_axis_crash = self._is_tensor_split_assert(
+                            "\n".join(self._stdout_lines[-50:])
+                        )
+                        if (
+                            _spawn_attempt == 0
+                            and _fit_retry_allowed
+                            and _startup_crashed
+                            and not _split_axis_crash
+                        ):
                             logger.warning(
                                 "llama-server crashed during startup (exit code %s) "
                                 "with the default memory-fit step enabled; Studio "
@@ -5904,6 +6047,21 @@ class LlamaCppBackend:
                 )
 
                 healthy = _spawn_and_wait(cmd)
+                # #6415 split-mode tensor warmup abort. Latch it on THIS first spawn:
+                # the flash-attn-off retry below can't run tensor (needs flash_attn),
+                # so its output drops the marker and recording later would miss it,
+                # looping every load. Record and raise to the route's layer fallback,
+                # skipping the futile flash-attn/MTP retries.
+                if not healthy and self._tensor_parallel and not self._cancel_event.is_set():
+                    _ts_out = "\n".join(self._stdout_lines[-50:])
+                    _ts_rc = self._process.poll() if self._process is not None else None
+                    if self._should_record_tensor_split_abort(_ts_rc, _ts_out):
+                        LlamaCppBackend._record_tensor_split_abort(binary, model_identifier)
+                        self._kill_process()
+                        raise RuntimeError(
+                            "llama-server aborted on --split-mode tensor "
+                            "(split-axis geometry); retrying with layer split."
+                        )
                 # Flash-attention kernels hard-crash at startup on some ROCm/GPU
                 # builds (frequently inside the vision tower). Disabling FA keeps
                 # both vision and MTP, so retry that way before dropping either.
@@ -6048,6 +6206,7 @@ class LlamaCppBackend:
                     # Read the crash code before _kill_process() clears _process.
                     _crash_rc = self._process.poll() if self._process is not None else None
                     self._kill_process()
+                    # The #6415 split-axis abort is latched earlier (first spawn).
                     # Skip if a cancel/unload is pending (mirrors the MTP guard).
                     if (
                         launched_with_mmproj
@@ -6479,6 +6638,7 @@ class LlamaCppBackend:
         spec_draft_n_max: Optional[int] = None,
         tensor_parallel: bool = False,
         mtp_draft_path: Optional[str] = None,
+        preserve_multi_gpu_on_layer: bool = False,
     ) -> bool:
         """True iff the live server already satisfies these load kwargs.
 
@@ -6521,6 +6681,17 @@ class LlamaCppBackend:
         # server. An identical request would downgrade the same way.
         if not _tensor_parallel_matches_loaded(extra_args, tensor_parallel, self._tensor_parallel):
             return False
+        # Preserved tensor->layer fallback + an EXPLICIT tensor drop: reload so
+        # placement re-selects instead of keeping the all-GPU mask (mirrors the route,
+        # #6659). preserve_multi_gpu_on_layer carries the route's carry-forward decision
+        # (True for an implicit same-settings reload), so those still dedupe -- the HF
+        # auto-pick / local-dir flows skip the route guard and only reach here.
+        if (
+            self._layer_preserves_tensor_intent
+            and not _effective_tensor_parallel(extra_args, tensor_parallel)
+            and not preserve_multi_gpu_on_layer
+        ):
+            return False
 
         # Compare on the canonical requested mode. With --spec-type in
         # extra_args the backend stores None; mirror that here.
@@ -6632,6 +6803,7 @@ class LlamaCppBackend:
             self._supports_tools = False
             self._cache_type_kv = None
             self._tensor_parallel = False
+            self._layer_preserves_tensor_intent = False
             self._speculative_type = None
             self._requested_spec_mode = None
             self._spec_draft_n_max = None
diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py
index b42be5ee0d..f400d2ae40 100644
--- a/studio/backend/core/inference/llama_server_args.py
+++ b/studio/backend/core/inference/llama_server_args.py
@@ -25,6 +25,11 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = (
     # Model identity: Studio resolves it from LoadRequest; a second -m would
     # load a different model than Studio thinks it loaded.
     frozenset({"-m", "--model"}),
+    # Public model id: Studio sets a sanitized --alias so the OpenAI API never
+    # exposes the local .gguf path. A user-supplied alias is appended after
+    # Studio's and, with llama.cpp's last-wins parsing, would reintroduce the
+    # path leak this is meant to prevent.
+    frozenset({"-a", "--alias"}),
     frozenset({"-mu", "--model-url"}),
     frozenset({"-dr", "--docker-repo"}),
     frozenset({"-hf", "-hfr", "--hf-repo"}),
diff --git a/studio/backend/core/inference/model_ids.py b/studio/backend/core/inference/model_ids.py
new file mode 100644
index 0000000000..548cc60f94
--- /dev/null
+++ b/studio/backend/core/inference/model_ids.py
@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Public model identifiers for the OpenAI-compatible API.
+
+The exposed API must report a stable, clean model id rather than the absolute
+on-disk path of a local GGUF. The internal identifier for a direct local load is
+the absolute ``.gguf`` path, which leaks the host filesystem layout and is
+awkward for clients to round-trip. ``public_model_id`` maps such an internal
+identifier to a clean name while leaving Hugging Face repo ids (``org/model``)
+and already-clean names untouched.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Optional
+
+_GGUF_SUFFIX = ".gguf"
+
+
+def _looks_like_path(identifier: str) -> bool:
+    """True when *identifier* is a local filesystem path, not a HF repo id.
+
+    A repo id is ``org/model`` (a single forward slash, no leading separator, no
+    drive, no ``.gguf``). Anything ending in ``.gguf``, starting with a path
+    separator or a relative/home prefix (``./``, ``../``, ``~``), carrying a
+    Windows drive, or with three or more ``/`` segments is treated as a local
+    path.
+    """
+    if identifier.lower().endswith(_GGUF_SUFFIX):
+        return True
+    if identifier.startswith(("/", "\\", "./", "../", ".\\", "..\\", "~")):
+        return True
+    if len(identifier) >= 2 and identifier[1] == ":":  # Windows drive, e.g. C:\
+        return True
+    if identifier.count("/") >= 2 or "\\" in identifier:
+        return True
+    return False
+
+
+def public_model_id(identifier: Optional[str]) -> Optional[str]:
+    """Return a clean, path-free public id for *identifier*.
+
+    - Local GGUF path -> the file stem with ``.gguf`` stripped, e.g.
+      ``/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf`` -> ``Qwen3-30B-A3B-Q4_K_M``.
+    - HF repo id (``org/model``) and already-clean names -> returned unchanged.
+    - ``None`` / empty -> returned unchanged.
+    """
+    if not identifier:
+        return identifier
+    if not _looks_like_path(identifier):
+        return identifier
+    name = os.path.basename(identifier.replace("\\", "/").rstrip("/"))
+    if name.lower().endswith(_GGUF_SUFFIX):
+        name = name[: -len(_GGUF_SUFFIX)]
+    return name or identifier
+
+
+def model_id_matches(requested: Optional[str], internal: Optional[str]) -> bool:
+    """Whether a client-supplied *requested* id refers to *internal*.
+
+    Accepts the clean public id (preferred) and, for backward compatibility, the
+    raw internal identifier (e.g. a legacy absolute path a client cached from an
+    older ``/v1/models`` response).
+    """
+    if requested is None or internal is None:
+        return False
+    if requested == internal:
+        return True
+    return public_model_id(internal) == requested
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 9d991c6512..f4233fcf04 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -299,6 +299,7 @@ class TrainingBackend:
         # Build config dict for the subprocess
         config = {
             "model_name": kwargs["model_name"],
+            "project_name": kwargs.get("project_name"),
             "training_type": kwargs.get("training_type", "LoRA/QLoRA"),
             "hf_token": kwargs.get("hf_token", ""),
             "load_in_4bit": kwargs.get("load_in_4bit", True),
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 3f020c8abc..610af2472e 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -44,6 +44,7 @@ if sys.platform.startswith("linux") and "HSA_ENABLE_DXG_DETECTION" not in os.env
 
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
+from utils.training_runs import build_default_output_dir_name
 from utils.wheel_utils import (
     direct_wheel_url,
     flash_attn_wheel_url,
@@ -1787,11 +1788,14 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     # ── 5. Build output dir ──
     # Resolve to ~/.unsloth/studio/outputs/ so the export page finds it
-    from utils.paths import resolve_output_dir, ensure_dir, default_run_dir_name
+    from utils.paths import resolve_output_dir, ensure_dir
 
     output_dir = config.get("output_dir", "")
     if not output_dir:
-        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        output_dir = build_default_output_dir_name(
+            model_name,
+            config.get("project_name"),
+        )
     output_dir = str(resolve_output_dir(output_dir))
     ensure_dir(Path(output_dir))
 
@@ -3019,7 +3023,10 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             resume_from_checkpoint
         )
         if not output_dir:
-            output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+            output_dir = build_default_output_dir_name(
+                model_name,
+                config.get("project_name"),
+            )
         output_dir = str(resolve_output_dir(output_dir))
         ensure_dir(Path(output_dir))
 
@@ -3500,7 +3507,10 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
         resume_from_checkpoint
     )
     if not output_dir:
-        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        output_dir = build_default_output_dir_name(
+            model_name,
+            config.get("project_name"),
+        )
     output_dir = str(resolve_output_dir(output_dir))
 
     num_epochs = config.get("num_epochs", 2)
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 731625ca74..0a5b775775 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -441,9 +441,30 @@ def _start_llama_cpp_probes_if_enabled(app: FastAPI) -> None:
     ).start()
 
 
+def _warm_rag_embedder() -> None:
+    """Warm RAG embeddings without blocking backend readiness."""
+    try:
+        from storage import rag_db
+
+        if not rag_db.RAG_AVAILABLE:
+            return
+        from core.rag import embeddings
+
+        embeddings.warm()
+    except Exception:
+        pass
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """Startup: detect hardware, seed default admin if needed. Shutdown: clean up compiled cache."""
+
+    import time as _time
+
+    _lifespan_started = _time.perf_counter()
+    import structlog as _structlog
+
+    _lifespan_log = _structlog.get_logger(__name__)
     clear_unsloth_compiled_cache()
 
     # Remove stale .venv_overlay from old versions; switching now uses .venv_t5/.
@@ -454,6 +475,11 @@ async def lifespan(app: FastAPI):
     # Detect hardware first — sets the DEVICE global used everywhere.
     detect_hardware()
 
+    _lifespan_log.info(
+        "lifespan hardware detection completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
+
     # Apple Silicon with MLX missing => Train/Export are greyed out (chat-only).
     # Reinstall mlx by name on a background thread (off the critical path) and
     # re-detect, so a reinstall/update that dropped mlx self-heals. No-op
@@ -465,7 +491,13 @@ async def lifespan(app: FastAPI):
         import structlog as _structlog
         _structlog.get_logger(__name__).debug("mlx autorepair skipped: %s", _mlx_exc)
 
-    # Reap download workers orphaned by a previous crash before new downloads start.
+    # Reap workers/runs orphaned by a previous crash before new work starts.
+    try:
+        from storage.studio_db import cleanup_orphaned_runs
+        cleanup_orphaned_runs()
+    except Exception as exc:
+        _lifespan_log.warning("cleanup_orphaned_runs failed at startup: %s", exc)
+
     reap_hub_orphan_workers()
 
     # llama.cpp probes: capability (MTP support) + freshness (release age).
@@ -479,45 +511,23 @@ async def lifespan(app: FastAPI):
     app.state.llama_cpp_freshness = None
     _start_llama_cpp_probes_if_enabled(app)
 
-    from storage.studio_db import cleanup_orphaned_runs
-
-    try:
-        cleanup_orphaned_runs()
-    except Exception as exc:
-        import structlog
-        structlog.get_logger(__name__).warning("cleanup_orphaned_runs failed at startup: %s", exc)
-
-    # Same for RAG: fail ingestion jobs stranded mid-ingest by a crash.
     try:
         from storage.rag_db import reconcile_orphaned_ingestion_jobs
         reconcile_orphaned_ingestion_jobs()
     except Exception as exc:
-        import structlog
-        structlog.get_logger(__name__).warning(
-            "reconcile_orphaned_ingestion_jobs failed at startup: %s", exc
-        )
+        _lifespan_log.warning("reconcile_orphaned_ingestion_jobs failed at startup: %s", exc)
 
     _start_helper_precache_if_enabled()
+    threading.Thread(target = _warm_rag_embedder, daemon = True, name = "rag-embedder-warm").start()
 
-    # Warm the RAG embedder so the first upload skips the cold load. Non-fatal.
-    def _warm_rag_embedder():
-        try:
-            from storage import rag_db
-
-            if not rag_db.RAG_AVAILABLE:
-                return
-            from core.rag import embeddings
-
-            embeddings.warm()
-        except Exception:
-            pass
-
-    threading.Thread(target = _warm_rag_embedder, daemon = True).start()
-
-    # Initialize RSA key pair for API key encryption (external providers)
+    # Initialize RSA key pair for API key encryption (external providers).
     from core.inference.key_exchange import init_key_pair
 
     init_key_pair()
+    _lifespan_log.info(
+        "lifespan pre-auth setup completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
 
     if storage.ensure_default_admin():
         bootstrap_pw = storage.get_bootstrap_password()
@@ -532,6 +542,11 @@ async def lifespan(app: FastAPI):
         print("=" * 60 + "\n")
     else:
         app.state.bootstrap_password = storage.get_bootstrap_password()
+
+    _lifespan_log.info(
+        "lifespan startup completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
     yield
 
     from core.inference.llama_http import aclose as _close_llama_http
@@ -919,6 +934,21 @@ install_api_error_handlers(app)
 # ============ Health and System Endpoints ============
 
 
+@app.get("/api/liveness")
+async def liveness_check():
+    """Cheap process liveness for desktop port validation."""
+    return {
+        "status": "alive",
+        "service": "Unsloth UI Backend",
+        "desktop_protocol_version": 1,
+        "desktop_manageability_version": 1,
+        "supports_desktop_auth": True,
+        "supports_desktop_backend_ownership": True,
+        "studio_root_id": _studio_root_id(),
+        **({"desktop_owner": owner} if (owner := _desktop_owner()) else {}),
+    }
+
+
 @app.get("/api/health")
 async def health_check(request: Request):
     """Liveness plus launcher capability bits; host fingerprint gated on a bearer.
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 26825a472e..4a3162b09e 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -106,8 +106,7 @@ class LoadRequest(BaseModel):
             "Extra arguments forwarded verbatim to llama-server for GGUF models. "
             "One token per list entry, e.g. ['--top-k', '20', '--seed', '42']. "
             "Studio-managed flags (model identity, port, context length, GPU placement, "
-            "auth, --flash-attn, --no-context-shift, --jinja) are rejected. Ignored for "
-            "non-GGUF models."
+            "auth, UI/server mode) are rejected. Ignored for non-GGUF models."
         ),
     )
 
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index e64b6f731a..ff815a2fa9 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -9,6 +9,8 @@ import re
 from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 from typing import Any, Optional, List, Dict, Literal
 
+from utils.training_runs import normalize_project_name
+
 
 # ASCII integer, optional single sign. Rejects "++512" and Unicode digits
 # ("５１２") that slip through str.isdigit() + int().
@@ -97,6 +99,11 @@ class TrainingStartRequest(BaseModel):
     model_name: str = Field(
         ..., description = "Model identifier (e.g., 'unsloth/llama-3-8b-bnb-4bit')"
     )
+    project_name: Optional[str] = Field(
+        None,
+        max_length = 80,
+        description = "Optional user-defined project name appended to run folders and shown in history",
+    )
     training_type: Literal["LoRA/QLoRA", "Full Finetuning", "Continued Pretraining"] = Field(
         ...,
         description = "Training type: 'LoRA/QLoRA', 'Full Finetuning', or 'Continued Pretraining'",
@@ -155,6 +162,11 @@ class TrainingStartRequest(BaseModel):
             values.setdefault("train_split", values.pop("split"))
         return values
 
+    @field_validator("project_name")
+    @classmethod
+    def _normalize_project_name(cls, value: Optional[str]) -> Optional[str]:
+        return normalize_project_name(value)
+
     # NOTE: pydantic runs all `mode="after"` validators in definition order. A
     # second one, `_check_steps_or_epochs`, is defined lower in this class; keep
     # these cross-field checks order-independent so the two stay decoupled.
@@ -588,6 +600,7 @@ class TrainingRunSummary(BaseModel):
     id: str
     status: Literal["running", "completed", "stopped", "error"]
     model_name: str
+    project_name: Optional[str] = None
     dataset_name: str
     display_name: Optional[str] = None
     started_at: str
diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index 8fb034ea4e..57a291291e 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -481,6 +481,37 @@ async def upload_unstructured_file(
                 error = "No extractable text found in file",
             )
         extracted_path.write_text(extracted_text, encoding = "utf-8")
+    except ImportError as e:
+        raw_path.unlink(missing_ok = True)
+        extracted_path.unlink(missing_ok = True)
+        missing = getattr(e, "name", None)
+        expected_missing = {".pdf": "pymupdf4llm", ".docx": "mammoth"}.get(ext)
+        if isinstance(e, ModuleNotFoundError) and missing == expected_missing:
+            logger.error(
+                "data_recipe.seed.text_extraction_dependency_missing",
+                error = str(e),
+                missing = missing,
+                exc_info = True,
+            )
+            return UnstructuredFileUploadResponse(
+                file_id = file_id,
+                filename = original_filename,
+                size_bytes = size_bytes,
+                status = "error",
+                error = f"Cannot read {ext} files: the '{missing}' package is not installed.",
+            )
+        logger.error(
+            "data_recipe.seed.text_extraction_failed",
+            error = str(e),
+            exc_info = True,
+        )
+        return UnstructuredFileUploadResponse(
+            file_id = file_id,
+            filename = original_filename,
+            size_bytes = size_bytes,
+            status = "error",
+            error = "Text extraction failed.",
+        )
     except Exception as e:
         raw_path.unlink(missing_ok = True)
         extracted_path.unlink(missing_ok = True)
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index a9c85fa99a..d3f56edec5 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -683,7 +683,9 @@ try:
         detect_reasoning_flags,
     )
     from core.inference.llama_server_args import (
+        _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
         validate_extra_args,
@@ -718,7 +720,9 @@ except ImportError:
         detect_reasoning_flags,
     )
     from core.inference.llama_server_args import (
+        _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
         validate_extra_args,
@@ -1107,6 +1111,7 @@ from auth.authentication import get_current_subject
 from state.tool_approvals import resolve_tool_decision
 
 from core.inference.key_exchange import decrypt_api_key
+from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
 from core.inference.providers import get_base_url
@@ -2077,6 +2082,32 @@ def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[
     )
 
 
+def _carry_preserved_tensor_intent(
+    *, preserved: bool, same_model: bool, explicit_drop: bool
+) -> bool:
+    """Carry a preserved multi-GPU layer fallback forward only for a reload of the
+    SAME loaded model that doesn't explicitly drop tensor intent, so a fitting model
+    isn't collapsed to one GPU on a ctx-only change -- but an unrelated model switch
+    (without /unload) or an explicit tensor-off doesn't inherit it (#6659)."""
+    return preserved and same_model and not explicit_drop
+
+
+def _is_explicit_tensor_drop(request: LoadRequest) -> bool:
+    """True only when the request explicitly selects a non-tensor --split-mode (e.g.
+    layer/row/none), a deliberate departure from a preserved tensor->layer fallback.
+
+    A bare tensor_parallel field is NOT a drop: the Studio UI always sends it and echoes
+    the /load response's resolved value back, so after a fallback every reload carries
+    tensor_parallel=false even though the user never changed it -- treating that as a drop
+    would collapse the preserved multi-GPU placement on the next ctx/settings reload. An
+    empty clear is not a drop either (a fallback always stores --split-mode layer, never a
+    tensor split mode, so a clear never wipes tensor intent), nor is an unrelated extra
+    (--top-k) or inherit (None). tensor_parallel=true / --split-mode tensor re-engage
+    tensor. Shared by the already-loaded dedup and the load carry-forward (#6659)."""
+    override = parse_split_mode_override(request.llama_extra_args)
+    return override is not None and override.strip().lower() != "tensor"
+
+
 def _request_matches_loaded_settings(
     request: LoadRequest,
     llama_backend: LlamaCppBackend,
@@ -2115,6 +2146,13 @@ def _request_matches_loaded_settings(
         effective_extra, request.tensor_parallel, llama_backend.tensor_parallel
     ):
         return False
+    # Preserved tensor->layer fallback (both report tensor=off, so the check above
+    # matches): if the user now explicitly drops tensor intent, reload so placement
+    # re-selects instead of keeping the all-GPU mask (#6659). The effective check
+    # includes the env, so an env-only tensor (LLAMA_ARG_SPLIT_MODE=tensor) that
+    # can't actually be dropped falls through to the env-downgrade match, not a loop.
+    if llama_backend.layer_preserves_tensor_intent and _is_explicit_tensor_drop(request):
+        return False
     # 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
@@ -2809,6 +2847,48 @@ async def load_model(
                     hf_variant = config.gguf_variant,
                 )
 
+            # Tensor intent for this load: the request itself, or a preserved
+            # multi-GPU layer fallback carried across a reload of the SAME model that
+            # doesn't drop it (e.g. a ctx-only change), so a fitting model doesn't
+            # silently collapse to one GPU. Only an explicit non-tensor --split-mode
+            # override counts as the drop -- the tensor field echo / unrelated extras keep
+            # the preserved placement; the same-model guard stops a switch-without-unload
+            # inheriting the prior model's intent.
+            _explicit_tensor_drop = _is_explicit_tensor_drop(request)
+            # Compare the resolved config.identifier (what load_model stores), not the
+            # raw request id: from_identifier normalizes shorthands (adds unsloth/, fixes
+            # case), so a reload with the shorthand would otherwise miss the match and
+            # drop the carry-forward. #6659
+            _same_model_loaded = (
+                llama_backend.is_loaded
+                and (llama_backend.model_identifier or "").lower()
+                == (config.identifier or "").lower()
+            )
+            # model_identifier is variant-agnostic for HF repos and dir-level for a
+            # local multi-variant directory, so also require the loaded quant to match
+            # (path else variant, mirroring _already_in_target_state) -- otherwise a
+            # different variant inherits the prior one's preserved intent. #6659
+            if _same_model_loaded:
+                if config.gguf_file and llama_backend.gguf_path:
+                    try:
+                        _same_model_loaded = (
+                            Path(llama_backend.gguf_path).resolve()
+                            == Path(config.gguf_file).resolve()
+                        )
+                    except OSError:
+                        _same_model_loaded = False
+                else:
+                    _same_model_loaded = (llama_backend.hf_variant or "").lower() == (
+                        config.gguf_variant or ""
+                    ).lower()
+            _tensor_intent_overall = _effective_tensor_parallel(
+                extra_llama_args, request.tensor_parallel
+            ) or _carry_preserved_tensor_intent(
+                preserved = llama_backend.layer_preserves_tensor_intent,
+                same_model = _same_model_loaded,
+                explicit_drop = _explicit_tensor_drop,
+            )
+
             # Run a single load attempt with the given tensor flag + extras.
             async def _attempt_gguf_load(
                 tensor_parallel: bool, attempt_extra_args: Optional[list[str]]
@@ -2822,6 +2902,12 @@ async def load_model(
                     **_source_load_kwargs,
                     **attempt_kwargs,
                     tensor_parallel = tensor_parallel,
+                    # True on the layer fallback retry (tensor wanted overall but not on
+                    # this attempt): keep multi-GPU. Mirrors the fallback's key.
+                    preserve_multi_gpu_on_layer = bool(
+                        _tensor_intent_overall
+                        and not _effective_tensor_parallel(attempt_extra_args, tensor_parallel)
+                    ),
                 )
 
             # Tensor parallelism is arch-gated in llama.cpp and crashes some loads
@@ -3703,7 +3789,7 @@ async def generate_audio(
     # Pick backend — both return (wav_bytes, sample_rate)
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded and getattr(llama_backend, "_is_audio", False):
-        model_name = llama_backend.model_identifier
+        model_name = public_model_id(llama_backend.model_identifier)
         gen = lambda: llama_backend.generate_audio_response(
             text = text,
             audio_type = llama_backend._audio_type,
@@ -3721,7 +3807,7 @@ async def generate_audio(
         model_info = backend.models.get(backend.active_model_name, {})
         if not model_info.get("is_audio"):
             raise HTTPException(status_code = 400, detail = "Active model is not an audio model.")
-        model_name = backend.active_model_name
+        model_name = public_model_id(backend.active_model_name)
         gen = lambda: backend.generate_audio_response(
             text = text,
             temperature = payload.temperature,
@@ -4837,7 +4923,8 @@ async def openai_chat_completions(
         return response
 
     if using_gguf:
-        model_name = llama_backend.model_identifier or payload.model
+        # Echo a clean public id in the response, never the absolute .gguf path.
+        model_name = public_model_id(llama_backend.model_identifier) or payload.model
         if getattr(llama_backend, "_is_audio", False):
             if _wants_multiple_choices(payload):
                 _raise_unsupported_n("GGUF audio chat completions")
@@ -4852,7 +4939,9 @@ async def openai_chat_completions(
                 status_code = 400,
                 detail = "No model loaded. Call POST /inference/load first.",
             )
-        model_name = backend.active_model_name or payload.model
+        # Clean public id so the response never echoes a local path; the audio
+        # branch below receives this sanitized label too.
+        model_name = public_model_id(backend.active_model_name) or payload.model
         if _wants_multiple_choices(payload):
             _raise_unsupported_n("non-GGUF chat completions")
 
@@ -6387,6 +6476,9 @@ async def serve_sandbox_file(
 # OpenAI-Compatible Models Listing  (/models → /v1/models)
 # =====================================================================
 
+# `owned_by` marker on every /v1/models entry (loaded and available alike).
+_OWNED_BY = "unsloth-studio"
+
 
 def _openai_model_objects() -> list[dict]:
     """The model objects GET /v1/models exposes (one per loaded local backend).
@@ -6401,10 +6493,12 @@ def _openai_model_objects() -> list[dict]:
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded:
         entry = {
-            "id": llama_backend.model_identifier,
+            # Public id, never the absolute .gguf path (which leaks the host
+            # filesystem layout); see core.inference.model_ids.public_model_id.
+            "id": public_model_id(llama_backend.model_identifier),
             "object": "model",
             "created": _created,
-            "owned_by": "local",
+            "owned_by": _OWNED_BY,
         }
         _ctx = _positive_int_or_none(getattr(llama_backend, "context_length", None))
         if _ctx is not None:
@@ -6422,10 +6516,10 @@ def _openai_model_objects() -> list[dict]:
     if backend.active_model_name:
         model_info = backend.models.get(backend.active_model_name, {})
         entry = {
-            "id": backend.active_model_name,
+            "id": public_model_id(backend.active_model_name),
             "object": "model",
             "created": _created,
-            "owned_by": "local",
+            "owned_by": _OWNED_BY,
         }
         _ctx = _positive_int_or_none(model_info.get("context_length"))
         if _ctx is None:
@@ -6443,15 +6537,86 @@ def _openai_model_objects() -> list[dict]:
     return models
 
 
+# Brief cache for the local-model filesystem scan so repeated /v1/models calls
+# don't rescan the HF cache and models dirs on every request.
+_CATALOG_CACHE: dict = {"at": 0.0, "models": []}
+_CATALOG_TTL_S = 30.0
+_CATALOG_LOCK = asyncio.Lock()
+
+
+async def _cached_local_catalog() -> list:
+    """Locally available models (models dir + HF caches + LM Studio + scan
+    folders), cached for a few seconds. Returns a list of LocalModelInfo.
+
+    The scan walks several directories and stats many files, so it runs in a
+    worker thread (asyncio.to_thread) -- calling it inline would block the event
+    loop and stall every concurrent request and in-flight inference stream. A
+    lock with a double-check collapses a burst of simultaneous /v1/models calls
+    into a single scan instead of one per request."""
+    # Validity is keyed on "at" (set only after a scan), not on list contents, so
+    # an empty/errored scan is still cached instead of rescanning on every poll.
+    now = time.monotonic()
+    if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
+        return _CATALOG_CACHE["models"]
+    async with _CATALOG_LOCK:
+        now = time.monotonic()
+        if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
+            return _CATALOG_CACHE["models"]
+        try:
+            from routes.models import collect_local_models
+            _CATALOG_CACHE["models"] = await asyncio.to_thread(
+                collect_local_models, Path("./models").resolve()
+            )
+        except Exception as exc:
+            logger.debug("model catalog scan failed: %s", exc)
+            _CATALOG_CACHE["models"] = []
+        # Stamp after the scan, not the pre-scan "now": a scan slower than the TTL
+        # would otherwise leave the cache already expired, so every waiter rescans.
+        _CATALOG_CACHE["at"] = time.monotonic()
+    return _CATALOG_CACHE["models"]
+
+
+async def _openai_catalog_objects() -> list[dict]:
+    """Every model the server knows about for ``GET /v1/models``: the loaded
+    model(s) plus locally available (downloaded/cached) models discovered by
+    scanning. Loaded entries keep their context fields and are marked
+    ``loaded: true``. All ids are clean public ids (never absolute paths)."""
+    _created = int(time.time())
+    # Loaded models first (clean ids + context fields), marked loaded.
+    by_id: dict[str, dict] = {}
+    for entry in _openai_model_objects():
+        by_id[entry["id"]] = {**entry, "loaded": True}
+
+    # Locally available (downloaded/cached) models that are not already loaded.
+    for info in await _cached_local_catalog():
+        cid = getattr(info, "model_id", None) or public_model_id(getattr(info, "id", None))
+        if not cid or cid in by_id:
+            continue
+        obj = {
+            "id": cid,
+            "object": "model",
+            "created": _created,
+            "owned_by": _OWNED_BY,
+            "loaded": False,
+        }
+        display = getattr(info, "display_name", None)
+        if display:
+            obj["display_name"] = display
+        by_id[cid] = obj
+
+    return list(by_id.values())
+
+
 @router.get("/models")
 async def openai_list_models(current_subject: str = Depends(get_current_subject)):
     """
-    OpenAI-compatible model listing endpoint.
+    OpenAI-compatible model listing endpoint (``GET /v1/models``).
 
-    Returns the currently loaded model in the format expected by
-    OpenAI-compatible clients (``GET /v1/models``).
+    Lists every model available on this server -- the loaded model(s) plus
+    locally available (downloaded/cached) models -- not only what is resident in
+    memory. Each entry carries a clean public id and a ``loaded`` flag.
     """
-    return {"object": "list", "data": _openai_model_objects()}
+    return {"object": "list", "data": await _openai_catalog_objects()}
 
 
 @router.get("/models/{model_id:path}")
@@ -6459,13 +6624,37 @@ async def openai_retrieve_model(model_id: str, current_subject: str = Depends(ge
     """
     OpenAI-compatible single-model retrieval endpoint (``GET /v1/models/{id}``).
 
-    Returns the bare model object when ``model_id`` matches a loaded local
-    model, or 404 model_not_found otherwise. Defined after the LIST route so
-    it does not shadow it; ``{model_id:path}`` keeps ids with slashes intact.
+    Returns the bare model object when ``model_id`` matches a known model
+    (loaded or locally available), or 404 model_not_found otherwise. Defined
+    after the LIST route so it does not shadow it; ``{model_id:path}`` keeps ids
+    with slashes intact.
     """
-    for model in _openai_model_objects():
+    from core.inference.model_ids import model_id_matches
+
+    # Loaded models resolve without a catalog scan (the common case); only build
+    # the full catalog -- which may hit the filesystem -- for unloaded ids.
+    for entry in _openai_model_objects():
+        if entry["id"] == model_id:
+            return {**entry, "loaded": True}
+
+    objects = await _openai_catalog_objects()
+    for model in objects:
         if model["id"] == model_id:
             return model
+    # Backward compatibility: a client may still send the legacy raw identifier
+    # (e.g. an absolute .gguf path cached from an older /v1/models). Resolve it to
+    # the clean object so it keeps working, without ever echoing the path back.
+    llama_backend = get_llama_cpp_backend()
+    backend = get_inference_backend()
+    for raw in (
+        llama_backend.model_identifier if llama_backend.is_loaded else None,
+        backend.active_model_name or None,
+    ):
+        if raw and model_id_matches(model_id, raw):
+            clean = public_model_id(raw)
+            for model in objects:
+                if model["id"] == clean:
+                    return model
     raise HTTPException(
         status_code = 404,
         detail = openai_error_body(
@@ -7402,6 +7591,15 @@ async def _responses_stream(
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
 
     async def event_generator():
+        # Clean public id for every response envelope. Prefer the loaded model's
+        # id so the stream agrees with /v1/models, chat/completions and the
+        # non-streaming twin; fall back to a sanitized payload.model (a legacy
+        # raw .gguf path is stripped, never echoed back).
+        _clean_model = (
+            public_model_id(getattr(llama_backend, "model_identifier", None))
+            or public_model_id(payload.model)
+            or payload.model
+        )
         full_text = ""
         full_reasoning = ""
         input_tokens = 0
@@ -7563,7 +7761,7 @@ async def _responses_stream(
                     "object": "response",
                     "created_at": created_at,
                     "status": "failed",
-                    "model": payload.model,
+                    "model": _clean_model,
                     "output": _snapshot_output(),
                     "usage": {
                         "input_tokens": input_tokens,
@@ -7587,7 +7785,7 @@ async def _responses_stream(
                     "object": "response",
                     "created_at": created_at,
                     "status": "in_progress",
-                    "model": payload.model,
+                    "model": _clean_model,
                     "output": [],
                     "usage": {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0},
                 },
@@ -7627,7 +7825,7 @@ async def _responses_stream(
                             "object": "response",
                             "created_at": created_at,
                             "status": "failed",
-                            "model": payload.model,
+                            "model": _clean_model,
                             "output": [],
                             "error": {"code": 502, "message": _friendly_error(e)},
                         },
@@ -7653,7 +7851,7 @@ async def _responses_stream(
                             "object": "response",
                             "created_at": created_at,
                             "status": "failed",
-                            "model": payload.model,
+                            "model": _clean_model,
                             "output": [],
                             "error": {
                                 "code": resp.status_code,
@@ -8002,7 +8200,7 @@ async def _responses_stream(
                 "object": "response",
                 "created_at": created_at,
                 "status": "completed",
-                "model": payload.model,
+                "model": _clean_model,
                 "output": _snapshot_output(),
                 "usage": {
                     "input_tokens": input_tokens,
@@ -8274,7 +8472,13 @@ async def anthropic_messages(
             ),
         )
 
-    model_name = getattr(llama_backend, "model_identifier", None) or payload.model
+    # Clean public id so /v1/messages never echoes the local .gguf path (and a
+    # legacy raw path sent as payload.model is sanitized rather than returned).
+    model_name = (
+        public_model_id(getattr(llama_backend, "model_identifier", None))
+        or public_model_id(payload.model)
+        or payload.model
+    )
     message_id = f"msg_{uuid.uuid4().hex[:24]}"
 
     # ── Translate Anthropic → OpenAI ──────────────────────────
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 951c2960f3..e22f65751c 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -722,6 +722,94 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca
     return found
 
 
+def collect_local_models(models_root: Path) -> List[LocalModelInfo]:
+    """Scan ``models_root``, the HF caches, LM Studio dirs, and user scan folders,
+    returning a deduplicated, hidden-filtered list of discovered local models.
+
+    Shared by ``GET /models/local`` (the model picker) and the OpenAI-compatible
+    catalog (``GET /v1/models``) so the UI and the API never drift. ``models_root``
+    must already be validated/trusted by the caller.
+    """
+    from storage.studio_db import list_scan_folders
+    from utils.paths import (
+        hf_default_cache_dir,
+        legacy_hf_cache_dir,
+        lmstudio_model_dirs,
+    )
+
+    hf_cache_dir = _resolve_hf_cache_dir()
+    legacy_hf = legacy_hf_cache_dir()
+    hf_default = hf_default_cache_dir()
+    lm_dirs = lmstudio_model_dirs()
+
+    local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir)
+
+    # Resolve once; an inaccessible aux cache must skip that scan, not 500.
+    hf_cache_real = _safe_resolve(hf_cache_dir)
+    legacy_real = _safe_resolve(legacy_hf)
+    default_real = _safe_resolve(hf_default)
+
+    # Scan legacy Unsloth HF cache for backward compatibility.
+    if _safe_is_dir(legacy_hf) and legacy_real != hf_cache_real:
+        local_models += _scan_hf_cache(legacy_hf)
+
+    # Scan HF system default cache (may differ under env overrides).
+    if _safe_is_dir(hf_default) and default_real != hf_cache_real and default_real != legacy_real:
+        local_models += _scan_hf_cache(hf_default)
+
+    # Scan LM Studio directories.
+    for lm_dir in lm_dirs:
+        local_models += _scan_lmstudio_dir(lm_dir)
+
+    # Scan user-added custom folders (per-folder cap).
+    _MAX_MODELS_PER_FOLDER = 200
+    try:
+        custom_folders = list_scan_folders()
+    except Exception as e:
+        logger.warning("Could not load custom scan folders: %s", e)
+        custom_folders = []
+    for folder in custom_folders:
+        folder_path = Path(folder["path"])
+        try:
+            # Filter Ollama .studio_links/ from generic scanners to
+            # avoid duplicates and leaking internal paths into the UI.
+            _generic = [
+                m
+                for m in (
+                    _scan_models_dir(folder_path, limit = _MAX_MODELS_PER_FOLDER)
+                    + _scan_hf_cache(folder_path)
+                    + _scan_lmstudio_dir(folder_path)
+                )
+                if not any(p in (".studio_links", "ollama_links") for p in Path(m.path).parts)
+            ]
+            custom_models = _generic
+            if len(custom_models) < _MAX_MODELS_PER_FOLDER:
+                custom_models += _scan_ollama_dir(
+                    folder_path,
+                    limit = _MAX_MODELS_PER_FOLDER - len(custom_models),
+                )
+        except OSError as e:
+            logger.warning("Skipping unreadable scan folder %s: %s", folder_path, e)
+            continue
+        local_models += [m.model_copy(update = {"source": "custom"}) for m in custom_models]
+
+    # Deduplicate, but always keep custom folder entries (keyed by
+    # (id, source)) so they show in the "Custom Folders" UI section
+    # even when the model is also in the HF cache.
+    deduped: dict[str, LocalModelInfo] = {}
+    for model in local_models:
+        key = f"{model.id}\x00custom" if model.source == "custom" else model.id
+        if key not in deduped:
+            deduped[key] = model
+
+    models = sorted(
+        deduped.values(),
+        key = lambda item: (item.updated_at or 0),
+        reverse = True,
+    )
+    return [m for m in models if not _is_hidden_model(m.id, m.path)]
+
+
 @router.get("/local", response_model = LocalModelListResponse)
 async def list_local_models(
     models_dir: str = Query(
@@ -770,78 +858,7 @@ async def list_local_models(
         )
 
     try:
-        local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir)
-
-        # Resolve once; an inaccessible aux cache must skip that scan, not 500.
-        hf_cache_real = _safe_resolve(hf_cache_dir)
-        legacy_real = _safe_resolve(legacy_hf)
-        default_real = _safe_resolve(hf_default)
-
-        # Scan legacy Unsloth HF cache for backward compatibility.
-        if _safe_is_dir(legacy_hf) and legacy_real != hf_cache_real:
-            local_models += _scan_hf_cache(legacy_hf)
-
-        # Scan HF system default cache (may differ under env overrides).
-        if (
-            _safe_is_dir(hf_default)
-            and default_real != hf_cache_real
-            and default_real != legacy_real
-        ):
-            local_models += _scan_hf_cache(hf_default)
-
-        # Scan LM Studio directories.
-        for lm_dir in lm_dirs:
-            local_models += _scan_lmstudio_dir(lm_dir)
-
-        # Scan user-added custom folders (per-folder cap).
-        from storage.studio_db import list_scan_folders
-
-        _MAX_MODELS_PER_FOLDER = 200
-        try:
-            custom_folders = list_scan_folders()
-        except Exception as e:
-            logger.warning("Could not load custom scan folders: %s", e)
-            custom_folders = []
-        for folder in custom_folders:
-            folder_path = Path(folder["path"])
-            try:
-                # Filter Ollama .studio_links/ from generic scanners to
-                # avoid duplicates and leaking internal paths into the UI.
-                _generic = [
-                    m
-                    for m in (
-                        _scan_models_dir(folder_path, limit = _MAX_MODELS_PER_FOLDER)
-                        + _scan_hf_cache(folder_path)
-                        + _scan_lmstudio_dir(folder_path)
-                    )
-                    if not any(p in (".studio_links", "ollama_links") for p in Path(m.path).parts)
-                ]
-                custom_models = _generic
-                if len(custom_models) < _MAX_MODELS_PER_FOLDER:
-                    custom_models += _scan_ollama_dir(
-                        folder_path,
-                        limit = _MAX_MODELS_PER_FOLDER - len(custom_models),
-                    )
-            except OSError as e:
-                logger.warning("Skipping unreadable scan folder %s: %s", folder_path, e)
-                continue
-            local_models += [m.model_copy(update = {"source": "custom"}) for m in custom_models]
-
-        # Deduplicate, but always keep custom folder entries (keyed by
-        # (id, source)) so they show in the "Custom Folders" UI section
-        # even when the model is also in the HF cache.
-        deduped: dict[str, LocalModelInfo] = {}
-        for model in local_models:
-            key = f"{model.id}\x00custom" if model.source == "custom" else model.id
-            if key not in deduped:
-                deduped[key] = model
-
-        models = sorted(
-            deduped.values(),
-            key = lambda item: (item.updated_at or 0),
-            reverse = True,
-        )
-        models = [m for m in models if not _is_hidden_model(m.id, m.path)]
+        models = collect_local_models(models_root)
 
         return LocalModelListResponse(
             models_dir = str(models_root),
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 3818fe9f73..4f131ad2f2 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -255,6 +255,7 @@ async def start_training(
         # Convert request to backend kwargs.
         training_kwargs = {
             "model_name": request.model_name,
+            "project_name": request.project_name,
             "training_type": request.training_type,
             "hf_token": request.hf_token or "",
             "load_in_4bit": request.load_in_4bit,
diff --git a/studio/backend/run.py b/studio/backend/run.py
index d4cbc26b41..6f36f0928a 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -933,6 +933,9 @@ def run_server(
     """
     global _server, _server_thread, _shutdown_event
 
+    boot_started = time.perf_counter()
+    logger.info("run_server startup begin api_only=%s host=%s port=%s", api_only, host, port)
+
     # Reap every child if the parent dies abnormally (terminal close, Task
     # Manager kill, SIGKILL); must run before any child can spawn.
     from utils.process_lifetime import initialize_parent_lifetime
@@ -984,7 +987,14 @@ def run_server(
     from threading import Thread, Event
     import uvicorn
 
+    import_started = time.perf_counter()
+
     from main import app, setup_frontend, _IS_COLAB
+
+    logger.info(
+        "Imported FastAPI app in %.1fms",
+        (time.perf_counter() - import_started) * 1000,
+    )
     from utils.paths import ensure_studio_directories
 
     # Allow local stdio MCP servers on a loopback bind (the user's own machine),
@@ -997,6 +1007,11 @@ def run_server(
     # Create all standard directories on startup.
     ensure_studio_directories()
 
+    logger.info(
+        "Ensured Studio directories in %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     # Auto-find a free port if the requested one is in use.
     if not _is_port_free(host, port):
         original_port = port
@@ -1060,6 +1075,11 @@ def run_server(
     display_host = _resolve_external_ip() if host == "0.0.0.0" else host
     _install_uvicorn_startup_log_rewrite(host, display_host)
 
+    logger.info(
+        "run_server pre-uvicorn setup completed in %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     ready_event = Event()
     startup_failed = Event()
     startup_errors = []
@@ -1068,6 +1088,10 @@ def run_server(
         async def startup(self, *args, **kwargs):
             await super().startup(*args, **kwargs)
             if getattr(self, "started", False) and not self.should_exit:
+                logger.info(
+                    "Uvicorn startup hook completed in %.1fms",
+                    (time.perf_counter() - boot_started) * 1000,
+                )
                 ready_event.set()
 
     # server_header=False suppresses uvicorn's "Server: uvicorn"; SecurityHeadersMiddleware sets its own.
@@ -1150,6 +1174,11 @@ def run_server(
         _shutdown_event.set()
         raise
 
+    logger.info(
+        "run_server uvicorn ready after %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     _write_pid_file()
     import atexit
 
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 7421b42b2f..23b90d7002 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -23,6 +23,16 @@ from typing import Any, Iterable, Optional
 
 
 from utils.paths import project_workspaces_root, studio_db_path, ensure_dir
+from utils.training_runs import extract_project_name
+
+
+def _extract_project_name_from_config_json(config_json: Optional[str]) -> Optional[str]:
+    if not config_json:
+        return None
+    try:
+        return extract_project_name(json.loads(config_json))
+    except (json.JSONDecodeError, TypeError):
+        return None
 
 
 def _denied_path_prefixes() -> list[str]:
@@ -680,6 +690,7 @@ def list_runs(limit: int = 50, offset: int = 0) -> dict:
         runs = []
         for row in rows:
             run = dict(row)
+            run["project_name"] = _extract_project_name_from_config_json(run.get("config_json"))
             sparkline = run.get("loss_sparkline")
             if sparkline:
                 try:
@@ -719,6 +730,7 @@ def get_run(id: str) -> Optional[dict]:
         if row is None:
             return None
         run = dict(row)
+        run["project_name"] = _extract_project_name_from_config_json(run.get("config_json"))
         sparkline = run.get("loss_sparkline")
         if sparkline:
             try:
diff --git a/studio/backend/tests/test_checkpoints_scan.py b/studio/backend/tests/test_checkpoints_scan.py
new file mode 100644
index 0000000000..6d473146f5
--- /dev/null
+++ b/studio/backend/tests/test_checkpoints_scan.py
@@ -0,0 +1,256 @@
+# 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 json
+import sqlite3
+import sys
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+sys.modules.setdefault("structlog", _types.ModuleType("structlog"))
+
+from utils.models import checkpoints as checkpoints_module
+from utils.training_runs import build_default_output_dir_name
+
+
+def _make_history_connection(db_path: Path) -> sqlite3.Connection:
+    conn = sqlite3.connect(str(db_path))
+    conn.row_factory = sqlite3.Row
+    return conn
+
+
+def _setup_training_runs_table(db_path: Path) -> None:
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            CREATE TABLE training_runs (
+                id TEXT PRIMARY KEY,
+                model_name TEXT NOT NULL,
+                config_json TEXT NOT NULL,
+                output_dir TEXT,
+                started_at TEXT NOT NULL
+            )
+            """
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def _make_outputs_dir(tmp_path, monkeypatch) -> Path:
+    studio_home = tmp_path / "studio-home"
+    outputs_dir = studio_home / "outputs"
+    outputs_dir.mkdir(parents = True)
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(studio_home))
+    return outputs_dir
+
+
+def test_scan_checkpoints_uses_output_dir_history_for_base_model(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "custom-run"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-1",
+                "unsloth/Llama-3.2-3B-Instruct",
+                "{}",
+                str(run_dir.resolve()),
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_matches_project_suffixed_default_dir_against_history(
+    tmp_path, monkeypatch
+):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-2",
+                "unsloth/Llama-3.2-3B-Instruct",
+                json.dumps({"project_name": "Customer Support"}),
+                None,
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_strips_project_suffix_without_history(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_preserves_project_marker_in_model_without_history(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "org/foo__project-bar",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "org/foo__project-bar"
+
+
+def test_scan_checkpoints_preserves_legacy_folder_name_fallback(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "unsloth_Llama-3.2-3B-Instruct_1771227800"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_prefers_exact_history_match_over_newer_suffix(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "unsloth_Test_1771227800"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    copied_dir = tmp_path / "copied" / run_dir.name
+    copied_dir.mkdir(parents = True)
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-exact",
+                "correct/base",
+                "{}",
+                str(run_dir.resolve()),
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-suffix",
+                "wrong/base",
+                "{}",
+                str(copied_dir.resolve()),
+                "2026-04-10T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "correct/base"
diff --git a/studio/backend/tests/test_data_recipe_seed.py b/studio/backend/tests/test_data_recipe_seed.py
index 601df8bbfe..09e22116ed 100644
--- a/studio/backend/tests/test_data_recipe_seed.py
+++ b/studio/backend/tests/test_data_recipe_seed.py
@@ -1,12 +1,126 @@
 # 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 asyncio
+import importlib.util
 from pathlib import Path
 
+import pytest
 
-def test_seed_inspect_load_kwargs_disables_remote_code_execution():
-    seed_route = (
+
+def _seed_route_source() -> str:
+    return (
         Path(__file__).resolve().parent.parent / "routes" / "data_recipe" / "seed.py"
     ).read_text()
 
-    assert '"trust_remote_code": False' in seed_route
+
+def test_seed_inspect_load_kwargs_disables_remote_code_execution():
+    assert '"trust_remote_code": False' in _seed_route_source()
+
+
+class _FakeUpload:
+    def __init__(self, filename: str, content: bytes):
+        self.filename = filename
+        self._content = content
+
+    async def read(self) -> bytes:
+        return self._content
+
+
+def _load_seed_route(monkeypatch: pytest.MonkeyPatch, tmp_path: Path):
+    pytest.importorskip("fastapi")
+    pytest.importorskip("multipart")
+    pytest.importorskip("structlog")
+
+    backend_root = Path(__file__).resolve().parent.parent
+    monkeypatch.syspath_prepend(str(backend_root))
+    route_path = backend_root / "routes" / "data_recipe" / "seed.py"
+    spec = importlib.util.spec_from_file_location("seed_under_test", route_path)
+    assert spec is not None and spec.loader is not None
+    seed_route = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(seed_route)
+    seed_route.UNSTRUCTURED_UPLOAD_ROOT = tmp_path / "unstructured-uploads"
+    return seed_route
+
+
+def _run_upload(
+    seed_route,
+    filename: str,
+    content: bytes,
+    block_id: str = "block",
+):
+    return asyncio.run(
+        seed_route.upload_unstructured_file(_FakeUpload(filename, content), block_id)
+    )
+
+
+def _block_files(seed_route, block_id: str = "block") -> list[str]:
+    block_dir = seed_route.UNSTRUCTURED_UPLOAD_ROOT / block_id
+    if not block_dir.exists():
+        return []
+    return sorted(path.name for path in block_dir.iterdir())
+
+
+def _raise(exc: BaseException):
+    def raise_exc(*args, **kwargs):
+        raise exc
+
+    return raise_exc
+
+
+@pytest.mark.parametrize(
+    ("filename", "package"),
+    [
+        ("paper.pdf", "pymupdf4llm"),
+        ("notes.docx", "mammoth"),
+    ],
+)
+def test_unstructured_upload_names_missing_extractor_dependency(
+    monkeypatch, tmp_path, filename, package
+):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(
+        seed_route,
+        "_extract_text_from_file",
+        _raise(ModuleNotFoundError(f"No module named {package!r}", name = package)),
+    )
+
+    result = _run_upload(seed_route, filename, b"%PDF-1.7")
+
+    assert result.status == "error"
+    assert (
+        result.error
+        == f"Cannot read {Path(filename).suffix} files: the '{package}' package is not installed."
+    )
+    assert _block_files(seed_route) == []
+
+
+def test_unstructured_upload_keeps_txt_path_working(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    result = _run_upload(seed_route, "notes.txt", b"hello")
+
+    assert result.status == "ok"
+    assert result.error is None
+    assert any(name.endswith(".txt") for name in _block_files(seed_route))
+    assert any(name.endswith(".extracted.txt") for name in _block_files(seed_route))
+
+
+@pytest.mark.parametrize(
+    "exc",
+    [
+        ImportError("cannot import internal symbol"),
+        ModuleNotFoundError(
+            "No module named 'missing_transitive_pkg'",
+            name = "missing_transitive_pkg",
+        ),
+    ],
+)
+def test_unstructured_upload_import_errors_stay_generic(monkeypatch, tmp_path, exc):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(seed_route, "_extract_text_from_file", _raise(exc))
+    result = _run_upload(seed_route, "paper.pdf", b"%PDF-1.7")
+
+    assert result.status == "error"
+    assert result.error == "Text extraction failed."
+    assert _block_files(seed_route) == []
diff --git a/studio/backend/tests/test_model_ids.py b/studio/backend/tests/test_model_ids.py
new file mode 100644
index 0000000000..f9116afec3
--- /dev/null
+++ b/studio/backend/tests/test_model_ids.py
@@ -0,0 +1,62 @@
+# 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 pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.model_ids import model_id_matches, public_model_id  # noqa: E402
+
+
+def test_local_gguf_path_becomes_clean_stem():
+    assert public_model_id("/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf") == "Qwen3-30B-A3B-Q4_K_M"
+    assert public_model_id("/home/u/.cache/models/llama.gguf") == "llama"
+
+
+def test_hf_repo_id_unchanged():
+    assert public_model_id("unsloth/Qwen3-30B-A3B-GGUF") == "unsloth/Qwen3-30B-A3B-GGUF"
+    assert public_model_id("Qwen3-30B-A3B") == "Qwen3-30B-A3B"
+
+
+def test_none_and_empty_passthrough():
+    assert public_model_id(None) is None
+    assert public_model_id("") == ""
+
+
+def test_windows_path():
+    assert public_model_id("C:\\models\\foo.gguf") == "foo"
+    assert public_model_id("models\\sub\\bar.gguf") == "bar"
+
+
+def test_directory_path_uses_basename():
+    assert public_model_id("/opt/models/MyModelDir") == "MyModelDir"
+    # A 3+ segment relative path is a local path, not an org/model repo id.
+    assert public_model_id("a/b/c") == "c"
+
+
+def test_relative_and_home_paths_are_sanitized():
+    # ./ ../ ~ prefixed paths are local and must not be echoed raw.
+    assert public_model_id("./model.gguf") == "model"
+    assert public_model_id("../models/foo.gguf") == "foo"
+    assert public_model_id("~/models/baz.gguf") == "baz"
+    assert public_model_id("./mistral") == "mistral"
+    assert public_model_id("~/mistral") == "mistral"
+    assert public_model_id(".\\models\\foo.gguf") == "foo"
+
+
+def test_dotted_repo_id_not_mistaken_for_relative_path():
+    # A leading dot that is not ./ or ../ is an ordinary clean name.
+    assert public_model_id(".hidden-model") == ".hidden-model"
+    assert public_model_id("org/.config") == "org/.config"
+
+
+def test_matches_clean_and_legacy():
+    path = "/srv/models/Qwen3-Q4.gguf"
+    assert model_id_matches("Qwen3-Q4", path)  # clean public id
+    assert model_id_matches(path, path)  # legacy raw path
+    assert not model_id_matches("other", path)
+    assert not model_id_matches(None, path)
+    assert not model_id_matches("x", None)
diff --git a/studio/backend/tests/test_openai_catalog.py b/studio/backend/tests/test_openai_catalog.py
new file mode 100644
index 0000000000..f9baf20a66
--- /dev/null
+++ b/studio/backend/tests/test_openai_catalog.py
@@ -0,0 +1,181 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""GET /v1/models lists the full server catalog (loaded + locally available)."""
+
+import asyncio
+import json
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import routes.inference as inf  # noqa: E402
+
+
+class _Info:
+    def __init__(
+        self,
+        id,
+        display_name,
+        model_id = None,
+    ):
+        self.id = id
+        self.display_name = display_name
+        self.model_id = model_id
+
+
+class _FakeLlama:
+    is_loaded = True
+    model_identifier = "/srv/models/Qwen3-Q4.gguf"
+    context_length = 4096
+    max_context_length = None
+    native_context_length = None
+
+    def __init__(self, loaded = True):
+        self.is_loaded = loaded
+
+
+class _FakeUnsloth:
+    active_model_name = None
+    models: dict = {}
+    context_length = None
+    max_seq_length = None
+
+
+def test_catalog_lists_loaded_and_available(monkeypatch):
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    async def _fake_catalog():
+        return [
+            _Info("/data/models/Qwen3-Q4.gguf", "Qwen3-Q4"),  # same as loaded -> dedup
+            _Info("/data/models/Llama-8B-Q8.gguf", "Llama-8B-Q8"),  # available, not loaded
+            _Info("models--org--Foo", "Foo", model_id = "org/Foo"),  # hf cache repo id
+        ]
+
+    monkeypatch.setattr(inf, "_cached_local_catalog", _fake_catalog)
+
+    data = asyncio.run(inf._openai_catalog_objects())
+    ids = {m["id"]: m for m in data}
+
+    # Loaded model is present, marked loaded, and keeps context fields.
+    assert ids["Qwen3-Q4"]["loaded"] is True
+    assert ids["Qwen3-Q4"]["context_length"] == 4096
+    # Available-but-not-loaded models are listed too.
+    assert ids["Llama-8B-Q8"]["loaded"] is False
+    assert ids["org/Foo"]["loaded"] is False
+    # The loaded gguf and the on-disk copy collapse to one clean id.
+    assert [m["id"] for m in data].count("Qwen3-Q4") == 1
+    # No absolute paths or .gguf suffixes leak anywhere.
+    blob = json.dumps(data)
+    assert ".gguf" not in blob
+    assert "/srv/" not in blob
+    assert "/data/" not in blob
+
+
+def test_empty_and_errored_scans_are_cached(monkeypatch):
+    # Cache validity is keyed on the timestamp, not list contents, so an empty
+    # (fresh install / no local models) or errored scan is still cached for the
+    # TTL instead of rescanning the filesystem on every /v1/models poll.
+    import routes.models as models_mod
+    for outcome in ("empty", "error"):
+        calls = {"n": 0}
+
+        def _scan(_root, _outcome = outcome):
+            calls["n"] += 1
+            if _outcome == "error":
+                raise RuntimeError("scan blew up")
+            return []
+
+        monkeypatch.setattr(models_mod, "collect_local_models", _scan)
+        monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+        async def _run():
+            return [await inf._cached_local_catalog() for _ in range(3)]
+
+        results = asyncio.run(_run())
+        assert results == [[], [], []], outcome
+        assert calls["n"] == 1, f"{outcome} scan ran {calls['n']}x (TTL not honored)"
+
+
+def test_catalog_ttl_starts_after_scan_completes(monkeypatch):
+    # The cache timestamp must be taken AFTER the scan, not before it. A scan that
+    # outlives the TTL would otherwise leave the cache born-expired, so the next
+    # caller rescans instead of reusing the just-computed catalog.
+    import routes.models as models_mod
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(inf.time, "monotonic", lambda: clock["t"])
+    monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+    calls = {"n": 0}
+
+    def _slow_scan(_root):
+        calls["n"] += 1
+        clock["t"] += inf._CATALOG_TTL_S + 10  # the scan itself outlives the TTL
+        return [_Info("/m/A.gguf", "A")]
+
+    monkeypatch.setattr(models_mod, "collect_local_models", _slow_scan)
+
+    async def _run():
+        first = await inf._cached_local_catalog()
+        second = await inf._cached_local_catalog()  # clock unchanged since scan end
+        return first, second
+
+    first, second = asyncio.run(_run())
+    assert [i.id for i in first] == ["/m/A.gguf"]
+    assert calls["n"] == 1, "TTL started before the scan -> cache born expired, rescanned"
+
+
+def test_retrieve_loaded_model_skips_catalog_scan(monkeypatch):
+    # Retrieving a loaded id must resolve from the loaded set alone, never paying
+    # for the filesystem scan that _cached_local_catalog drives.
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    async def _boom():
+        raise AssertionError("catalog scan must not run for a loaded id")
+
+    monkeypatch.setattr(inf, "_cached_local_catalog", _boom)
+
+    model = asyncio.run(inf.openai_retrieve_model("Qwen3-Q4", current_subject = "t"))
+    assert model["id"] == "Qwen3-Q4"
+    assert model["loaded"] is True
+
+
+def test_cached_local_catalog_offloads_and_caches(monkeypatch):
+    # The filesystem scan must run off the event loop (asyncio.to_thread) and be
+    # cached, so a burst of /v1/models calls does not re-scan or block.
+    calls = {"scan": 0, "threaded": 0}
+
+    def _fake_collect(_root):
+        calls["scan"] += 1
+        return [_Info("/data/models/A.gguf", "A")]
+
+    import routes.models as models_mod
+
+    monkeypatch.setattr(models_mod, "collect_local_models", _fake_collect)
+
+    real_to_thread = inf.asyncio.to_thread
+
+    async def _counting_to_thread(fn, *a, **k):
+        calls["threaded"] += 1
+        return await real_to_thread(fn, *a, **k)
+
+    monkeypatch.setattr(inf.asyncio, "to_thread", _counting_to_thread)
+    # Fresh cache for a deterministic count.
+    monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+    async def _run():
+        first = await inf._cached_local_catalog()
+        second = await inf._cached_local_catalog()  # within TTL -> cached
+        return first, second
+
+    first, second = asyncio.run(_run())
+    assert [i.id for i in first] == ["/data/models/A.gguf"]
+    assert second is first or [i.id for i in second] == [i.id for i in first]
+    assert calls["scan"] == 1  # cached: scanned once for two calls
+    assert calls["threaded"] == 1  # offloaded to a worker thread
diff --git a/studio/backend/tests/test_openai_models_path_leak.py b/studio/backend/tests/test_openai_models_path_leak.py
new file mode 100644
index 0000000000..a84a33f840
--- /dev/null
+++ b/studio/backend/tests/test_openai_models_path_leak.py
@@ -0,0 +1,45 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""GET /v1/models must report a clean public id, never the on-disk .gguf path."""
+
+import json
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import routes.inference as inf  # noqa: E402
+
+
+class _FakeLlama:
+    is_loaded = True
+    model_identifier = "/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf"
+    context_length = 4096
+    max_context_length = None
+    native_context_length = None
+
+
+class _FakeUnsloth:
+    active_model_name = None
+    models: dict = {}
+    context_length = None
+    max_seq_length = None
+
+
+def test_openai_models_returns_clean_id_without_path(monkeypatch):
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    objs = inf._openai_model_objects()
+
+    assert len(objs) == 1
+    assert objs[0]["id"] == "Qwen3-30B-A3B-Q4_K_M"
+    # The serialized payload must not leak the absolute path or the .gguf suffix.
+    blob = json.dumps(objs)
+    assert "/srv/models" not in blob
+    assert ".gguf" not in blob
+    # Context fields still flow through.
+    assert objs[0]["context_length"] == 4096
diff --git a/studio/backend/tests/test_tp_vision_regression.py b/studio/backend/tests/test_tp_vision_regression.py
new file mode 100644
index 0000000000..09af876da6
--- /dev/null
+++ b/studio/backend/tests/test_tp_vision_regression.py
@@ -0,0 +1,805 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression guards for silent tensor-parallel downgrades in load_model.
+
+PR #6416 blanket-disabled tensor parallelism for vision models to dodge a
+--split-mode tensor + --mmproj GGML_ASSERT (#6415), which silently single-GPU'd
+any mmproj/MTP GGUF that fit on one card. The fix makes the skip self-healing:
+tensor is tried by default and recorded per (binary, model) only on a real abort.
+
+load_model is too entangled to drive end-to-end, so these tests inspect the
+source / drive the pure helpers. The headline test pins the set of TP-drop
+conditions, so a new silent drop fails CI. No GPU; fully deterministic.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib.util
+import inspect
+import os
+import sys
+import textwrap
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# External-dep stubs so importing the backend doesn't require structlog / httpx /
+# loggers -- but only when the real module is missing, so a lightweight stub never
+# shadows the real package (or `loggers.handlers` submodule) for tests collected
+# later in the same pytest process.
+try:
+    import structlog  # noqa: F401
+except ImportError:
+    _structlog_stub = _types.ModuleType("structlog")
+    _structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+    sys.modules["structlog"] = _structlog_stub
+try:
+    import loggers  # noqa: F401
+except ImportError:
+    _loggers_stub = _types.ModuleType("loggers")
+    _loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+    sys.modules["loggers"] = _loggers_stub
+try:
+    import httpx as _httpx_real  # noqa: F401
+except ImportError:
+    _httpx_stub = _types.ModuleType("httpx")
+    for _exc in (
+        "ConnectError",
+        "TimeoutException",
+        "ReadTimeout",
+        "ReadError",
+        "RemoteProtocolError",
+        "CloseError",
+        "HTTPError",
+        "RequestError",
+    ):
+        setattr(_httpx_stub, _exc, type(_exc, (Exception,), {}))
+    _httpx_stub.Timeout = type("T", (), {"__init__": lambda s, *a, **k: None})
+    _httpx_stub.Response = type("Response", (), {})
+    _httpx_stub.Client = type(
+        "C",
+        (),
+        {
+            "__init__": lambda s, **kw: None,
+            "__enter__": lambda s: s,
+            "__exit__": lambda s, *a: None,
+        },
+    )
+    sys.modules["httpx"] = _httpx_stub
+
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+_GB = 1024**3
+
+
+def _load_inference_routes_module():
+    """Load routes/inference.py directly, bypassing routes/__init__.py (which imports
+    every router, dragging in unrelated deps like python-multipart) (Codex #6659)."""
+    route_path = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    spec = importlib.util.spec_from_file_location(
+        "tp_vision_regression_inference_routes", route_path
+    )
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def _load_model_ast() -> ast.FunctionDef:
+    """Parse load_model into an AST FunctionDef (no import side effects)."""
+    src = textwrap.dedent(inspect.getsource(LlamaCppBackend.load_model))
+    return ast.parse(src).body[0]
+
+
+def _tensor_parallel_false_drop_guards() -> list[str]:
+    """Source of the guard expression for every `if ...: tensor_parallel = False`
+    (the LOCAL variable, not self._tensor_parallel) inside load_model."""
+    fn = _load_model_ast()
+
+    def _body_drops_tp(body) -> bool:
+        for n in body:
+            if (
+                isinstance(n, ast.Assign)
+                and any(isinstance(t, ast.Name) and t.id == "tensor_parallel" for t in n.targets)
+                and isinstance(n.value, ast.Constant)
+                and n.value.value is False
+            ):
+                return True
+        return False
+
+    return [
+        ast.unparse(node.test)
+        for node in ast.walk(fn)
+        if isinstance(node, ast.If) and _body_drops_tp(node.body)
+    ]
+
+
+# Every condition that may flip a requested tensor_parallel back to False. Adding
+# one must be conscious: update this allowlist and keep multi-GPU where possible.
+_ALLOWED_TP_DROP_GUARDS = {
+    # Capability: --split-mode tensor aborted for this (binary, model) (#6415).
+    # Self-healing -- tried by default, skipped only after a real abort (vs #6416).
+    "tensor_parallel and self._tensor_split_aborts(binary, model_identifier)",
+    # Capacity: tensor needs >= 2 GPUs clearing the compute-buffer reserve.
+    "tensor_parallel and len(tp_gpus) < 2",
+    # Capacity: pooled usable VRAM can't hold weights + MTP reserve -> layer split.
+    "_tp_weight_budget_mib <= _tp_required_mib",
+}
+
+
+def test_tensor_parallel_drop_sites_match_allowlist():
+    """The set of reasons a requested TP can be dropped is fixed and reviewed: a new
+    drop site fails this set-equality until consciously allowlisted (would catch #6416)."""
+    found = set(_tensor_parallel_false_drop_guards())
+    assert found == _ALLOWED_TP_DROP_GUARDS, (
+        "tensor_parallel drop sites changed.\n"
+        f"  unexpected (new) : {sorted(found - _ALLOWED_TP_DROP_GUARDS)}\n"
+        f"  missing (removed): {sorted(_ALLOWED_TP_DROP_GUARDS - found)}\n"
+        "A new drop means a user's TP request is ignored for a new reason -- "
+        "review it, keep multi-GPU where possible, surface it, then update "
+        "_ALLOWED_TP_DROP_GUARDS."
+    )
+
+
+def test_every_tp_drop_is_logged_not_silent():
+    """Each tensor_parallel downgrade must log why, so it never disappears silently."""
+    fn = _load_model_ast()
+
+    def _body_drops_tp(body):
+        return any(
+            isinstance(n, ast.Assign)
+            and any(isinstance(t, ast.Name) and t.id == "tensor_parallel" for t in n.targets)
+            and isinstance(n.value, ast.Constant)
+            and n.value.value is False
+            for n in body
+        )
+
+    def _body_logs(body) -> bool:
+        for n in ast.walk(ast.Module(body = list(body), type_ignores = [])):
+            if (
+                isinstance(n, ast.Call)
+                and isinstance(n.func, ast.Attribute)
+                and isinstance(n.func.value, ast.Name)
+                and n.func.value.id == "logger"
+            ):
+                return True
+        return False
+
+    for node in ast.walk(fn):
+        if isinstance(node, ast.If) and _body_drops_tp(node.body):
+            assert _body_logs(node.body), (
+                f"TP drop under `{ast.unparse(node.test)}` has no logger call -- "
+                "downgrades must explain themselves."
+            )
+
+
+def test_tensor_split_gate_is_self_healing_not_blanket():
+    """Skip is conditional on a recorded (binary, model) abort, not a blanket
+    is_vision disable (the #6416 regression)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert "self._tensor_split_aborts(binary, model_identifier)" in src
+    assert "if tensor_parallel and is_vision:" not in src
+    assert "if tensor_parallel and effective_is_vision:" not in src
+
+
+def test_tensor_split_skip_documents_layer_split_fallback():
+    """When the skip fires (known-bad binary+model), it states the fallback."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    gate = src.find("self._tensor_split_aborts(binary, model_identifier)")
+    assert gate != -1
+    block = src[gate : gate + 600]
+    assert "layer split" in block, "the skip should state it falls back to layer split"
+
+
+def test_tensor_split_abort_recorded_early_on_first_spawn():
+    """Recorded on the first spawn showing the marker, before the flash-attn-off
+    retry (which can't run tensor so drops the marker) -- else it loops (oobabooga, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    idx = src.find("_record_tensor_split_abort(binary, model_identifier)")
+    assert idx != -1, "load_model must record a (binary, model) tensor-split abort"
+    guard = src[max(0, idx - 600) : idx]
+    assert "self._tensor_parallel" in guard
+    assert (
+        "_should_record_tensor_split_abort" in guard
+    ), "record must be gated on the marker-plus-hard-crash decision helper"
+    # Recorded before the flash-attn-off retry, not after the full ladder.
+    fa_off = src.find("_with_flash_attn_off")
+    assert 0 <= idx < fa_off, "recording must latch on the first spawn, before flash-off"
+
+
+def test_vision_downgrade_preserves_multi_gpu_intent():
+    """The vision downgrade raises _layer_min_gpus and threads it into both the
+    _select_gpus and auto-context layer paths, so a fitting model still spreads."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert "_layer_min_gpus = max(_layer_min_gpus, len(gpus))" in src
+    assert src.count("min_gpus = _layer_min_gpus") >= 2
+    assert "range(_auto_min_gpus, len(ranked) + 1)" in src
+    auto = src.find("_auto_min_gpus = max(")
+    assert auto != -1 and "_layer_min_gpus" in src[auto : auto + 200]
+
+
+# ── per-binary capability cache (pure) ───────────────────────────────
+
+
+def test_tensor_attempted_by_default_for_unknown_binary():
+    """A (binary, model) not seen to abort -> tensor is attempted (not skipped)."""
+    assert LlamaCppBackend._tensor_split_aborts("/never/seen/llama-server", "m") is False
+    assert LlamaCppBackend._tensor_split_aborts(None, "m") is False
+    assert LlamaCppBackend._tensor_split_aborts("/x", None) is False
+
+
+def test_recorded_tensor_abort_is_per_model():
+    """A recorded (binary, model) abort trips the gate for that model only -- a
+    different model on the same binary still attempts tensor (oobabooga, #6659)."""
+    b = f"/tmp/llama-server-{id(object())}"
+    try:
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-a") is False
+        LlamaCppBackend._record_tensor_split_abort(b, "model-a")
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-a") is True
+        # a different model on the same binary is unaffected
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-b") is False
+    finally:
+        LlamaCppBackend._tensor_split_abort_keys.discard(
+            LlamaCppBackend._tensor_split_cache_key(b, "model-a")
+        )
+
+
+# ── _select_gpus: single-GPU collapse vs honored multi-GPU intent (pure) ──
+
+
+def test_select_gpus_collapses_to_single_gpu_when_model_fits():
+    """Default (min_gpus=1): a 39 GB model on four 183 GB GPUs pins ONE GPU -- the
+    'single GPU' symptom once TP drops, and why the downgrade needs min_gpus."""
+    gpus = [(0, 180000), (1, 180000), (2, 180000), (3, 180000)]  # (idx, free MiB)
+    gpu_indices, _use_fit = LlamaCppBackend._select_gpus(int(39 * _GB), gpus)
+    assert gpu_indices is not None and len(gpu_indices) == 1
+
+
+def test_select_gpus_min_gpus_keeps_multi_gpu_for_fitting_model():
+    """min_gpus>=2 must NOT collapse to one GPU for a model that fits on one."""
+    gpus = [(0, 180000), (1, 180000), (2, 180000), (3, 180000)]
+    gpu_indices, _ = LlamaCppBackend._select_gpus(int(39 * _GB), gpus, min_gpus = 2)
+    assert gpu_indices is not None and len(gpu_indices) >= 2
+
+
+def test_select_gpus_min_gpus_capped_to_available():
+    """min_gpus larger than the GPU count is capped, not an error."""
+    gpus = [(0, 180000), (1, 180000)]
+    gi, _ = LlamaCppBackend._select_gpus(int(10 * _GB), gpus, min_gpus = 8)
+    assert gi is not None and len(gi) == 2
+
+
+def test_select_gpus_uses_multiple_gpus_when_model_does_not_fit():
+    """Sanity: selection spreads across GPUs when one card can't hold the model."""
+    gpus = [(0, 40000), (1, 40000), (2, 40000), (3, 40000)]  # 40 GB free each
+    gpu_indices, _use_fit = LlamaCppBackend._select_gpus(int(120 * _GB), gpus)
+    assert gpu_indices is not None and len(gpu_indices) >= 2
+
+
+def test_select_gpus_min_gpus_excludes_unusable_gpu():
+    """min_gpus caps to usable cards: 2 free + 1 nearly-full -> 2-GPU split, not
+    forcing the full card (OOM) or tripping --fit (#6659)."""
+    gpus = [(0, 180000), (1, 180000), (2, 500)]  # GPU 2 is nearly full
+    total = {0: 180000, 1: 180000, 2: 180000}
+    gi, _ = LlamaCppBackend._select_gpus(
+        int(39 * _GB),
+        gpus,
+        min_gpus = 3,
+        total_by_idx = total,
+        per_device_overhead_bytes = int(1 * _GB),
+    )
+    assert gi is not None
+    assert 2 not in gi, "a nearly-full GPU must not be forced in to satisfy min_gpus"
+    assert len(gi) == 2
+
+
+def test_tensor_abort_cache_invalidated_on_binary_mtime_change(tmp_path):
+    """Cache keys on (path, mtime, model), so a binary swapped in place (in-app
+    update, no restart) is re-probed instead of inheriting the old abort (#6659)."""
+    binp = tmp_path / "llama-server"
+    binp.write_text("v1")
+    p = str(binp)
+    try:
+        LlamaCppBackend._record_tensor_split_abort(p, "m")
+        assert LlamaCppBackend._tensor_split_aborts(p, "m") is True
+        # Simulate an in-place update bumping the binary's mtime.
+        st = binp.stat()
+        os.utime(p, (st.st_atime, st.st_mtime + 10))
+        assert (
+            LlamaCppBackend._tensor_split_aborts(p, "m") is False
+        ), "a binary swapped in place (new mtime) must be re-probed"
+        # A same-second replacement (sub-second mtime bump) must also re-probe:
+        # second-resolution mtime would inherit the stale abort (reviewer.py P2).
+        sec_ns = (binp.stat().st_mtime_ns // 1_000_000_000) * 1_000_000_000
+        os.utime(p, ns = (sec_ns, sec_ns))
+        LlamaCppBackend._record_tensor_split_abort(p, "m")
+        binp.write_text("v2")
+        os.utime(p, ns = (sec_ns, sec_ns + 1))
+        assert (
+            LlamaCppBackend._tensor_split_aborts(p, "m") is False
+        ), "a same-second in-place swap (ns mtime bump) must be re-probed"
+    finally:
+        for key in list(LlamaCppBackend._tensor_split_abort_keys):
+            if key and key[0] == p:
+                LlamaCppBackend._tensor_split_abort_keys.discard(key)
+
+
+def test_tensor_split_abort_raises_early_to_layer_fallback():
+    """The first-spawn abort raises to the route's layer fallback (not the text-only
+    mmproj strip), before the flash-attn-off retry, preserving the projector (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    raise_idx = src.find("(split-axis geometry); retrying with layer split")
+    assert raise_idx != -1, "the split-axis abort must raise to trigger a layer retry"
+    # raises before both the flash-attn-off retry and the text-only mmproj strip
+    assert raise_idx < src.find("_with_flash_attn_off")
+    assert raise_idx < src.find("_strip_mmproj_args(_last_spawn_cmd)")
+    # gated on the marker-plus-crash helper, which also drives the record just above
+    guard = src[max(0, raise_idx - 600) : raise_idx]
+    assert "_should_record_tensor_split_abort" in guard
+    rec_idx = src.find("_record_tensor_split_abort(binary, model_identifier)")
+    assert rec_idx != -1 and rec_idx < raise_idx
+
+
+def test_budget_downgrade_preserves_multi_gpu_intent():
+    """The pooled-VRAM downgrade raises _layer_min_gpus from the usable tensor GPUs
+    too, symmetric with the vision downgrade (reviewer.py asymmetric fix, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    budget = src.find("_tp_weight_budget_mib <= _tp_required_mib")
+    assert budget != -1
+    block = src[budget : budget + 1000]
+    assert "tensor_parallel = False" in block
+    assert (
+        "_layer_min_gpus = max(_layer_min_gpus, len(tp_gpus))" in block
+    ), "the budget downgrade must preserve multi-GPU intent like the vision gate"
+
+
+def test_compute_buffer_downgrade_preserves_multi_gpu_intent():
+    """The len(tp_gpus) < 2 compute-buffer downgrade raises _layer_min_gpus from the
+    full GPU set too, so it is symmetric with the budget/geometry downgrades and
+    doesn't collapse a multi-GPU layer load to one card (reviewer.py P1 on #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    gate = src.find("tensor_parallel and len(tp_gpus) < 2")
+    assert gate != -1
+    # Bound to exactly this block: from its gate to the next (budget) downgrade.
+    nxt = src.find("_tp_weight_budget_mib <= _tp_required_mib", gate)
+    assert nxt != -1
+    block = src[gate:nxt]
+    assert "tensor_parallel = False" in block
+    assert (
+        "_layer_min_gpus = max(_layer_min_gpus, len(gpus))" in block
+    ), "the compute-buffer downgrade must preserve multi-GPU intent like the others"
+
+
+def test_tensor_split_layer_min_gpus_bump_requires_tensor_request():
+    """Every guard that bumps _layer_min_gpus off the abort cache also tests
+    tensor_parallel, so a non-tensor load on a known-bad binary doesn't grab every
+    GPU for a fitting model (#6659)."""
+    fn = _load_model_ast()
+    checked = 0
+    for node in ast.walk(fn):
+        if isinstance(node, ast.If):
+            test_src = ast.unparse(node.test)
+            if "self._tensor_split_aborts(binary, model_identifier)" not in test_src:
+                continue
+            body = "\n".join(ast.unparse(n) for n in node.body)
+            if "_layer_min_gpus" in body:
+                checked += 1
+                assert "tensor_parallel" in test_src, (
+                    "the cached _layer_min_gpus bump must require a current tensor "
+                    f"request, but fires under `{test_src}`"
+                )
+    assert checked >= 1, "expected an abort-cache guard that bumps _layer_min_gpus"
+
+
+# ── round-2 follow-up: route-fallback retry + auto-context cap + assert marker ──
+
+
+def test_layer_fallback_retry_preserves_multi_gpu_intent():
+    """load_model takes a preserve_multi_gpu_on_layer hint and raises _layer_min_gpus
+    for it, so the tensor-off fallback retry still spreads a fitting model (#6659)."""
+    sig = inspect.signature(LlamaCppBackend.load_model)
+    assert "preserve_multi_gpu_on_layer" in sig.parameters
+    assert sig.parameters["preserve_multi_gpu_on_layer"].default is False
+    fn = _load_model_ast()
+    found = any(
+        isinstance(n, ast.If)
+        and "preserve_multi_gpu_on_layer" in ast.unparse(n.test)
+        and "_layer_min_gpus" in "\n".join(ast.unparse(b) for b in n.body)
+        for n in ast.walk(fn)
+    )
+    assert found, "preserve_multi_gpu_on_layer must raise _layer_min_gpus"
+
+
+def test_auto_context_layer_loops_capped_to_usable_gpus():
+    """The auto-context loops bypass _select_gpus, so they apply its cap: a card
+    counts only if usable VRAM clears the per-device layer overhead (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert (
+        "range(max(1, _layer_min_gpus), len(ranked) + 1)" not in src
+    ), "auto-context loops must cap _layer_min_gpus to usable GPUs, not use it raw"
+    assert "_auto_min_gpus" in src
+    assert "range(_auto_min_gpus, len(ranked) + 1)" in src
+    # the eligibility threshold is the per-device layer overhead, not bare > 0
+    auto = src.find("_auto_min_gpus = max(")
+    assert auto != -1
+    block = src[auto : auto + 400]
+    assert "_pipeline_overhead_mib" in block, (
+        "a card must clear the per-device layer overhead to count, mirroring "
+        "_select_gpus, so a nearly-full GPU is not exposed and OOMs"
+    )
+
+
+def test_fallback_hint_uses_effective_tensor_request_not_just_toggle():
+    """Tensor intent keys off _effective_tensor_parallel (toggle + extras + env), not
+    just the toggle, so extra/env-driven tensor users keep multi-GPU (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_tensor_intent_overall = _effective_tensor_parallel(")
+    assert idx != -1, "the GGUF load closure must compute tensor intent"
+    block = src[idx : idx + 300]
+    assert "extra_llama_args, request.tensor_parallel" in block
+    pres = src.find("preserve_multi_gpu_on_layer = bool(")
+    assert (
+        "_effective_tensor_parallel(attempt_extra_args, tensor_parallel)" in src[pres : pres + 200]
+    )
+    # not the toggle-only form this replaced
+    assert (
+        "bool(\n                        request.tensor_parallel and not tensor_parallel" not in src
+    )
+
+
+def test_carry_preserved_tensor_intent_truth_table():
+    """Behavioral check of the carry-forward decision: carried only for the SAME
+    model, preserved, and not an explicit drop. Catches a `not` inversion (ctx-only
+    collapse) and a missing same-model guard (cross-model leak) (#6659)."""
+    inference_routes = _load_inference_routes_module()
+    f = inference_routes._carry_preserved_tensor_intent
+    assert f(preserved = True, same_model = True, explicit_drop = False) is True
+    assert f(preserved = True, same_model = True, explicit_drop = True) is False  # explicit drop
+    assert f(preserved = True, same_model = False, explicit_drop = False) is False  # model switch
+    assert f(preserved = False, same_model = True, explicit_drop = False) is False  # not a fallback
+
+
+def test_preserved_fallback_carried_across_non_drop_reload():
+    """The hint carries the preserved fallback via _carry_preserved_tensor_intent,
+    gated on the same model loaded, so a ctx-only reload keeps multi-GPU but a model
+    switch / explicit drop doesn't inherit it (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_tensor_intent_overall = _effective_tensor_parallel(")
+    assert idx != -1
+    block = src[idx : idx + 400]
+    assert "_carry_preserved_tensor_intent(" in block
+    assert "preserved = llama_backend.layer_preserves_tensor_intent" in block
+    assert "same_model = _same_model_loaded" in block
+    assert "explicit_drop = _explicit_tensor_drop" in block
+
+
+def test_same_model_guard_checks_path_and_variant():
+    """The same-model guard matches the resolved config.identifier (what load_model
+    stores, after from_identifier normalizes shorthands) -- not the raw request id --
+    and also matches the loaded quant by path (local multi-variant dir) else variant (HF
+    repo), so a reload keeps the carry-forward and a different variant doesn't inherit
+    the prior one's preserved tensor intent (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_same_model_loaded = (")
+    assert idx != -1
+    block = src[idx : idx + 1300]
+    # Identity compares the normalized config.identifier, not the raw model_identifier.
+    head = src[idx : idx + 200]
+    assert "config.identifier" in head and "== (model_identifier" not in head
+    assert "llama_backend.gguf_path" in block and "config.gguf_file" in block
+    assert "llama_backend.hf_variant" in block and "config.gguf_variant" in block
+
+
+def test_diffusion_load_clears_preserved_tensor_flag():
+    """The diffusion early-return path (skips the command builder) clears the
+    preserved-fallback flag, so a prior tensor fallback doesn't churn it (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    diff = src.find("if self._is_diffusion:")
+    assert diff != -1
+    start = src.find("return self._start_diffusion_server", diff)
+    assert start != -1
+    assert "self._layer_preserves_tensor_intent = False" in src[diff:start]
+
+
+def test_is_tensor_split_assert_marker():
+    """Matches the specific #6415 split-axis assert, not any ggml assert/abort, so
+    an unrelated invariant a corrupt GGUF/projector trips isn't cached (#6659)."""
+    f = LlamaCppBackend._is_tensor_split_assert
+    # the real #6415 warmup assert (split-axis enum, in ggml-backend-meta)
+    assert (
+        f(
+            "ggml-backend-meta.cpp:541: GGML_ASSERT(src_ss[0].axis != "
+            "GGML_BACKEND_SPLIT_AXIS_0) failed"
+        )
+        is True
+    )
+    # the split-axis token alone (file path elided / reworded) still matches
+    assert f("GGML_ASSERT(x.axis != GGML_BACKEND_SPLIT_AXIS_1) failed") is True
+    # UNRELATED asserts must NOT match -- including a different invariant from the
+    # same multi-assert source file (matched on the token, not the file name).
+    assert f("ggml-backend-meta.cpp:99: GGML_ASSERT(buf != NULL) failed") is False
+    assert f("/x/ggml.c:1234: GGML_ASSERT(ne == 1) failed") is False
+    assert f("ggml_abort: something else entirely") is False
+    assert f("Segmentation fault (core dumped)") is False
+    assert f("") is False
+    assert f(None) is False
+
+
+def test_layer_preserve_hint_replayed_on_respawn():
+    """The preserve hint is in the replay snapshot (_pending_load_kwargs), so a
+    respawn keeps the downgraded model multi-GPU (Codex review on #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    pend = src.find("_pending_load_kwargs = {")
+    assert pend != -1
+    block = src[pend : src.find("}", pend) + 1]
+    assert '"preserve_multi_gpu_on_layer": preserve_multi_gpu_on_layer' in block, (
+        "the layer-preserve hint must be in the replay snapshot so _respawn_if_dead "
+        "keeps the multi-GPU placement"
+    )
+
+
+def test_should_record_tensor_split_abort_decision():
+    """Behavioral check of marker AND (signal crash OR Windows abort), so an
+    or->and typo or caching a generic crash fails here, not just the source pins."""
+    f = LlamaCppBackend._should_record_tensor_split_abort
+    marker = "ggml-backend-meta.cpp:541: GGML_ASSERT(x.axis != GGML_BACKEND_SPLIT_AXIS_0) failed"
+    # marker + a hard crash records, across every platform's abort encoding
+    assert f(-6, marker) is True  # POSIX SIGABRT
+    assert f(-11, marker) is True  # POSIX SIGSEGV
+    assert f(3, marker) is True  # Windows CRT abort() exit (not a signal)
+    assert f(0xC0000005, marker) is True  # Windows NTSTATUS access violation
+    # marker present but no hard crash -> not recorded
+    assert f(0, marker) is False  # clean exit
+    assert f(-9, marker) is False  # SIGKILL (OOM / unload), not a fault
+    assert f(None, marker) is False  # still running
+    # hard crash but not the split-axis marker -> not recorded (no over-caching)
+    assert f(3, "some other failure") is False
+    assert f(-6, "GGML_ASSERT(buf != NULL) failed") is False
+    assert f(0xC0000005, "") is False
+
+
+def test_fit_off_retry_skipped_on_split_axis_abort():
+    """The fit-independent --fit off retry is skipped on the split-axis marker, else
+    the model crashes a second time before the latch records it (reviewer.py, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    retry = src.find('run_cmd = [*run_cmd, "--fit", "off"]')
+    assert retry != -1
+    guard = src[max(0, retry - 1000) : retry]
+    assert "_fit_retry_allowed" in guard and "_startup_crashed" in guard
+    assert (
+        "not _split_axis_crash" in guard
+    ), "the fit-off retry must be skipped when the crash is a split-axis abort"
+
+
+def test_is_abort_exit_recognizes_windows_crt_abort():
+    """exit code 3 (MSVC abort()) counts as a crash; signals / clean exits do not."""
+    f = LlamaCppBackend._is_abort_exit
+    assert f(3) is True
+    assert f(0) is False
+    assert f(-6) is False  # POSIX SIGABRT is handled by _is_signal_crash, not here
+    assert f(None) is False
+
+
+# ── tensor-off after a multi-GPU fallback forces a reload (route dedup) ─
+
+
+class _NoopProcess:
+    """Stand-in for Popen so is_loaded is True and atexit cleanup doesn't crash."""
+
+    def terminate(self):
+        pass
+
+    def wait(self, timeout = None):
+        return 0
+
+    def kill(self):
+        pass
+
+    def poll(self):
+        return 0
+
+
+def _fallback_loaded_backend(layer_preserves_tensor_intent: bool) -> LlamaCppBackend:
+    """A loaded backend in the tensor->layer fallback state (tensor off, --split-mode
+    layer stored), differing only in the preserved-multi-GPU flag."""
+    b = LlamaCppBackend()
+    b._model_identifier = "owner/repo"
+    b._requested_n_ctx = 0
+    b._cache_type_kv = None
+    b._tensor_parallel = False
+    b._layer_preserves_tensor_intent = layer_preserves_tensor_intent
+    b._extra_args = ["--split-mode", "layer"]
+    b._requested_spec_mode = "auto"
+    b._chat_template_override = None
+    b._gguf_path = None
+    return b
+
+
+def test_tensor_off_echo_preserves_multi_gpu_fallback():
+    """The Studio UI always sends tensor_parallel and echoes the /load response's
+    resolved value, so after a fallback a ctx/settings reload carries tensor_parallel=
+    false even though the user never changed it. That echo must NOT collapse the
+    preserved multi-GPU placement -- it dedupes (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo", tensor_parallel = False)
+    assert "tensor_parallel" in req.model_fields_set, "the UI always sends the field"
+
+    # Preserved fallback + bare tensor=false echo: dedupe, keep multi-GPU (no collapse).
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+    # A genuine layer load (no preserved intent): tensor-off also dedupes, no churn.
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = False)
+        )
+        is True
+    )
+
+
+def test_explicit_split_mode_layer_extras_reloads_after_multi_gpu_fallback():
+    """Tensor intent can be dropped via extras too: an explicit --split-mode layer
+    matches the stored fallback extras but must still reload (reviewer.py P1, #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "layer"])
+    assert "llama_extra_args" in req.model_fields_set
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is False
+    )
+
+
+def test_tensor_off_reload_requires_explicit_toggle():
+    """An Apply that doesn't touch the toggle (e.g. a context change) isn't churned
+    by the preserved-fallback reload -- the working server is kept (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo")  # tensor_parallel left unset
+    assert "tensor_parallel" not in req.model_fields_set
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+
+
+def test_tensor_off_under_env_tensor_does_not_reload_loop(monkeypatch):
+    """With LLAMA_ARG_SPLIT_MODE=tensor set, a tensor-off request can't drop tensor
+    intent, so the env-aware guard dedupes instead of reload-looping (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+    monkeypatch.setenv("LLAMA_ARG_SPLIT_MODE", "tensor")
+
+    req = LoadRequest(model_path = "owner/repo", tensor_parallel = False)
+    assert "tensor_parallel" in req.model_fields_set
+    # env still forces tensor -> not a real drop -> dedupe (no reload loop).
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+
+
+def test_is_explicit_tensor_drop_truth_table():
+    """Only an explicit non-tensor --split-mode override is a drop. A bare
+    tensor_parallel field (the UI always sends it and echoes the fallback's false), an
+    empty clear, an unrelated extra (--top-k), or inherit (None) must NOT collapse a
+    preserved fallback; --split-mode tensor / tensor_parallel=true re-engage (Codex
+    #6659)."""
+    from models.inference import LoadRequest
+
+    f = _load_inference_routes_module()._is_explicit_tensor_drop
+    # A non-tensor split-mode override is the one deliberate departure -> drop.
+    assert (
+        f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "layer"])) is True
+    )
+    # tensor / retry re-engages, never a drop.
+    assert (
+        f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "tensor"]))
+        is False
+    )
+    # A bare tensor_parallel field is the UI echo, not a drop (would collapse on reload).
+    assert f(LoadRequest(model_path = "owner/repo", tensor_parallel = False)) is False
+    assert f(LoadRequest(model_path = "owner/repo", tensor_parallel = True)) is False
+    # Unrelated extra / empty clear / inherit all keep the preserved placement.
+    assert f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--top-k", "20"])) is False
+    assert f(LoadRequest(model_path = "owner/repo", llama_extra_args = [])) is False
+    assert f(LoadRequest(model_path = "owner/repo")) is False
+
+
+def test_explicit_tensor_drop_uses_shared_helper_in_both_readers():
+    """Both the already-loaded dedup and the load carry-forward derive the drop from
+    _is_explicit_tensor_drop, so they agree on what counts as a drop -- a reload for
+    an unrelated extra still carries the preserved intent rather than collapsing to one
+    GPU (Codex #6659)."""
+    src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text()
+    # Dedup reader (the preserved-fallback reload guard).
+    assert "layer_preserves_tensor_intent and _is_explicit_tensor_drop(request)" in src
+    # Load carry-forward reader feeds the same decision into the carry-forward.
+    assert "_explicit_tensor_drop = _is_explicit_tensor_drop(request)" in src
+
+
+def test_layer_preserves_tensor_intent_set_only_on_preserved_downgrade():
+    """load_model latches the flag from _layer_min_gpus (raised only when a tensor
+    request is downgraded but kept multi-GPU), and clears it when tensor stays on."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    on = src.find("self._tensor_parallel = True")
+    off = src.find("self._tensor_parallel = False")
+    assert 0 <= on and 0 <= off
+    assert "self._layer_preserves_tensor_intent = False" in src[on : on + 120]
+    assert "self._layer_preserves_tensor_intent = _layer_min_gpus > 1" in src[off : off + 400]
+
+
+def test_layer_min_gpus_bound_before_gpu_selection_try():
+    """_layer_min_gpus is bound before the GPU-selection try, so the --fit-on except
+    path can't UnboundLocalError when the command builder reads it (Codex #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert src.count("_layer_min_gpus = 1") == 1, "exactly one init, before the try"
+    init = src.find("_layer_min_gpus = 1")
+    try_body = src.find("gguf_size = self._get_gguf_size_bytes")
+    fit_except = src.find("GPU selection failed")
+    use_after = src.find("self._layer_preserves_tensor_intent = _layer_min_gpus > 1")
+    assert (
+        -1 < init < try_body < fit_except < use_after
+    ), "the init must precede the try body, the except, and the command-builder use"
+
+
+def test_already_in_target_state_reloads_on_tensor_off_after_fallback():
+    """The backend fast path mirrors the route dedup: a preserved fallback reloads on
+    an EXPLICIT tensor-off request, but an implicit same-settings reload (carry-forward
+    preserve_multi_gpu_on_layer=True) still dedupes (Codex #6659)."""
+
+    def _backend(layer_preserves: bool) -> LlamaCppBackend:
+        b = _fallback_loaded_backend(layer_preserves_tensor_intent = layer_preserves)
+        b._process = _NoopProcess()
+        b._healthy = True
+        return b
+
+    kwargs = dict(
+        gguf_path = None,
+        mtp_draft_path = None,
+        model_identifier = "owner/repo",
+        hf_variant = None,
+        n_ctx = 0,
+        cache_type_kv = None,
+        speculative_type = None,
+        spec_draft_n_max = None,
+        tensor_parallel = False,
+        chat_template_override = None,
+        extra_args = ["--split-mode", "layer"],
+        is_vision = False,
+    )
+    # Preserved fallback + EXPLICIT tensor drop -> reload (not already in target state).
+    assert _backend(True)._already_in_target_state(**kwargs) is False
+    # Same preserved fallback but an implicit reload that carries the intent forward
+    # (HF auto-pick / local-dir flows skip the route guard and reach here) -> dedupe.
+    assert (
+        _backend(True)._already_in_target_state(**kwargs, preserve_multi_gpu_on_layer = True) is True
+    )
+    # A genuine layer load (no preserved intent) -> dedupe, no churn.
+    assert _backend(False)._already_in_target_state(**kwargs) is True
diff --git a/studio/backend/tests/test_training_runs.py b/studio/backend/tests/test_training_runs.py
new file mode 100644
index 0000000000..fd0d6d380f
--- /dev/null
+++ b/studio/backend/tests/test_training_runs.py
@@ -0,0 +1,103 @@
+# 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 json
+
+from storage.studio_db import _extract_project_name_from_config_json
+from utils.training_runs import (
+    build_default_output_dir_name,
+    model_segment_from_default_output_dir_name,
+    normalize_project_name,
+    slugify_project_name,
+)
+
+
+def test_normalize_project_name_trims_and_collapses_whitespace():
+    assert normalize_project_name("  Customer   Support   LoRA  ") == "Customer Support LoRA"
+
+
+def test_normalize_project_name_returns_none_for_empty_or_invalid_values():
+    assert normalize_project_name("   ") is None
+    assert normalize_project_name(None) is None
+
+
+def test_slugify_project_name_makes_safe_suffix():
+    assert slugify_project_name("Customer Support / LoRA v2") == "customer-support-lora-v2"
+
+
+def test_slugify_project_name_rejects_path_only_or_separator_only_values():
+    assert slugify_project_name("..") is None
+    assert slugify_project_name("///") is None
+
+
+def test_build_default_output_dir_name_appends_project_slug():
+    output_dir = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "unsloth_Llama-3.2-3B-Instruct__project-customer-support_1771227800"
+
+
+def test_build_default_output_dir_name_caps_final_component(tmp_path):
+    output_dir = build_default_output_dir_name(
+        "a" * 240,
+        "b" * 80,
+        timestamp = 1771227800,
+    )
+
+    assert len(output_dir.encode()) <= 255
+    (tmp_path / output_dir).mkdir()
+
+
+def test_build_default_output_dir_name_skips_invalid_project_slug():
+    output_dir = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "..",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "unsloth_Llama-3.2-3B-Instruct_1771227800"
+
+
+def test_model_segment_from_default_output_dir_name_strips_project_slug():
+    assert (
+        model_segment_from_default_output_dir_name(
+            "unsloth_Llama-3.2-3B-Instruct__project-customer-support_1771227800"
+        )
+        == "unsloth_Llama-3.2-3B-Instruct"
+    )
+
+
+def test_model_segment_preserves_project_marker_text_in_model_name():
+    output_dir = build_default_output_dir_name(
+        "org/foo__project-bar",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "org_foo__project--bar_1771227800"
+    assert model_segment_from_default_output_dir_name(output_dir) == "org_foo__project-bar"
+
+
+def test_model_segment_strips_project_slug_after_escaped_model_marker():
+    output_dir = build_default_output_dir_name(
+        "org/foo__project-bar",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "org_foo__project--bar__project-customer-support_1771227800"
+    assert model_segment_from_default_output_dir_name(output_dir) == "org_foo__project-bar"
+
+
+def test_extract_project_name_from_config_json_returns_normalized_name():
+    config_json = json.dumps({"project_name": "  Sales   Assistant  "})
+
+    assert _extract_project_name_from_config_json(config_json) == "Sales Assistant"
+
+
+def test_extract_project_name_from_config_json_handles_missing_or_invalid_payload():
+    assert _extract_project_name_from_config_json(None) is None
+    assert _extract_project_name_from_config_json("not-json") is None
+    assert _extract_project_name_from_config_json(json.dumps({"project_name": "   "})) is None
diff --git a/studio/backend/tests/test_training_streaming.py b/studio/backend/tests/test_training_streaming.py
index 8ff016d3bf..70b2d6fdcc 100644
--- a/studio/backend/tests/test_training_streaming.py
+++ b/studio/backend/tests/test_training_streaming.py
@@ -195,6 +195,16 @@ def test_hf_dataset_rejects_unsafe_values(bad_hf_dataset):
         )
 
 
+def test_project_name_rejects_values_over_ui_limit():
+    with pytest.raises(ValidationError):
+        TrainingStartRequest(
+            model_name = "unsloth/test",
+            project_name = "x" * 81,
+            training_type = "LoRA/QLoRA",
+            format_type = "alpaca",
+        )
+
+
 # --- Start-route streaming compatibility guards ---
 
 
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index d174f6677b..90e26d45d0 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -9,6 +9,12 @@ import structlog
 from loggers import get_logger
 from pathlib import Path
 from typing import List, Optional, Tuple
+from storage.studio_db import get_connection
+from utils.training_runs import (
+    build_default_output_dir_name,
+    extract_project_name,
+    model_segment_from_default_output_dir_name,
+)
 from utils.paths import outputs_root, resolve_output_dir
 
 logger = get_logger(__name__)
@@ -30,6 +36,93 @@ def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]:
     return (1, 0, str(checkpoint_path))
 
 
+def _infer_base_model_from_history(checkpoint_dir: Path) -> Optional[str]:
+    """Best-effort base-model lookup using persisted Studio run metadata."""
+    checkpoint_name = checkpoint_dir.name
+    resolved_checkpoint_dir = str(checkpoint_dir.resolve())
+
+    try:
+        conn = get_connection()
+    except Exception:
+        return None
+
+    try:
+        exact_rows = conn.execute(
+            """
+            SELECT model_name
+            FROM training_runs
+            WHERE output_dir IN (?, ?)
+            ORDER BY started_at DESC
+            """,
+            (
+                resolved_checkpoint_dir,
+                str(checkpoint_dir),
+            ),
+        ).fetchall()
+        for row in exact_rows:
+            model_name = row["model_name"]
+            if model_name:
+                return model_name
+
+        suffix_rows = conn.execute(
+            """
+            SELECT model_name, output_dir
+            FROM training_runs
+            WHERE output_dir IS NOT NULL
+            ORDER BY started_at DESC
+            """
+        ).fetchall()
+        for row in suffix_rows:
+            output_dir = str(row["output_dir"] or "").rstrip("/\\")
+            if not (
+                output_dir.endswith(f"/{checkpoint_name}")
+                or output_dir.endswith(f"\\{checkpoint_name}")
+            ):
+                continue
+            model_name = row["model_name"]
+            if model_name:
+                return model_name
+
+        parts = checkpoint_name.rsplit("_", 1)
+        if len(parts) != 2 or not parts[1].isdigit():
+            return None
+
+        timestamp = int(parts[1])
+        generated_rows = conn.execute(
+            """
+            SELECT model_name, config_json
+            FROM training_runs
+            ORDER BY started_at DESC
+            """
+        ).fetchall()
+        for row in generated_rows:
+            model_name = row["model_name"]
+            if not model_name:
+                continue
+
+            project_name = None
+            config_json = row["config_json"]
+            if config_json:
+                try:
+                    project_name = extract_project_name(json.loads(config_json))
+                except (TypeError, json.JSONDecodeError):
+                    project_name = None
+
+            expected_dir_name = build_default_output_dir_name(
+                model_name,
+                project_name,
+                timestamp = timestamp,
+            )
+            if expected_dir_name == checkpoint_name:
+                return model_name
+    except Exception:
+        return None
+    finally:
+        conn.close()
+
+    return None
+
+
 def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]:
     """Read loss from the last log_history entry of trainer_state.json, or None."""
     trainer_state = checkpoint_path / "trainer_state.json"
@@ -106,9 +199,11 @@ def scan_checkpoints(
             # Fallback: extract base model name from the folder name, e.g.
             # "unsloth_Llama-3.2-3B-Instruct_1771227800" → "unsloth/Llama-3.2-3B-Instruct"
             if not metadata.get("base_model"):
-                parts = item.name.rsplit("_", 1)
-                if len(parts) == 2 and parts[1].isdigit():
-                    name_part = parts[0]
+                metadata["base_model"] = _infer_base_model_from_history(item)
+
+            if not metadata.get("base_model"):
+                name_part = model_segment_from_default_output_dir_name(item.name)
+                if name_part:
                     idx = name_part.find("_")
                     if idx > 0:
                         metadata["base_model"] = name_part[:idx] + "/" + name_part[idx + 1 :]
diff --git a/studio/backend/utils/training_runs.py b/studio/backend/utils/training_runs.py
new file mode 100644
index 0000000000..dc2535e570
--- /dev/null
+++ b/studio/backend/utils/training_runs.py
@@ -0,0 +1,104 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Helpers for naming and describing Studio training runs."""
+
+from __future__ import annotations
+
+import re
+import time
+from typing import Any, Optional
+
+_INVALID_SEGMENT_CHARS = re.compile(r"[^A-Za-z0-9._-]+")
+_MAX_RUN_DIR_NAME_CHARS = 255
+_PROJECT_MARKER = "__project-"
+_PROJECT_MARKER_ESCAPE = f"{_PROJECT_MARKER}-"
+
+
+def _trim_segment(segment: str, max_chars: int) -> str:
+    if max_chars <= 0:
+        return ""
+    return segment[:max_chars].strip("._-")
+
+
+def _escape_project_marker(segment: str) -> str:
+    return segment.replace(_PROJECT_MARKER, _PROJECT_MARKER_ESCAPE)
+
+
+def _unescape_project_marker(segment: str) -> str:
+    return segment.replace(_PROJECT_MARKER_ESCAPE, _PROJECT_MARKER)
+
+
+def _appended_project_marker_index(segment: str) -> int:
+    marker_index = segment.rfind(_PROJECT_MARKER)
+    while marker_index >= 0 and segment.startswith(_PROJECT_MARKER_ESCAPE, marker_index):
+        marker_index = segment.rfind(_PROJECT_MARKER, 0, marker_index)
+    return marker_index
+
+
+def normalize_project_name(project_name: Any) -> Optional[str]:
+    """Return a trimmed project name, or None when empty/invalid."""
+    if not isinstance(project_name, str):
+        return None
+    normalized = " ".join(project_name.strip().split())
+    return normalized or None
+
+
+def slugify_project_name(project_name: Any) -> Optional[str]:
+    """Convert a project name into a filesystem-safe suffix."""
+    normalized = normalize_project_name(project_name)
+    if normalized is None:
+        return None
+
+    slug = _INVALID_SEGMENT_CHARS.sub("-", normalized).strip("-._")
+    if not slug:
+        return None
+    return slug.lower()
+
+
+def build_default_output_dir_name(
+    model_name: str,
+    project_name: Any = None,
+    *,
+    timestamp: Optional[int] = None,
+) -> str:
+    """Build the default training output folder name."""
+    from utils.paths import default_run_dir_name
+
+    timestamp_part = str(int(time.time() if timestamp is None else timestamp))
+    timestamp_suffix = f"_{timestamp_part}"
+    model_segment = _escape_project_marker(default_run_dir_name(model_name))
+    project_slug = slugify_project_name(project_name)
+    if not project_slug:
+        max_model_chars = _MAX_RUN_DIR_NAME_CHARS - len(timestamp_suffix)
+        model_segment = _trim_segment(model_segment, max_model_chars) or "model"
+        return f"{model_segment}{timestamp_suffix}"
+
+    max_project_chars = (
+        _MAX_RUN_DIR_NAME_CHARS - len("model") - len(_PROJECT_MARKER) - len(timestamp_suffix)
+    )
+    project_slug = _trim_segment(project_slug, max_project_chars) or "project"
+    project_suffix = f"{_PROJECT_MARKER}{project_slug}{timestamp_suffix}"
+    max_model_chars = _MAX_RUN_DIR_NAME_CHARS - len(project_suffix)
+    model_segment = _trim_segment(model_segment, max_model_chars) or "model"
+    return f"{model_segment}{project_suffix}"
+
+
+def model_segment_from_default_output_dir_name(output_dir_name: str) -> Optional[str]:
+    """Return the encoded model segment from a default run folder name."""
+    parts = str(output_dir_name or "").rsplit("_", 1)
+    if len(parts) != 2 or not parts[1].isdigit():
+        return None
+    model_segment = parts[0]
+    marker_index = _appended_project_marker_index(model_segment)
+    if marker_index >= 0:
+        model_segment = model_segment[:marker_index]
+    model_segment = _unescape_project_marker(model_segment)
+    return model_segment or None
+
+
+def extract_project_name(config: Any) -> Optional[str]:
+    """Read and normalize a project name from a stored config dict."""
+    if not isinstance(config, dict):
+        return None
+    return normalize_project_name(config.get("project_name"))
diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index 6202d3ce22..80db64553a 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -1704,6 +1704,7 @@
       "os": [
         "android"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1724,6 +1725,7 @@
       "os": [
         "darwin"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1744,6 +1746,7 @@
       "os": [
         "darwin"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1764,6 +1767,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1784,6 +1788,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1804,6 +1809,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1824,6 +1830,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1844,6 +1851,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1864,6 +1872,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1884,6 +1893,7 @@
       "os": [
         "win32"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1904,6 +1914,7 @@
       "os": [
         "win32"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -5669,9 +5680,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5688,9 +5696,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5707,9 +5712,6 @@
       "cpu": [
         "ppc64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5726,9 +5728,6 @@
       "cpu": [
         "s390x"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5745,9 +5744,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5764,9 +5760,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -10282,9 +10275,9 @@
       }
     },
     "node_modules/hono": {
-      "version": "4.12.21",
-      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.21.tgz",
-      "integrity": "sha512-uV63apnb0kyPtAUwoWgaGh9HyIFcv8lgmzPZSiTBQAFOFGIzka5EZ1dZocmGnn0XdX0+XTqJ6Tqv7selMuGLRQ==",
+      "version": "4.12.25",
+      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.25.tgz",
+      "integrity": "sha512-2NFaIyNVgJmBs/ecmtGzlmluTFs5cHEWGTdu0t1HBwYzoGXOL5nUQBRMXsXWla5i4KkG//QMzVP88m1+I3fdAQ==",
       "license": "MIT",
       "engines": {
         "node": ">=16.9.0"
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index 0956c710f5..a2eddecda3 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": "4.12.21",
+    "hono": "4.12.25",
     "qs": "6.15.2",
     "ip-address": "10.1.1",
     "brace-expansion@5.0.5": "5.0.6"
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 914abbbf1d..176665769d 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -354,12 +354,11 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     );
   }
 
-  const showApp = status === "running" && desktopAuthReady;
+  const showApp = status === "running";
+  const desktopBooting = status === "running" && !desktopAuthReady;
+  const showInteractiveApp = showApp && desktopAuthReady;
   const startupStatus = status === "running" ? "starting" : status;
-  const startupProgressDetail =
-    status === "running" && !desktopAuthReady
-      ? "Signing in to desktop session..."
-      : progressDetail;
+  const startupProgressDetail = progressDetail;
   const usesCustomTitlebar = shouldUseCustomWindowTitlebar();
   const usesNativeMacTitlebar = shouldUseNativeMacWindowTitlebar();
   const hidesTitlebarSidebar = HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname);
@@ -369,12 +368,23 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       <TauriUpdateLayer isExternalServer={isExternalServer}>
         <LlamaUpdateBanner
           positioned={false}
-          enabled={!hidesTitlebarSidebar}
+          enabled={showInteractiveApp && !hidesTitlebarSidebar}
         />
-        <DownloadManagerPanel positioned={false} />
+        {showInteractiveApp ? <DownloadManagerPanel positioned={false} /> : null}
       </TauriUpdateLayer>
-      <NativeIntentDrain />
-      {children}
+      {showInteractiveApp ? <NativeIntentDrain /> : null}
+      {showInteractiveApp ? children : null}
+      {desktopBooting ? (
+        <div className="pointer-events-none fixed inset-x-0 bottom-5 z-[9999] flex justify-center px-4">
+          <div className="absolute inset-x-4 bottom-16 mx-auto flex max-w-[520px] flex-col items-center gap-2 rounded-2xl border border-border/70 bg-background/95 px-6 py-5 text-center shadow-xl">
+            <div className="font-medium text-sm">Preparing Studio</div>
+            <div className="text-muted-foreground text-xs">The local backend is ready. Signing in to your desktop session before loading chats.</div>
+          </div>
+          <div className="rounded-full border border-border/70 bg-background/95 px-4 py-2 text-xs text-muted-foreground shadow-lg">
+            Signing in to desktop session...
+          </div>
+        </div>
+      ) : null}
     </>
   ) : (
     <StartupScreen
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 06f2701a16..ef4178ea42 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -123,6 +123,7 @@ import {
   deleteTrainingRun,
   emitTrainingRunDeleted,
   emitTrainingRunUpdated,
+  getTrainingRunDisplayTitle,
   removeTrainingUnloadGuard,
   renameTrainingRun,
   useTrainingCompletionWatch,
@@ -261,19 +262,6 @@ function NavItem({
   );
 }
 
-// TEMP DEV override: preview the update card on installs with no real update
-// (e.g. an editable checkout). In the browser console run
-// `localStorage.setItem("unsloth_force_update_card", "1")` and reload. Remove
-// before merge.
-function devForceUpdateCard(): boolean {
-  if (typeof window === "undefined") return false;
-  try {
-    return window.localStorage.getItem("unsloth_force_update_card") === "1";
-  } catch {
-    return false;
-  }
-}
-
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
@@ -290,13 +278,10 @@ export function AppSidebar() {
 
   // Web update detection: `webUpdate` is non-null only when the installed
   // (PyPI) version is behind the latest release, so the card is hidden by
-  // default. `forceUpdateCard` is a TEMP dev override to preview it on installs
-  // with no real update (e.g. an editable checkout); remove before merge.
+  // default.
   const { status: webUpdate } = useWebUpdateCheck();
-  const [forceUpdateCard] = useState(devForceUpdateCard);
-  const showUpdateCard = Boolean(webUpdate) || forceUpdateCard;
-  const updateVersion =
-    webUpdate?.latestVersion ?? (forceUpdateCard ? "0.0.0" : null);
+  const showUpdateCard = Boolean(webUpdate);
+  const updateVersion = webUpdate?.latestVersion ?? null;
 
   // Auto-close mobile Sheet after navigation
   const closeMobileIfOpen = () => {
@@ -592,7 +577,7 @@ export function AppSidebar() {
     setRenamingTarget({ kind: "chat", item, current: item.title });
   }
   function openRenameRun(run: TrainingRunSummary) {
-    const current = run.display_name ?? run.model_name;
+    const current = getTrainingRunDisplayTitle(run);
     setRenameDraft(current);
     setRenamingTarget({ kind: "run", run, current });
   }
@@ -1377,7 +1362,7 @@ export function AppSidebar() {
                               aria-hidden
                             />
                             <span className="truncate">
-                              {run.display_name ?? run.model_name}
+                              {getTrainingRunDisplayTitle(run)}
                             </span>
                             <span className="ml-auto mr-0.5 shrink-0 text-[10px] text-muted-foreground">
                               {formatRelativeShort(run.started_at)}
@@ -1653,8 +1638,7 @@ export function AppSidebar() {
               renderEmphasizedTranslation(
                 t,
                 "shell.dialog.deleteRun.description",
-                confirmingDelete.run.display_name ??
-                  confirmingDelete.run.model_name,
+                getTrainingRunDisplayTitle(confirmingDelete.run),
               )
             ) : confirmingDelete?.kind === "chat" ? (
               renderEmphasizedTranslation(
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 386f233c7b..5f5d112204 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -1114,6 +1114,17 @@ function localModelIsGguf(m: LocalModelInfo): boolean {
   );
 }
 
+function localPathTooltip(name: string, path: string): ReactNode {
+  return (
+    <>
+      <span className="block break-words">{name}</span>
+      <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+        {path}
+      </span>
+    </>
+  );
+}
+
 /** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
  * callers gate visibility on the host being a Mac. */
 function localModelIsMlx(m: LocalModelInfo): boolean {
@@ -2928,6 +2939,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
@@ -3017,6 +3032,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
@@ -3098,6 +3117,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
index cc96665df4..24f0edc784 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
@@ -64,38 +64,19 @@ export function matchesFormatFilter(
   }
 }
 
-// Model-size extraction from repo id, matching the backend's 3-regex priority:
-//   active params (MoE "A3B") > effective params (Gemma "E4B") > total ("8B").
-// Bounded by separators so we never read "16" from "bf16" or "2" from "Kimi-K2".
-// Examples: "Qwen3.5-35B-A3B" -> 3, "gemma-4-E4B" -> 4, "Llama-3-8B" -> 8.
-const ACTIVE_PARAM_RE = /(?:^|[-_/. ])a(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
-const EFFECTIVE_PARAM_RE = /(?:^|[-_/. ])e(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
-const TOTAL_PARAM_RE = /(?:^|[-_/. ])(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/;
-
-function paramsFromMatch(match: RegExpExecArray | null): number | undefined {
-  if (!match) return undefined;
-  const billions = parseFloat(match[1]);
-  return Number.isFinite(billions) && billions > 0
-    ? billions * 1e9
-    : undefined;
-}
-
-/** Active/effective parameter count parsed from a repo id, if it uses explicit
- * MoE/Gemma-style notation such as A3B or E4B. */
-export function activeOrEffectiveParamsFromId(id: string): number | undefined {
-  return (
-    paramsFromMatch(ACTIVE_PARAM_RE.exec(id)) ??
-    paramsFromMatch(EFFECTIVE_PARAM_RE.exec(id))
-  );
-}
+// First "<n>B" token in a repo id, e.g. "Qwen3-4B-GGUF" -> 4, "gpt-oss-20b" ->
+// 20, "Qwen3-30B-A3B" -> 30 (MoE total), "gemma-4-E4B" -> 4 (effective-param
+// "E" series). The digits must be bounded by a separator so we never read "16"
+// from "bf16" or the "2" in "Kimi-K2".
+const PARAM_RE = /(?:^|[-_/. ])[eE]?(\d+(?:\.\d+)?)\s*[bB](?=$|[-_./ ])/;
 
 /** Parameter count (absolute, e.g. 4e9) parsed from a repo id, or undefined
- * when the id has no size token (so callers can treat the size as unknown).
- * Prefers MoE active-param notation (A3B) over effective (E4B) over total. */
+ * when the id has no size token (so callers can treat the size as unknown). */
 export function paramsFromId(id: string): number | undefined {
-  return (
-    activeOrEffectiveParamsFromId(id) ?? paramsFromMatch(TOTAL_PARAM_RE.exec(id))
-  );
+  const match = PARAM_RE.exec(id);
+  if (!match) return undefined;
+  const billions = parseFloat(match[1]);
+  return Number.isFinite(billions) && billions > 0 ? billions * 1e9 : undefined;
 }
 
 // Smallest practical GGUF/MLX quant (~Q2_K, low-bit). The fit check asks whether
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 0106980871..ca7ac174fc 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -47,6 +47,7 @@ import {
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
 import { useExternalProvidersStore } from "../stores/external-providers-store";
+import type { ModelType } from "../types";
 import { isMultimodalResponse } from "../types/api";
 import type {
   GgufVariantDetail,
@@ -142,6 +143,11 @@ interface ServerTimings {
 type RunMessages = Parameters<ChatModelAdapter["run"]>[0]["messages"];
 type RunMessage = RunMessages[number];
 
+type OpenAIStreamAdapterOptions = {
+  modelType?: ModelType;
+  pairId?: string;
+};
+
 /** Tracks which user messages were sent with an audio file (messageId → filename). */
 export const sentAudioNames = new Map<string, string>();
 
@@ -1182,7 +1188,17 @@ export function findLatestUserAudioBase64(
 
 async function resolveUseAdapter(
   threadId: string | undefined,
+  options: OpenAIStreamAdapterOptions = {},
 ): Promise<boolean | undefined> {
+  if (options.modelType === "model1" || options.modelType === "model2") {
+    return undefined;
+  }
+  if (
+    options.pairId &&
+    (options.modelType === "base" || options.modelType === "lora")
+  ) {
+    return options.modelType === "lora";
+  }
   if (!threadId) {
     return undefined;
   }
@@ -1629,7 +1645,9 @@ async function autoLoadSmallestModel(): Promise<{
   }
 }
 
-export function createOpenAIStreamAdapter(): ChatModelAdapter {
+export function createOpenAIStreamAdapter(
+  options: OpenAIStreamAdapterOptions = {},
+): ChatModelAdapter {
   return {
     async *run({ messages, abortSignal, unstable_threadId }) {
       await useChatRuntimeStore.getState().hydratePersistedSettings();
@@ -2076,7 +2094,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         }
         runtime.clearPendingAudio();
       }
-      const useAdapter = await resolveUseAdapter(resolvedThreadId);
+      const useAdapter = await resolveUseAdapter(resolvedThreadId, options);
 
       // ── Audio model path (non-streaming) ─────────────────────
       const activeModel = runtime.models.find(
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index dfd96a1e86..cd7cfc77fc 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -606,7 +606,7 @@ const LoraCompareContent = memo(function LoraCompareContent({
           handleName="lora"
           borderClassName="border-t border-border/60 md:border-t-0 md:border-l"
           header={
-            <div className="shrink-0 px-3 py-1.5 text-start md:text-end">
+            <div className="shrink-0 px-3 py-1.5 text-start md:text-end md:pr-[calc(4rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]">
               <span className="text-[10px] font-semibold uppercase tracking-wider text-primary">
                 Fine-tuned
               </span>
@@ -654,7 +654,7 @@ function GeneralCompareHeader({
   return (
     <div
       className={cn(
-        "pointer-events-none relative z-[65] flex h-[48px] shrink-0 items-start gap-2 bg-background pt-[var(--studio-chat-header-padding-top,11px)]",
+        "pointer-events-none relative z-40 flex h-[48px] shrink-0 items-start gap-2 bg-background pt-[var(--studio-chat-header-padding-top,11px)]",
         side === "left"
           ? pinned
             ? "pl-12 pr-3 md:pl-2"
@@ -2388,7 +2388,7 @@ export function ChatPage({
         )}
         <div
           className={cn(
-            "pointer-events-none absolute top-[var(--studio-content-top-inset,0px)] left-0 right-[10px] z-[66] flex h-[var(--studio-chat-header-height,48px)] shrink-0 items-start bg-background pt-[var(--studio-chat-header-padding-top,11px)] pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
+            "pointer-events-none absolute top-[var(--studio-content-top-inset,0px)] left-0 right-[10px] z-40 flex h-[var(--studio-chat-header-height,48px)] shrink-0 items-start bg-background pt-[var(--studio-chat-header-padding-top,11px)] pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
             isMobile
               ? "pl-12"
               : pinned
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 360e081f0f..67980b94c6 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -1045,16 +1045,17 @@ function useStudioRuntimeAdapters(
   return adapters;
 }
 
-const chatAdapter = createOpenAIStreamAdapter();
-
 function useRuntimeHook(
   modelType: ModelType,
   pairId?: string,
 ): ReturnType<typeof useLocalRuntime> {
   const adapters = useStudioRuntimeAdapters(modelType, pairId);
   const persistedChatAdapter = useMemo(
-    () => createPersistedRunAdapter(chatAdapter),
-    [],
+    () =>
+      createPersistedRunAdapter(
+        createOpenAIStreamAdapter({ modelType, pairId }),
+      ),
+    [modelType, pairId],
   );
   return useLocalRuntime(persistedChatAdapter, { adapters });
 }
diff --git a/studio/frontend/src/features/chat/utils/delete-thread-message.ts b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
index 3b439eb779..bea5a63ea7 100644
--- a/studio/frontend/src/features/chat/utils/delete-thread-message.ts
+++ b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
@@ -112,7 +112,26 @@ export async function deleteThreadMessage(args: {
   const exported = thread.export();
   const repo = new MessageRepository();
   repo.import(exported);
+
+  const target = exported.messages.find(
+    ({ message }) => message.id === messageId,
+  );
+  const assistantReplyIds =
+    target?.message.role === "user"
+      ? exported.messages
+          .filter(
+            ({ parentId, message }) =>
+              parentId === messageId && message.role === "assistant",
+          )
+          .map(({ message }) => message.id)
+      : [];
+
+  // Delete the prompt first; that relinks its replies up to the prompt's parent
   repo.deleteMessage(messageId);
+  for (const replyId of assistantReplyIds) {
+    repo.deleteMessage(replyId);
+  }
+
   const next = repo.export();
   if (remoteId) {
     await syncExportedRepositoryToBackend(remoteId, next, {
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 7d38d9b222..4037dbe079 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -75,16 +75,35 @@ import { exportTourSteps } from "./tour";
 const SEARCH_INPUT_REASONS = new Set(["input-change", "input-paste", "input-clear"]);
 
 type SourceTab = "local" | "checkpoint" | "hf";
+type SourceMode = "checkpoint" | "model";
+
+
+function safePathSegment(
+  value: string | null | undefined,
+  fallback = "model",
+  maxLength = 250,
+): string {
+  const safe = (value ?? "")
+    .replace(/[^a-zA-Z0-9._-]/g, "-")
+    .replace(/^[._-]+|[._-]+$/g, "")
+    .slice(0, maxLength)
+    .replace(/[._-]+$/g, "");
+  return safe || fallback;
+}
 
 function buildRelativeSaveDirectory(
   exportMethod: ExportMethod | null,
+  sourceMode: SourceMode,
   sourceBaseModelName: string,
   selectedModelIdx: string | null,
   checkpoint: string | null,
 ): string {
   if (exportMethod === "gguf") {
-    return `${(sourceBaseModelName.split("/").pop() ?? selectedModelIdx ?? "model")
-      .replace(/[^a-zA-Z0-9._-]/g, "-")}-GGUF`;
+    const rawName =
+      sourceMode === "checkpoint"
+        ? checkpoint ?? selectedModelIdx ?? sourceBaseModelName
+        : sourceBaseModelName;
+    return `${safePathSegment(rawName)}-GGUF`;
   }
   return `${selectedModelIdx ?? "model"}/${checkpoint}`;
 }
@@ -125,9 +144,7 @@ export function ExportPage() {
 
   const [selectedModelIdx, setSelectedModelIdx] = useState<string | null>(null);
   const [checkpoint, setCheckpoint] = useState<string | null>(null);
-  const [sourceMode, setSourceMode] = useState<"checkpoint" | "model">(
-    "checkpoint",
-  );
+  const [sourceMode, setSourceMode] = useState<SourceMode>("checkpoint");
   const [modelSource, setModelSource] = useState<"hf" | "local">("hf");
   const [modelInput, setModelInput] = useState("");
   const [selectedSourceModel, setSelectedSourceModel] = useState<string | null>(
@@ -449,6 +466,7 @@ export function ExportPage() {
   const defaultSaveDirectory = useMemo(() => {
     const relative = buildRelativeSaveDirectory(
       exportMethod,
+      sourceMode,
       sourceBaseModelName,
       selectedModelIdx,
       checkpoint,
diff --git a/studio/frontend/src/features/hub/catalog/model-card.tsx b/studio/frontend/src/features/hub/catalog/model-card.tsx
index c3c526cac3..358a3409b5 100644
--- a/studio/frontend/src/features/hub/catalog/model-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-card.tsx
@@ -11,7 +11,7 @@ import { ownerPaletteColor } from "@/features/hub/lib/avatar-theme";
 import { buildAdaptiveCardAccentStyle } from "@/features/hub/lib/card-accent";
 import { useDominantColor } from "@/features/hub/lib/use-dominant-color";
 import { formatModelParamLabel } from "@/features/hub/lib/view-models";
-import { cn, formatCompact } from "@/lib/utils";
+import { formatCompact } from "@/lib/utils";
 import { Download01Icon, FavouriteIcon } from "@hugeicons/core-free-icons";
 import { type CSSProperties, memo, useMemo } from "react";
 import type { DiscoverRow } from "../types";
@@ -339,25 +339,11 @@ export const ModelCard = memo(function ModelCard({
             value={formatCompact(row.result.likes)}
           />
         </div>
-        <div className="flex shrink-0 items-center gap-1.5">
-          {row.fitLevel && (
-            <span
-              className={cn(
-                "text-[9px] font-medium px-1.5 py-0.5 rounded leading-none uppercase shrink-0",
-                row.fitLevel === "comfortable" && "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15",
-                row.fitLevel === "fits" && "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15",
-                row.fitLevel === "oom" && "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15"
-              )}
-            >
-              {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-            </span>
-          )}
-          {hasSize ? (
-            <span className="hub-chip shrink-0">{sizeLabel}</span>
-          ) : topCapability ? (
-            <CapabilityPill capability={topCapability} iconOnly={true} />
-          ) : null}
-        </div>
+        {hasSize ? (
+          <span className="hub-chip shrink-0">{sizeLabel}</span>
+        ) : topCapability ? (
+          <CapabilityPill capability={topCapability} iconOnly={true} />
+        ) : null}
       </div>
     </button>
   );
diff --git a/studio/frontend/src/features/hub/catalog/models-table.tsx b/studio/frontend/src/features/hub/catalog/models-table.tsx
index 35f2f966e5..a62806dcb6 100644
--- a/studio/frontend/src/features/hub/catalog/models-table.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-table.tsx
@@ -570,22 +570,6 @@ export const ResultCard = memo(function ResultCard({
       node: <span className="shrink-0">{sizeLabel}</span>,
     });
   }
-  if (row.fitLevel) {
-    const label = row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM";
-    const toneClass = row.fitLevel === "comfortable"
-      ? "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15"
-      : row.fitLevel === "fits"
-        ? "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15"
-        : "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15";
-    textParts.push({
-      key: "gpuFit",
-      node: (
-        <span className={cn("shrink-0 uppercase text-[9px] font-medium px-1.5 py-0.5 rounded leading-none", toneClass)}>
-          {label}
-        </span>
-      ),
-    });
-  }
   if (row.result.updatedAt) {
     textParts.push({
       key: "updated",
@@ -745,21 +729,6 @@ export const ResultGridRow = memo(function ResultGridRow({
             </div>
             <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
               <VerifiedOwner owner={row.owner} />
-              {row.fitLevel && (
-                <>
-                  <span aria-hidden="true" className="text-muted-foreground/35">•</span>
-                  <span
-                    className={cn(
-                      "text-[9px] font-semibold uppercase shrink-0",
-                      row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
-                      row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
-                      row.fitLevel === "oom" && "text-red-500"
-                    )}
-                  >
-                    {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-                  </span>
-                </>
-              )}
             </span>
           </div>
         </div>
@@ -882,21 +851,6 @@ export const ResultSplitRow = memo(function ResultSplitRow({
         </div>
         <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[10.5px] leading-[14px] text-muted-foreground/80">
           <VerifiedOwner owner={row.owner} />
-          {row.fitLevel && (
-            <>
-              <span aria-hidden="true" className="text-muted-foreground/35">•</span>
-              <span
-                className={cn(
-                  "text-[9px] font-semibold uppercase shrink-0",
-                  row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
-                  row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
-                  row.fitLevel === "oom" && "text-red-500"
-                )}
-              >
-                {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-              </span>
-            </>
-          )}
         </span>
       </div>
       <div className="flex shrink-0 flex-col items-end gap-0.5 text-[10.5px] tabular-nums text-muted-foreground/70">
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 9cb17736af..48f7fcffaa 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -21,7 +21,6 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import type { HfSortKey } from "@/features/hub/hooks/use-hub-model-search";
 import type {
   CapabilityFilter,
-  GpuFitFilter,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
@@ -29,7 +28,6 @@ import type {
 import {
   CAPABILITY_FILTER_OPTIONS,
   FORMAT_FILTER_OPTIONS,
-  GPU_FIT_FILTER_OPTIONS,
 } from "../lib/view-models";
 import { HubOptionMenu, type HubOption } from "./hub-option-menu";
 import {
@@ -70,8 +68,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange,
   capabilityFilter,
   onCapabilityFilterChange,
-  gpuFitFilter,
-  onGpuFitFilterChange,
   onManageLocalFolders,
   onOpenFineTune,
 }: {
@@ -88,8 +84,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange: (value: ModelFormatFilter) => void;
   capabilityFilter: CapabilityFilter;
   onCapabilityFilterChange: (value: CapabilityFilter) => void;
-  gpuFitFilter: GpuFitFilter;
-  onGpuFitFilterChange: (value: GpuFitFilter) => void;
   onManageLocalFolders: () => void;
   /** Opens the curated "Fine-tune ready" channel (discover only). Exposed as a
    *  format-dropdown option rather than a standalone feed section. */
@@ -159,14 +153,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
       })),
     [],
   );
-  const gpuFitOptions = useMemo<HubOption<GpuFitFilter>[]>(
-    () =>
-      GPU_FIT_FILTER_OPTIONS.map((option) => ({
-        value: option.value,
-        label: option.label,
-      })),
-    [],
-  );
   const sortOptions = useMemo<HubOption<HfSortKey>[]>(
     () =>
       SORT_OPTIONS.map((option) => ({
@@ -357,16 +343,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
           />
         )}
 
-        {tab === "discover" && !isDataset && (
-          <HubOptionMenu
-            value={gpuFitFilter}
-            options={gpuFitOptions}
-            onValueChange={onGpuFitFilterChange}
-            ariaLabel="GPU fit filter"
-            className={cn(triggerBase, "w-[128px]")}
-          />
-        )}
-
         {tab === "discover" && (
           <HubOptionMenu
             value={sortBy}
diff --git a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
index 20a4eca4dd..caa89db196 100644
--- a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
+++ b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
@@ -180,7 +180,7 @@ export function OnDeviceFoldersDialog({
     <>
       <Dialog open={open} onOpenChange={onOpenChange}>
         <DialogContent
-          className="max-w-[600px] gap-0 overflow-hidden p-0"
+          className="gap-0 overflow-hidden p-0 sm:max-w-[620px] lg:max-w-[660px] xl:max-w-[680px] [&_[data-slot=dialog-close]]:right-3 [&_[data-slot=dialog-close]]:top-3"
           overlayClassName="bg-black/20 backdrop-blur-none"
         >
           <DialogHeader className="border-b border-border/60 px-5 py-4">
@@ -319,7 +319,12 @@ export function OnDeviceFoldersDialog({
                     return (
                       <div
                         key={folder.id}
-                        className="flex min-h-12 items-center gap-3 border-b border-border/50 px-3 py-2 last:border-b-0"
+                        className={cn(
+                          "grid min-h-12 w-full items-center gap-3 border-b border-border/50 px-3 py-2 last:border-b-0",
+                          isTauri
+                            ? "grid-cols-[2rem_minmax(0,1fr)_2rem_2rem]"
+                            : "grid-cols-[2rem_minmax(0,1fr)_2rem]",
+                        )}
                       >
                         <div className="flex size-8 shrink-0 items-center justify-center rounded-[9px] bg-muted text-muted-foreground">
                           <HugeiconsIcon
@@ -328,13 +333,18 @@ export function OnDeviceFoldersDialog({
                             className="size-4"
                           />
                         </div>
-                        <div className="min-w-0 flex-1">
-                          <p className="truncate text-[12.5px] font-medium text-foreground">
+                        <div className="min-w-0 overflow-hidden">
+                          <p
+                            className="block w-full truncate text-[12.5px] font-medium text-foreground"
+                            title={pathTail(folder.path)}
+                          >
                             {pathTail(folder.path)}
                           </p>
                           <Tooltip>
                             <TooltipTrigger asChild={true}>
-                              <p className="truncate font-mono text-[10.5px] text-muted-foreground">
+                              <p
+                                className="block w-full truncate font-mono text-[10.5px] text-muted-foreground"
+                              >
                                 {folder.path}
                               </p>
                             </TooltipTrigger>
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index 4c79b607d9..c696862283 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -85,17 +85,12 @@ import type {
   CachedInventoryRow,
   CapabilityFilter,
   DiscoverRow,
-  GpuFitFilter,
   LocalInventoryRow,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
   SelectedModelView,
 } from "./types";
-import {
-  classifyGpuFit,
-  matchesGpuFitFilter,
-} from "./lib/gpu-fit-filter";
 
 const MODELS_TAB_STORAGE_KEY = "unsloth.hub.modelsTab";
 const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
@@ -418,7 +413,6 @@ export function ModelsPage() {
   );
   const [capabilityFilter, setCapabilityFilter] =
     useState<CapabilityFilter>("all");
-  const [gpuFitFilter, setGpuFitFilter] = useState<GpuFitFilter>("all");
   const [allModelsView, setAllModelsViewState] = useState<AllModelsView>(
     readAllModelsViewPreference,
   );
@@ -567,7 +561,6 @@ export function ModelsPage() {
   const apiHfToken = hfApiToken(debouncedHfToken);
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
-  const deferredGpuFitFilter = useDeferredValue(gpuFitFilter);
 
   const hasQuery = deferredDebouncedQuery.trim() !== "";
   const mode: DiscoverMode = !isModelDiscover
@@ -697,59 +690,20 @@ export function ModelsPage() {
 
   const discoverRows = isDatasetMode ? datasetDiscoverRows : modelDiscoverRows;
 
-  // Pre-compute GPU fit level for every discover row so filteredDiscoverRows
-  // and model cards can both consume the same classification.
-  const gpuFitLevelById = useMemo(() => {
-    const map = new Map<string, ReturnType<typeof classifyGpuFit>>();
-    for (const row of discoverRows) {
-      map.set(
-        row.id,
-        classifyGpuFit({
-          totalParams: row.result.totalParams,
-          estimatedSizeBytes: row.result.estimatedSizeBytes,
-          repoId: row.id,
-          gpu,
-        }),
-      );
-    }
-    return map;
-  }, [discoverRows, gpu]);
-
-  const addGpuFitLevel = useCallback(
-    (row: DiscoverRow): DiscoverRow => ({
-      ...row,
-      fitLevel: classifyGpuFit({
-        totalParams: row.result.totalParams,
-        estimatedSizeBytes: row.result.estimatedSizeBytes,
-        repoId: row.id,
-        gpu,
-      }),
-    }),
-    [gpu],
-  );
-
   const filteredDiscoverRows = useMemo(() => {
     if (isDatasetMode) return discoverRows;
-    return discoverRows
-      .filter(
-        (row) =>
-          !isHiddenModelId(row.id) &&
-          matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
-          matchesCapability(row.capabilities, deferredCapabilityFilter) &&
-          matchesGpuFitFilter(gpuFitLevelById.get(row.id) ?? null, deferredGpuFitFilter) &&
-          (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
-      )
-      .map((row) => ({
-        ...row,
-        fitLevel: gpuFitLevelById.get(row.id) ?? null,
-      }));
+    return discoverRows.filter(
+      (row) =>
+        !isHiddenModelId(row.id) &&
+        matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
+        matchesCapability(row.capabilities, deferredCapabilityFilter) &&
+        (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
+    );
   }, [
     discoverRows,
     isDatasetMode,
     effectiveDiscoverFormat,
     deferredCapabilityFilter,
-    deferredGpuFitFilter,
-    gpuFitLevelById,
     activeChannel,
   ]);
 
@@ -770,17 +724,8 @@ export function ModelsPage() {
         effectiveLocalRows,
       )
         .filter((row) => !isHiddenModelId(row.id))
-        .filter((row) => matchesFormat(row.result.isGguf, "gguf"))
-        .map(addGpuFitLevel)
-        .filter((row) =>
-          matchesGpuFitFilter(row.fitLevel ?? null, deferredGpuFitFilter),
-        ),
-    [
-      hubFeed.trending.results,
-      modelDiscoveryInventorySignature,
-      addGpuFitLevel,
-      deferredGpuFitFilter,
-    ],
+        .filter((row) => matchesFormat(row.result.isGguf, "gguf")),
+    [hubFeed.trending.results, modelDiscoveryInventorySignature],
   );
   const feedRows = useMemo(() => {
     if (!isFeedMode) return [];
@@ -897,7 +842,6 @@ export function ModelsPage() {
         resourceType,
         deferredFormatFilter,
         deferredCapabilityFilter,
-        deferredGpuFitFilter,
         effectiveSort,
         effectiveDirection,
         activeChannelId,
@@ -908,7 +852,6 @@ export function ModelsPage() {
       resourceType,
       deferredFormatFilter,
       deferredCapabilityFilter,
-      deferredGpuFitFilter,
       effectiveSort,
       effectiveDirection,
       activeChannelId,
@@ -929,7 +872,6 @@ export function ModelsPage() {
       setDownloadedFormat("all");
     }
     setCapabilityFilter("all");
-    setGpuFitFilter("all");
   }, [isDiscoverTab, urlSection, navigate]);
   const handleDiscoverFetchIntent = useCallback(() => {
     setDiscoverFetchIntent((value) => value + 1);
@@ -1295,10 +1237,8 @@ export function ModelsPage() {
       hasMore,
       manualFetchAvailable: discoverManualFetchAvailable,
       hasActiveFilters:
-        deferredGpuFitFilter !== "all" ||
-        (!isFeedMode &&
-          (deferredFormatFilter !== "all" ||
-            deferredCapabilityFilter !== "all")),
+        !isFeedMode &&
+        (deferredFormatFilter !== "all" || deferredCapabilityFilter !== "all"),
     }),
     [
       tab,
@@ -1324,7 +1264,6 @@ export function ModelsPage() {
       discoverManualFetchAvailable,
       deferredFormatFilter,
       deferredCapabilityFilter,
-      deferredGpuFitFilter,
     ],
   );
 
@@ -1509,8 +1448,6 @@ export function ModelsPage() {
           onFormatFilterChange={setFormatFilter}
           capabilityFilter={capabilityFilter}
           onCapabilityFilterChange={setCapabilityFilter}
-          gpuFitFilter={gpuFitFilter}
-          onGpuFitFilterChange={setGpuFitFilter}
           onManageLocalFolders={handleManageLocalFolders}
           onOpenFineTune={() => handleOpenList("finetune")}
         />
diff --git a/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts b/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
deleted file mode 100644
index 7eca4c315a..0000000000
--- a/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
+++ /dev/null
@@ -1,80 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// GPU-aware model-fit filtering: classifies whether a model fits the device
-// and provides a filter predicate for the Hub page and model selector.
-
-import type { GpuInfo } from "@/hooks/use-gpu-info";
-import {
-  activeOrEffectiveParamsFromId,
-  estimateQuantBytes,
-  paramsFromId,
-} from "@/components/assistant-ui/model-selector/recommended-fit";
-
-/** The three filter states exposed in the toolbar dropdown. */
-export type GpuFitFilter = "all" | "fits" | "comfortable";
-
-/** Per-model fit classification. */
-export type GpuFitLevel = "comfortable" | "fits" | "oom";
-
-/**
- * Classify whether a model fits the device.
- *
- * - "comfortable": estimated size ≤ 70% of GPU VRAM (runs fully in VRAM)
- * - "fits": estimated size ≤ 70% GPU + 70% system RAM (runs with CPU offload)
- * - "oom": exceeds both budgets
- *
- * Returns null when we can't determine the size (unknown → no badge).
- */
-export function classifyGpuFit(opts: {
-  totalParams?: number;
-  estimatedSizeBytes?: number;
-  repoId: string;
-  gpu: GpuInfo;
-}): GpuFitLevel | null {
-  const { totalParams, estimatedSizeBytes, repoId, gpu } = opts;
-  const gpuGb = gpu.memoryTotalGb;
-  const ramGb = gpu.systemRamAvailableGb;
-  if (gpuGb <= 0 && ramGb <= 0) return null; // no budget info
-
-  // Active/effective model tokens (for example MoE A3B) describe runnable size
-  // better than HF total-parameter metadata; otherwise prefer exact metadata.
-  const activeOrEffectiveParams = activeOrEffectiveParamsFromId(repoId);
-  const params = activeOrEffectiveParams ?? totalParams ?? paramsFromId(repoId);
-  const sizeBytes =
-    activeOrEffectiveParams
-      ? estimateQuantBytes(activeOrEffectiveParams)
-      : estimatedSizeBytes ?? (params ? estimateQuantBytes(params) : undefined);
-
-  if (!sizeBytes || sizeBytes <= 0) return null; // can't determine
-
-  const sizeGb = sizeBytes / 1024 ** 3;
-  let comfortBudget: number;
-  let fitBudget: number;
-
-  if (!gpu.available || gpuGb <= 0) {
-    // Unified memory system (no discrete GPU)
-    comfortBudget = ramGb * 0.7;
-    fitBudget = ramGb * 0.7;
-  } else {
-    // Discrete GPU
-    comfortBudget = gpuGb * 0.7;
-    fitBudget = gpuGb * 0.7 + ramGb * 0.7;
-  }
-
-  if (sizeGb <= comfortBudget) return "comfortable";
-  if (sizeGb <= fitBudget) return "fits";
-  return "oom";
-}
-
-/** Whether a row passes the given GPU fit filter. */
-export function matchesGpuFitFilter(
-  level: GpuFitLevel | null,
-  filter: GpuFitFilter,
-): boolean {
-  if (filter === "all") return true;
-  if (level === null) return false;
-  if (filter === "comfortable") return level === "comfortable";
-  // "fits" shows both comfortable and fits
-  return level === "comfortable" || level === "fits";
-}
diff --git a/studio/frontend/src/features/hub/lib/view-models.ts b/studio/frontend/src/features/hub/lib/view-models.ts
index cdd17855db..9ee6c5de5d 100644
--- a/studio/frontend/src/features/hub/lib/view-models.ts
+++ b/studio/frontend/src/features/hub/lib/view-models.ts
@@ -10,7 +10,6 @@ import type {
 import type {
   CapabilityFilter,
   DiscoverRow,
-  GpuFitFilter,
   ModelFormatFilter,
 } from "../types";
 import {
@@ -52,15 +51,6 @@ export const FORMAT_FILTER_OPTIONS: ReadonlyArray<{
   { value: "mlx", label: "MLX" },
 ];
 
-export const GPU_FIT_FILTER_OPTIONS: ReadonlyArray<{
-  value: GpuFitFilter;
-  label: string;
-}> = [
-  { value: "all", label: "All sizes" },
-  { value: "fits", label: "Fits GPU" },
-  { value: "comfortable", label: "Comfortable" },
-];
-
 const BILLION = 1_000_000_000;
 
 export function formatParamCount(totalParams: number | undefined): string {
diff --git a/studio/frontend/src/features/hub/types.ts b/studio/frontend/src/features/hub/types.ts
index 9fabdb2a8b..ae9ddf5d4b 100644
--- a/studio/frontend/src/features/hub/types.ts
+++ b/studio/frontend/src/features/hub/types.ts
@@ -28,9 +28,6 @@ export type ModelFormatFilter = "all" | "gguf" | "checkpoint" | "mlx";
 
 export type CapabilityFilter = "all" | CapabilityKey;
 
-import type { GpuFitFilter, GpuFitLevel } from "./lib/gpu-fit-filter";
-export type { GpuFitFilter, GpuFitLevel };
-
 export interface DiscoverRow {
   id: string;
   owner: string;
@@ -40,7 +37,6 @@ export interface DiscoverRow {
   isPartialOnDevice: boolean;
   summary: string;
   capabilities: Capability[];
-  fitLevel?: GpuFitLevel | null;
 }
 
 export type SelectedResourceSource = "huggingface" | "hub_cache" | LocalSource;
diff --git a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
index a1174186af..2fb2f7adeb 100644
--- a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
@@ -7,6 +7,7 @@ import {
   FieldLegend,
   FieldSet,
 } from "@/components/ui/field";
+import { Input } from "@/components/ui/input";
 import {
   Select,
   SelectContent,
@@ -59,6 +60,8 @@ function stepLR(value: number, direction: 1 | -1): number {
 export function HyperparametersStep() {
   const {
     trainingMethod,
+    projectName,
+    setProjectName,
     maxSteps,
     setMaxSteps,
     epochs,
@@ -79,6 +82,8 @@ export function HyperparametersStep() {
   } = useTrainingConfigStore(
     useShallow((s) => ({
       trainingMethod: s.trainingMethod,
+      projectName: s.projectName,
+      setProjectName: s.setProjectName,
       maxSteps: s.maxSteps,
       setMaxSteps: s.setMaxSteps,
       epochs: s.epochs,
@@ -125,6 +130,26 @@ export function HyperparametersStep() {
       <FieldSet>
         <FieldLegend variant="label">Choose your training parameters</FieldLegend>
         <div className="flex flex-col gap-4">
+          <div className="flex flex-col gap-2">
+            <div className="flex items-center justify-between">
+              <FieldLabel className="flex items-center gap-1.5 !text-sm text-muted-foreground">
+                Project Name
+                <span className="text-xs font-normal text-muted-foreground/70">
+                  Optional
+                </span>
+              </FieldLabel>
+            </div>
+            <Input
+              value={projectName || ""}
+              onChange={(e) => setProjectName(e.target.value)}
+              placeholder="customer-support-lora"
+              maxLength={80}
+            />
+            <p className="text-xs text-muted-foreground">
+              Used in training output folder names, export defaults, and history.
+            </p>
+          </div>
+
           <div
             key={useEpochs ? "epochs" : "steps"}
             className="flex flex-col gap-2 animate-in fade-in-1 slide-in-from-bottom-1 duration-200"
diff --git a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
index 1988cfe970..6b6b11bf1f 100644
--- a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
@@ -50,6 +50,7 @@ export function SummaryStep() {
   const {
     modelType,
     selectedModel,
+    projectName,
     trainingMethod,
     datasetSource,
     datasetFormat,
@@ -68,6 +69,7 @@ export function SummaryStep() {
       ({
         modelType,
         selectedModel,
+        projectName,
         trainingMethod,
         datasetSource,
         datasetFormat,
@@ -84,6 +86,7 @@ export function SummaryStep() {
       }) => ({
         modelType,
         selectedModel,
+        projectName,
         trainingMethod,
         datasetSource,
         datasetFormat,
@@ -152,6 +155,7 @@ export function SummaryStep() {
           <Separator className="my-2" />
           <div className="space-y-1 text-sm">
             <Row label="Type" value={modelType} capitalize />
+            <Row label="Project" value={projectName || "--"} />
             <Row label="Method" value={trainingMethodLabel} />
           </div>
         </CardContent>
diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx
index 0bca090413..2f80fc29ca 100644
--- a/studio/frontend/src/features/studio/historical-training-view.tsx
+++ b/studio/frontend/src/features/studio/historical-training-view.tsx
@@ -78,6 +78,7 @@ function mapToViewData(
     error: run.status === "error" ? run.error_message : null,
     isTrainingRunning: false,
     modelName: run.display_name ?? run.model_name,
+    projectName: run.project_name,
     trainingMethod: parseBackendTrainingMethod(
       detail.config?.training_type,
       detail.config?.load_in_4bit,
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index 5ff9f3bf82..bd76171ce2 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -15,6 +15,8 @@ import { Button } from "@/components/ui/button";
 import type { TrainingRunSummary } from "@/features/training";
 import {
   deleteTrainingRun,
+  getTrainingRunDisplayTitle,
+  getTrainingRunModelSubtitle,
   emitTrainingRunDeleted,
   listTrainingRuns,
   onTrainingRunDeleted,
@@ -387,6 +389,12 @@ export function HistoryCardGrid({
           const isRunning = run.status === "running";
           const canResume = run.can_resume && !wasContinued;
           const isResuming = resumeTarget === run.id;
+
+          const title = getTrainingRunDisplayTitle(run);
+          const modelSubtitle = getTrainingRunModelSubtitle(run);
+
+          const projectSubtitle =
+            run.project_name && title !== run.project_name ? run.project_name : null;
           // Backend /p ref + its capability token. Both are required: the link
           // is useless (404s) without the signature, so don't offer to copy it.
           const canCopyPreview = !!run.preview_ref && !!run.preview_sig;
@@ -476,16 +484,16 @@ export function HistoryCardGrid({
               <div className="min-w-0">
                 <p
                   className="truncate text-sm font-medium"
-                  title={run.display_name ?? run.model_name}
+                  title={title}
                 >
-                  {run.display_name ?? run.model_name}
+                  {title}
                 </p>
-                {run.display_name && (
+                {modelSubtitle && (
                   <p
                     className="truncate text-xs text-muted-foreground"
-                    title={run.model_name}
+                    title={modelSubtitle}
                   >
-                    {run.model_name}
+                    {modelSubtitle}
                   </p>
                 )}
                 <p
@@ -494,6 +502,14 @@ export function HistoryCardGrid({
                 >
                   {run.dataset_name}
                 </p>
+                {projectSubtitle && (
+                  <p
+                    className="truncate text-xs text-muted-foreground/80"
+                    title={projectSubtitle}
+                  >
+                    {projectSubtitle}
+                  </p>
+                )}
               </div>
               {run.loss_sparkline && run.loss_sparkline.length >= 2 && (
                 <div className={cn((canResume || canCopyPreview) && "h-7 overflow-hidden")}>
diff --git a/studio/frontend/src/features/studio/live-training-view.tsx b/studio/frontend/src/features/studio/live-training-view.tsx
index e355655d54..cce39adbf4 100644
--- a/studio/frontend/src/features/studio/live-training-view.tsx
+++ b/studio/frontend/src/features/studio/live-training-view.tsx
@@ -33,6 +33,8 @@ export function LiveTrainingView(): ReactElement {
       evalEnabled: state.evalEnabled,
       outputDir: state.outputDir,
       isTrainingRunning: state.isTrainingRunning,
+      startModelName: state.startModelName,
+      startProjectName: state.startProjectName,
       lossHistory: state.lossHistory,
       lrHistory: state.lrHistory,
       gradNormHistory: state.gradNormHistory,
@@ -45,10 +47,16 @@ export function LiveTrainingView(): ReactElement {
   const config = useTrainingConfigStore(
     useShallow((state) => ({
       selectedModel: state.selectedModel,
+      projectName: state.projectName,
       trainingMethod: state.trainingMethod,
     })),
   );
 
+  const activeProjectName =
+    runtime.startProjectName !== null
+      ? runtime.startProjectName.trim() || null
+      : (config.projectName || "").trim() || null;
+
   const viewData: TrainingViewData = {
     phase: runtime.phase,
     currentStep: runtime.currentStep,
@@ -66,7 +74,8 @@ export function LiveTrainingView(): ReactElement {
     message: runtime.message,
     error: runtime.error,
     isTrainingRunning: runtime.isTrainingRunning,
-    modelName: config.selectedModel ?? "",
+    modelName: runtime.startModelName ?? config.selectedModel ?? "",
+    projectName: activeProjectName,
     trainingMethod: config.trainingMethod ?? "",
     lossHistory: runtime.lossHistory,
     lrHistory: runtime.lrHistory,
diff --git a/studio/frontend/src/features/studio/sections/params-section.tsx b/studio/frontend/src/features/studio/sections/params-section.tsx
index 059d665122..2609558145 100644
--- a/studio/frontend/src/features/studio/sections/params-section.tsx
+++ b/studio/frontend/src/features/studio/sections/params-section.tsx
@@ -229,6 +229,24 @@ export function ParamsSection(): ReactElement {
           : "h-studio-config-column"} duration-150`}
       >
         <div className="flex flex-col gap-4">
+          <div className="flex flex-col gap-2">
+            <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
+              {t("studio.params.projectName")}
+              <span className="text-[10px] font-normal text-muted-foreground/70">
+                {t("studio.params.optional")}
+              </span>
+            </span>
+            <Input
+              value={store.projectName || ""}
+              onChange={(event) => store.setProjectName(event.target.value)}
+              placeholder="customer-support-lora"
+              maxLength={80}
+            />
+            <p className="text-[10px] text-muted-foreground">
+              {t("studio.params.projectNameDescription")}
+            </p>
+          </div>
+
           {/* Max Steps / Epochs */}
           <div className="flex flex-col gap-2">
             <div
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index de8605e021..6aaca86e0b 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -270,6 +270,11 @@ export function ProgressSection({
             >
               {t(phaseLabelKeys[data.phase])}
             </span>
+            {data.projectName && (
+              <span className="rounded-full border border-border/60 px-2.5 py-1 text-[10px] font-medium text-foreground/80">
+                {data.projectName}
+              </span>
+            )}
             <span className="text-[10px] tabular-nums text-muted-foreground">
               {t("studio.progress.epoch", {
                 value: formatNumber(data.currentEpoch, 2),
@@ -290,7 +295,7 @@ export function ProgressSection({
               </span>
               <span>{pct}%</span>
             </div>
-            <Progress value={pct} className="h-2 bg-foreground/[0.05]" />
+            <Progress value={pct} className="h-2 bg-foreground/5" />
           </div>
 
           {!isHistorical && (
@@ -307,7 +312,12 @@ export function ProgressSection({
             </p>
           )}
 
-          <div className="grid gap-x-4 gap-y-3 pt-1 sm:grid-cols-2 xl:grid-cols-5">
+          <div
+            className={cn(
+              "grid gap-x-4 gap-y-3 pt-1 sm:grid-cols-2",
+              data.projectName ? "xl:grid-cols-6" : "xl:grid-cols-5",
+            )}
+          >
             <MetricStat
               label={t("studio.progress.loss")}
               valueClassName="text-2xl font-bold tracking-tight"
@@ -318,6 +328,11 @@ export function ProgressSection({
             <MetricStat label={t("studio.progress.gradNorm")}>
               {formatNumber(stoppedGradNorm, 3)}
             </MetricStat>
+            {data.projectName && (
+              <MetricStat label={t("studio.progress.project")} valueClassName="truncate">
+                {data.projectName}
+              </MetricStat>
+            )}
             <MetricStat label={t("studio.progress.model")} valueClassName="truncate">
               {data.modelName || "--"}
             </MetricStat>
diff --git a/studio/frontend/src/features/training/api/mappers.ts b/studio/frontend/src/features/training/api/mappers.ts
index d4c800afe4..c2e23dbf3c 100644
--- a/studio/frontend/src/features/training/api/mappers.ts
+++ b/studio/frontend/src/features/training/api/mappers.ts
@@ -73,6 +73,7 @@ export function buildTrainingStartPayload(
 
   return {
     model_name: config.selectedModel ?? "",
+    project_name: (config.projectName || "").trim() || null,
     training_type: toBackendTrainingType(config.trainingMethod),
     hf_token: config.hfToken.trim() || null,
     load_in_4bit: (adapterMethod && isQloraMethod) || (isCpt && isFourBitModel),
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 8c2b3e9e77..2f04656c23 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -60,6 +60,7 @@ export function useTrainingActions() {
       config.selectedModel ?? null,
       getHfDatasetName(config),
       false,
+      config.projectName || "",
     );
     runtimeStore.setStarting(true);
 
@@ -152,7 +153,12 @@ export function useTrainingActions() {
 
       // Re-read config after potential store updates from dataset check
       const payload = buildTrainingStartPayload(useTrainingConfigStore.getState());
-      runtimeStore.setStartResources(payload.model_name, payload.hf_dataset, false);
+      runtimeStore.setStartResources(
+        payload.model_name,
+        payload.hf_dataset,
+        false,
+        payload.project_name ?? "",
+      );
       const response = await startTraining(payload);
 
       if (response.status === "error") {
@@ -196,7 +202,7 @@ export function useTrainingActions() {
   const resumeTrainingRunFromHistory = useCallback(async (runId: string): Promise<boolean> => {
     const runtimeStore = useTrainingRuntimeStore.getState();
     runtimeStore.setStartError(null);
-    runtimeStore.setStartResources(null, null, true);
+    runtimeStore.setStartResources(null, null, true, null);
     runtimeStore.setStarting(true);
 
     try {
@@ -220,7 +226,12 @@ export function useTrainingActions() {
         resume_from_checkpoint: outputDir,
       } as TrainingStartRequest;
 
-      runtimeStore.setStartResources(payload.model_name, payload.hf_dataset, true);
+      runtimeStore.setStartResources(
+        payload.model_name,
+        payload.hf_dataset,
+        true,
+        payload.project_name ?? "",
+      );
 
       // Resume goes straight to startTraining, so it runs the same consent gate as a
       // fresh start; otherwise a resumed custom-code run hits the worker block with no dialog.
diff --git a/studio/frontend/src/features/training/index.ts b/studio/frontend/src/features/training/index.ts
index 5157d89582..553dcc2af5 100644
--- a/studio/frontend/src/features/training/index.ts
+++ b/studio/frontend/src/features/training/index.ts
@@ -7,6 +7,11 @@ export {
   useTrainingRuntimeStore,
 } from "./stores/training-runtime-store";
 export { useTrainingActions } from "./hooks/use-training-actions";
+
+export {
+  getTrainingRunDisplayTitle,
+  getTrainingRunModelSubtitle,
+} from "./lib/run-display";
 export { useTrainingHistorySidebarItems } from "./hooks/use-training-history-sidebar";
 export { useTrainingRuntimeLifecycle } from "./hooks/use-training-runtime-lifecycle";
 export { useTrainingCompletionWatch } from "./hooks/use-training-completion-watch";
diff --git a/studio/frontend/src/features/training/lib/run-display.ts b/studio/frontend/src/features/training/lib/run-display.ts
new file mode 100644
index 0000000000..b691b257c4
--- /dev/null
+++ b/studio/frontend/src/features/training/lib/run-display.ts
@@ -0,0 +1,24 @@
+// 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 { TrainingRunSummary } from "../types/history";
+
+type TrainingRunTitleFields = Pick<
+  TrainingRunSummary,
+  "display_name" | "project_name" | "model_name"
+>;
+
+function nonEmpty(value: string | null | undefined): string | null {
+  const trimmed = value?.trim();
+  return trimmed ? trimmed : null;
+}
+
+export function getTrainingRunDisplayTitle(run: TrainingRunTitleFields): string {
+  return nonEmpty(run.display_name) ?? nonEmpty(run.project_name) ?? run.model_name;
+}
+
+export function getTrainingRunModelSubtitle(
+  run: TrainingRunTitleFields,
+): string | null {
+  return getTrainingRunDisplayTitle(run) === run.model_name ? null : run.model_name;
+}
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 85ce25aa97..07be8a247b 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -58,6 +58,7 @@ const initialState: TrainingConfigState = {
   currentStep: MIN_STEP,
   modelType: null,
   selectedModel: null,
+  projectName: "",
   trainingMethod: "qlora",
   hfToken: "",
   datasetSource: "huggingface",
@@ -613,6 +614,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           if (state.modelDefaultsAppliedFor === state.selectedModel) return;
           void loadAndApplyModelDefaults(state.selectedModel);
         },
+        setProjectName: (projectName) => set({ projectName }),
         setTrainingMethod: (trainingMethod) => {
           const state = get();
           set(
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 9eaaa98c0e..acc80a3be2 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -24,6 +24,7 @@ const initialState: TrainingRuntimeState = {
   startError: null,
   startModelName: null,
   startDatasetName: null,
+  startProjectName: null,
   startFromResume: false,
   sseConnected: false,
   firstStepReceived: false,
@@ -125,8 +126,12 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
   setHasHydrated: (value) => set({ hasHydrated: value }),
   setStarting: (value) => set({ isStarting: value }),
   setStartError: (value) => set({ startError: value }),
-  setStartResources: (startModelName, startDatasetName, startFromResume = false) =>
-    set({ startModelName, startDatasetName, startFromResume }),
+  setStartResources: (
+    startModelName,
+    startDatasetName,
+    startFromResume = false,
+    startProjectName = null,
+  ) => set({ startModelName, startDatasetName, startProjectName, startFromResume }),
   setSseConnected: (value) => set({ sseConnected: value }),
   setLastEventId: (value) => set({ lastEventId: value }),
 
diff --git a/studio/frontend/src/features/training/types/api.ts b/studio/frontend/src/features/training/types/api.ts
index 4f7a41bdea..ecd29a2ff0 100644
--- a/studio/frontend/src/features/training/types/api.ts
+++ b/studio/frontend/src/features/training/types/api.ts
@@ -5,6 +5,7 @@ import type { S3Config } from "@/types/training";
 
 export interface TrainingStartRequest {
   model_name: string;
+  project_name: string | null;
   training_type: string;
   hf_token: string | null;
   load_in_4bit: boolean;
diff --git a/studio/frontend/src/features/training/types/config.ts b/studio/frontend/src/features/training/types/config.ts
index d24ce31a6a..5658dfc7a1 100644
--- a/studio/frontend/src/features/training/types/config.ts
+++ b/studio/frontend/src/features/training/types/config.ts
@@ -21,6 +21,7 @@ export interface TrainingConfigState {
   currentStep: StepNumber;
   modelType: ModelType | null;
   selectedModel: string | null;
+  projectName: string;
   trainingMethod: TrainingMethod;
   hfToken: string;
   datasetSource: DatasetSource;
@@ -95,6 +96,7 @@ export interface TrainingConfigActions {
   prevStep: () => void;
   setModelType: (type: ModelType) => void;
   setSelectedModel: (model: string | null) => void;
+  setProjectName: (value: string) => void;
   ensureModelDefaultsLoaded: () => void;
   ensureDatasetChecked: () => void;
   setTrainingMethod: (method: TrainingMethod) => void;
diff --git a/studio/frontend/src/features/training/types/history.ts b/studio/frontend/src/features/training/types/history.ts
index 45d83a31d2..99e9bdcb17 100644
--- a/studio/frontend/src/features/training/types/history.ts
+++ b/studio/frontend/src/features/training/types/history.ts
@@ -5,6 +5,7 @@ export interface TrainingRunSummary {
   id: string;
   status: "running" | "completed" | "stopped" | "error";
   model_name: string;
+  project_name: string | null;
   dataset_name: string;
   display_name: string | null;
   started_at: string;
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index c27a2f2bed..8ed0ce0037 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -83,6 +83,7 @@ export interface TrainingRuntimeState {
   startError: string | null;
   startModelName: string | null;
   startDatasetName: string | null;
+  startProjectName: string | null;
   startFromResume: boolean;
   sseConnected: boolean;
   firstStepReceived: boolean;
@@ -121,6 +122,7 @@ export interface TrainingRuntimeActions {
     modelName: string | null,
     datasetName: string | null,
     fromResume?: boolean,
+    projectName?: string | null,
   ) => void;
   setSseConnected: (value: boolean) => void;
   setLastEventId: (value: number | null) => void;
@@ -160,6 +162,7 @@ export interface TrainingViewData {
 
   // Config summary
   modelName: string;
+  projectName: string | null;
   trainingMethod: string;
 
   // Time-series (for ChartsSection)
diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts
index 086cfc8090..db996768a6 100644
--- a/studio/frontend/src/hooks/use-tauri-backend.ts
+++ b/studio/frontend/src/hooks/use-tauri-backend.ts
@@ -75,8 +75,7 @@ function externalConflictMessage(preflight: DesktopPreflightResult) {
     : "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(
-  invoke: TauriInvoke,
+async function waitForManagedServerPort(
   getPort: () => number | null,
   shouldContinue: () => boolean,
 ): Promise<ManagedStartupResult> {
@@ -91,15 +90,7 @@ async function waitForManagedServerReady(
       continue;
     }
 
-    const healthy = await invoke<boolean>("check_health", { port });
-    if (!shouldContinue()) {
-      return { status: "aborted" };
-    }
-    if (healthy && getPort() === port) {
-      return { status: "ready", port };
-    }
-
-    await wait(MANAGED_STARTUP_POLL_MS);
+    return { status: "ready", port };
   }
 }
 
@@ -280,10 +271,9 @@ export function useTauriBackend() {
       // backend/run.py keeps the 8888-8908 fallback via server-port/TAURI_PORT.
       await invoke("start_managed_server", { port: 8888 });
 
-      // Wait for the owned backend's server-port event. Don't attach to an
-      // external backend if the managed start doesn't report a port.
-      const startupResult = await waitForManagedServerReady(
-        invoke,
+      // Rust emits server-port only after validating the desktop-owned process.
+      // Treat that as the UI handoff point instead of doing a second health poll.
+      const startupResult = await waitForManagedServerPort(
         () => portRef.current,
         () => startingRef.current,
       );
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 019a23a891..f3c9eef44e 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -609,6 +609,10 @@ export const en = {
     params: {
       title: "Parameters",
       description: "Configure training hyperparameters",
+      projectName: "Project Name",
+      optional: "Optional",
+      projectNameDescription:
+        "Used in training output folder names, export defaults, and history.",
       loraSettings: "LoRA Settings",
       trainingHyperparameters: "Training Hyperparameters",
       maxSteps: "Max Steps",
@@ -850,6 +854,7 @@ export const en = {
       loss: "Loss",
       lr: "LR",
       gradNorm: "Grad Norm",
+      project: "Project",
       model: "Model",
       method: "Method",
       elapsed: "Elapsed: {value}",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index bc239294bd..5fd31dc7cb 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -541,6 +541,9 @@ export const zhCN = {
     params: {
       title: "参数",
       description: "配置训练超参数",
+      projectName: "项目名称",
+      optional: "可选",
+      projectNameDescription: "用于训练输出文件夹名称、导出默认值和历史记录。",
       loraSettings: "LoRA 设置",
       trainingHyperparameters: "训练超参数",
       maxSteps: "最大步数",
@@ -766,6 +769,7 @@ export const zhCN = {
       loss: "Loss",
       lr: "LR",
       gradNorm: "梯度范数",
+      project: "项目",
       model: "模型",
       method: "方法",
       elapsed: "已用时间：{value}",
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index c7f34e39f2..e40cb3083e 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -7,6 +7,7 @@
 from __future__ import annotations
 
 import argparse
+import atexit
 import errno
 import fnmatch
 import hashlib
@@ -5203,7 +5204,8 @@ def ldconfig_runtime_dirs(required_libraries: Iterable[str]) -> list[str]:
 
 
 def linux_runtime_dirs(binary_path: Path) -> list[str]:
-    missing = linux_missing_libraries(binary_path)
+    # ldd may execute the binary, so probe it with a secret-free env.
+    missing = linux_missing_libraries(binary_path, env = scrubbed_environ())
     if not missing:
         return []
     return linux_runtime_dirs_for_required_libraries(missing)
@@ -5499,6 +5501,140 @@ def _wsl_system_rocm_lib_dirs() -> list[str]:
     return out
 
 
+# Secrets a downloaded llama.cpp binary never needs; keep them out of binary_env().
+# The installer's own API calls read os.environ directly, so auth is unaffected.
+_SECRET_ENV_EXACT_NAMES = frozenset(
+    {
+        "HF_TOKEN",
+        "HUGGING_FACE_HUB_TOKEN",
+        "GH_TOKEN",
+        "GITHUB_TOKEN",
+        "WANDB_API_KEY",
+        "OPENAI_API_KEY",
+        "ANTHROPIC_API_KEY",
+        "AWS_ACCESS_KEY_ID",
+        "AWS_SECRET_ACCESS_KEY",
+        "AWS_SESSION_TOKEN",
+        "GOOGLE_APPLICATION_CREDENTIALS",
+        "AZURE_CLIENT_SECRET",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN",
+        "ACTIONS_ID_TOKEN_REQUEST_URL",
+        "ACTIONS_RUNTIME_TOKEN",
+        # Credential pointers (cluster / remote-host access).
+        "KUBECONFIG",
+        "SSH_AUTH_SOCK",
+    }
+)
+# Case-insensitive substring markers for names we do not enumerate (no bare "KEY",
+# which would hit benign runtime vars).
+_SECRET_ENV_MARKERS = (
+    "TOKEN",
+    "SECRET",
+    "PASSWORD",
+    "PASSWD",
+    "PASSPHRASE",
+    "CREDENTIAL",
+    "PRIVATE_KEY",
+    "API_KEY",
+)
+# Proxy / index URLs embed creds in their value; the offline binaries never need them.
+_SECRET_ENV_URL_NAMES = frozenset(
+    {
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "ALL_PROXY",
+        "FTP_PROXY",
+        "RSYNC_PROXY",
+        "PIP_INDEX_URL",
+        "PIP_EXTRA_INDEX_URL",
+        "UV_INDEX_URL",
+        "UV_DEFAULT_INDEX",
+        "UV_EXTRA_INDEX_URL",
+    }
+)
+# Also drop values with URL userinfo creds (scheme://user:secret@host or token@host).
+_URL_USERINFO_CREDENTIAL_RE = re.compile(r"://[^/@\s]+@")
+
+
+def is_secret_env_name(name: str) -> bool:
+    upper = name.upper()
+    return (
+        upper in _SECRET_ENV_EXACT_NAMES
+        or upper in _SECRET_ENV_URL_NAMES
+        or any(marker in upper for marker in _SECRET_ENV_MARKERS)
+    )
+
+
+def scrub_env(env: dict[str, str]) -> dict[str, str]:
+    """Drop secret-bearing variables before handing an env to a downloaded binary."""
+    return {
+        key: value
+        for key, value in env.items()
+        if not is_secret_env_name(key) and not _URL_USERINFO_CREDENTIAL_RE.search(value or "")
+    }
+
+
+# Home / cache pointers to on-disk token stores (~/.cache/huggingface/token,
+# ~/.aws/credentials, ...). Stripping env tokens is not enough; point these at an
+# empty home so the binary cannot read those files via $HOME.
+_RUNTIME_HOME_POINTER_VARS = (
+    "HOME",
+    "USERPROFILE",
+    "APPDATA",
+    "LOCALAPPDATA",
+    "XDG_CACHE_HOME",
+    "XDG_CONFIG_HOME",
+    "XDG_DATA_HOME",
+    "HF_HOME",
+    "HUGGINGFACE_HUB_CACHE",
+    "HF_HUB_CACHE",
+)
+# Credential / config file pointers outside HOME; drop so lookups fall back to the
+# empty home.
+_CREDENTIAL_FILE_POINTER_VARS = (
+    "NETRC",
+    "PIP_CONFIG_FILE",
+    "DOCKER_CONFIG",
+    "GIT_CONFIG_GLOBAL",
+)
+# GitHub Actions command files: appending to these injects PATH/env into later steps.
+_CI_COMMAND_FILE_VARS = (
+    "GITHUB_ENV",
+    "GITHUB_PATH",
+    "GITHUB_OUTPUT",
+    "GITHUB_STEP_SUMMARY",
+    "BASH_ENV",
+)
+
+_isolated_runtime_home_dir: str | None = None
+
+
+def isolated_runtime_home() -> str:
+    # Empty dir, created lazily and removed at exit. (A binary resolving the real
+    # home via getpwuid is out of scope; that needs OS sandboxing.)
+    global _isolated_runtime_home_dir
+    if _isolated_runtime_home_dir is None:
+        path = tempfile.mkdtemp(prefix = "unsloth-prebuilt-home-")
+        atexit.register(shutil.rmtree, path, ignore_errors = True)
+        _isolated_runtime_home_dir = path
+    return _isolated_runtime_home_dir
+
+
+def scrubbed_environ() -> dict[str, str]:
+    # os.environ minus secrets, with home / credential pointers neutralised. Used for
+    # the binary env and any probe (e.g. ldd) that runs the untrusted binary.
+    env = scrub_env(os.environ.copy())
+    runtime_home = isolated_runtime_home()
+    for pointer in _RUNTIME_HOME_POINTER_VARS:
+        env[pointer] = runtime_home
+    # Windows rebuilds the profile from %HOMEDRIVE%%HOMEPATH% (no-op pair on POSIX).
+    drive, tail = os.path.splitdrive(runtime_home)
+    env["HOMEDRIVE"], env["HOMEPATH"] = drive, tail or runtime_home
+    for pointer in (*_CREDENTIAL_FILE_POINTER_VARS, *_CI_COMMAND_FILE_VARS):
+        env.pop(pointer, None)
+    return env
+
+
 def binary_env(
     binary_path: Path,
     install_dir: Path,
@@ -5506,7 +5642,7 @@ def binary_env(
     *,
     runtime_line: str | None = None,
 ) -> dict[str, str]:
-    env = os.environ.copy()
+    env = scrubbed_environ()
     if host.is_windows:
         path_dirs = [
             str(binary_path.parent),
diff --git a/studio/src-tauri/src/commands.rs b/studio/src-tauri/src/commands.rs
index 7c1ce611b8..6bc2116786 100644
--- a/studio/src-tauri/src/commands.rs
+++ b/studio/src-tauri/src/commands.rs
@@ -65,10 +65,18 @@ pub async fn desktop_preflight(
     shutdown: tauri::State<'_, ShutdownFlag>,
     diagnostics: tauri::State<'_, DiagnosticsState>,
 ) -> Result<crate::preflight::DesktopPreflightResult, String> {
+    let started = Instant::now();
     let (result, adopted_watchdog_generation) =
         crate::preflight::desktop_preflight_result_with_state(state.inner()).await?;
     diagnostics::record_preflight(&diagnostics, &result);
 
+    info!(
+        "desktop_preflight completed disposition={:?} port={:?} in {}ms",
+        result.disposition,
+        result.port,
+        started.elapsed().as_millis()
+    );
+
     if let Some((generation, newly_adopted)) = adopted_watchdog_generation {
         if newly_adopted {
             if let Some(port) = result.port {
@@ -205,9 +213,17 @@ pub async fn start_managed_server(
     port: u16,
 ) -> Result<(), String> {
     info!("start_managed_server command called with port {}", port);
+
+    let started = Instant::now();
     let diagnostics_state = diagnostics.inner().clone();
     let generation = process::start_backend(&app, &state, port, &shutdown, &diagnostics_state)?;
 
+    info!(
+        "start_managed_server spawned generation={} in {}ms",
+        generation,
+        started.elapsed().as_millis()
+    );
+
     let watchdog_state = state.inner().clone();
     let watchdog_shutdown = shutdown.inner().clone();
     let watchdog_app = app.clone();
diff --git a/studio/src-tauri/src/desktop_backend_owner.rs b/studio/src-tauri/src/desktop_backend_owner.rs
index ed4b5f7fd9..c7d0a7b309 100644
--- a/studio/src-tauri/src/desktop_backend_owner.rs
+++ b/studio/src-tauri/src/desktop_backend_owner.rs
@@ -93,7 +93,7 @@ enum PreviousAppPidStatus {
     Uncertain,
 }
 
-#[derive(Debug, Deserialize)]
+#[derive(Clone, Debug, Deserialize)]
 struct HealthDesktopOwner {
     kind: Option<String>,
     token_sha256: Option<String>,
@@ -101,15 +101,7 @@ struct HealthDesktopOwner {
 
 #[derive(Debug, Deserialize)]
 struct HealthResponse {
-    status: Option<String>,
-    service: Option<String>,
     version: Option<String>,
-    desktop_protocol_version: Option<u16>,
-    desktop_manageability_version: Option<u16>,
-    supports_desktop_auth: Option<bool>,
-    supports_desktop_backend_ownership: Option<bool>,
-    studio_root_id: Option<String>,
-    desktop_owner: Option<HealthDesktopOwner>,
 }
 
 #[derive(Debug)]
@@ -123,6 +115,18 @@ struct DesktopLoginPayload<'a> {
     secret: &'a str,
 }
 
+#[derive(Clone, Debug, Deserialize)]
+pub(crate) struct DesktopLiveness {
+    status: Option<String>,
+    service: Option<String>,
+    desktop_protocol_version: Option<u16>,
+    desktop_manageability_version: Option<u16>,
+    supports_desktop_auth: Option<bool>,
+    supports_desktop_backend_ownership: Option<bool>,
+    studio_root_id: Option<String>,
+    desktop_owner: Option<HealthDesktopOwner>,
+}
+
 #[derive(Deserialize)]
 struct TokenResponse {
     access_token: String,
@@ -290,10 +294,10 @@ impl BackendOwnerState {
     }
 
     pub(crate) fn verifies_exact_port_blocking(&self, port: u16) -> bool {
-        match fetch_health_blocking(port) {
-            Ok(Some(health)) => {
-                health_verifies_metadata(&health, &self.metadata)
-                    && lifecycle_control_block_reason(&health).is_none()
+        match fetch_liveness_blocking(port) {
+            Ok(Some(liveness)) => {
+                liveness_verifies_metadata(&liveness, &self.metadata)
+                    && lifecycle_control_block_reason(&liveness).is_none()
             }
             _ => false,
         }
@@ -498,66 +502,110 @@ pub(crate) fn test_owner_state(root_id: &str, token: &str, port: u16) -> Backend
     }
 }
 
-fn health_verifies_metadata(health: &HealthResponse, metadata: &DesktopBackendMetadata) -> bool {
-    let healthy = health.status.as_deref() == Some("healthy")
-        && health.service.as_deref() == Some("Unsloth UI Backend");
-    let Some(owner) = health.desktop_owner.as_ref() else {
+fn liveness_verifies_metadata(
+    liveness: &DesktopLiveness,
+    metadata: &DesktopBackendMetadata,
+) -> bool {
+    let alive = matches!(liveness.status.as_deref(), Some("alive") | Some("healthy"))
+        && liveness.service.as_deref() == Some("Unsloth UI Backend");
+    let Some(owner) = liveness.desktop_owner.as_ref() else {
         return false;
     };
-    healthy
+    alive
         && owner_matches_metadata(
             metadata,
-            health.studio_root_id.as_deref(),
+            liveness.studio_root_id.as_deref(),
             owner.kind.as_deref(),
             owner.token_sha256.as_deref(),
         )
 }
 
-fn lifecycle_control_block_reason(health: &HealthResponse) -> Option<String> {
-    if health.desktop_protocol_version != Some(crate::preflight::DESKTOP_PROTOCOL_VERSION) {
+fn lifecycle_control_block_reason(liveness: &DesktopLiveness) -> Option<String> {
+    if liveness.desktop_protocol_version != Some(crate::preflight::DESKTOP_PROTOCOL_VERSION) {
         return Some("desktop_protocol_incompatible".to_string());
     }
-    if health.supports_desktop_auth != Some(true) {
+    if liveness.supports_desktop_auth != Some(true) {
         return Some("desktop_auth_unsupported".to_string());
     }
-    if health.desktop_manageability_version.unwrap_or(0)
+    if liveness.desktop_manageability_version.unwrap_or(0)
         < crate::preflight::DESKTOP_MANAGEABILITY_VERSION
     {
         return Some("desktop_manageability_unsupported".to_string());
     }
-    if health.supports_desktop_backend_ownership != Some(true) {
+    if liveness.supports_desktop_backend_ownership != Some(true) {
         return Some("desktop_backend_ownership_unsupported".to_string());
     }
     None
 }
 
-fn ready_for_use_status(health: &HealthResponse) -> OwnedBackendReadiness {
-    match crate::preflight::backend_version_stale_reason(health.version.as_deref()) {
+fn ready_for_use_status(health: Option<&HealthResponse>) -> OwnedBackendReadiness {
+    let version = health
+        .and_then(|h| h.version.as_deref())
+        .filter(|v| !v.is_empty());
+    match crate::preflight::backend_version_stale_reason(version) {
         Some(reason) => OwnedBackendReadiness::Stale { reason },
         None => OwnedBackendReadiness::Ready,
     }
 }
 
-async fn fetch_health(port: u16) -> Result<Option<HealthResponse>, reqwest::Error> {
+async fn health_ready_status(port: u16) -> OwnedBackendReadiness {
+    match fetch_health(port).await {
+        Ok(health) => ready_for_use_status(health.as_ref()),
+        Err(reason) => OwnedBackendReadiness::Stale { reason },
+    }
+}
+
+async fn fetch_liveness(port: u16) -> Result<Option<DesktopLiveness>, reqwest::Error> {
     let client = reqwest::Client::builder()
         .timeout(LOCAL_HTTP_TIMEOUT)
         .build()?;
+    for path in ["/api/liveness", "/api/health"] {
+        let response = client
+            .get(format!("http://127.0.0.1:{port}{path}"))
+            .send()
+            .await?;
+        if response.status() == reqwest::StatusCode::NOT_FOUND && path == "/api/liveness" {
+            continue;
+        }
+        if !response.status().is_success() {
+            return Ok(None);
+        }
+        return response.json::<DesktopLiveness>().await.map(Some);
+    }
+    Ok(None)
+}
+
+fn fetch_liveness_blocking(port: u16) -> Result<Option<DesktopLiveness>, String> {
+    for path in ["/api/liveness", "/api/health"] {
+        let response = http_request_blocking(port, "GET", path, &[], &[])?;
+        if response.status == 404 && path == "/api/liveness" {
+            continue;
+        }
+        if !(200..300).contains(&response.status) {
+            return Ok(None);
+        }
+        return serde_json::from_slice::<DesktopLiveness>(&response.body)
+            .map(Some)
+            .map_err(|e| e.to_string());
+    }
+    Ok(None)
+}
+async fn fetch_health(port: u16) -> Result<Option<HealthResponse>, String> {
+    let client = reqwest::Client::builder()
+        .timeout(LOCAL_HTTP_TIMEOUT)
+        .build()
+        .map_err(|e| e.to_string())?;
     let response = client
         .get(format!("http://127.0.0.1:{port}/api/health"))
         .send()
-        .await?;
+        .await
+        .map_err(|e| e.to_string())?;
     if !response.status().is_success() {
         return Ok(None);
     }
-    response.json::<HealthResponse>().await.map(Some)
-}
-
-fn fetch_health_blocking(port: u16) -> Result<Option<HealthResponse>, String> {
-    let response = http_request_blocking(port, "GET", "/api/health", &[], &[])?;
-    if !(200..300).contains(&response.status) {
-        return Ok(None);
-    }
-    serde_json::from_slice::<HealthResponse>(&response.body)
+    response
+        .json::<HealthResponse>()
+        .await
         .map(Some)
         .map_err(|e| e.to_string())
 }
@@ -618,21 +666,21 @@ pub(crate) async fn probe_owned_backend_state(
     };
     let mut verified = Vec::new();
     for port in ports {
-        let health = match fetch_health(port).await {
-            Ok(Some(health)) => health,
+        let liveness = match fetch_liveness(port).await {
+            Ok(Some(liveness)) => liveness,
             Ok(None) => continue,
             Err(error) => {
                 warn!(
-                    "Desktop-owned backend probe skipped port {} after health error: {}",
+                    "Desktop-owned backend probe skipped port {} after liveness error: {}",
                     port, error
                 );
                 continue;
             }
         };
-        if !health_verifies_metadata(&health, &owner.metadata) {
+        if !liveness_verifies_metadata(&liveness, &owner.metadata) {
             continue;
         }
-        if let Some(reason) = lifecycle_control_block_reason(&health) {
+        if let Some(reason) = lifecycle_control_block_reason(&liveness) {
             return OwnedBackendProbe::Unmanageable { port, reason };
         }
         if !desktop_login_route_compatible(port).await {
@@ -646,7 +694,7 @@ pub(crate) async fn probe_owned_backend_state(
                 return OwnedBackendProbe::Unmanageable { port, reason };
             }
         }
-        verified.push((port, ready_for_use_status(&health)));
+        verified.push((port, health_ready_status(port).await));
     }
 
     if verified.len() != 1 {
@@ -967,12 +1015,11 @@ mod tests {
     }
 
     #[test]
-    fn health_verification_requires_root_kind_and_token_sha() {
+    fn liveness_verification_requires_root_kind_and_token_sha() {
         let metadata = metadata(1, Some(8888));
-        let health = HealthResponse {
-            status: Some("healthy".to_string()),
+        let liveness = DesktopLiveness {
+            status: Some("alive".to_string()),
             service: Some("Unsloth UI Backend".to_string()),
-            version: Some("2026.5.2".to_string()),
             desktop_protocol_version: Some(1),
             desktop_manageability_version: Some(1),
             supports_desktop_auth: Some(true),
@@ -983,12 +1030,12 @@ mod tests {
                 token_sha256: Some(token_sha256(TOKEN)),
             }),
         };
-        assert!(health_verifies_metadata(&health, &metadata));
+        assert!(liveness_verifies_metadata(&liveness, &metadata));
 
-        let mut wrong_root = health;
+        let mut wrong_root = liveness;
         wrong_root.studio_root_id =
             Some("bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb".to_string());
-        assert!(!health_verifies_metadata(&wrong_root, &metadata));
+        assert!(!liveness_verifies_metadata(&wrong_root, &metadata));
     }
 
     #[tokio::test]
diff --git a/studio/src-tauri/src/preflight.rs b/studio/src-tauri/src/preflight.rs
index 9138f30a58..5a48d26632 100644
--- a/studio/src-tauri/src/preflight.rs
+++ b/studio/src-tauri/src/preflight.rs
@@ -193,6 +193,9 @@ pub async fn desktop_preflight_result_with_state(
 
     if let Some(snapshot) = crate::process::owned_backend_snapshot(state)? {
         let Some(owner) = snapshot.owner.clone() else {
+            // TAURI_PORT is emitted only after uvicorn lifespan completes; keep
+            // this ownerless path on full health so auth/bootstrap are ready.
+
             let probe = match snapshot.port {
                 Some(port) => backend::probe_ownerless_spawned_backend(port).await,
                 None => backend,
@@ -494,9 +497,22 @@ mod tests {
         FakeCli { bin, dir }
     }
 
+    #[cfg(unix)]
+    fn remove_managed_capability_cache() {
+        let _ = std::fs::remove_file(
+            dirs::home_dir()
+                .unwrap()
+                .join(".unsloth")
+                .join("studio")
+                .join("desktop_capability_cache.json"),
+        );
+    }
+
     #[cfg(unix)]
     #[tokio::test]
     async fn managed_cli_capability_probe_classifies_core_cases() {
+        remove_managed_capability_cache();
+
         for (name, script, stale_reason) in [
             (
                 "cap-missing",
diff --git a/studio/src-tauri/src/preflight/backend.rs b/studio/src-tauri/src/preflight/backend.rs
index 0277ef5149..5a58142667 100644
--- a/studio/src-tauri/src/preflight/backend.rs
+++ b/studio/src-tauri/src/preflight/backend.rs
@@ -3,7 +3,10 @@ use super::version::{
     backend_version_stale_reason, DESKTOP_MANAGEABILITY_VERSION, DESKTOP_PROTOCOL_VERSION,
 };
 use serde::{Deserialize, Serialize};
+
+use log::info;
 use std::time::Duration;
+use std::time::Instant;
 
 #[derive(Debug, Deserialize)]
 struct DesktopOwnerHealth {
@@ -24,6 +27,7 @@ pub(super) struct BackendHealth {
 }
 
 pub(super) async fn backend_health(client: &reqwest::Client, port: u16) -> Option<BackendHealth> {
+    let started = Instant::now();
     let url = format!("http://127.0.0.1:{port}/api/health");
     let response = client.get(url).send().await.ok()?;
     if !response.status().is_success() {
@@ -40,6 +44,14 @@ pub(super) async fn backend_health(client: &reqwest::Client, port: u16) -> Optio
         .and_then(|v| v.as_str())
         .map(|s| s == "Unsloth UI Backend")
         .unwrap_or(false);
+    info!(
+        "Desktop preflight: health probe on port {} healthy={} service={} in {}ms",
+        port,
+        healthy,
+        service,
+        started.elapsed().as_millis()
+    );
+
     if !healthy || !service {
         return None;
     }
diff --git a/studio/src-tauri/src/preflight/managed.rs b/studio/src-tauri/src/preflight/managed.rs
index 51fde51896..57b8365ec5 100644
--- a/studio/src-tauri/src/preflight/managed.rs
+++ b/studio/src-tauri/src/preflight/managed.rs
@@ -2,14 +2,30 @@ use super::types::ManagedProbe;
 use super::version::{
     backend_version_stale_reason, DESKTOP_MANAGEABILITY_VERSION, DESKTOP_PROTOCOL_VERSION,
 };
-use serde::Deserialize;
+use log::{info, warn};
+use serde::{Deserialize, Serialize};
+use std::fs;
 use std::path::{Path, PathBuf};
 use std::process::Stdio;
-use std::time::Duration;
+use std::time::{Duration, Instant, UNIX_EPOCH};
 use tokio::io::AsyncReadExt;
 use tokio::process::Command;
 
-#[derive(Debug, Deserialize)]
+const MANAGED_CAPABILITY_CACHE_SCHEMA: u16 = 2;
+
+const FNV64_OFFSET_BASIS: u64 = 0xcbf29ce484222325;
+const FNV64_PRIME: u64 = 0x100000001b3;
+const HASHED_MARKER_MAX_BYTES: u64 = 64 * 1024;
+
+const FALLBACK_MARKER_NAMES: &[&str] = &[
+    "pyvenv.cfg",
+    "uv.lock",
+    "requirements.txt",
+    "python.exe",
+    "python",
+];
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
 struct DesktopCapability {
     desktop_protocol_version: Option<u16>,
     desktop_manageability_version: Option<u16>,
@@ -20,7 +36,225 @@ struct DesktopCapability {
     version: Option<String>,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+struct ManagedCapabilityCache {
+    schema: u16,
+    bin_path: String,
+    bin_size: u64,
+    bin_mtime_ms: u64,
+    studio_root_id: Option<String>,
+    marker_path: Option<String>,
+    marker_size: Option<u64>,
+    marker_mtime_ms: Option<u64>,
+    desktop_protocol_version: u16,
+    desktop_manageability_version: u16,
+    capability: DesktopCapability,
+}
+
+#[derive(Debug, Clone)]
+struct MarkerFingerprint {
+    path: String,
+    size: u64,
+    mtime_ms: u64,
+    content_hash: Option<u64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ManagedBinFingerprint {
+    bin_path: String,
+    bin_size: u64,
+    bin_mtime_ms: u64,
+    studio_root_id: Option<String>,
+    marker_path: Option<String>,
+    marker_size: Option<u64>,
+    marker_mtime_ms: Option<u64>,
+}
+
+fn modified_ms(metadata: &fs::Metadata) -> Option<u64> {
+    metadata
+        .modified()
+        .ok()?
+        .duration_since(UNIX_EPOCH)
+        .ok()
+        .and_then(|duration| u64::try_from(duration.as_millis()).ok())
+}
+fn hash_bytes(hash: u64, bytes: &[u8]) -> u64 {
+    bytes.iter().fold(hash, |mut next, byte| {
+        next ^= u64::from(*byte);
+        next.wrapping_mul(FNV64_PRIME)
+    })
+}
+
+fn marker_content_hash(path: &Path, metadata: &fs::Metadata) -> Option<u64> {
+    if metadata.len() > HASHED_MARKER_MAX_BYTES {
+        return None;
+    }
+    fs::read(path)
+        .ok()
+        .map(|bytes| hash_bytes(FNV64_OFFSET_BASIS, &bytes))
+}
+
+fn marker_candidates_for_bin(bin: &Path) -> Vec<PathBuf> {
+    let Some(scripts_dir) = bin.parent() else {
+        return Vec::new();
+    };
+    let Some(venv_dir) = scripts_dir.parent() else {
+        return Vec::new();
+    };
+    let mut out = Vec::new();
+
+    #[cfg(unix)]
+    {
+        if let Ok(lib_dir) = fs::read_dir(venv_dir.join("lib")) {
+            for entry in lib_dir.flatten() {
+                out.push(
+                    entry
+                        .path()
+                        .join("site-packages")
+                        .join("unsloth_cli")
+                        .join("commands")
+                        .join("studio.py"),
+                );
+            }
+        }
+    }
+    for marker_name in FALLBACK_MARKER_NAMES {
+        out.push(venv_dir.join(marker_name));
+        out.push(scripts_dir.join(marker_name));
+    }
+
+    out.push(
+        venv_dir
+            .join("Lib")
+            .join("site-packages")
+            .join("unsloth_cli")
+            .join("commands")
+            .join("studio.py"),
+    );
+    out
+}
+
+fn managed_bin_fingerprint(bin: &Path) -> Option<ManagedBinFingerprint> {
+    let bin_metadata = fs::metadata(bin).ok()?;
+    let bin_path = bin
+        .canonicalize()
+        .unwrap_or_else(|_| bin.to_path_buf())
+        .to_string_lossy()
+        .into_owned();
+
+    let studio_root_id = crate::desktop_backend_owner::read_expected_studio_root_id();
+    let mut marker_entries: Vec<MarkerFingerprint> = marker_candidates_for_bin(bin)
+        .into_iter()
+        .filter_map(|path| {
+            let metadata = fs::metadata(&path).ok()?;
+            Some(MarkerFingerprint {
+                path: path
+                    .canonicalize()
+                    .unwrap_or(path.clone())
+                    .to_string_lossy()
+                    .into_owned(),
+                size: metadata.len(),
+                mtime_ms: modified_ms(&metadata)?,
+                content_hash: marker_content_hash(&path, &metadata),
+            })
+        })
+        .collect();
+    marker_entries.sort_by(|left, right| left.path.cmp(&right.path));
+    let marker_hash = marker_entries
+        .iter()
+        .fold(FNV64_OFFSET_BASIS, |hash, marker| {
+            let next = hash_bytes(hash, marker.path.as_bytes());
+            let next = hash_bytes(next, &marker.size.to_le_bytes());
+            let next = hash_bytes(next, &marker.mtime_ms.to_le_bytes());
+            if let Some(content_hash) = marker.content_hash {
+                hash_bytes(next, &content_hash.to_le_bytes())
+            } else {
+                next
+            }
+        });
+    let marker_path = (!marker_entries.is_empty()).then(|| "markers".to_string());
+    let marker_size = (!marker_entries.is_empty()).then(|| marker_entries.len() as u64);
+    let marker_mtime_ms = (!marker_entries.is_empty()).then_some(marker_hash);
+
+    Some(ManagedBinFingerprint {
+        bin_path,
+        bin_size: bin_metadata.len(),
+        bin_mtime_ms: modified_ms(&bin_metadata)?,
+        studio_root_id,
+        marker_path,
+        marker_size,
+        marker_mtime_ms,
+    })
+}
+
+fn capability_cache_path() -> Option<PathBuf> {
+    dirs::home_dir().map(|home| {
+        home.join(".unsloth")
+            .join("studio")
+            .join("desktop_capability_cache.json")
+    })
+}
+
+fn cache_matches(cache: &ManagedCapabilityCache, fingerprint: &ManagedBinFingerprint) -> bool {
+    cache.schema == MANAGED_CAPABILITY_CACHE_SCHEMA
+        && cache.desktop_protocol_version == DESKTOP_PROTOCOL_VERSION
+        && cache.desktop_manageability_version == DESKTOP_MANAGEABILITY_VERSION
+        && cache.bin_path == fingerprint.bin_path
+        && cache.bin_size == fingerprint.bin_size
+        && cache.bin_mtime_ms == fingerprint.bin_mtime_ms
+        && cache.studio_root_id == fingerprint.studio_root_id
+        && cache.marker_path == fingerprint.marker_path
+        && cache.marker_size == fingerprint.marker_size
+        && cache.marker_mtime_ms == fingerprint.marker_mtime_ms
+        && desktop_capability_ready(&cache.capability)
+}
+
+fn read_cached_capability(fingerprint: &ManagedBinFingerprint) -> Option<DesktopCapability> {
+    let path = capability_cache_path()?;
+    let bytes = fs::read(path).ok()?;
+    let cache = serde_json::from_slice::<ManagedCapabilityCache>(&bytes).ok()?;
+    if cache_matches(&cache, fingerprint) {
+        Some(cache.capability)
+    } else {
+        None
+    }
+}
+
+fn write_cached_capability(fingerprint: &ManagedBinFingerprint, capability: &DesktopCapability) {
+    let Some(path) = capability_cache_path() else {
+        return;
+    };
+    let cache = ManagedCapabilityCache {
+        schema: MANAGED_CAPABILITY_CACHE_SCHEMA,
+        bin_path: fingerprint.bin_path.clone(),
+        bin_size: fingerprint.bin_size,
+        bin_mtime_ms: fingerprint.bin_mtime_ms,
+        studio_root_id: fingerprint.studio_root_id.clone(),
+        marker_path: fingerprint.marker_path.clone(),
+        marker_size: fingerprint.marker_size,
+        marker_mtime_ms: fingerprint.marker_mtime_ms,
+        desktop_protocol_version: DESKTOP_PROTOCOL_VERSION,
+        desktop_manageability_version: DESKTOP_MANAGEABILITY_VERSION,
+        capability: capability.clone(),
+    };
+    if let Some(parent) = path.parent() {
+        if fs::create_dir_all(parent).is_err() {
+            return;
+        }
+    }
+    let Ok(bytes) = serde_json::to_vec_pretty(&cache) else {
+        return;
+    };
+    if let Err(error) = fs::write(&path, bytes) {
+        warn!(
+            "Managed preflight: could not write capability cache: {}",
+            error
+        );
+    }
+}
+
 async fn run_cli_probe(bin: &Path, args: &[&str]) -> bool {
+    let started = Instant::now();
     let mut cmd = Command::new(bin);
     cmd.args(args).stdout(Stdio::null()).stderr(Stdio::null());
 
@@ -43,20 +277,33 @@ async fn run_cli_probe(bin: &Path, args: &[&str]) -> bool {
     }
 
     let Ok(mut child) = cmd.spawn() else {
+        info!(
+            "Managed preflight probe {:?} failed to spawn in {}ms",
+            args,
+            started.elapsed().as_millis()
+        );
         return false;
     };
 
-    match tokio::time::timeout(Duration::from_secs(10), child.wait()).await {
+    let ok = match tokio::time::timeout(Duration::from_secs(10), child.wait()).await {
         Ok(Ok(status)) => status.success(),
         _ => {
             let _ = child.kill().await;
             let _ = child.wait().await;
             false
         }
-    }
+    };
+    info!(
+        "Managed preflight probe {:?} finished ok={} in {}ms",
+        args,
+        ok,
+        started.elapsed().as_millis()
+    );
+    ok
 }
 
 async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
+    let started = Instant::now();
     let mut cmd = Command::new(bin);
     cmd.args(["studio", "desktop-capabilities", "--json"])
         .stdout(Stdio::piped())
@@ -81,6 +328,10 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
     }
 
     let Ok(mut child) = cmd.spawn() else {
+        info!(
+            "Managed desktop-capabilities probe failed to spawn in {}ms",
+            started.elapsed().as_millis()
+        );
         return None;
     };
     let Some(mut stdout) = child.stdout.take() else {
@@ -92,9 +343,19 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
         Err(_) => {
             let _ = child.kill().await;
             let _ = child.wait().await;
+            info!(
+                "Managed desktop-capabilities probe timed out in {}ms",
+                started.elapsed().as_millis()
+            );
+            return None;
+        }
+        _ => {
+            info!(
+                "Managed desktop-capabilities probe exited unsuccessfully in {}ms",
+                started.elapsed().as_millis()
+            );
             return None;
         }
-        _ => return None,
     }
 
     let mut output = Vec::new();
@@ -102,7 +363,13 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
         return None;
     }
 
-    serde_json::from_slice::<DesktopCapability>(&output).ok()
+    let capability = serde_json::from_slice::<DesktopCapability>(&output).ok();
+    info!(
+        "Managed desktop-capabilities probe finished ok={} in {}ms",
+        capability.is_some(),
+        started.elapsed().as_millis()
+    );
+    capability
 }
 
 fn desktop_capability_stale_reason(capability: &DesktopCapability) -> Option<String> {
@@ -132,18 +399,48 @@ fn desktop_capability_ready(capability: &DesktopCapability) -> bool {
 }
 
 pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
+    let started = Instant::now();
     if !run_cli_probe(&bin, &["-h"]).await {
+        info!(
+            "Managed preflight: cli unusable for {:?} in {}ms",
+            bin,
+            started.elapsed().as_millis()
+        );
         return ManagedProbe::Stale {
             bin,
             reason: "cli_unusable".to_string(),
         };
     }
 
-    let capability = probe_cli_capability(&bin).await;
-    if let Some(capability) = capability {
-        if desktop_capability_ready(&capability) {
+    if let Some(fingerprint) = managed_bin_fingerprint(&bin) {
+        if read_cached_capability(&fingerprint).is_some() {
+            info!(
+                "Managed preflight: using cached desktop capability for {:?} in {}ms",
+                bin,
+                started.elapsed().as_millis()
+            );
             return ManagedProbe::Ready { bin };
         }
+    }
+
+    let capability = probe_cli_capability(&bin).await;
+    if let Some(capability) = capability {
+        if let Some(fingerprint) = managed_bin_fingerprint(&bin) {
+            write_cached_capability(&fingerprint, &capability);
+        }
+        if desktop_capability_ready(&capability) {
+            info!(
+                "Managed preflight: cli ready for {:?} in {}ms",
+                bin,
+                started.elapsed().as_millis()
+            );
+            return ManagedProbe::Ready { bin };
+        }
+        info!(
+            "Managed preflight: cli stale for {:?} in {}ms",
+            bin,
+            started.elapsed().as_millis()
+        );
         return ManagedProbe::Stale {
             bin,
             reason: desktop_capability_stale_reason(&capability)
@@ -151,6 +448,11 @@ pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
         };
     }
 
+    info!(
+        "Managed preflight: desktop capability probe failed for {:?} in {}ms",
+        bin,
+        started.elapsed().as_millis()
+    );
     ManagedProbe::Stale {
         bin,
         reason: "desktop_capability_probe_failed".to_string(),
@@ -158,10 +460,17 @@ pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
 }
 
 pub(super) async fn probe_managed_install() -> ManagedProbe {
-    match crate::process::find_unsloth_binary() {
+    let started = Instant::now();
+    let result = match crate::process::find_unsloth_binary() {
         Some(bin) => probe_managed_bin(bin).await,
         None => ManagedProbe::Missing,
-    }
+    };
+    info!(
+        "Managed preflight: install probe result {:?} in {}ms",
+        result,
+        started.elapsed().as_millis()
+    );
+    result
 }
 
 pub async fn managed_install_ready() -> bool {
diff --git a/studio/src-tauri/src/process.rs b/studio/src-tauri/src/process.rs
index 01ef77f098..56d9dd2e21 100644
--- a/studio/src-tauri/src/process.rs
+++ b/studio/src-tauri/src/process.rs
@@ -735,6 +735,7 @@ pub fn start_backend(
 }
 
 async fn generic_backend_health_ok(port: u16) -> bool {
+    let started = std::time::Instant::now();
     let client = match reqwest::Client::builder()
         .timeout(Duration::from_secs(2))
         .build()
@@ -745,49 +746,75 @@ async fn generic_backend_health_ok(port: u16) -> bool {
             return false;
         }
     };
-    let response = match client
-        .get(format!("http://127.0.0.1:{port}/api/health"))
-        .send()
-        .await
-    {
-        Ok(response) => response,
-        Err(error) => {
+    let mut last_status = None;
+    let mut json = None;
+    for path in ["/api/liveness", "/api/health"] {
+        let response = match client
+            .get(format!("http://127.0.0.1:{port}{path}"))
+            .send()
+            .await
+        {
+            Ok(response) => response,
+            Err(error) => {
+                warn!(
+                    "Backend port candidate {} failed health request: {}",
+                    port, error
+                );
+                return false;
+            }
+        };
+        if response.status() == reqwest::StatusCode::NOT_FOUND && path == "/api/liveness" {
+            last_status = Some(response.status());
+            continue;
+        }
+        if !response.status().is_success() {
             warn!(
-                "Backend port candidate {} failed health request: {}",
-                port, error
+                "Backend port candidate {} returned HTTP {} from health",
+                port,
+                response.status()
             );
             return false;
         }
-    };
-    if !response.status().is_success() {
+        json = match response.json::<serde_json::Value>().await {
+            Ok(json) => Some(json),
+            Err(error) => {
+                warn!(
+                    "Backend port candidate {} returned invalid health JSON: {}",
+                    port, error
+                );
+                return false;
+            }
+        };
+        break;
+    }
+    let Some(json) = json else {
         warn!(
             "Backend port candidate {} returned HTTP {} from health",
             port,
-            response.status()
+            last_status
+                .map(|status| status.to_string())
+                .unwrap_or_else(|| "unknown".to_string())
         );
         return false;
-    }
-    let json = match response.json::<serde_json::Value>().await {
-        Ok(json) => json,
-        Err(error) => {
-            warn!(
-                "Backend port candidate {} returned invalid health JSON: {}",
-                port, error
-            );
-            return false;
-        }
     };
-    let healthy = json
+    let live = json
         .get("status")
         .and_then(|v| v.as_str())
-        .map(|s| s == "healthy")
+        .map(|s| s == "alive" || s == "healthy")
         .unwrap_or(false);
     let service = json
         .get("service")
         .and_then(|v| v.as_str())
         .map(|s| s == "Unsloth UI Backend")
         .unwrap_or(false);
-    healthy && service
+    info!(
+        "Backend port candidate {} liveness live={} service={} in {}ms",
+        port,
+        live,
+        service,
+        started.elapsed().as_millis()
+    );
+    live && service
 }
 
 async fn validate_candidate_port(
@@ -798,6 +825,7 @@ async fn validate_candidate_port(
     generation: u64,
     port: u16,
 ) {
+    let started = std::time::Instant::now();
     let owner = {
         let proc = match state.lock() {
             Ok(proc) => proc,
@@ -852,6 +880,14 @@ async fn validate_candidate_port(
         }
     };
 
+    info!(
+        "Validated backend port candidate {} valid={} emit={} in {}ms",
+        port,
+        valid,
+        should_emit,
+        started.elapsed().as_millis()
+    );
+
     if should_emit {
         diagnostics::record_backend_port(&diagnostics_state, &session_id, port);
         info!("Validated backend port: {}", port);
diff --git a/tests/python/test_change_system_message.py b/tests/python/test_change_system_message.py
new file mode 100644
index 0000000000..1ccd6547b2
--- /dev/null
+++ b/tests/python/test_change_system_message.py
@@ -0,0 +1,76 @@
+import ast
+import re
+import types
+from pathlib import Path
+
+import pytest
+
+
+def _load_change_system_message():
+    # Extract just _change_system_message from chat_templates.py so the test runs
+    # without importing unsloth (which needs unsloth_zoo / a GPU). Same pattern as
+    # tests/saving/test_is_gpt_oss_detection.py.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "_change_system_message"
+    ]
+    namespace = {
+        "re": re,
+        "logger": types.SimpleNamespace(warning_once = lambda *a, **k: None),
+        "DEFAULT_SYSTEM_MESSAGE": {"unsloth": "You are a helpful assistant to the user"},
+    }
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_change_system_message"]
+
+
+CUSTOM = "mycustom"  # not in DEFAULT_SYSTEM_MESSAGE -> no predefined default
+
+
+def test_custom_template_fills_placeholder():
+    # A custom template with a {system_message} placeholder must be filled, not
+    # left with the literal placeholder.
+    fn = _load_change_system_message()
+    template, used = fn("System: {system_message}\nUser:", CUSTOM, "You are a pirate")
+    assert template == "System: You are a pirate\nUser:"
+    assert "{system_message}" not in template
+    assert used == "You are a pirate"
+
+
+def test_custom_template_preserves_backslashes():
+    # Why str.replace and not re.sub: a system message with backslashes (Windows
+    # paths, LaTeX, group-like text) must be inserted verbatim. re.sub treats the
+    # replacement specially -- r"C:\Users" raises bad-escape, r"\1" is a group ref.
+    fn = _load_change_system_message()
+    for msg in (r"C:\Users\me", r"\frac{a}{b}", r"see \1 here"):
+        template, used = fn("System: {system_message}", CUSTOM, msg)
+        assert template == f"System: {msg}"
+        assert used == msg
+
+
+def test_custom_template_requires_system_message():
+    # A custom template with a placeholder but no system message must raise,
+    # rather than silently leaving the placeholder in.
+    fn = _load_change_system_message()
+    with pytest.raises(ValueError):
+        fn("System: {system_message}", CUSTOM, None)
+
+
+def test_custom_template_without_placeholder_unchanged():
+    fn = _load_change_system_message()
+    template, used = fn("System: fixed", CUSTOM, "ignored")
+    assert template == "System: fixed"
+
+
+def test_predefined_template_uses_default_then_override():
+    # Predefined templates with a default are unaffected by the change.
+    fn = _load_change_system_message()
+    t1, u1 = fn("System: {system_message}", "unsloth", None)
+    assert t1 == "System: You are a helpful assistant to the user"
+    t2, u2 = fn("System: {system_message}", "unsloth", "Custom override")
+    assert t2 == "System: Custom override"
+    assert u2 == "Custom override"
diff --git a/tests/python/test_v100_fullft_precision.py b/tests/python/test_v100_fullft_precision.py
new file mode 100644
index 0000000000..c8ca769d45
--- /dev/null
+++ b/tests/python/test_v100_fullft_precision.py
@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression tests for full finetuning precision on no-bf16 GPUs (V100/T4).
+
+Full finetuning upcasts trainable weights to float32, so the model dtype is
+float32 (not bfloat16). The SFTTrainer mixed-precision template in
+unsloth/models/rl.py must then:
+  - run the forward pass under float16 autocast for normal models,
+  - keep FORCE_FLOAT32 models (Gemma3, gpt_oss, ...) in pure float32,
+  - never select bf16 on hardware without bf16.
+
+We execute the REAL template block extracted from rl.py source (no heavy unsloth
+import) against mocked inputs. See issue #4082.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+torch = pytest.importorskip("torch")
+
+RL_PY = Path(__file__).resolve().parents[2] / "unsloth" / "models" / "rl.py"
+
+
+def _extract_mixed_precision_code() -> str:
+    lines = RL_PY.read_text().split("\n")
+    try:
+        start = next(i for i, l in enumerate(lines) if "mixed_precision = (" in l)
+    except StopIteration:
+        pytest.skip("mixed_precision template not found in rl.py")
+    body, k = [], start + 1
+    while lines[k].strip() != ")":
+        body.append(lines[k])
+        k += 1
+    return eval("(\n" + "\n".join(body) + "\n)")  # only string literals + comments
+
+
+CODE = _extract_mixed_precision_code()
+
+
+def _restore(mapping, saved):
+    """Restore a dict-like to its saved snapshot: pop keys that were absent."""
+    for k, v in saved.items():
+        if v is None:
+            mapping.pop(k, None)
+        else:
+            mapping[k] = v
+
+
+def _decide(dtype, *, bf16_supported, force_float32, full_finetuning, mixed_precision, fp16, bf16):
+    """Run the template block; return (args.fp16, args.bf16, ACCELERATE_MP, raised).
+
+    Stubs (sys.modules, env vars, torch.cuda.is_bf16_supported) are restored on
+    exit so a decision can't leak into later tests in the same process.
+    """
+    uzu = types.ModuleType("unsloth_zoo.utils")
+    uzu._get_dtype = lambda x: x
+    uzd = types.ModuleType("unsloth_zoo.device_type")
+    uzd.device_is_bf16_supported = lambda: bf16_supported  # device-aware signal stub
+
+    env_keys = (
+        "UNSLOTH_FORCE_FLOAT32",
+        "UNSLOTH_ENABLE_FULL_FINETUNING",
+        "UNSLOTH_MIXED_PRECISION",
+        "ACCELERATE_MIXED_PRECISION",
+    )
+    mod_keys = ("unsloth_zoo", "unsloth_zoo.utils", "unsloth_zoo.device_type")
+    saved_env = {k: os.environ.get(k) for k in env_keys}
+    saved_mods = {k: sys.modules.get(k) for k in mod_keys}
+    orig_bf16 = torch.cuda.is_bf16_supported
+    try:
+        sys.modules.setdefault("unsloth_zoo", types.ModuleType("unsloth_zoo"))
+        sys.modules["unsloth_zoo.utils"] = uzu
+        sys.modules["unsloth_zoo.device_type"] = uzd
+        for k in env_keys:
+            os.environ.pop(k, None)
+        os.environ["UNSLOTH_FORCE_FLOAT32"] = "1" if force_float32 else "0"
+        os.environ["UNSLOTH_ENABLE_FULL_FINETUNING"] = "1" if full_finetuning else "0"
+        os.environ["UNSLOTH_MIXED_PRECISION"] = mixed_precision
+        torch.cuda.is_bf16_supported = lambda *a, **k: bf16_supported
+        args = types.SimpleNamespace(fp16 = fp16, bf16 = bf16, mixed_precision = None)
+        emb = types.SimpleNamespace(weight = types.SimpleNamespace(dtype = dtype))
+        model = types.SimpleNamespace(
+            config = types.SimpleNamespace(dtype = dtype, torch_dtype = dtype),
+            get_input_embeddings = lambda: emb,
+        )
+        raised = None
+        try:
+            exec(CODE, {"torch": torch, "os": os}, {"args": args, "model": model})
+        except TypeError:
+            raised = "TypeError"
+        return args.fp16, args.bf16, os.environ.get("ACCELERATE_MIXED_PRECISION"), raised
+    finally:
+        torch.cuda.is_bf16_supported = orig_bf16
+        _restore(os.environ, saved_env)
+        _restore(sys.modules, saved_mods)
+
+
+def test_v100_normal_fullft_fp16_explicit():
+    # Normal model, full FT (weights upcast to float32), V100, fp16=True.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (True, False)  # float32 weights + fp16 forward
+
+
+def test_v100_normal_fullft_precision_unset():
+    # Same, but user left precision unset -> must pick fp16, never bf16.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (True, False)
+    assert amp == "fp16"
+
+
+def test_force_float32_model_fullft_is_pure_float32():
+    # FORCE_FLOAT32 model (Gemma3, gpt_oss, ...) in full FT -> pure float32, no autocast.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = True,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (False, False)
+    assert amp in (None, "no")
+
+
+def test_no_bf16_on_volta_in_auto_branch():
+    # bf16 model dtype but no bf16 HW, precision unset -> fp16, never bf16.
+    fp16, bf16, amp, raised = _decide(
+        torch.bfloat16,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = False,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert bf16 is False
+
+
+def test_bf16_gpu_unchanged_auto_branch():
+    # Regression guard: on a bf16 GPU, a float32 model with unset precision
+    # still selects bf16 autocast (behavior must not change for bf16 hardware).
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = True,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (False, True)
+
+
+def test_genuine_bf16_model_with_fp16_still_raises():
+    # A real bfloat16 model on bf16 HW with fp16 requested is a genuine mismatch.
+    _, _, _, raised = _decide(
+        torch.bfloat16,
+        bf16_supported = True,
+        force_float32 = False,
+        full_finetuning = False,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised == "TypeError"
diff --git a/tests/saving/test_is_gpt_oss_detection.py b/tests/saving/test_is_gpt_oss_detection.py
new file mode 100644
index 0000000000..c8790e4da1
--- /dev/null
+++ b/tests/saving/test_is_gpt_oss_detection.py
@@ -0,0 +1,52 @@
+import ast
+import types
+from pathlib import Path
+
+
+def _load_is_gpt_oss():
+    # Extract just the helper from save.py so the test runs without importing
+    # unsloth (which requires unsloth_zoo / a GPU), matching the pattern used by
+    # test_qwen3_5_vlm_full_finetune_key_remap.py.
+    source = Path(__file__).parents[2] / "unsloth" / "save.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    helpers = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "_is_gpt_oss"
+    ]
+    module = ast.Module(body = helpers, type_ignores = [])
+    ast.fix_missing_locations(module)
+    namespace = {}
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_is_gpt_oss"]
+
+
+def _model(architectures = None, model_type = None):
+    config = types.SimpleNamespace()
+    if architectures is not None:
+        config.architectures = architectures
+    if model_type is not None:
+        config.model_type = model_type
+    return types.SimpleNamespace(config = config)
+
+
+def test_detects_gpt_oss_by_architecture():
+    # config.architectures is a list, so detection must use membership, not ==.
+    # A model that declares GptOssForCausalLM but has no matching model_type must
+    # still be routed to the mxfp4 save path.
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"])) is True
+    assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"], model_type = "gpt_oss")) is True
+
+
+def test_detects_gpt_oss_by_model_type():
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["SomethingElse"], model_type = "gpt-oss")) is True
+    assert is_gpt_oss(_model(architectures = ["SomethingElse"], model_type = "gpt_oss")) is True
+
+
+def test_non_gpt_oss_is_false():
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["LlamaForCausalLM"], model_type = "llama")) is False
+    assert is_gpt_oss(_model()) is False
+    assert is_gpt_oss(types.SimpleNamespace()) is False
diff --git a/tests/studio/install/test_install_llama_prebuilt_logic.py b/tests/studio/install/test_install_llama_prebuilt_logic.py
index 9ee8759bb4..c852d7c495 100644
--- a/tests/studio/install/test_install_llama_prebuilt_logic.py
+++ b/tests/studio/install/test_install_llama_prebuilt_logic.py
@@ -22,6 +22,9 @@ SPEC.loader.exec_module(INSTALL_LLAMA_PREBUILT)
 PrebuiltFallback = INSTALL_LLAMA_PREBUILT.PrebuiltFallback
 extract_archive = INSTALL_LLAMA_PREBUILT.extract_archive
 binary_env = INSTALL_LLAMA_PREBUILT.binary_env
+is_secret_env_name = INSTALL_LLAMA_PREBUILT.is_secret_env_name
+scrub_env = INSTALL_LLAMA_PREBUILT.scrub_env
+isolated_runtime_home = INSTALL_LLAMA_PREBUILT.isolated_runtime_home
 HostInfo = INSTALL_LLAMA_PREBUILT.HostInfo
 AssetChoice = INSTALL_LLAMA_PREBUILT.AssetChoice
 ApprovedArtifactHash = INSTALL_LLAMA_PREBUILT.ApprovedArtifactHash
@@ -779,6 +782,259 @@ def test_binary_env_linux_includes_binary_parent_in_ld_library_path(
     assert str(install_dir) in ld_dirs
 
 
+def test_scrub_env_drops_secrets_and_keeps_runtime_vars():
+    raw = {
+        # secrets
+        "HF_TOKEN": "hf_x",
+        "HUGGING_FACE_HUB_TOKEN": "hf_y",
+        "GH_TOKEN": "gh_x",
+        "GITHUB_TOKEN": "gh_y",
+        "WANDB_API_KEY": "wandb_x",
+        "AWS_SECRET_ACCESS_KEY": "aws_x",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN": "oidc_x",
+        "ACTIONS_ID_TOKEN_REQUEST_URL": "https://oidc",
+        "SOME_VENDOR_API_KEY": "vendor_x",
+        "DB_PASSWORD": "pw",
+        "MY_PRIVATE_KEY": "pk",
+        "KUBECONFIG": "/home/runner/.kube/config",
+        "SSH_AUTH_SOCK": "/tmp/ssh-agent.sock",
+        "SSH_PASSPHRASE": "ssh_pass",
+        # runtime vars to keep
+        "PATH": "/usr/bin",
+        "LD_LIBRARY_PATH": "/opt/lib",
+        "DYLD_LIBRARY_PATH": "/opt/dyld",
+        "HOME": "/home/runner",
+        "TMPDIR": "/tmp",
+        "CUDA_VISIBLE_DEVICES": "0",
+        "HSA_OVERRIDE_GFX_VERSION": "11.0.0",
+    }
+
+    cleaned = scrub_env(raw)
+
+    for secret in (
+        "HF_TOKEN",
+        "HUGGING_FACE_HUB_TOKEN",
+        "GH_TOKEN",
+        "GITHUB_TOKEN",
+        "WANDB_API_KEY",
+        "AWS_SECRET_ACCESS_KEY",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN",
+        "ACTIONS_ID_TOKEN_REQUEST_URL",
+        "SOME_VENDOR_API_KEY",
+        "DB_PASSWORD",
+        "MY_PRIVATE_KEY",
+        "KUBECONFIG",
+        "SSH_AUTH_SOCK",
+        "SSH_PASSPHRASE",
+    ):
+        assert secret not in cleaned, f"{secret} must be stripped from binary env"
+
+    for keep in (
+        "PATH",
+        "LD_LIBRARY_PATH",
+        "DYLD_LIBRARY_PATH",
+        "HOME",
+        "TMPDIR",
+        "CUDA_VISIBLE_DEVICES",
+        "HSA_OVERRIDE_GFX_VERSION",
+    ):
+        assert cleaned[keep] == raw[keep], f"{keep} must be preserved for the binary"
+
+    # no bare "KEY" marker: benign KEY-containing names survive
+    assert is_secret_env_name("API_KEY") is True
+    assert is_secret_env_name("SSH_KEYFILE_PATH") is False
+    assert is_secret_env_name("PATH") is False
+
+
+def test_scrub_env_drops_proxy_index_and_embedded_url_credentials():
+    raw = {
+        # proxy / package-index URLs whose values commonly embed credentials
+        "HTTPS_PROXY": "https://user:secret@proxy:8080",
+        "https_proxy": "https://user:secret@proxy:8080",  # lower-case variant
+        "ALL_PROXY": "socks5://user:secret@proxy:1080",
+        "PIP_INDEX_URL": "https://u:p@pypi.internal/simple",
+        "UV_INDEX_URL": "https://u:p@index.internal/simple",
+        # credentials embedded in an otherwise benign-named variable's value
+        "MY_DB_DSN": "postgres://admin:secret@db:5432/app",
+        # benign vars the binary needs, including a URL with no userinfo
+        "PATH": "/usr/bin",
+        "CUDA_VISIBLE_DEVICES": "0",
+        "NO_PROXY": "localhost,127.0.0.1",
+        "SOME_ENDPOINT": "https://example.com:8080/v1",
+    }
+
+    cleaned = scrub_env(raw)
+
+    for secret in (
+        "HTTPS_PROXY",
+        "https_proxy",
+        "ALL_PROXY",
+        "PIP_INDEX_URL",
+        "UV_INDEX_URL",
+        "MY_DB_DSN",
+    ):
+        assert secret not in cleaned, f"{secret} must be stripped from binary env"
+    for keep in ("PATH", "CUDA_VISIBLE_DEVICES", "NO_PROXY", "SOME_ENDPOINT"):
+        assert cleaned[keep] == raw[keep], f"{keep} must be preserved for the binary"
+
+    assert is_secret_env_name("HTTPS_PROXY") is True
+    assert is_secret_env_name("https_proxy") is True
+    assert is_secret_env_name("NO_PROXY") is False
+
+
+def test_binary_env_strips_secrets_from_downloaded_binary_environment(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "llama.cpp"
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    binary_path = bin_dir / "llama-server"
+    binary_path.write_bytes(b"fake")
+
+    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 = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+
+    monkeypatch.setenv("HF_TOKEN", "hf_secret_from_ci")
+    monkeypatch.setenv("GITHUB_TOKEN", "gh_secret_from_ci")
+    monkeypatch.setenv("GH_TOKEN", "gh_secret_from_ci")
+    monkeypatch.setenv("WANDB_API_KEY", "wandb_secret_from_ci")
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "1")
+
+    env = binary_env(binary_path, install_dir, host)
+
+    assert "HF_TOKEN" not in env
+    assert "GITHUB_TOKEN" not in env
+    assert "GH_TOKEN" not in env
+    assert "WANDB_API_KEY" not in env
+    # library/runtime resolution unaffected
+    assert str(bin_dir) in env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert env["CUDA_VISIBLE_DEVICES"] == "1"
+
+
+def test_binary_env_redirects_home_away_from_real_credential_stores(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "llama.cpp"
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    binary_path = bin_dir / "llama-server"
+    binary_path.write_bytes(b"fake")
+
+    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 = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+
+    real_home = str(tmp_path / "real_home")
+    monkeypatch.setenv("HOME", real_home)
+    monkeypatch.setenv("HF_HOME", real_home + "/.cache/huggingface")
+
+    env = binary_env(binary_path, install_dir, host)
+
+    # HOME and the cache pointers are redirected to a single empty, existing dir.
+    assert env["HOME"] != real_home
+    assert env["HF_HOME"] == env["HOME"]
+    assert env["HOME"] == isolated_runtime_home()
+    assert os.path.isdir(env["HOME"])
+    assert os.listdir(env["HOME"]) == []
+    # Windows reconstructs the profile from HOMEDRIVE + HOMEPATH.
+    assert env["HOMEDRIVE"] + env["HOMEPATH"] == env["HOME"]
+
+
+def test_scrub_env_drops_token_only_url_userinfo():
+    raw = {
+        "GENERIC_REPO": "https://ghp_tokenonly@github.com/org/repo",
+        "GENERIC_OK": "https://example.com:8080/v1",
+    }
+    cleaned = scrub_env(raw)
+    assert "GENERIC_REPO" not in cleaned
+    assert cleaned["GENERIC_OK"] == raw["GENERIC_OK"]
+
+
+def test_binary_env_drops_explicit_credential_file_pointers(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    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 = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+    dropped = (
+        "NETRC",
+        "PIP_CONFIG_FILE",
+        "DOCKER_CONFIG",
+        "GIT_CONFIG_GLOBAL",
+        "GITHUB_ENV",
+        "GITHUB_PATH",
+        "GITHUB_OUTPUT",
+        "GITHUB_STEP_SUMMARY",
+        "BASH_ENV",
+    )
+    for var in dropped:
+        monkeypatch.setenv(var, "/home/realuser/secret")
+
+    env = binary_env(tmp_path / "llama-server", tmp_path, host)
+
+    for var in dropped:
+        assert var not in env
+
+
+def test_linux_runtime_dirs_probes_with_secret_free_env(monkeypatch: pytest.MonkeyPatch):
+    captured: dict[str, object] = {}
+
+    def fake_missing(binary_path, *, env = None):
+        captured["env"] = env
+        return []
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_missing_libraries", fake_missing)
+    monkeypatch.setenv("HF_TOKEN", "hf_secret")
+    monkeypatch.setenv("GITHUB_TOKEN", "gh_secret")
+
+    INSTALL_LLAMA_PREBUILT.linux_runtime_dirs(Path("/fake/llama-server"))
+
+    probe_env = captured["env"]
+    assert probe_env is not None
+    assert "HF_TOKEN" not in probe_env
+    assert "GITHUB_TOKEN" not in probe_env
+
+
 def test_install_prebuilt_falls_back_to_older_release_plan(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index a892d3414d..a53534acc0 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1225,15 +1225,17 @@ with sync_playwright() as p:
     # ─────────────────────────────────────────────────────
     step("Shutdown via account menu")
     # Re-login with NEW2 for a valid /api/shutdown token (CLI rotation
-    # invalidated the old one). The stale token can make the SPA auth
-    # guard abort this goto with ERR_ABORTED; resolve on
-    # domcontentloaded and tolerate it -- the pw-field wait confirms /login.
+    # invalidated the old one). The stale token can make the SPA auth guard
+    # abort this goto with ERR_ABORTED, or redirect to the same /login URL
+    # ("interrupted by another navigation"); resolve on domcontentloaded and
+    # tolerate either -- the pw-field wait below confirms we are on /login.
+    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
     try:
         page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
     except Exception as exc:
-        if "ERR_ABORTED" not in str(exc):
+        if not any(t in str(exc) for t in _tolerated_nav):
             raise
-        info(f"goto /login aborted ({exc!r}); password-field wait will confirm /login")
+        info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
     pw_field = page.locator("#password")
     pw_field.wait_for(state = "visible", timeout = 60_000)
     pw_field.fill(NEW2)
diff --git a/tests/test_vllm_broken_detection.py b/tests/test_vllm_broken_detection.py
new file mode 100644
index 0000000000..ee89ddbedf
--- /dev/null
+++ b/tests/test_vllm_broken_detection.py
@@ -0,0 +1,166 @@
+# 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.
+
+"""Regression test for #6590: modern vLLM lazy-loads its compiled extensions, so
+a bare ``import vllm`` succeeds even when ``vllm._C`` (or a sibling) is ABI-broken
+and ``disable_broken_vllm`` missed it. GPU-free, via a synthetic vLLM."""
+
+from __future__ import annotations
+
+import contextlib
+import importlib.abc
+import importlib.machinery
+import importlib.util
+import sys
+import types
+
+import pytest
+
+
+_LIBCUDART_ERROR = "libcudart.so.13: cannot open shared object file: No such file or directory"
+
+
+class _ExtensionLoader(importlib.abc.Loader):
+    """A compiled extension that loads cleanly or fails on dlopen."""
+
+    def __init__(self, broken, error):
+        self.broken = broken
+        self.error = error
+
+    def create_module(self, spec):
+        return None
+
+    def exec_module(self, module):
+        if self.broken:
+            raise ImportError(self.error)
+
+
+class _FakeVllmFinder(importlib.abc.MetaPathFinder):
+    """Lazy vLLM: ``import vllm`` succeeds; each ``vllm._*`` ext is healthy,
+    ABI-broken, or absent, as real vLLM only loads ``_C`` & friends on use."""
+
+    def __init__(self, present, broken, error):
+        self.present = present
+        self.broken = broken
+        self.error = error
+
+    def find_spec(
+        self,
+        fullname,
+        path = None,
+        target = None,
+    ):
+        if fullname in self.present:
+            return importlib.machinery.ModuleSpec(
+                name = fullname,
+                loader = _ExtensionLoader(broken = fullname in self.broken, error = self.error),
+                is_package = False,
+            )
+        return None  # absent -> ModuleNotFoundError, which the guard ignores
+
+
+@contextlib.contextmanager
+def _fake_vllm(
+    present,
+    broken,
+    error = _LIBCUDART_ERROR,
+):
+    """Install a synthetic lazy vLLM, restoring VLLM_BROKEN, find_spec,
+    meta_path, and the vllm* sys.modules entries on exit."""
+    from unsloth import import_fixes
+
+    submodules = import_fixes._VLLM_COMPILED_EXTENSIONS
+    saved_meta_path = list(sys.meta_path)
+    saved_find_spec = importlib.util.find_spec
+    saved_broken = import_fixes.VLLM_BROKEN
+    saved_modules = {n: sys.modules.get(n) for n in ("vllm", *submodules)}
+    try:
+        import_fixes.VLLM_BROKEN = False
+        fake_vllm = types.ModuleType("vllm")
+        fake_vllm.__path__ = []
+        fake_vllm.__spec__ = importlib.machinery.ModuleSpec("vllm", loader = None, is_package = True)
+        sys.modules["vllm"] = fake_vllm
+        for name in submodules:
+            sys.modules.pop(name, None)
+        sys.meta_path.insert(0, _FakeVllmFinder(present, broken, error))
+        yield import_fixes
+    finally:
+        import_fixes.VLLM_BROKEN = saved_broken
+        sys.meta_path[:] = saved_meta_path
+        importlib.util.find_spec = saved_find_spec
+        for name, module in saved_modules.items():
+            if module is None:
+                sys.modules.pop(name, None)
+            else:
+                sys.modules[name] = module
+
+
+@pytest.mark.parametrize(
+    "broken_ext",
+    ["vllm._C", "vllm._C_stable_libtorch"],
+    ids = ["core_C", "sibling_C_stable_libtorch"],
+)
+def test_disable_broken_vllm_detects_lazy_loaded_broken_extension(broken_ext):
+    # A CUDA-major mismatch breaks every ext; whichever one loads first must trip detection.
+    present = {"vllm._C", "vllm._C_stable_libtorch"}
+    with _fake_vllm(present = present, broken = {broken_ext}) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is True, (
+            f"disable_broken_vllm missed an ABI-broken {broken_ext} behind a "
+            "lazily-importable vllm package — issue #6590 would resurface."
+        )
+        assert import_fixes.VLLM_BROKEN is True
+        # Once disabled, vLLM must look absent so callers fall back cleanly.
+        assert importlib.util.find_spec("vllm") is None
+
+
+@pytest.mark.parametrize(
+    "error",
+    [
+        "libnccl.so.2: cannot open shared object file: No such file or directory",
+        "libcuda.so.1: cannot open shared object file: No such file or directory",
+    ],
+    ids = ["libnccl", "libcuda"],
+)
+def test_disable_broken_vllm_detects_non_cudart_so_failure(error):
+    # A CUDA mismatch can surface through a non-libcudart .so (libnccl, libcuda),
+    # which the old libcudart/libcublas/libnvrtc allow-list let slip through.
+    with _fake_vllm(present = {"vllm._C"}, broken = {"vllm._C"}, error = error) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is True, (
+            f"disable_broken_vllm missed a present-but-broken vllm._C raising "
+            f"{error!r} — vLLM would be left enabled and crash later."
+        )
+        assert import_fixes.VLLM_BROKEN is True
+
+
+@pytest.mark.parametrize(
+    "present",
+    [{"vllm._C"}, {"vllm._C", "vllm._C_stable_libtorch", "vllm._moe_C"}],
+    ids = ["core_only", "all_present"],
+)
+def test_disable_broken_vllm_keeps_healthy_vllm_enabled(present):
+    # Healthy install: an absent sibling (ModuleNotFoundError) or an extra present
+    # ext that loads cleanly must NOT be mistaken for an ABI break.
+    with _fake_vllm(present = present, broken = set()) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is False
+        assert import_fixes.VLLM_BROKEN is False
+        assert importlib.util.find_spec("vllm") is not None
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index 917717e08e..7f080336aa 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -36,16 +36,15 @@ from .import_fixes import (
     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.
+# Redirect a read-only Hugging Face cache before anything below imports
+# huggingface_hub / transformers / vllm (disable_broken_vllm probes `import vllm`
+# and its compiled extensions, check_fbgemm_gpu_version imports transformers,
+# fix_huggingface_hub imports huggingface_hub) -- any of which would freeze Hub's
+# cache constants with the un-redirected paths. unsloth_zoo runs the same redirect
+# at import, but only after these probes. hf_cache.py is stdlib-only, so load it
+# straight from its file without triggering the full unsloth_zoo init this early;
+# the zoo's later call is an idempotent no-op. Older unsloth_zoo without it is
+# skipped silently.
 try:
     import importlib.util as _importlib_util
     from pathlib import Path as _Path
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 7f453bff82..8c4606f418 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -1804,10 +1804,19 @@ CHAT_TEMPLATES["yi-chat"] = (yi_chat_template, yi_chat_template_eos_token, False
 DEFAULT_SYSTEM_MESSAGE["yi-chat"] = None
 
 def _change_system_message(template: str, type_chat_template: str, system_message: str = None):
-    system_message_pattern = r"\{system_message\}"
-
     # For predefined templates, check if default system message exists
     default_system_message = DEFAULT_SYSTEM_MESSAGE.get(f"{type_chat_template}", None)
+
+    # Custom templates have no predefined default, but may still carry a
+    # {system_message} placeholder. Handle it before the no-default early return
+    # below, which would otherwise leave the literal "{system_message}" in the
+    # template. A placeholder with no system message is an error, not a no-op.
+    if default_system_message is None and "{system_message}" in template:
+        if system_message is None:
+            raise ValueError("Unsloth: You need to provide a system message for custom templates.")
+        new_template = template.replace("{system_message}", system_message)
+        return new_template, system_message
+
     if default_system_message is None:
         if system_message is not None:
             logger.warning_once(
@@ -1817,21 +1826,9 @@ def _change_system_message(template: str, type_chat_template: str, system_messag
             )
         return template, system_message
 
-    # For custom templates
-    if type_chat_template is None:
-        has_placeholder = re.search(system_message_pattern, template) is not None
-
-        if has_placeholder:
-            if system_message is None:
-                raise ValueError("Unsloth: You need to provide a system message for custom templates.")
-            new_template = re.sub(system_message_pattern, system_message, template)
-            return new_template, system_message
-
-        return template, system_message
-
     # For predefined templates with default system message
     message_to_use = system_message if system_message is not None else default_system_message
-    new_template = re.sub(system_message_pattern, message_to_use, template)
+    new_template = template.replace("{system_message}", message_to_use)
 
     return new_template, message_to_use
 
@@ -2279,28 +2276,28 @@ def get_ollama_eos_tokens(tokenizer, extra_eos_tokens = []):
     if getattr(tokenizer, "bos_token", None) is not None:
         added_tokens_decoder = [x for x in added_tokens_decoder if x != tokenizer.bos_token]
 
-    repeatted_tokens = []
+    repeated_tokens = []
     # Join all vocab
     joined_text = "\x01\x00".join(added_tokens_decoder)
     for token in added_tokens_decoder:
         n = len(token)
-        repeatted_counts = joined_text.count(token[:n//2])
+        repeated_counts = joined_text.count(token[:n//2])
         # Try finding longer than 1/2 of the token in the rest
         # For eg <|reserved_special_token_0|>, <|reserved_special_token_1|>
-        if repeatted_counts > 2:
+        if repeated_counts > 2:
             for j in range(n//2+1, n):
-                if joined_text.count(token[:j]) < repeatted_counts:
+                if joined_text.count(token[:j]) < repeated_counts:
                     j -= 1
-                    # Remove repeatted tokens to reduce search space
+                    # Remove repeated tokens to reduce search space
                     joined_text = joined_text.replace(token[:j], "")
-                    repeatted_tokens.append(token[:j])
+                    repeated_tokens.append(token[:j])
                     break
 
     # Remove duplicates
-    splitted = joined_text.split("\x01\x00")
-    final_eos_tokens = [old for old, new in zip(added_tokens_decoder, splitted) if old == new]
+    split = joined_text.split("\x01\x00")
+    final_eos_tokens = [old for old, new in zip(added_tokens_decoder, split) if old == new]
     final_eos_tokens += extra_eos_tokens
-    final_eos_tokens += repeatted_tokens
+    final_eos_tokens += repeated_tokens
 
     # Remove new lines, spaces and HTML tags
     filtered_eos_tokens = []
diff --git a/unsloth/dataprep/raw_text.py b/unsloth/dataprep/raw_text.py
index 076a781d4b..7d18b1ff29 100644
--- a/unsloth/dataprep/raw_text.py
+++ b/unsloth/dataprep/raw_text.py
@@ -223,48 +223,63 @@ class RawTextDataLoader:
                 return "\n\n".join(texts)
         return ""
 
+    # Cache text fields/columns for better performance
+    _TEXT_FIELDS = ("text", "content", "message", "body", "description", "prompt")
+    _TEXT_COLUMNS = _TEXT_FIELDS
+
     def _extract_text_from_json(self, data):
         """Extract text from JSON object using common field names."""
-        text_fields = ["text", "content", "message", "body", "description", "prompt"]
-        for field in text_fields:
+        for field in self._TEXT_FIELDS:
             if field in data and isinstance(data[field], str):
                 return data[field]
         return ""
 
     def _extract_text_from_csv_row(self, row):
         """Extract text from CSV row using common column names."""
-        text_columns = ["text", "content", "message", "body", "description", "prompt"]
-        for column in text_columns:
+        for column in self._TEXT_COLUMNS:
             if column in row and row[column]:
                 return row[column]
         return ""
 
 
 class TextPreprocessor:
+    # Compile regex patterns once for better performance
+    _WHITESPACE_PATTERN = re.compile(r"[^\S\n]+")
+    _INVALID_CHARS_PATTERN = re.compile(r"[^\x20-\x7E\n]")
+    _MULTIPLE_SPACES_PATTERN = re.compile(r"[ ]{2,}")
+    _NEWLINE_SPACES_PATTERN = re.compile(r" *\n *")
+    _MULTIPLE_NEWLINES_PATTERN = re.compile(r"\n{3,}")
+    _CHAPTER_PATTERN = re.compile(r"^# (.+)$", re.MULTILINE)
+    _SECTION_PATTERN = re.compile(r"^## (.+)$", re.MULTILINE)
+    _SUBSECTION_PATTERN = re.compile(r"^### (.+)$", re.MULTILINE)
+    _CODE_BLOCK_PATTERN = re.compile(r"```(\w*)\n(.*?)\n```", re.DOTALL)
+
     def clean_text(self, text):
         """Remove unwanted characters, normalize whitespace"""
         text = text.replace("\r\n", "\n").replace("\r", "\n")
-        text = re.sub(r"[^\S\n]+", " ", text)
-        text = re.sub(r"[^\x20-\x7E\n]", "", text)
-        text = re.sub(r"[ ]{2,}", " ", text)
-        text = re.sub(r" *\n *", "\n", text)
-        text = re.sub(r"\n{3,}", "\n\n", text)
+        text = self._WHITESPACE_PATTERN.sub(" ", text)
+        text = self._INVALID_CHARS_PATTERN.sub("", text)
+        text = self._MULTIPLE_SPACES_PATTERN.sub(" ", text)
+        text = self._NEWLINE_SPACES_PATTERN.sub("\n", text)
+        text = self._MULTIPLE_NEWLINES_PATTERN.sub("\n\n", text)
         return text.strip()
 
     def extract_sections(self, text, patterns):
         """Extract specific sections (e.g., code blocks, quotes)"""
         sections = []
         for pattern in patterns:
+            # Compile pattern on first use and cache? Well, patterns are user-provided,
+            # so just use re.findall with compiled flags
             matches = re.findall(pattern, text, re.MULTILINE | re.DOTALL)
             sections.extend(matches)
         return sections
 
     def add_structure_tokens(self, text):
         """Add special tokens for structure (chapters, sections)"""
-        text = re.sub(r"^# (.+)$", r"<|chapter|>\1<|/chapter|>", text, flags = re.MULTILINE)
-        text = re.sub(r"^## (.+)$", r"<|section|>\1<|/section|>", text, flags = re.MULTILINE)
-        text = re.sub(r"^### (.+)$", r"<|subsection|>\1<|/subsection|>", text, flags = re.MULTILINE)
-        text = re.sub(r"```(\w*)\n(.*?)\n```", r"<|code|\1|>\2<|/code|>", text, flags = re.DOTALL)
+        text = self._CHAPTER_PATTERN.sub(r"<|chapter|>\1<|/chapter|>", text)
+        text = self._SECTION_PATTERN.sub(r"<|section|>\1<|/section|>", text)
+        text = self._SUBSECTION_PATTERN.sub(r"<|subsection|>\1<|/subsection|>", text)
+        text = self._CODE_BLOCK_PATTERN.sub(r"<|code|\1|>\2<|/code|>", text)
         return text
 
     def validate_dataset(self, dataset):
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 5afd6f4b37..d979e5f22f 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -158,6 +158,11 @@ if not UNSLOTH_ENABLE_LOGGING:
     logging.getLogger("torchao").addFilter(
         HideLoggingMessage("Skipping import of cpp extensions due to incompatible torch version")
     )
+    # torch >= 2.11 path: torchao dlopens each prebuilt _C*.so and logs "Failed to load
+    # .../_C*.so" when one can't (ABI tag mismatch in the wheel, e.g. a cp310 .so under a
+    # cp312 runtime on Colab, or an arch-specific kernel the GPU lacks). It falls back to
+    # non-cpp paths and Unsloth doesn't use these kernels, so drop the cosmetic record.
+    logging.getLogger("torchao").addFilter(HideLoggingMessage("Failed to load "))
     # SyntaxWarning: invalid escape sequence '\.'
     warnings.filterwarnings("ignore", message = "invalid escape sequence", category = SyntaxWarning)
     # PYTORCH_CUDA_ALLOC_CONF is deprecated warning from torch
@@ -2349,11 +2354,9 @@ def _is_broken_vllm_error(error) -> bool:
             )
         ) or ("vllm" in message and "undefined symbol" in message):
             return True
-        # Also catch CUDA shared library mismatches during vllm import
-        # e.g. "libcudart.so.12: cannot open shared object file"
-        if (
-            "libcudart" in message or "libcublas" in message or "libnvrtc" in message
-        ) and "cannot open shared object file" in message:
+        # Forced extension load raises the bare loader error (no "vllm._C"
+        # wrapper); match any .so failure as callers feed only vLLM imports.
+        if "cannot open shared object file" in message:
             return True
         current = getattr(current, "__cause__", None) or getattr(current, "__context__", None)
     return False
@@ -2545,6 +2548,16 @@ def _clear_vllm_modules():
             sys.modules.pop(module_name, None)
 
 
+# vLLM's compiled extensions. A CUDA-major ABI break hits all of them, so
+# probing the eagerly-loaded _C and its siblings reliably trips it.
+_VLLM_COMPILED_EXTENSIONS = (
+    "vllm._C",
+    "vllm._C_stable_libtorch",
+    "vllm._moe_C",
+    "vllm._rocm_C",
+)
+
+
 def disable_broken_vllm(error = None):
     """Disable vLLM dynamically when its shared library is ABI-broken."""
     global VLLM_BROKEN
@@ -2562,6 +2575,15 @@ def disable_broken_vllm(error = None):
 
         try:
             import vllm  # noqa: F401
+
+            # Lazy vLLM lets a bare `import vllm` succeed even when an extension
+            # is ABI-broken; force-load each to surface the .so failure here.
+            # A missing one raises ModuleNotFoundError (skipped below).
+            for _ext in _VLLM_COMPILED_EXTENSIONS:
+                try:
+                    importlib.import_module(_ext)
+                except ModuleNotFoundError:
+                    pass
             return False
         except Exception as import_error:
             failure = import_error
diff --git a/unsloth/kernels/geglu.py b/unsloth/kernels/geglu.py
index 3a628f105c..0ed325afa0 100644
--- a/unsloth/kernels/geglu.py
+++ b/unsloth/kernels/geglu.py
@@ -97,7 +97,7 @@ def _exact_backward_kernel(
     e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
     g_row = tl.load(g + offsets, mask = mask, other = 0)  # .to(tl.float32)
 
-    # Break e_row away for re-use
+    # Break e_row away for reuse
     # f = 1/2 * e * (1 + erf(1/sqrt(2) * e))
     f_partial_row = 0.5 * (tl.math.erf(tl.math.rsqrt(2.0) * e_row) + 1.0)
     f_row = f_partial_row * e_row
diff --git a/unsloth/kernels/moe/grouped_gemm/kernels/forward.py b/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
index e95179dd9f..cb1f22d0b8 100644
--- a/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
+++ b/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
@@ -111,7 +111,7 @@ def _grouped_gemm_forward_kernel(
             while tidx >= processed_tiles and tidx < processed_tiles + num_tiles_per_expert:
                 tile_idx = tidx - processed_tiles
 
-                # Check if L2 cache re-use for this order is optimal
+                # Check if L2 cache reuse for this order is optimal
                 tile_m_idx = tile_idx % num_m_tiles
                 tile_n_idx = tile_idx // num_m_tiles
 
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 7a056cef82..599a5c0262 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -86,6 +86,8 @@ __all__ = [
     "is_moe_model",
     "get_moe_target_parameters",
     "make_fast_generate_wrapper",
+    "_mark_unsloth_disable_data_parallel",
+    "_patch_transformers_trainer_data_parallel",
 ]
 
 import torch
@@ -160,6 +162,85 @@ from unsloth_zoo.training_utils import (
 )
 
 
+def _iter_wrapped_models(model):
+    seen = set()
+    current = model
+    while current is not None and id(current) not in seen:
+        yield current
+        seen.add(id(current))
+        next_model = getattr(current, "model", None)
+        if next_model is None:
+            next_model = getattr(current, "base_model", None)
+        if next_model is None:
+            next_model = getattr(current, "module", None)
+        current = next_model
+
+
+def _patch_transformers_trainer_data_parallel():
+    try:
+        from transformers.trainer import Trainer
+    except (ImportError, ModuleNotFoundError):
+        return False
+
+    original_wrap_model = getattr(Trainer, "_wrap_model", None)
+    if original_wrap_model is None:
+        return False
+    if getattr(original_wrap_model, "_unsloth_data_parallel_patched", False):
+        return True
+    try:
+        supports_dataloader = "dataloader" in inspect.signature(original_wrap_model).parameters
+    except (TypeError, ValueError):
+        supports_dataloader = True
+
+    def _call_original_wrap_model(self, model, wrap_args, wrap_kwargs):
+        if supports_dataloader:
+            return original_wrap_model(self, model, *wrap_args, **wrap_kwargs)
+
+        if "dataloader" in wrap_kwargs:
+            wrap_kwargs = {k: v for k, v in wrap_kwargs.items() if k != "dataloader"}
+        return original_wrap_model(self, model, *wrap_args, **wrap_kwargs)
+
+    @functools.wraps(original_wrap_model)
+    def _unsloth_wrap_model(self, model, *wrap_args, **wrap_kwargs):
+        args = getattr(self, "args", None)
+        disable_data_parallel = getattr(model, "_unsloth_disable_data_parallel", False)
+        is_real_8bit = getattr(model, "is_loaded_in_8bit", False)
+        if (
+            args is None
+            or not disable_data_parallel
+            or is_real_8bit
+            or getattr(args, "n_gpu", 0) <= 1
+        ):
+            return _call_original_wrap_model(self, model, wrap_args, wrap_kwargs)
+
+        had_n_gpu = hasattr(args, "_n_gpu")
+        old_n_gpu = getattr(args, "_n_gpu", None)
+        args._n_gpu = 1
+        try:
+            return _call_original_wrap_model(self, model, wrap_args, wrap_kwargs)
+        finally:
+            if had_n_gpu:
+                args._n_gpu = old_n_gpu
+            else:
+                try:
+                    delattr(args, "_n_gpu")
+                except AttributeError:
+                    pass
+
+    _unsloth_wrap_model._unsloth_data_parallel_patched = True
+    _unsloth_wrap_model._unsloth_original_wrap_model = original_wrap_model
+    Trainer._wrap_model = _unsloth_wrap_model
+    return True
+
+
+def _mark_unsloth_disable_data_parallel(model, disable = True):
+    if disable:
+        _patch_transformers_trainer_data_parallel()
+    for module in _iter_wrapped_models(model):
+        setattr(module, "_unsloth_disable_data_parallel", bool(disable))
+    return model
+
+
 def resolve_hip_gpu_stats_name(gpu_stats):
     name = str(getattr(gpu_stats, "name", "") or "").strip()
     name = re.sub(r"\s*\([^)]*\)\s*$", "", name).strip()
@@ -2997,7 +3078,7 @@ class TorchAOConfig:
 def _untie_input_output_embeddings(model: torch.nn.Module) -> None:
     """
     Utility to untie input/output embeddings in a HuggingFace model.
-    This is useful if we want to quantize the input/ouput embeddings differently.
+    This is useful if we want to quantize the input/output embeddings differently.
     Model is modified in-place.
     """
 
diff --git a/unsloth/models/falcon_h1.py b/unsloth/models/falcon_h1.py
index 05bfd2ebb3..e3e04e4fdc 100644
--- a/unsloth/models/falcon_h1.py
+++ b/unsloth/models/falcon_h1.py
@@ -38,7 +38,6 @@ try:
         FalconH1Model,
         FalconH1ForCausalLM,
         FalconH1RMSNorm,
-        FalconH1RMSNormGated,
         FalconHybridMambaAttentionDynamicCache,
     )
 except:
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index d5eca8d6df..14ee5ee24e 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2832,13 +2832,11 @@ class FastLlamaModel:
         internal_model = model
         while hasattr(internal_model, "model"):
             internal_model._saved_temp_tokenizer = tokenizer
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            internal_model.is_loaded_in_8bit = True
 
             internal_model = internal_model.model
         internal_model._saved_temp_tokenizer = tokenizer
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        internal_model.is_loaded_in_8bit = True
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model)
 
         # For transformers > 4.47.1, we need to add rotary_emb to all attention layers
         if IS_ATTENTION_REFACTOR or hasattr(model.model, "rotary_emb"):
@@ -3379,13 +3377,11 @@ class FastLlamaModel:
         while hasattr(internal_model, "model"):
             if hasattr(internal_model, "_saved_temp_tokenizer"):
                 internal_model._saved_temp_tokenizer.padding_side = "right"
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            internal_model.is_loaded_in_8bit = True
             internal_model = internal_model.model
         if hasattr(internal_model, "_saved_temp_tokenizer"):
             internal_model._saved_temp_tokenizer.padding_side = "right"
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        internal_model.is_loaded_in_8bit = True
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model)
 
         # Clear deleted GPU items
         for _ in range(3):
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index a85c1d08a4..53668d14d8 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -994,8 +994,18 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "use_fp16 = getattr(args, 'fp16', False)\n"
             "if type(use_fp16) is not bool: use_fp16 = False\n"
             "force_float32 = False\n"
+            # device-aware bf16 check (CUDA/XPU/HIP), so V100/T4 never pick bf16
+            # but AMD/Intel are unaffected; fall back on older unsloth_zoo.
+            "try:\n"
+            "    from unsloth_zoo.device_type import device_is_bf16_supported as _bf16_supported\n"
+            "except Exception:\n"
+            "    _bf16_supported = torch.cuda.is_bf16_supported\n"
+            # FORCE_FLOAT32 models (Gemma3, gpt_oss, ...) cannot use float16. On a GPU without
+            # bf16 (V100/T4) keep them in float32 so they never autocast to fp16. On a bf16 GPU,
+            # full finetuning can still use bf16 autocast (master weights stay float32), which is
+            # faster and uses less memory; LoRA/QLoRA keep float32 when forced.
             "full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'\n"
-            "if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):\n"
+            "if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1' and not (full_finetuning and _bf16_supported()):\n"
             "    print('Unsloth: Switching to float32 training since model cannot work with float16')\n"
             "    force_float32 = True\n"
             "mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')\n"
@@ -1004,8 +1014,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "from unsloth_zoo.utils import _get_dtype\n"
             "dtype = _get_dtype(dtype)\n"
             "float16 = dtype == torch.float16\n"
+            "bfloat16 = dtype == torch.bfloat16\n"
             "if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')\n"
-            "if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
+            "if not force_float32 and (bfloat16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
             "if force_float32:\n"
             "    # Forced float32 training\n"
             "    args.fp16 = False\n"
@@ -1014,11 +1025,12 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'\n"
             "    # args.mixed_precision is a new argument which needs to be set now\n"
             "elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':\n"
-            "    # Mixed precision training\n"
-            "    args.fp16 = float16\n"
-            "    args.bf16 = not float16\n"
-            "    os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'\n"
-            "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'\n"
+            "    # Mixed precision training. bf16 only if the GPU supports it; V100/T4 use fp16.\n"
+            "    use_bf16_amp = (not float16) and _bf16_supported()\n"
+            "    args.fp16 = not use_bf16_amp\n"
+            "    args.bf16 = use_bf16_amp\n"
+            "    os.environ['ACCELERATE_MIXED_PRECISION'] = 'bf16' if use_bf16_amp else 'fp16'\n"
+            "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'bf16' if use_bf16_amp else 'fp16'\n"
             "    # args.mixed_precision is a new argument which needs to be set now\n"
             "elif mixed_precision_dtype == 'bfloat16':\n"
             "    # Both False since bfloat16 full finetuning doesn't do any autocasting.\n"
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 39004b4d45..bdc2bd9ef6 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1440,16 +1440,14 @@ class FastBaseModel:
         while hasattr(m, "model"):
             m.max_seq_length = max_seq_length
             m._saved_temp_tokenizer = tokenizer
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            m.is_loaded_in_8bit = True if not full_finetuning else False
             m = m.model
         m.max_seq_length = max_seq_length
         # Save to modules as well
         for module in model.modules():
             module.max_seq_length = max_seq_length
         m._saved_temp_tokenizer = tokenizer
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        m.is_loaded_in_8bit = True if not full_finetuning else False
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model, disable = not full_finetuning)
 
         # Patch generate
         if os.environ.get("UNSLOTH_DISABLE_FAST_GENERATION", "0") == "0" and hasattr(
@@ -1826,14 +1824,12 @@ class FastBaseModel:
             if hasattr(m, "_saved_temp_tokenizer"):
                 if hasattr(m._saved_temp_tokenizer, "tokenizer"):
                     m._saved_temp_tokenizer.tokenizer.padding_side = "left"
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            m.is_loaded_in_8bit = True if not full_finetuning else False
             m = m.model
         if hasattr(m, "_saved_temp_tokenizer"):
             if hasattr(m._saved_temp_tokenizer, "tokenizer"):
                 m._saved_temp_tokenizer.tokenizer.padding_side = "left"
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        m.is_loaded_in_8bit = True if not full_finetuning else False
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model, disable = not full_finetuning)
 
         # Clear deleted GPU items
         for _ in range(3):
diff --git a/unsloth/save.py b/unsloth/save.py
index 8595c19dee..76bc6aa733 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -226,6 +226,29 @@ def _normalize_compressed_method(save_method):
     return None
 
 
+def _is_cmake_only_llama_cpp(llama_cpp_dir: str = "llama.cpp") -> bool:
+    """
+    True if llama.cpp's Makefile is the post-CMake-migration deprecation stub,
+    so `make` cannot build it. A genuinely missing/empty checkout returns False
+    so it isn't treated as CMake-only: the caller then probes make and fails
+    loudly on a real error rather than silently assuming a CMake build.
+    """
+    makefile_path = os.path.join(llama_cpp_dir, "Makefile")
+    if not os.path.exists(makefile_path):
+        # No Makefile: only CMake-only if a real CMake project is present
+        return os.path.exists(os.path.join(llama_cpp_dir, "CMakeLists.txt"))
+    try:
+        with open(makefile_path, "r", encoding = "utf-8", errors = "ignore") as f:
+            content = f.read(4096).lower()
+            if "cmake" in content and "deprecated" in content:
+                return True
+            if "build system changed" in content:
+                return True
+    except (IOError, OSError):
+        pass
+    return False
+
+
 def print_quantization_methods():
     for key, value in ALLOWED_QUANTS.items():
         print(f'"{key}"  ==> {value}')
@@ -565,6 +588,17 @@ def _is_qwen3_5_vlm(model):
     ) or getattr(config, "model_type", None) in ("qwen3_5", "qwen3_5_moe")
 
 
+def _is_gpt_oss(model):
+    config = getattr(model, "config", None)
+    if config is None:
+        return False
+    architectures = getattr(config, "architectures", None) or ()
+    return "GptOssForCausalLM" in architectures or getattr(config, "model_type", None) in (
+        "gpt-oss",
+        "gpt_oss",
+    )
+
+
 def _qwen3_5_vlm_state_dict_for_save(state_dict):
     remapped_state_dict = {}
     for key, value in state_dict.items():
@@ -1269,14 +1303,27 @@ def install_llama_cpp_make_non_blocking():
     # https://github.com/ggerganov/llama.cpp/issues/7062
     # Weirdly GPU conversion for GGUF breaks??
     # env = { **os.environ, "LLAMA_CUDA": "1", }
-    # Force make clean
-    check = os.system("make clean -C llama.cpp")
-    IS_CMAKE = False
-    if check == 0:
+
+    # Skip the make-clean probe on CMake-only checkouts (its error output is misleading)
+    IS_CMAKE = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not IS_CMAKE:
+        # Confirm make still works, silently
+        try:
+            result = subprocess.run(
+                ["make", "clean", "-C", "llama.cpp"],
+                stdout = subprocess.DEVNULL,
+                stderr = subprocess.DEVNULL,
+            )
+            IS_CMAKE = result.returncode != 0
+        except FileNotFoundError:
+            # No make executable; use CMake
+            IS_CMAKE = True
+
+    if not IS_CMAKE:
         # Uses old MAKE
         n_jobs = max(int((psutil.cpu_count() or 1) * 1.5), 1)
         full_command = ["make", "all", "-j" + str(n_jobs), "-C", "llama.cpp"]
-        IS_CMAKE = False
     else:
         # Uses new CMAKE
         n_jobs = max(int(psutil.cpu_count() or 1), 1)  # Use less CPUs since 1.5x faster
@@ -1299,7 +1346,6 @@ def install_llama_cpp_make_non_blocking():
             "--clean-first",
             "--target",
         ] + LLAMA_CPP_TARGETS
-        IS_CMAKE = True
     # https://github.com/ggerganov/llama.cpp/issues/7062
     # Weirdly GPU conversion for GGUF breaks??
     # run_installer = subprocess.Popen(full_command, env = env, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
@@ -1468,20 +1514,25 @@ def install_llama_cpp_old(version = -10):
     ]
     try_execute(commands)
 
-    # Try using MAKE
-    commands = [
-        "make clean -C llama.cpp",
-        f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-    ]
-    if try_execute(commands) == "CMAKE":
-        # Instead use CMAKE
+    # Detect CMake-only build system before trying make
+    use_cmake = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not use_cmake:
+        # Try using MAKE
+        commands = [
+            "make clean -C llama.cpp",
+            f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+        ]
+        use_cmake = try_execute(commands) == "CMAKE"
+
+    if use_cmake:
+        # Use CMAKE
         commands = [
             f"cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=OFF {CURL_FLAG}",
             f"cmake --build llama.cpp/build --config Release -j{(psutil.cpu_count() or 1)*2} --clean-first --target {' '.join(LLAMA_CPP_TARGETS)}",
             "cp llama.cpp/build/bin/llama-* llama.cpp",
             "rm -rf llama.cpp/build",
         ]
-
         try_execute(commands)
 
     # Check if successful
@@ -1513,15 +1564,21 @@ def install_llama_cpp_blocking(use_cuda = False):
         return
     try_execute(commands)
 
-    commands = [
-        "make clean -C llama.cpp",
-        # https://github.com/ggerganov/llama.cpp/issues/7062
-        # Weirdly GPU conversion for GGUF breaks??
-        # f"{use_cuda} make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-        f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-    ]
-    if try_execute(commands) == "CMAKE":
-        # Instead use CMAKE
+    # Detect CMake-only build system before trying make
+    use_cmake = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not use_cmake:
+        commands = [
+            "make clean -C llama.cpp",
+            # https://github.com/ggerganov/llama.cpp/issues/7062
+            # Weirdly GPU conversion for GGUF breaks??
+            # f"{use_cuda} make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+            f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+        ]
+        use_cmake = try_execute(commands) == "CMAKE"
+
+    if use_cmake:
+        # Use CMAKE
         commands = [
             f"cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=OFF {CURL_FLAG}",
             f"cmake --build llama.cpp/build --config Release -j{(psutil.cpu_count() or 1)*2} --clean-first --target {' '.join(LLAMA_CPP_TARGETS)}",
@@ -2541,15 +2598,7 @@ def unsloth_save_pretrained_gguf(
 
     is_processor = is_vlm and isinstance(tokenizer, ProcessorMixin)
 
-    is_gpt_oss = (
-        True
-        if (
-            hasattr(self.config, "architectures")
-            and self.config.architectures == "GptOssForCausalLM"
-        )
-        or (hasattr(self.config, "model_type") and self.config.model_type in ["gpt-oss", "gpt_oss"])
-        else False
-    )
+    is_gpt_oss = _is_gpt_oss(self)
     # Step 2: Prepare arguments for model saving
     arguments = dict(locals())
     arguments["model"] = self
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
index 0baeb2ffe5..e8a3414d73 100644
--- a/unsloth_cli/_inference.py
+++ b/unsloth_cli/_inference.py
@@ -3,11 +3,12 @@
 
 """Model loading and streaming shared by `inference` and `chat`."""
 
+import asyncio
 import os
 import re
 import sys
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional
 
 import typer
 
@@ -211,28 +212,65 @@ def resolve_model_config(model: str, *, hf_token: Optional[str]):
     return model_config
 
 
-def _load_gguf_backend(model_config, *, hf_token, max_seq_length):
+def _validate_llama_extra_args_or_exit(llama_extra_args: Optional[List[str]]) -> list[str]:
+    from core.inference.llama_server_args import validate_extra_args
+    try:
+        return validate_extra_args(llama_extra_args)
+    except ValueError as exc:
+        typer.echo(f"Error: {exc}", err = True)
+        raise typer.Exit(code = 1)
+
+
+def _load_gguf_backend(
+    model_config,
+    *,
+    hf_token,
+    max_seq_length,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
+):
     ensure_studio_backend_path()
     from core.inference.llama_cpp import LlamaCppBackend
+    from core.inference.tensor_fallback import load_with_tensor_fallback
 
     llama_backend = LlamaCppBackend()
+    extra_args = _validate_llama_extra_args_or_exit(llama_extra_args)
     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
+
+    async def _attempt_gguf_load(
+        requested_tensor_parallel: bool, attempt_extra_args: Optional[List[str]]
+    ) -> bool:
+        attempt_common = dict(
+            common,
+            tensor_parallel = requested_tensor_parallel,
+            extra_args = attempt_extra_args,
         )
-    else:
-        loaded = llama_backend.load_model(
+        if model_config.gguf_hf_repo:
+            return llama_backend.load_model(
+                hf_repo = model_config.gguf_hf_repo,
+                hf_token = hf_token,
+                **attempt_common,
+            )
+        return 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,
+            **attempt_common,
         )
+
+    loaded = asyncio.run(
+        load_with_tensor_fallback(
+            _attempt_gguf_load,
+            requested_tensor = tensor_parallel,
+            extra_args = extra_args,
+            label = model_config.identifier,
+        )
+    )
     if not loaded:
         typer.echo("Model load failed", err = True)
         raise typer.Exit(code = 1)
@@ -245,6 +283,8 @@ def load_chat_backend(
     hf_token: Optional[str],
     max_seq_length: int,
     load_in_4bit: bool,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
     model_config = None,
     fresh_backend: bool = False,
 ):
@@ -259,7 +299,13 @@ def load_chat_backend(
     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)
+        return _load_gguf_backend(
+            model_config,
+            hf_token = hf_token,
+            max_seq_length = max_seq_length,
+            tensor_parallel = tensor_parallel,
+            llama_extra_args = llama_extra_args,
+        )
 
     if fresh_backend:
         ensure_studio_backend_path()
@@ -447,18 +493,31 @@ class HttpChatBackend:
         # No redirects: this carries a bearer token (see urlopen_no_redirect).
         return urlopen_no_redirect(request, timeout = timeout)
 
-    def ensure_loaded(self, model: str, *, hf_token, max_seq_length, load_in_4bit) -> None:
+    def ensure_loaded(
+        self,
+        model: str,
+        *,
+        hf_token,
+        max_seq_length,
+        load_in_4bit,
+        tensor_parallel: bool = False,
+        llama_extra_args: Optional[List[str]] = None,
+    ) -> None:
         typer.echo(f"Loading {model} on the Studio server", err = True)
+        payload = {
+            "model_path": model,
+            "hf_token": hf_token,
+            "max_seq_length": max_seq_length,
+            "load_in_4bit": load_in_4bit,
+            "tensor_parallel": tensor_parallel,
+        }
+        if llama_extra_args:
+            payload["llama_extra_args"] = llama_extra_args
         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,
-                },
+                payload,
             ).close()
         except Exception as exc:
             typer.echo(f"Model load failed: {exc}", err = True)
@@ -538,7 +597,15 @@ class HttpChatBackend:
         pass
 
 
-def connect_studio_server(model: str, *, hf_token, max_seq_length, load_in_4bit):
+def connect_studio_server(
+    model: str,
+    *,
+    hf_token,
+    max_seq_length,
+    load_in_4bit,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
+):
     """Backend on a running Studio server, or None (caller loads locally)."""
     base_url = find_studio_server()
     if not base_url:
@@ -576,6 +643,11 @@ def connect_studio_server(model: str, *, hf_token, max_seq_length, load_in_4bit)
         return _refuse("couldn't self-issue a Studio token (is Studio set up here?).")
     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
+        model,
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
     )
     return backend
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index c62916bc75..a483aeeb6c 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.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
 
-from typing import Optional
+from typing import List, Optional
 
 import typer
 from rich.console import Console
@@ -153,6 +153,22 @@ def chat(
     ),
     max_seq_length: int = typer.Option(4096, "--max-seq-length"),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
+    tensor_parallel: bool = typer.Option(
+        False,
+        "--tensor-parallel/--no-tensor-parallel",
+        help = (
+            "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
+            "of by layer. Ignored for non-GGUF models."
+        ),
+    ),
+    llama_extra_args: Optional[List[str]] = typer.Option(
+        None,
+        "--llama-extra-arg",
+        help = (
+            "Extra llama-server arg for GGUF models. Repeat for multiple "
+            "tokens, e.g. --llama-extra-arg=--top-k --llama-extra-arg 20."
+        ),
+    ),
     think: bool = typer.Option(
         False,
         "--think/--no-think",
@@ -190,7 +206,13 @@ def chat(
         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)
+    load_opts = dict(
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
+    )
 
     # 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)
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 5dbc32c7d2..1401fa9e9e 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.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
 
-from typing import Optional
+from typing import List, Optional
 
 import typer
 
@@ -31,6 +31,22 @@ 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"),
+    tensor_parallel: bool = typer.Option(
+        False,
+        "--tensor-parallel/--no-tensor-parallel",
+        help = (
+            "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
+            "of by layer. Ignored for non-GGUF models."
+        ),
+    ),
+    llama_extra_args: Optional[List[str]] = typer.Option(
+        None,
+        "--llama-extra-arg",
+        help = (
+            "Extra llama-server arg for GGUF models. Repeat for multiple "
+            "tokens, e.g. --llama-extra-arg=--top-k --llama-extra-arg 20."
+        ),
+    ),
     think: bool = typer.Option(
         False,
         "--think/--no-think",
@@ -55,7 +71,13 @@ def inference(
 
     # 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)
+    load_opts = dict(
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
+    )
     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)
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
index 770bd0db29..013b7c5a81 100644
--- a/unsloth_cli/tests/test_inference_chat.py
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -9,6 +9,7 @@ import inspect
 import sys
 import types
 from pathlib import Path
+from types import SimpleNamespace
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
 if str(_REPO_ROOT) not in sys.path:
@@ -16,6 +17,7 @@ if str(_REPO_ROOT) not in sys.path:
 
 
 import typer
+import pytest
 from rich.console import Console
 from typer.testing import CliRunner
 
@@ -43,6 +45,14 @@ def _chat_app():
     return cli
 
 
+def _inference_app():
+    from unsloth_cli.commands.inference import inference
+
+    cli = typer.Typer()
+    cli.command()(inference)
+    return cli
+
+
 def test_visible_text_passthrough_when_shown():
     text = "<think>reasoning</think>answer"
     assert visible_text(text, show_thinking = True) == text
@@ -81,6 +91,16 @@ def test_inference_think_defaults_off():
     assert "--think/--no-think" in (getattr(opt, "param_decls", None) or [])
 
 
+def test_inference_exposes_gguf_runtime_options():
+    from unsloth_cli.commands.inference import inference
+
+    tensor = _option(inference, "tensor_parallel")
+    assert "--tensor-parallel/--no-tensor-parallel" in (getattr(tensor, "param_decls", None) or [])
+
+    extra = _option(inference, "llama_extra_args")
+    assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
+
+
 def test_chat_command_is_registered_with_options():
     params = inspect.signature(chatmod.chat).parameters
     assert "model" in params
@@ -94,6 +114,12 @@ def test_chat_command_is_registered_with_options():
     verbose = _option(chatmod.chat, "verbose")
     assert {"--verbose", "-v"} <= set(getattr(verbose, "param_decls", None) or [])
 
+    tensor = _option(chatmod.chat, "tensor_parallel")
+    assert "--tensor-parallel/--no-tensor-parallel" in (getattr(tensor, "param_decls", None) or [])
+
+    extra = _option(chatmod.chat, "llama_extra_args")
+    assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
+
 
 class _FakeBackend:
     def __init__(self):
@@ -324,6 +350,243 @@ def test_http_backend_streams_cumulative_text(monkeypatch):
     assert out == ["He", "Hello"]
 
 
+def test_http_backend_load_forwards_gguf_runtime_options(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    requests = []
+
+    class _OK:
+        def close(self):
+            pass
+
+    def fake_request(
+        method,
+        path,
+        payload = None,
+        timeout = None,
+    ):
+        requests.append((method, path, payload, timeout))
+        return _OK()
+
+    monkeypatch.setattr(backend, "_request", fake_request)
+
+    backend.ensure_loaded(
+        "org/model-GGUF",
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        load_in_4bit = False,
+        tensor_parallel = True,
+        llama_extra_args = ["--top-k", "20"],
+    )
+
+    assert requests == [
+        (
+            "POST",
+            "/api/inference/load",
+            {
+                "model_path": "org/model-GGUF",
+                "hf_token": "hf_x",
+                "max_seq_length": 8192,
+                "load_in_4bit": False,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+            None,
+        )
+    ]
+
+
+def test_http_backend_load_sends_explicit_false_tensor_parallel(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    requests = []
+
+    class _OK:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(
+        backend,
+        "_request",
+        lambda method, path, payload = None, timeout = None: (
+            requests.append((method, path, payload, timeout)),
+            _OK(),
+        )[1],
+    )
+
+    backend.ensure_loaded(
+        "org/model-GGUF",
+        hf_token = None,
+        max_seq_length = 4096,
+        load_in_4bit = True,
+        tensor_parallel = False,
+    )
+
+    assert requests[0][2]["tensor_parallel"] is False
+
+
+def test_load_gguf_backend_forwards_local_runtime_options(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+
+    class _FakeLlamaCppBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return True
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = _FakeLlamaCppBackend
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+    fake_args.validate_extra_args = lambda args: list(args or [])
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+
+    async def _passthrough(
+        attempt_load,
+        *,
+        requested_tensor,
+        extra_args,
+        label = "",
+        cancelled = None,
+    ):
+        return await attempt_load(requested_tensor, extra_args)
+
+    fake_tensor_fallback.load_with_tensor_fallback = _passthrough
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    backend = inference._load_gguf_backend(
+        config,
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        tensor_parallel = True,
+        llama_extra_args = ["--top-k", "20"],
+    )
+
+    assert isinstance(backend, ChatBackend)
+    assert calls == [
+        {
+            "hf_repo": "org/model-GGUF",
+            "hf_token": "hf_x",
+            "hf_variant": "Q4_K_M",
+            "model_identifier": "org/model-GGUF",
+            "is_vision": False,
+            "n_ctx": 8192,
+            "tensor_parallel": True,
+            "extra_args": ["--top-k", "20"],
+        }
+    ]
+
+
+def test_load_gguf_backend_exits_cleanly_on_invalid_extra_args(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = object
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+
+    def _raise(_args):
+        raise ValueError("llama-server flag '--model' is managed by Unsloth Studio")
+
+    fake_args.validate_extra_args = _raise
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+    fake_tensor_fallback.load_with_tensor_fallback = None
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    with pytest.raises(typer.Exit) as excinfo:
+        inference._load_gguf_backend(
+            config,
+            hf_token = "hf_x",
+            max_seq_length = 8192,
+            llama_extra_args = ["--model"],
+        )
+
+    assert excinfo.value.exit_code == 1
+
+
+def test_load_gguf_backend_uses_tensor_fallback(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+    fallback_calls = []
+
+    class _FakeLlamaCppBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return kwargs["tensor_parallel"] is False
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = _FakeLlamaCppBackend
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+    fake_args.validate_extra_args = lambda args: list(args or [])
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+
+    async def _fallback(
+        attempt_load,
+        *,
+        requested_tensor,
+        extra_args,
+        label = "",
+        cancelled = None,
+    ):
+        fallback_calls.append((requested_tensor, extra_args, label))
+        ok = await attempt_load(requested_tensor, extra_args)
+        if ok:
+            return True
+        return await attempt_load(False, ["--split-mode", "layer"])
+
+    fake_tensor_fallback.load_with_tensor_fallback = _fallback
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    backend = inference._load_gguf_backend(
+        config,
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        tensor_parallel = True,
+    )
+
+    assert isinstance(backend, ChatBackend)
+    assert fallback_calls == [(True, [], "org/model-GGUF")]
+    assert [call["tensor_parallel"] for call in calls] == [True, False]
+    assert calls[1]["extra_args"] == ["--split-mode", "layer"]
+
+
 def test_http_backend_merges_emoji_split_across_deltas(monkeypatch):
     backend = HttpChatBackend("http://localhost:8888", "token")
     response = _FakeSSEResponse(
@@ -365,6 +628,98 @@ def test_chat_prefers_running_studio_server(monkeypatch):
     assert closed == ["http"]
 
 
+def test_chat_forwards_gguf_runtime_options_to_loader(monkeypatch):
+    loads = []
+
+    class _FakeHttpBackend:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeHttpBackend())[1],
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: None)
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        [
+            "fake-model",
+            "--tensor-parallel",
+            "--llama-extra-arg=--top-k",
+            "--llama-extra-arg",
+            "20",
+        ],
+        input = "/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert loads == [
+        (
+            "fake-model",
+            {
+                "hf_token": None,
+                "max_seq_length": 4096,
+                "load_in_4bit": True,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+        )
+    ]
+
+
+def test_inference_forwards_gguf_runtime_options_to_loader(monkeypatch):
+    from unsloth_cli.commands import inference as infermod
+
+    loads, streams, closed = [], [], []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            streams.append((messages, kwargs))
+            return iter(["answer"])
+
+        def close(self):
+            closed.append(True)
+
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeBackend())[1],
+    )
+    monkeypatch.setattr(infermod, "load_chat_backend", lambda *a, **k: None)
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        [
+            "fake-model",
+            "hello",
+            "--tensor-parallel",
+            "--llama-extra-arg=--top-k",
+            "--llama-extra-arg",
+            "20",
+        ],
+    )
+
+    assert result.exit_code == 0, result.output
+    assert loads == [
+        (
+            "fake-model",
+            {
+                "hf_token": None,
+                "max_seq_length": 2048,
+                "load_in_4bit": True,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+        )
+    ]
+    assert streams[0][0] == [{"role": "user", "content": "hello"}]
+    assert closed == [True]
+
+
 def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
     streamed, closed, base_loads = [], [], []