diff --git a/.gitattributes b/.gitattributes index 5f04b5e9d1..0025f2a697 100644 --- a/.gitattributes +++ b/.gitattributes @@ -6,7 +6,7 @@ # them when run in WSL/Linux (e.g. `set -e` -> "set: Illegal option -"). *.sh text eol=lf -# Normalize Studio frontend sources to LF. Scoped to the frontend tree (rather +# Normalize Unsloth frontend sources to LF. Scoped to the frontend tree (rather # than repo-wide *.ts/*.tsx/... rules) so the policy can't force LF on files # elsewhere. text=auto lets Git detect and leave binary assets (logos, fonts) # untouched while text files (.ts/.tsx/.json/.html/.svg/...) are stored as LF. diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh index f4189a159e..b63ac94b93 100755 --- a/.github/scripts/agent-guides-drive.sh +++ b/.github/scripts/agent-guides-drive.sh @@ -36,6 +36,23 @@ AGENT="${2:?usage: agent-guides-drive.sh }" # Determinism (seed/temp) is applied at the server level by # serve-unsloth-run.sh --extra; agents inherit it through the API. TIMEOUT="${AGENT_INVOKE_TIMEOUT:-180}" +# opencode is the slow outlier. Unlike the print-mode agents (claude -p, codex +# exec) it runs a full turn AND a separate small_model call to name the session, +# so one connection reply takes ~8 min on a CPU-served 4B -- right at the shared +# 600s cap, so the cell flaked when a run drifted past a ~480s success. Give it +# headroom (still well under the 40-min job budget); the fast agents keep the +# tight cap that still catches a real headless-TTY hang. +case "$AGENT" in + opencode) + # Double it, but only for a bare-integer seconds value. A GNU timeout(1) + # duration suffix (s/m/h/d, including floats like 0.5s) is left unchanged so + # the arithmetic never sees a non-number; timeout(1) parses it directly. + case "$TIMEOUT" in + *[!0-9]*) ;; + *) TIMEOUT=$(( TIMEOUT * 2 )) ;; + esac + ;; +esac # Claude refuses --dangerously-skip-permissions outside a sandbox; the CI runner # IS the sandbox, so declare it (mirrors unslothai/scripts launcher.sh). Harmless @@ -166,8 +183,8 @@ parse_connect() { echo "[$AGENT] connect --no-launch printed:"; cat_redacted "$raw" CONNECT_ENV="$(grep -E '^(export |unset )' "$raw" || true)" # The launch command is the last non-export, non-status line. start.py - # prints "Studio · model " and "Updated ..." status lines first. - CONNECT_CMD="$(grep -vE '^(export |unset |Studio |Updated |Disabled |Warning|Loading)' "$raw" \ + # prints "Unsloth · model " and "Updated ..." status lines first. + CONNECT_CMD="$(grep -vE '^(export |unset |Unsloth |Updated |Disabled |Warning|Loading)' "$raw" \ | grep -E '[^[:space:]]' | tail -1)" [ -n "$CONNECT_CMD" ] || guide_fail "could not parse a launch command from connect --no-launch output" redact "$raw" diff --git a/.github/scripts/assert-llama-loads.sh b/.github/scripts/assert-llama-loads.sh index c2ffe27469..62ef80d364 100755 --- a/.github/scripts/assert-llama-loads.sh +++ b/.github/scripts/assert-llama-loads.sh @@ -2,7 +2,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. # -# Assert Studio installed a llama.cpp that loads and runs on THIS macOS. Tests +# Assert Unsloth installed a llama.cpp that loads and runs on THIS macOS. Tests # the contract that matters (binaries load and their minimum-OS is <= this host) # instead of the old "did install.sh fall back to a source build?" grep, since a # source build with a correct deployment target is a valid outcome. diff --git a/.github/scripts/assert-prompt-cache.sh b/.github/scripts/assert-prompt-cache.sh index f5b6b075eb..8c28569f77 100755 --- a/.github/scripts/assert-prompt-cache.sh +++ b/.github/scripts/assert-prompt-cache.sh @@ -31,7 +31,7 @@ # (llama_cpp.py:337-340). So default: ~/.unsloth/studio/logs/llama-server/. # #

is the INTERNAL llama-server port (self._find_free_port(), -# llama_cpp.py:3489 / :4641) -- a RANDOM port, NOT the Studio port. So we must +# llama_cpp.py:3489 / :4641) -- a RANDOM port, NOT the Unsloth port. So we must # NOT filter the log glob by STUDIO_PORT (the brief's `port-` # glob would never match). We pick the newest llama-*.log instead. # diff --git a/.github/scripts/hf-download-with-retry.sh b/.github/scripts/hf-download-with-retry.sh index 013a459f46..6dec93356a 100755 --- a/.github/scripts/hf-download-with-retry.sh +++ b/.github/scripts/hf-download-with-retry.sh @@ -3,7 +3,7 @@ # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 # # Download a single file from a Hugging Face repo with a stall-retry -# watchdog. Used by the Studio CI workflows so a hung hf-xet transfer +# watchdog. Used by the Unsloth CI workflows so a hung hf-xet transfer # kills + retries instead of silently consuming the job's timeout. # # Usage: hf-download-with-retry.sh REPO FILE LOCAL_DIR @@ -35,7 +35,7 @@ REPO="${1:?usage: hf-download-with-retry.sh REPO FILE [LOCAL_DIR]}" FILE="${2:?usage: hf-download-with-retry.sh REPO FILE [LOCAL_DIR]}" # LOCAL_DIR is optional. If empty, hf falls back to HF_HUB_CACHE # (~/.cache/huggingface/hub) which is the desired path for callers -# that populate HF_HOME for a downstream Studio model load. +# that populate HF_HOME for a downstream Unsloth model load. LOCAL_DIR="${3:-}" # Stall threshold per attempt, in seconds. Override with diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml index fa84471d36..d1bea819eb 100644 --- a/.github/workflows/consolidated-tests-ci.yml +++ b/.github/workflows/consolidated-tests-ci.yml @@ -268,6 +268,7 @@ jobs: tests/saving/test_save_shell_injection.py \ tests/saving/test_patch_saving_none_tokenizer.py \ tests/saving/test_fix_sentencepiece_gguf_robustness.py \ + tests/saving/test_fix_sentencepiece_tokenizer_guard.py \ tests/saving/test_compressed_export_schemes.py \ tests/saving/test_export_api_surface.py \ tests/saving/test_export_dispatch.py \ @@ -358,6 +359,7 @@ jobs: tests/saving/test_save_shell_injection.py \ tests/saving/test_patch_saving_none_tokenizer.py \ tests/saving/test_fix_sentencepiece_gguf_robustness.py \ + tests/saving/test_fix_sentencepiece_tokenizer_guard.py \ tests/saving/test_compressed_export_schemes.py \ tests/saving/test_export_api_surface.py \ tests/saving/test_export_dispatch.py \ diff --git a/.github/workflows/lint-ci.yml b/.github/workflows/lint-ci.yml index bd859a6e9e..e1f0afd299 100644 --- a/.github/workflows/lint-ci.yml +++ b/.github/workflows/lint-ci.yml @@ -13,10 +13,10 @@ # committed YAML / JSON config. # # TypeScript and Rust are NOT duplicated here on purpose: -# - Studio Frontend CI runs `npm run typecheck` (= `tsc --noEmit`) +# - Unsloth Frontend CI runs `npm run typecheck` (= `tsc --noEmit`) # and `npm run build` (vite/swc) on every studio/frontend/** # change, which is a full TS AST + type check. -# - Studio Tauri CI runs `tauri build --debug --no-bundle` on +# - Unsloth Tauri CI runs `tauri build --debug --no-bundle` on # every studio/src-tauri/** or studio/frontend/** change, which # compiles the Rust crate (= cargo check + cargo build). # Each is a stricter check than a parse-only step would be, so a diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml index 25796bd5cf..c48328e90f 100644 --- a/.github/workflows/local-agent-guides-ci.yml +++ b/.github/workflows/local-agent-guides-ci.yml @@ -154,7 +154,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Gated off PR (see note above); public GGUF still downloads. @@ -256,7 +256,7 @@ jobs: done fi - - name: Stop Studio + - name: Stop Unsloth if: always() run: | # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make @@ -359,7 +359,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Gated off PR (see note above); public GGUF still downloads. @@ -448,7 +448,7 @@ jobs: done fi - - name: Stop Studio + - name: Stop Unsloth if: always() run: | # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make @@ -543,7 +543,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} HF_TOKEN: ${{ secrets.HF_TOKEN }} @@ -620,7 +620,7 @@ jobs: done fi - - name: Stop Studio + - name: Stop Unsloth if: always() run: | if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then @@ -706,7 +706,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Gated off PR (see note above); public GGUF still downloads. @@ -764,7 +764,7 @@ jobs: done fi - - name: Stop Studio + - name: Stop Unsloth if: always() run: | # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml index a2f716a93c..aadf0b54e6 100644 --- a/.github/workflows/mlx-ci.yml +++ b/.github/workflows/mlx-ci.yml @@ -130,7 +130,7 @@ jobs: # MLX support landed after the most recent unsloth-zoo PyPI # release; the wheel still raises NotImplementedError on # Apple Silicon when device_type.get_device_type() runs - # unguarded. Studio's own install.sh overlays unsloth-zoo + # unguarded. Unsloth's own install.sh overlays unsloth-zoo # from git main for the same reason. Pulling deps lets pip # resolve the platform-conditional MLX-only wheels (mlx, # mlx-lm, mlx-vlm gated on darwin+arm64 in unsloth-zoo's @@ -317,13 +317,13 @@ jobs: echo done - # Validates the macOS prebuilt path Studio's setup.sh uses (#5963): install the + # Validates the macOS prebuilt path Unsloth'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) + - name: Unsloth prebuilt llama.cpp install + GGUF download (Mac M1) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -344,12 +344,12 @@ jobs: # Final step: runs the downloaded binaries with no secrets present, and clears # the GitHub Actions command files so a tampered prebuilt cannot influence the job. - - name: Studio prebuilt llama.cpp GGUF inference smoke (Mac M1) + - name: Unsloth prebuilt llama.cpp GGUF inference smoke (Mac M1) run: | set -euo pipefail unset GITHUB_ENV GITHUB_PATH GITHUB_OUTPUT GITHUB_STEP_SUMMARY INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp" - # Studio bundles only llama-server + llama-quantize (not llama-cli); + # Unsloth 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" @@ -400,4 +400,4 @@ jobs: tail -40 /tmp/llama-server.log exit 1 fi - echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works" + echo "OK: Unsloth prebuilt llama.cpp on Mac M1 + GGUF /completion works" diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml index 4daafae35d..081eda4e32 100644 --- a/.github/workflows/release-desktop.yml +++ b/.github/workflows/release-desktop.yml @@ -4,7 +4,7 @@ on: workflow_dispatch: inputs: studio_version: - description: 'Studio version tag to release (for example, v0.1.39-beta)' + description: 'Unsloth version tag to release (for example, v0.1.39-beta)' type: string required: true pypi_version: @@ -69,7 +69,7 @@ jobs: if not studio_version: sys.exit('studio_version is required, for example v0.1.39-beta') if re.fullmatch(r'v?20\d{2}\.\d+\.\d+(?:[-+][0-9A-Za-z.-]+)?', studio_version): - sys.exit(f'studio_version must be a Studio SemVer tag, not a date-style backend version: {studio_version}') + sys.exit(f'studio_version must be an Unsloth SemVer tag, not a date-style backend version: {studio_version}') semver_tag = re.compile( r'^v(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)' @@ -146,7 +146,7 @@ jobs: print(f'pypi_version={pypi_version}', file=output) PY - - name: Verify PyPI package and Studio stamp + - name: Verify PyPI package and Unsloth stamp shell: bash env: STUDIO_VERSION: ${{ steps.prepare.outputs.studio_version }} @@ -211,7 +211,7 @@ jobs: fi python3 scripts/stamp_studio_release.py --verify-dist "$RUNNER_TEMP/pypi-unsloth-dist" --expected "$STUDIO_VERSION" else - echo "scripts/stamp_studio_release.py not found; release-desktop requires #5308 to verify the PyPI Studio stamp." >&2 + echo "scripts/stamp_studio_release.py not found; release-desktop requires #5308 to verify the PyPI Unsloth stamp." >&2 exit 1 fi diff --git a/.github/workflows/security-audit.yml b/.github/workflows/security-audit.yml index 1275d12216..27eafbedea 100644 --- a/.github/workflows/security-audit.yml +++ b/.github/workflows/security-audit.yml @@ -36,8 +36,8 @@ # - unsloth `huggingfacenotorch` extras (the canonical install path # for fine-tuning users; pulls transformers / peft / accelerate / # trl / datasets / diffusers / sentence-transformers / etc.) -# - all six Studio backend requirements files -# - Studio frontend (npm) and Tauri shell (cargo) +# - all six Unsloth backend requirements files +# - Unsloth frontend (npm) and Tauri shell (cargo) # Each Python step builds a filtered dep list from pyproject.toml + # requirements/*.txt before auditing. We do NOT install any of these # -- pip-audit resolves through PyPI metadata, scan_packages.py @@ -218,7 +218,7 @@ jobs: # on the runner). A comment line is left in place so the # skipped specs are obvious in the artifact. # The `huggingface` extra is `huggingfacenotorch` plus torch / - # torchvision / triton, deliberately skipped: Studio backend + # torchvision / triton, deliberately skipped: Unsloth backend # already pins a torch and the +cu* / +cpu local-version tags # trip up the PyPI resolver in `-r` mode. run: | @@ -253,7 +253,7 @@ jobs: # `-r requirements.txt` resolves the requirements through pip's # dependency resolver against PyPI metadata and audits the # resolved tree without ever executing setup.py / install - # hooks. Way faster than installing the full Studio runtime + # hooks. Way faster than installing the full Unsloth runtime # and -- critically -- safer: an attacker who has compromised # a transitive dep cannot run code in this job. # @@ -326,9 +326,9 @@ jobs: } >> "$GITHUB_STEP_SUMMARY" # ───────────────────────────────────────────────────────────── - # npm: Studio frontend + # npm: Unsloth frontend # ───────────────────────────────────────────────────────────── - - name: npm audit (Studio frontend) + - name: npm audit (Unsloth frontend) # `npm audit` resolves the lockfile through the npmjs.com # advisory DB. `--audit-level=high` filters the noise floor # to only HIGH and CRITICAL. We do NOT pass --omit=dev: a @@ -342,7 +342,7 @@ jobs: # Always also write the full JSON for grep-ability. npm audit --json > ../../logs-npm-audit.json || true { - echo "## npm audit (Studio frontend)" + echo "## npm audit (Unsloth frontend)" echo echo '```' tail -200 ../../logs-npm-audit.txt @@ -350,9 +350,9 @@ jobs: } >> "$GITHUB_STEP_SUMMARY" # ───────────────────────────────────────────────────────────── - # cargo: Studio Tauri shell + # cargo: Unsloth Tauri shell # ───────────────────────────────────────────────────────────── - - name: cargo audit (Studio Tauri) + - name: cargo audit (Unsloth Tauri) # `--deny warnings` would make the job fail on any advisory. # Keep non-blocking initially; drop continue-on-error after # the baseline closes. @@ -362,7 +362,7 @@ jobs: set +e cargo audit | tee ../../logs-cargo-audit.txt { - echo "## cargo audit (Studio Tauri)" + echo "## cargo audit (Unsloth Tauri)" echo echo '```' tail -200 ../../logs-cargo-audit.txt @@ -559,7 +559,7 @@ jobs: # ───────────────────────────────────────────────────────────── # CycloneDX SBOM. Lets downstream consumers audit what's - # actually shipped in unsloth wheels and the Studio backend + # actually shipped in unsloth wheels and the Unsloth backend # runtime. Generates one JSON file per requirements input plus # a combined SBOM keyed off pyproject.toml; uploads as a build # artifact (and a future step can attest it via SLSA). @@ -740,7 +740,7 @@ jobs: # `--with-deps` makes the scan transitive: every package the # declared set resolves to gets fetched and pattern-scanned, not # just the top-level pins. Resolving the full transitive closure - # of the unsloth + Studio dep tree downloads several hundred + # of the unsloth + Unsloth dep tree downloads several hundred # archives, hence the longer timeout. # # Sharded across runners for wall-clock parallelism. Each shard @@ -749,7 +749,7 @@ jobs: # composition tries to balance load: # - hf-stack: pyproject extras + no-torch-runtime # (~150 archives, transformers/peft/accelerate/...) - # - studio: FastAPI/Studio backend + overrides + extras-no-deps + # - studio: FastAPI/Unsloth backend + overrides + extras-no-deps # (~150 archives, smaller scientific stack) # - extras: the heavy openai-whisper / scikit-learn / librosa # stack (~250 archives, dominant cost) @@ -964,7 +964,7 @@ jobs: # documented at scripts/scan_npm_packages.py top-of-file. The # script is stdlib-only so adding it does not increase the # transitive supply-chain surface. - name: npm scan-packages (Studio frontend tarballs) + name: npm scan-packages (Unsloth frontend tarballs) runs-on: ubuntu-latest timeout-minutes: 30 needs: [] @@ -1173,7 +1173,7 @@ jobs: with: python-version: '3.12' - - name: Install Studio frontend deps (--ignore-scripts) + - name: Install Unsloth frontend deps (--ignore-scripts) # `npm audit signatures` requires node_modules to be populated. # `--ignore-scripts` is mandatory: this is exactly the lever the # new-install-script gate below protects against, and we must diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml index 15efee382e..cdf1f6bf12 100644 --- a/.github/workflows/studio-api-smoke.yml +++ b/.github/workflows/studio-api-smoke.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# Studio API & Auth Tests -- HTTP-level integration tests for the +# Unsloth API & Auth Tests -- HTTP-level integration tests for the # FastAPI surface. No Playwright, no model UI; tests/studio/test_studio_api_smoke.py # runs ~30 s and asserts: # - CORS hardening (no wildcard + credentials, no bootstrap leak) @@ -15,7 +15,7 @@ # Reuses the GGUF cache key from studio-ui-smoke.yml so the model # download is one cache-hit on the second job. -name: Studio API CI +name: Unsloth API CI on: pull_request: @@ -40,7 +40,7 @@ permissions: jobs: api-smoke: - name: Studio API & Auth Tests + name: Unsloth API & Auth Tests runs-on: ubuntu-latest timeout-minutes: 12 env: @@ -98,7 +98,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -111,7 +111,7 @@ jobs: - name: Install pyjwt for the JWT-expiry forge test run: pip install 'pyjwt>=2.6' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -144,7 +144,7 @@ jobs: echo "STUDIO_NEW_PW=$NEW" >> "$GITHUB_ENV" echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV" - - name: Run Studio API & Auth tests + - name: Run Unsloth API & Auth tests # The script is named WITHOUT a `test_` prefix so it isn't # auto-collected by pytest in Backend CI's `tests/` walk # (which doesn't set BASE_URL and would crash at import). @@ -153,7 +153,7 @@ jobs: STUDIO_AUTH_DIR: /home/runner/.unsloth/studio/auth run: python tests/studio/studio_api_smoke.py - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml index a2db26e125..243e295318 100644 --- a/.github/workflows/studio-backend-ci.yml +++ b/.github/workflows/studio-backend-ci.yml @@ -69,7 +69,7 @@ jobs: - name: Install backend test dependencies (CPU only) run: | python -m pip install --upgrade pip - # Studio's declared backend deps: + # Unsloth's declared backend deps: pip install -r studio/backend/requirements/studio.txt # Extras that studio.txt does not list but the import chain needs # (python-multipart for FastAPI form/file uploads, sqlalchemy/cryptography diff --git a/.github/workflows/studio-export-capability-ci.yml b/.github/workflows/studio-export-capability-ci.yml index 1ee6489209..83df3ed476 100644 --- a/.github/workflows/studio-export-capability-ci.yml +++ b/.github/workflows/studio-export-capability-ci.yml @@ -9,7 +9,7 @@ # export is validated separately. No GPU / model / llama.cpp: the tests mock the probes and block # torch/unsloth, so the job installs only a CPU PyTorch plus import deps. -name: Studio export capability +name: Unsloth export capability on: pull_request: diff --git a/.github/workflows/studio-frontend-ci.yml b/.github/workflows/studio-frontend-ci.yml index b42086f191..3a9e373915 100644 --- a/.github/workflows/studio-frontend-ci.yml +++ b/.github/workflows/studio-frontend-ci.yml @@ -136,7 +136,7 @@ jobs: - name: Build run: npm run build - - name: Built bundle must not contain Studio's unstable_Provider call site + - name: Built bundle must not contain Unsloth's unstable_Provider call site run: | set -e JS=$(ls dist/assets/index-*.js | head -1) @@ -144,7 +144,7 @@ jobs: echo "main bundle: $JS" echo "unstable_Provider: hits=$HITS (assistant-ui internals contribute up to 3)" if [ "$HITS" -gt 3 ]; then - echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Studio bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead." + echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Unsloth bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead." exit 1 fi diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml index 58ef2558f3..c2d52eac22 100644 --- a/.github/workflows/studio-inference-smoke.yml +++ b/.github/workflows/studio-inference-smoke.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# Three end-to-end smoke jobs that boot a freshly-installed Studio and +# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and # exercise the surfaces real users hit through the OpenAI / Anthropic # SDKs and curl. Each job picks the smallest model that exercises the # behaviour under test, primes HF_HOME via actions/cache, and shares @@ -27,7 +27,7 @@ # All three jobs run in parallel. Total wall time is dominated by job 3 # on a cold cache; warm cache cuts that to ~3 min. -name: Studio GGUF CI +name: Unsloth GGUF CI on: pull_request: @@ -112,7 +112,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -125,7 +125,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -142,7 +142,7 @@ jobs: fi sleep 1 done - echo "Studio did not become healthy in 180s" + echo "Unsloth did not become healthy in 180s" tail -200 logs/studio.log exit 1 @@ -229,11 +229,11 @@ jobs: return replies def run_anthropic(): - # Two SDK quirks vs. Studio: + # Two SDK quirks vs. Unsloth: # 1. base_url must NOT include /v1 -- the SDK appends # /v1/messages itself; otherwise the request hits # /v1/v1/messages and 405s. - # 2. The SDK sends `x-api-key` by default, but Studio's + # 2. The SDK sends `x-api-key` by default, but Unsloth's # auth layer is HTTPBearer-only. Override via # default_headers so Authorization: Bearer ... is # sent instead. @@ -276,7 +276,7 @@ jobs: print( f"[{label}] WARN non-determinism at temperature=0.0 across " f"{len(determinism_failures)} of {len(first)} turn(s); " - f"small-quant model drift, not a Studio regression. " + f"small-quant model drift, not an Unsloth regression. " f"Details: " + " | ".join(determinism_failures) ) # Sanity: turn-2 reply should mention the earlier question, and @@ -290,7 +290,7 @@ jobs: print(f"[{label}] {status_word} -- 4 turns, history grounded ('paris' present)") PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true @@ -323,7 +323,7 @@ jobs: # store xet chunks + blobs + snapshots = ~4 GiB compressed -- # 4-5x file-size inflation, dominated by xet chunks. Use main's # `--local-dir gguf-cache` pattern to cache the flat .gguf only. - # Studio's /api/inference/load accepts either a HF repo (which + # Unsloth's /api/inference/load accepts either a HF repo (which # uses HF_HOME) or an absolute file path; passing the absolute # path keeps the test off HF_HOME entirely so the cache size # tracks the GGUF file 1:1. The OpenAI/Anth and JSON+images @@ -380,7 +380,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -390,7 +390,7 @@ jobs: set -o pipefail bash install.sh --local --no-torch 2>&1 | tee logs/install.log - - name: Reset auth + boot Studio (API-only, default tool policy) + - name: Reset auth + boot Unsloth (API-only, default tool policy) # We deliberately use the API-only mode rather than # `unsloth studio run` because the latter calls # `set_tool_policy(...)` with a resolved bool: on loopback the @@ -503,7 +503,7 @@ jobs: that the tool path executed. A shared CI runner can stall the stream transport (the - connection opening, or a mid-stream read) even when Studio + connection opening, or a mid-stream read) even when Unsloth is healthy, so retry a stall once with a fresh request capped at 300s. A stall means the stream did NOT complete, so partial events are normally NOT returned (an early @@ -575,11 +575,11 @@ jobs: def _tool_invoked(events): """Structural check: True iff some SSE payload is a real - tool envelope (Studio tool_start/tool_end, Anthropic + tool envelope (Unsloth tool_start/tool_end, Anthropic tool_use/tool_result, OpenAI non-empty delta.tool_calls / message.tool_calls / finish_reason='tool_calls' / role:'tool' / function_call). tool_status is NOT - evidence: Studio emits empty tool_status events on + evidence: Unsloth emits empty tool_status events on iteration boundaries even when no tool ran. """ for raw in events: @@ -698,7 +698,7 @@ jobs: attempt has structural invocation evidence. WARN (not FAIL) if invoked but no attempt produces the expected literal in tool_end.result -- small-quant Qwen3.5-2B can - emit OpenAI tool_calls deltas without Studio's GGUF + emit OpenAI tool_calls deltas without Unsloth's GGUF agentic loop intercepting them, and that GGUF-vs-OpenAI format mismatch is out of scope for #5642. """ @@ -811,7 +811,7 @@ jobs: # because (a) the search may legitimately return no results, # and (b) DuckDuckGo upstream blocks GHA IP ranges often # enough that requiring a tool_call marker would create - # red-herring failures from infra rather than from Studio. + # red-herring failures from infra rather than from Unsloth. try: # Best-effort and bounded: a single 180s attempt keeps a stall # from eating the job's timeout-minutes (it already WARNs, so a @@ -834,7 +834,7 @@ jobs: print(f"[tools] WARN web_search probe failed (non-blocking): {exc}") # ── 5. Thinking on / off ───────────────────────────────────── - # Studio strips think blocks from message.content for tools-mode + # Unsloth strips think blocks from message.content for tools-mode # responses, so we toggle plain chat (no enable_tools) and look # at the surfaced reasoning_content / message.thinking field. def thinking_call(enable): @@ -848,7 +848,7 @@ jobs: }) assert status == 200 msg = data["choices"][0]["message"] - # Studio surfaces thinking via reasoning_content (OpenAI + # Unsloth surfaces thinking via reasoning_content (OpenAI # extension). Fall back to inline markers for # robustness across template versions. raw = (msg.get("content") or "") + (msg.get("reasoning_content") or "") @@ -868,7 +868,7 @@ jobs: print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)") PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true @@ -960,7 +960,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -973,7 +973,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) # See Job 2's comment: API-only mode keeps tool_policy=None so # response_format requests aren't routed through the agentic # tool loop. @@ -1076,13 +1076,13 @@ jobs: # llama.cpp's HTTP server supports OpenAI-compatible JSON # mode: `response_format: {"type": "json_object"}` constrains # the model to emit syntactically-valid JSON. We use raw HTTP - # rather than the OpenAI SDK so that the field shape Studio + # rather than the OpenAI SDK so that the field shape Unsloth # forwards to llama-server is unambiguous (the SDK rewrites # response_format depending on which variant it recognises). # We deliberately do NOT pass a strict JSON schema -- on # small Gemma-4 quants the GBNF-from-schema path occasionally # produces empty output, and JSON mode is the surface we care - # about exposing through Studio. + # about exposing through Unsloth. status, data = post("/v1/chat/completions", { "model": "default", "messages": [ @@ -1112,7 +1112,7 @@ jobs: print(f"[json] PASS json_object -> {parsed}") # ── 2. OpenAI image_url (data URI base64) ─────────────────── - # 64x64 solid-red PNG. stb_image (used by Studio's image + # 64x64 solid-red PNG. stb_image (used by Unsloth's image # normaliser at routes/inference.py:3410) rejects 4x4 or # smaller PNGs as truncated, so we go up to 64x64 -- still # tiny in token cost. The assertion is loose: any non-empty @@ -1148,9 +1148,9 @@ jobs: print("[image/openai] PASS image_url accepted, non-empty response") # ── 3. Anthropic source/base64 image ──────────────────────── - # Two SDK quirks vs. Studio: base_url must NOT include /v1 + # Two SDK quirks vs. Unsloth: base_url must NOT include /v1 # (the SDK appends it itself; otherwise /v1/v1/messages -> 405), - # and Studio's auth is HTTPBearer-only so the SDK's default + # and Unsloth's auth is HTTPBearer-only so the SDK's default # x-api-key header is ignored -- send Authorization: Bearer # via default_headers. anthropic = Anthropic( @@ -1184,7 +1184,7 @@ jobs: print("[image/anthropic] PASS source/base64 accepted, non-empty response") PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-load-orchestrator-ci.yml b/.github/workflows/studio-load-orchestrator-ci.yml index 93d1a7742d..8710efc2bd 100644 --- a/.github/workflows/studio-load-orchestrator-ci.yml +++ b/.github/workflows/studio-load-orchestrator-ci.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. # -# Event-loop regression test for the Studio model-load orchestrator. +# Event-loop regression test for the Unsloth model-load orchestrator. # Pins down issue #5642 (Win10 UI freeze on model load): the /load # route calls LlamaCppBackend.detect_audio_type synchronously, blocking # the FastAPI event loop on a chain of sync httpx.Client.post() probes. @@ -14,7 +14,7 @@ # danielhanchen/unsloth-staging-2 (Ubuntu / macOS / Windows all # green at PR time). -name: Studio load-orchestrator CI +name: Unsloth load-orchestrator CI on: pull_request: diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml index 617ce189dc..1968885a1d 100644 --- a/.github/workflows/studio-mac-api-smoke.yml +++ b/.github/workflows/studio-mac-api-smoke.yml @@ -33,7 +33,7 @@ permissions: jobs: api-smoke: - name: Studio API & Auth Tests + name: Unsloth API & Auth Tests runs-on: macos-14 timeout-minutes: 25 env: @@ -83,7 +83,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -99,7 +99,7 @@ jobs: - name: Install pyjwt for the JWT-expiry forge test run: pip install 'pyjwt>=2.6' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -129,13 +129,13 @@ jobs: echo "STUDIO_NEW_PW=$NEW" >> "$GITHUB_ENV" echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV" - - name: Run Studio API & Auth tests + - name: Run Unsloth API & Auth tests env: BASE_URL: http://127.0.0.1:18895 STUDIO_AUTH_DIR: /Users/runner/.unsloth/studio/auth run: python tests/studio/studio_api_smoke.py - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml index 946681706a..ce15eed5c8 100644 --- a/.github/workflows/studio-mac-inference-smoke.yml +++ b/.github/workflows/studio-mac-inference-smoke.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# Three end-to-end smoke jobs that boot a freshly-installed Studio and +# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and # exercise the surfaces real users hit through the OpenAI / Anthropic # SDKs and curl. Each job picks the smallest model that exercises the # behaviour under test, primes a model cache via actions/cache, and @@ -108,7 +108,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -124,7 +124,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -141,7 +141,7 @@ jobs: fi sleep 1 done - echo "Studio did not become healthy in 180s" + echo "Unsloth did not become healthy in 180s" tail -200 logs/studio.log exit 1 @@ -228,11 +228,11 @@ jobs: return replies def run_anthropic(): - # Two SDK quirks vs. Studio: + # Two SDK quirks vs. Unsloth: # 1. base_url must NOT include /v1 -- the SDK appends # /v1/messages itself; otherwise the request hits # /v1/v1/messages and 405s. - # 2. The SDK sends `x-api-key` by default, but Studio's + # 2. The SDK sends `x-api-key` by default, but Unsloth's # auth layer is HTTPBearer-only. Override via # default_headers so Authorization: Bearer ... is # sent instead. @@ -283,7 +283,7 @@ jobs: print(f"[{label}] OK -- 4 turns, run1 == run2, history grounded") PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true @@ -363,7 +363,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -376,7 +376,7 @@ jobs: - name: Assert llama.cpp loads on this macOS run: bash .github/scripts/assert-llama-loads.sh - - name: Reset auth + boot Studio (API-only, default tool policy) + - name: Reset auth + boot Unsloth (API-only, default tool policy) # We deliberately use the API-only mode rather than # `unsloth studio run` because the latter calls # `set_tool_policy(...)` with a resolved bool: on loopback the @@ -478,7 +478,7 @@ jobs: call with enable_tools=true must use this helper. A shared CI runner can stall the stream transport (the - connection opening, or a mid-stream read) even when Studio + connection opening, or a mid-stream read) even when Unsloth is healthy, so harden the read three ways: retry a stall once with a fresh request capped at 300s; return any text already streamed before a stall (a stall on the trailing @@ -574,11 +574,11 @@ jobs: assert status == 200, f"tool call status {status}: {data}" choice = data["choices"][0] tool_calls = (choice.get("message") or {}).get("tool_calls") or [] - # Studio's contract: when tool_choice='required', llama.cpp's + # Unsloth's contract: when tool_choice='required', llama.cpp's # grammar should force a tool_calls payload. On Mac that # contract is sometimes broken by the underlying quant; the # PASS path is "tool_calls present + correct schema", the - # WARN path documents Studio still returned 200 with a + # WARN path documents Unsloth still returned 200 with a # well-formed choices[] envelope. if tool_calls: tc = tool_calls[0] @@ -660,7 +660,7 @@ jobs: print(f"[tools] WARN web_search probe failed (non-blocking): {exc}") # ── 4. Thinking on / off ───────────────────────────────────── - # Studio strips think blocks from message.content for tools-mode + # Unsloth strips think blocks from message.content for tools-mode # responses, so we toggle plain chat (no enable_tools) and look # at the surfaced reasoning_content / message.thinking field. def thinking_call(enable): @@ -678,7 +678,7 @@ jobs: }, timeout = 180) assert status == 200 msg = data["choices"][0]["message"] - # Studio surfaces thinking via reasoning_content (OpenAI + # Unsloth surfaces thinking via reasoning_content (OpenAI # extension). Fall back to inline markers for # robustness across template versions. raw = (msg.get("content") or "") + (msg.get("reasoning_content") or "") @@ -704,7 +704,7 @@ jobs: print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)") PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true @@ -810,7 +810,7 @@ jobs: path: gguf-cache key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-${{ env.MMPROJ_FILE }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -826,7 +826,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) # See Job 2's comment: API-only mode keeps tool_policy=None so # response_format requests aren't routed through the agentic # tool loop. @@ -929,13 +929,13 @@ jobs: # llama.cpp's HTTP server supports OpenAI-compatible JSON # mode: `response_format: {"type": "json_object"}` constrains # the model to emit syntactically-valid JSON. We use raw HTTP - # rather than the OpenAI SDK so that the field shape Studio + # rather than the OpenAI SDK so that the field shape Unsloth # forwards to llama-server is unambiguous (the SDK rewrites # response_format depending on which variant it recognises). # We deliberately do NOT pass a strict JSON schema -- on # small Gemma-4 quants the GBNF-from-schema path occasionally # produces empty output, and JSON mode is the surface we care - # about exposing through Studio. + # about exposing through Unsloth. status, data = post("/v1/chat/completions", { "model": "default", "messages": [ @@ -1007,7 +1007,7 @@ jobs: ) # ── 2. OpenAI image_url (data URI base64) ─────────────────── - # 64x64 solid-red PNG. stb_image (used by Studio's image + # 64x64 solid-red PNG. stb_image (used by Unsloth's image # normaliser at routes/inference.py:3410) rejects 4x4 or # smaller PNGs as truncated, so we go up to 64x64 -- still # tiny in token cost. The assertion is loose: any non-empty @@ -1023,11 +1023,11 @@ jobs: # The Mac prebuilt llama.cpp server has a known crash when # processing image inputs alongside the gemma-4-E2B mmproj # (server disconnects mid-completion). This is upstream - # llama.cpp behaviour, not Studio. Wrap both SDK calls in + # llama.cpp behaviour, not Unsloth. Wrap both SDK calls in # try/except so an upstream crash registers as a WARN rather - # than failing the whole job. Studio's contract (OpenAI/ + # than failing the whole job. Unsloth's contract (OpenAI/ # Anthropic image fields are accepted and forwarded) is - # validated by the request body Studio constructs, not by + # validated by the request body Unsloth constructs, not by # whether llama.cpp can decode it on Mac Metal. client = OpenAI(base_url = f"{BASE}/v1", api_key = KEY) try: @@ -1053,14 +1053,14 @@ jobs: except Exception as exc: print( f"[image/openai] WARN image_url SDK call raised: {type(exc).__name__}: " - f"{exc}. Likely upstream llama.cpp Mac+vision crash, NOT a Studio " - f"regression. Studio successfully forwarded the request." + f"{exc}. Likely upstream llama.cpp Mac+vision crash, NOT an Unsloth " + f"regression. Unsloth successfully forwarded the request." ) # ── 3. Anthropic source/base64 image ──────────────────────── - # Two SDK quirks vs. Studio: base_url must NOT include /v1 + # Two SDK quirks vs. Unsloth: base_url must NOT include /v1 # (the SDK appends it itself; otherwise /v1/v1/messages -> 405), - # and Studio's auth is HTTPBearer-only so the SDK's default + # and Unsloth's auth is HTTPBearer-only so the SDK's default # x-api-key header is ignored -- send Authorization: Bearer # via default_headers. anthropic = Anthropic( @@ -1099,11 +1099,11 @@ jobs: print( f"[image/anthropic] WARN anthropic image SDK call raised: " f"{type(exc).__name__}: {exc}. Likely upstream llama.cpp Mac+vision " - f"crash, NOT a Studio regression." + f"crash, NOT an Unsloth regression." ) PY - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-mac-install-matrix.yml b/.github/workflows/studio-mac-install-matrix.yml index 362305cdd4..e990f752d4 100644 --- a/.github/workflows/studio-mac-install-matrix.yml +++ b/.github/workflows/studio-mac-install-matrix.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# Proves Studio's llama.cpp install loads on every supported macOS. The heavy +# Proves Unsloth's llama.cpp install loads on every supported macOS. The heavy # app smokes stay single-OS; this matrix covers the OS-version dimension cheaply # (install.sh + binary-load assert). Regression guard for the macOS-version # selection in studio/install_llama_prebuilt.py. @@ -60,7 +60,7 @@ jobs: with: python-version: '3.12' - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml index 20ca247b9f..378e8ee5a6 100644 --- a/.github/workflows/studio-mac-ui-smoke.yml +++ b/.github/workflows/studio-mac-ui-smoke.yml @@ -83,7 +83,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -143,7 +143,7 @@ jobs: print(f"pipeTransport.js: patched JSON.parse calls in {path}") PY - - name: Reset auth + boot Studio + - name: Reset auth + boot Unsloth run: | unsloth studio reset-password mkdir -p logs @@ -188,7 +188,7 @@ jobs: # dies mid-test, (2) Chromium net::ERR_NO_BUFFER_SPACE when the # runner's kernel briefly runs out of socket buffers, and (3) a # goto 'interrupted by another navigation' when the SPA auth - # guard redirects mid-navigation. The retry FULLY resets Studio + # guard redirects mid-navigation. The retry FULLY resets Unsloth # (kill, reset-password, reboot, wait /api/health, re-export # bootstrap pw) before re-running the script. A real test failure # (assertion / timeout) does NOT match any pattern so it bypasses @@ -209,7 +209,7 @@ jobs: || grep -q "ERR_NO_BUFFER_SPACE" logs/playwright_attempt_${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..." + echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..." kill "${STUDIO_PID}" 2>/dev/null || true sleep 2 unsloth studio reset-password @@ -238,13 +238,13 @@ jobs: exit "$rc" done - - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders) + - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders) if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true sleep 2 - - name: Reset auth + boot Studio for extra UI tests (port 18897) + - name: Reset auth + boot Unsloth for extra UI tests (port 18897) run: | unsloth studio reset-password mkdir -p logs @@ -271,7 +271,7 @@ jobs: echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV" echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV" - - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright + - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright env: BASE_URL: http://127.0.0.1:18897 STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }} @@ -300,7 +300,7 @@ jobs: || grep -q "ERR_NO_BUFFER_SPACE" logs/playwright_extra_attempt_${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..." + echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..." kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true sleep 2 unsloth studio reset-password @@ -327,7 +327,7 @@ jobs: exit "$rc" done - - name: Stop second Studio + - name: Stop second Unsloth if: always() run: | kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-mac-update-smoke.yml b/.github/workflows/studio-mac-update-smoke.yml index d104306c7e..fe9880f3ca 100644 --- a/.github/workflows/studio-mac-update-smoke.yml +++ b/.github/workflows/studio-mac-update-smoke.yml @@ -4,15 +4,15 @@ # Mac counterpart to studio-update-smoke.yml. Verifies that on a real # Apple Silicon (macos-14, M1) runner: # -# 1. install.sh --local --no-torch installs Studio AND auto-fetches +# 1. install.sh --local --no-torch installs Unsloth AND auto-fetches # the prebuilt llama.cpp Mac binary (llama-bNNNN-bin-macos-arm64 # from ggml-org/llama.cpp). Hitting the source-build fallback is -# treated as an Unsloth bug -- Studio must always pick the +# treated as an Unsloth bug -- Unsloth must always pick the # prebuilt on Mac. # 2. unsloth studio update --local is idempotent. Two consecutive # runs both report "prebuilt up to date and validated", no # source-build fallback. -# 3. The installed Studio still boots and /api/health returns +# 3. The installed Unsloth still boots and /api/health returns # healthy after the update path. name: Mac Studio Update CI @@ -42,7 +42,7 @@ permissions: jobs: update-idempotency: - name: Studio Updating Tests + name: Unsloth Updating Tests runs-on: macos-14 timeout-minutes: 30 steps: @@ -59,7 +59,7 @@ jobs: python-version: '3.12' cache: 'pip' - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -106,7 +106,7 @@ jobs: grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log echo "second update was clean" - - name: Boot Studio briefly to confirm the install is still usable + - name: Boot Unsloth briefly to confirm the install is still usable run: | mkdir -p logs UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18891 \ @@ -123,13 +123,13 @@ jobs: sleep 1 done if [ -z "$HEALTHY" ]; then - echo "Studio failed to come up after \`update\`" + echo "Unsloth failed to come up after \`update\`" tail -200 logs/studio.log kill "$PID" 2>/dev/null || true exit 1 fi kill "$PID" 2>/dev/null || true - echo "post-update Studio /api/health OK" + echo "post-update Unsloth /api/health OK" - name: Uninstall and verify clean # Round-trip through scripts/uninstall.sh on real macOS. As a side diff --git a/.github/workflows/studio-tauri-smoke.yml b/.github/workflows/studio-tauri-smoke.yml index 018857de68..8e26b9fd0c 100644 --- a/.github/workflows/studio-tauri-smoke.yml +++ b/.github/workflows/studio-tauri-smoke.yml @@ -12,7 +12,7 @@ # stay in release-desktop.yml (manual `workflow_dispatch`) because they need # code-signing secrets and ~30 min of runner time each. -name: Studio Tauri CI +name: Unsloth Tauri CI on: pull_request: diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml index 297a585430..b6d6d7d6e2 100644 --- a/.github/workflows/studio-ui-smoke.yml +++ b/.github/workflows/studio-ui-smoke.yml @@ -1,8 +1,8 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# End-to-end Studio chat UI smoke via Playwright + Chromium against a -# headless Linux runner. Boots Studio with the smallest GGUF +# End-to-end Unsloth chat UI smoke via Playwright + Chromium against a +# headless Linux runner. Boots Unsloth with the smallest GGUF # (gemma-3-270m-it UD-Q4_K_XL, ~254 MiB), drives the actual frontend # bundle, and asserts the full bootstrap-password / change-password / # send-message / persist-on-reload journey works end to end. @@ -14,7 +14,7 @@ # frontend-only CI happily pass while the actual user-visible UI is # broken (cf. the 2026.5.1 chat-history release). -name: Studio UI CI +name: Unsloth UI CI on: pull_request: @@ -97,7 +97,7 @@ jobs: path: hf-cache key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2 - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads. @@ -115,7 +115,7 @@ jobs: # warm runner. python -m playwright install --with-deps chromium - - name: Reset auth + boot Studio + - name: Reset auth + boot Unsloth run: | unsloth studio reset-password mkdir -p logs @@ -147,7 +147,7 @@ jobs: # NEW + NEW2 are generated freshly per CI run via secrets.token_urlsafe # rather than hardcoded. If a workflow gets compromised, the # attacker can't replay a known-good rotated password against - # any future / parallel Studio install -- the rotated value + # any future / parallel Unsloth install -- the rotated value # only ever exists for the lifetime of this single job, masked # in the log via ::add-mask::. run: | @@ -165,18 +165,18 @@ jobs: env: BASE_URL: http://127.0.0.1:18892 # The test file lives in the repo so it can be run locally - # against a freshly-installed Studio (BASE_URL=...; STUDIO_OLD_PW= + # against a freshly-installed Unsloth (BASE_URL=...; STUDIO_OLD_PW= # $(cat ~/.unsloth/studio/auth/.bootstrap_password); python ...). PW_ART_DIR: logs/playwright # Strict mode: in CI a missing button / nav / dialog must # FAIL the test. Locally the test still runs against partial - # Studio installs without STUDIO_UI_STRICT. + # Unsloth installs without STUDIO_UI_STRICT. STUDIO_UI_STRICT: '1' run: | mkdir -p logs/playwright python tests/studio/playwright_chat_ui.py - - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders) + - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders) if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true @@ -184,10 +184,10 @@ jobs: # The chat UI test ends by clicking the Shutdown menuitem, which # leaves the server dead. The extra UI test (Compare / Recipes / - # Export / Studio / Settings) needs a fresh Studio, so we boot a + # Export / Unsloth / Settings) needs a fresh Unsloth, so we boot a # second one on a different port. Boot is fast (~3-5s on the # warm install we already did) so this adds little wall time. - - name: Reset auth + boot Studio for extra UI tests (port 18894) + - name: Reset auth + boot Unsloth for extra UI tests (port 18894) run: | unsloth studio reset-password mkdir -p logs @@ -214,7 +214,7 @@ jobs: echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV" echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV" - - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright + - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright env: BASE_URL: http://127.0.0.1:18894 STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }} @@ -227,16 +227,16 @@ jobs: mkdir -p logs/playwright_extra python tests/studio/playwright_extra_ui.py - - name: Stop second Studio + - name: Stop second Unsloth if: always() run: | kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true sleep 2 # IME + multilingual paste regression (issue #5318 / PR #5327). - # Third Studio on its own port so a hang here cannot poison the + # Third Unsloth on its own port so a hang here cannot poison the # earlier UI tests. No GGUF -- the bug surface is the composer. - - name: Reset auth + boot Studio for IME / i18n tests (port 18896) + - name: Reset auth + boot Unsloth for IME / i18n tests (port 18896) run: | unsloth studio reset-password mkdir -p logs @@ -256,7 +256,7 @@ jobs: - name: Pass bootstrap pw for IME / i18n test # IME smoke does the change-password against the bootstrap that - # Studio's frontend injects into the page, so it only needs the + # Unsloth's frontend injects into the page, so it only needs the # NEW password. run: | NEW="CIIme-$(python -c 'import secrets; print(secrets.token_urlsafe(16))')" @@ -273,7 +273,7 @@ jobs: mkdir -p logs/playwright_ime python tests/studio/playwright_chat_ime_i18n.py - - name: Stop third Studio + - name: Stop third Unsloth if: always() run: | kill "${STUDIO_IME_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-update-smoke.yml b/.github/workflows/studio-update-smoke.yml index 08a79afacd..625c2c7811 100644 --- a/.github/workflows/studio-update-smoke.yml +++ b/.github/workflows/studio-update-smoke.yml @@ -9,7 +9,7 @@ # This catches regressions in setup.sh's update path that the existing # GGUF / wheel jobs would miss because they only invoke install.sh once. -name: Studio Update CI +name: Unsloth Update CI on: pull_request: @@ -36,7 +36,7 @@ permissions: jobs: update-idempotency: - name: Studio Updating Tests + name: Unsloth Updating Tests runs-on: ubuntu-latest timeout-minutes: 15 steps: @@ -63,7 +63,7 @@ jobs: # post-step then fatal-errors with "Cache folder path is # retrieved for pip but doesn't exist on disk". - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) # Pass the workflow token so the llama.cpp prebuilt installer's # GitHub-API call to list releases isn't rate-limited (60/hr # unauthenticated). Without this, three consecutive install + @@ -122,7 +122,7 @@ jobs: grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log echo "second update was clean" - - name: Boot Studio briefly to confirm the install is still usable + - name: Boot Unsloth briefly to confirm the install is still usable # If `update --local` accidentally broke the venv or wiped the # llama-server binary, the server would fail to start here. run: | @@ -138,13 +138,13 @@ jobs: sleep 1 done if ! jq -e '.status == "healthy"' /tmp/health.json 2>/dev/null; then - echo "Studio failed to come up after `update`" + echo "Unsloth failed to come up after `update`" tail -200 logs/studio.log kill "$PID" 2>/dev/null || true exit 1 fi kill "$PID" 2>/dev/null || true - echo "post-update Studio /api/health OK" + echo "post-update Unsloth /api/health OK" - name: Uninstall and verify clean # Round-trip the installer through scripts/uninstall.sh: confirms the diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml index e9abd2d669..6dbcceebbd 100644 --- a/.github/workflows/studio-windows-api-smoke.yml +++ b/.github/workflows/studio-windows-api-smoke.yml @@ -9,7 +9,7 @@ # (Section 6) is Linux-only and short-circuits on non-POSIX; the rest # is platform-portable. -name: Windows Studio API CI +name: Windows Unsloth API CI on: pull_request: @@ -34,7 +34,7 @@ permissions: jobs: api-smoke: - name: Studio API & Auth Tests + name: Unsloth API & Auth Tests runs-on: windows-latest timeout-minutes: 30 defaults: @@ -105,7 +105,7 @@ jobs: # studio-windows-update-smoke.yml for the full rationale -- # creating an empty studio/frontend/dist trips setup.ps1's # mtime-based staleness check into "frontend up to date, skip - # rebuild" and Studio boots with an empty dist directory. + # rebuild" and Unsloth boots with an empty dist directory. # Add-MpPreference accepts paths that do not yet exist. foreach ($p in @( "$env:USERPROFILE\.unsloth", @@ -121,7 +121,7 @@ jobs: } } - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -161,7 +161,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH # install.ps1's User-PATH update doesn't propagate to a # running Git Bash session; export the shim dir so the # next `unsloth ...` invocation finds it. @@ -177,7 +177,7 @@ jobs: - name: Install pyjwt for the JWT-expiry forge test run: python -m pip install 'pyjwt>=2.6' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -207,7 +207,7 @@ jobs: echo "STUDIO_NEW_PW=$NEW" >> "$GITHUB_ENV" echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV" - - name: Run Studio API & Auth tests + - name: Run Unsloth API & Auth tests # Do NOT pin STUDIO_AUTH_DIR here. The Mac/Linux mirrors # hardcode runner-specific paths (/Users/runner/..., # /home/runner/...), but on Windows the path is @@ -219,7 +219,7 @@ jobs: BASE_URL: http://127.0.0.1:18895 run: python tests/studio/studio_api_smoke.py - - name: Stop Studio + - name: Stop Unsloth if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml index 63a7e9dc8f..3ebe442f52 100644 --- a/.github/workflows/studio-windows-inference-smoke.yml +++ b/.github/workflows/studio-windows-inference-smoke.yml @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -# Three end-to-end smoke jobs that boot a freshly-installed Studio and +# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and # exercise the surfaces real users hit through the OpenAI / Anthropic # SDKs and curl, on the FREE windows-latest runner. Each job picks the # smallest model that exercises the behaviour under test, primes @@ -16,7 +16,7 @@ # Qwen3-VL-2B-Instruct UD-IQ2_XXS + mmproj-F16 (~1.4 GiB total). # Within the 14 GB windows-latest SSD budget. -name: Windows Studio GGUF CI +name: Windows Unsloth GGUF CI on: pull_request: @@ -57,7 +57,7 @@ jobs: STUDIO_PORT: '18888' HF_HOME: ${{ github.workspace }}/hf-cache # Force UTF-8 for stdio (Windows defaults to cp1252; hf - # download / Studio CLI print "✓" checkmarks and crash + # download / Unsloth CLI print "✓" checkmarks and crash # otherwise). PYTHONIOENCODING: utf-8 PYTHONUTF8: '1' @@ -160,7 +160,7 @@ jobs: # studio-windows-update-smoke.yml for the full rationale -- # creating an empty studio/frontend/dist trips setup.ps1's # mtime-based staleness check into "frontend up to date, skip - # rebuild" and Studio boots with an empty dist directory. + # rebuild" and Unsloth boots with an empty dist directory. # Add-MpPreference accepts paths that do not yet exist. foreach ($p in @( "$env:USERPROFILE\.unsloth", @@ -176,7 +176,7 @@ jobs: } } - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -214,7 +214,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH run: | SHIM_DIR=~/.unsloth/studio/bin if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then @@ -227,7 +227,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: python -m pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -244,7 +244,7 @@ jobs: fi sleep 1 done - echo "Studio did not become healthy in 180s" + echo "Unsloth did not become healthy in 180s" tail -200 logs/studio.log exit 1 @@ -281,7 +281,7 @@ jobs: # Retry the load step a few times so a transient TCP RST during # llama-server warm-up (Windows runner image churn, # windows-latest -> windows-2025-vs2026 rollout) doesn't fail - # the whole job. The Studio backend's _wait_for_health now + # the whole job. The Unsloth backend's _wait_for_health now # catches httpx.ReadError too; this retry layer covers the # cases the backend can't recover from on its own. LOAD_OK=0 @@ -382,15 +382,15 @@ jobs: print(f"[{label}] OK -- 4 turns, run1 == run2, history grounded") PY - - name: Stop Studio + - name: Stop Unsloth if: always() # Run as cmd so we are not running through the Git Bash shell; # Git Bash on windows-latest has been observed to exit 143 # (SIGTERM) from any inline kill/sleep block, masking a green - # test run. The runner reclaims the Studio child process at + # test run. The runner reclaims the Unsloth child process at # job end either way, so just emit a marker and exit 0. shell: cmd - run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) + run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) - name: Collect llama-server logs if: always() @@ -398,10 +398,10 @@ jobs: # copy must not fail an otherwise-green job. continue-on-error: true shell: bash - # Copy llama-server's own stdout/stderr (teed by Studio under + # Copy llama-server's own stdout/stderr (teed by Unsloth under # ~/.unsloth/studio/logs/llama-server/) into the workspace so # upload-artifact can pick it up. Crucial for diagnosing a - # subprocess crash where Studio's traceback only shows the + # subprocess crash where Unsloth's traceback only shows the # symptom (httpx ReadError) but not the cause. run: | mkdir -p logs/llama-server @@ -439,14 +439,14 @@ jobs: # (211 s on first run; subsequent runs hit the cache, but the # one-time cost recurs every time the cache key bumps). Use # main's `--local-dir gguf-cache` pattern: cache the flat .gguf - # only, pass an absolute path to Studio's /api/inference/load. + # only, pass an absolute path to Unsloth's /api/inference/load. # The OpenAI/Anth and JSON+images jobs still cover the # gguf_variant resolution path. GGUF_REPO: unsloth/Qwen3.5-2B-GGUF GGUF_FILE: Qwen3.5-2B-UD-Q4_K_XL.gguf STUDIO_PORT: '18898' # Force UTF-8 for stdio (Windows defaults to cp1252; hf - # download / Studio CLI print "✓" checkmarks and crash + # download / Unsloth CLI print "✓" checkmarks and crash # otherwise). PYTHONIOENCODING: utf-8 PYTHONUTF8: '1' @@ -507,7 +507,7 @@ jobs: # studio-windows-update-smoke.yml for the full rationale -- # creating an empty studio/frontend/dist trips setup.ps1's # mtime-based staleness check into "frontend up to date, skip - # rebuild" and Studio boots with an empty dist directory. + # rebuild" and Unsloth boots with an empty dist directory. # Add-MpPreference accepts paths that do not yet exist. foreach ($p in @( "$env:USERPROFILE\.unsloth", @@ -523,7 +523,7 @@ jobs: } } - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -561,7 +561,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH run: | SHIM_DIR=~/.unsloth/studio/bin if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then @@ -571,7 +571,7 @@ jobs: fi cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH" - - name: Reset auth + boot Studio (API-only, default tool policy) + - name: Reset auth + boot Unsloth (API-only, default tool policy) run: | unsloth studio reset-password mkdir -p logs @@ -607,7 +607,7 @@ jobs: # raw string, but we cannot embed `\a` etc. in JSON without # JSON-string-escaping every backslash. Replace `\` with `/` # via bash parameter expansion -- pathlib.Path on Windows - # accepts forward slashes natively, so Studio's loader sees + # accepts forward slashes natively, so Unsloth's loader sees # a normal path. GGUF_PATH="${GITHUB_WORKSPACE//\\//}/gguf-cache/${GGUF_FILE}" ls -lh "$GGUF_PATH" @@ -680,7 +680,7 @@ jobs: def post_sse(path, body, *, timeout = 600, retries = 1, soft = False): # The server-side agentic loop always answers over SSE. A # shared CI runner can stall the stream transport (the - # connection opening, or a mid-stream read) even when Studio + # connection opening, or a mid-stream read) even when Unsloth # is healthy, so harden the read three ways: # * retry a transport stall once with a fresh request, # capped at 300s (a healthy server answers a retry @@ -882,15 +882,15 @@ jobs: print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)") PY - - name: Stop Studio + - name: Stop Unsloth if: always() # Run as cmd so we are not running through the Git Bash shell; # Git Bash on windows-latest has been observed to exit 143 # (SIGTERM) from any inline kill/sleep block, masking a green - # test run. The runner reclaims the Studio child process at + # test run. The runner reclaims the Unsloth child process at # job end either way, so just emit a marker and exit 0. shell: cmd - run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) + run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) - name: Collect llama-server logs if: always() @@ -898,10 +898,10 @@ jobs: # copy must not fail an otherwise-green job. continue-on-error: true shell: bash - # Copy llama-server's own stdout/stderr (teed by Studio under + # Copy llama-server's own stdout/stderr (teed by Unsloth under # ~/.unsloth/studio/logs/llama-server/) into the workspace so # upload-artifact can pick it up. Crucial for diagnosing a - # subprocess crash where Studio's traceback only shows the + # subprocess crash where Unsloth's traceback only shows the # symptom (httpx ReadError) but not the cause. run: | mkdir -p logs/llama-server @@ -939,7 +939,7 @@ jobs: STUDIO_PORT: '18899' HF_HOME: ${{ github.workspace }}/hf-cache # Force UTF-8 for stdio (Windows defaults to cp1252; hf - # download / Studio CLI print "✓" checkmarks and crash + # download / Unsloth CLI print "✓" checkmarks and crash # otherwise). PYTHONIOENCODING: utf-8 PYTHONUTF8: '1' @@ -1005,7 +1005,7 @@ jobs: # studio-windows-update-smoke.yml for the full rationale -- # creating an empty studio/frontend/dist trips setup.ps1's # mtime-based staleness check into "frontend up to date, skip - # rebuild" and Studio boots with an empty dist directory. + # rebuild" and Unsloth boots with an empty dist directory. # Add-MpPreference accepts paths that do not yet exist. foreach ($p in @( "$env:USERPROFILE\.unsloth", @@ -1021,7 +1021,7 @@ jobs: } } - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -1059,7 +1059,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH run: | SHIM_DIR=~/.unsloth/studio/bin if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then @@ -1072,7 +1072,7 @@ jobs: - name: Install OpenAI + Anthropic Python SDKs run: python -m pip install 'openai>=1.50' 'anthropic>=0.40' - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -1262,7 +1262,7 @@ jobs: except Exception as exc: print( f"[image/openai] WARN image_url SDK call raised: {type(exc).__name__}: " - f"{exc}. Studio successfully forwarded the request; failure here is " + f"{exc}. Unsloth successfully forwarded the request; failure here is " f"upstream llama.cpp vision behaviour." ) @@ -1303,19 +1303,19 @@ jobs: print( f"[image/anthropic] WARN anthropic image SDK call raised: " f"{type(exc).__name__}: {exc}. Likely upstream llama.cpp vision " - f"behaviour, NOT a Studio regression." + f"behaviour, NOT an Unsloth regression." ) PY - - name: Stop Studio + - name: Stop Unsloth if: always() # Run as cmd so we are not running through the Git Bash shell; # Git Bash on windows-latest has been observed to exit 143 # (SIGTERM) from any inline kill/sleep block, masking a green - # test run. The runner reclaims the Studio child process at + # test run. The runner reclaims the Unsloth child process at # job end either way, so just emit a marker and exit 0. shell: cmd - run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) + run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) - name: Collect llama-server logs if: always() @@ -1323,10 +1323,10 @@ jobs: # copy must not fail an otherwise-green job. continue-on-error: true shell: bash - # Copy llama-server's own stdout/stderr (teed by Studio under + # Copy llama-server's own stdout/stderr (teed by Unsloth under # ~/.unsloth/studio/logs/llama-server/) into the workspace so # upload-artifact can pick it up. Crucial for diagnosing a - # subprocess crash where Studio's traceback only shows the + # subprocess crash where Unsloth's traceback only shows the # symptom (httpx ReadError) but not the cause. run: | mkdir -p logs/llama-server @@ -1348,7 +1348,7 @@ jobs: # ── folded from studio-windows-no-vs-smoke.yml: install + run with no Visual Studio ── no-vs-cpu: - name: Studio install + inference without Visual Studio + name: Unsloth install + inference without Visual Studio runs-on: windows-latest timeout-minutes: 35 defaults: @@ -1502,7 +1502,7 @@ jobs: 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 + - name: Install Unsloth (--local, --no-torch) with no build tools present shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -1538,13 +1538,13 @@ jobs: echo "Prebuilt installed with no build tools:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH run: | SHIM_DIR=~/.unsloth/studio/bin [ -f "$SHIM_DIR/unsloth.exe" ] || { echo "::error::unsloth.exe shim not found"; ls -la ~/.unsloth/studio/ || true; exit 1; } cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH" - - name: Reset auth + boot Studio (API-only) + - name: Reset auth + boot Unsloth (API-only) run: | unsloth studio reset-password mkdir -p logs @@ -1613,10 +1613,10 @@ jobs: } Remove-Item -LiteralPath $root -Recurse -Force -ErrorAction SilentlyContinue - - name: Stop Studio + - name: Stop Unsloth if: always() shell: cmd - run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) + run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end) - name: Collect llama-server logs if: always() diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml index 405309916a..12d7475b53 100644 --- a/.github/workflows/studio-windows-ui-smoke.yml +++ b/.github/workflows/studio-windows-ui-smoke.yml @@ -4,11 +4,11 @@ # Windows counterpart to studio-ui-smoke.yml / studio-mac-ui-smoke.yml. # Same Playwright + Chromium end-to-end chat UI flow + extra UI flow, # but on the FREE windows-latest runner so we catch Windows-specific -# regressions in the install path (install.ps1), the Studio CLI's +# regressions in the install path (install.ps1), the Unsloth CLI's # Windows process-management branches, and the llama.cpp prebuilt's # Windows HTTP layer. -name: Windows Studio UI CI +name: Windows Unsloth UI CI on: pull_request: @@ -49,7 +49,7 @@ jobs: GGUF_FILE: gemma-3-270m-it-UD-Q4_K_XL.gguf STUDIO_PORT: '18896' HF_HOME: ${{ github.workspace }}/hf-cache - # Force UTF-8 for stdio so Python tools (hf download, Studio + # Force UTF-8 for stdio so Python tools (hf download, Unsloth # CLI, etc.) can print Unicode characters like the success # checkmark "✓". Windows defaults to cp1252 / charmap and # any tool that prints "OK ✓" hits a UnicodeEncodeError. @@ -121,7 +121,7 @@ jobs: # studio-windows-update-smoke.yml for the full rationale -- # creating an empty studio/frontend/dist trips setup.ps1's # mtime-based staleness check into "frontend up to date, skip - # rebuild" and Studio boots with an empty dist directory. + # rebuild" and Unsloth boots with an empty dist directory. # Add-MpPreference accepts paths that do not yet exist. foreach ($p in @( "$env:USERPROFILE\.unsloth", @@ -148,7 +148,7 @@ jobs: Set-Content -LiteralPath (Join-Path $appDir 'launch-studio.vbs') -Value 'WScript.Echo "legacy"' -Encoding Unicode Write-Host "seeded legacy launch-studio.vbs at $appDir" - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) # install.ps1 is the supported Windows installer. install.sh # has no Windows branch (apt-get / brew calls). The PS1 # script's `Install-UnslothStudio @args` line at the bottom @@ -205,7 +205,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Assert Studio launcher chain (no VBS, hidden PowerShell shortcut) + - name: Assert Unsloth launcher chain (no VBS, hidden PowerShell shortcut) # The shortcut launch path is otherwise untested here (the steps below # boot `unsloth studio` directly). Guard against re-introducing the VBS # that tripped Kaspersky HEUR:Trojan.VBS.Agent.gen and against the .lnk @@ -234,7 +234,7 @@ jobs: } Write-Host "launcher chain OK (no VBS; hidden powershell over launch-studio.ps1)" - - name: Launch Studio via the shortcut and assert health + - name: Launch Unsloth via the shortcut and assert health # Run the exact command the .lnk stores (hidden PowerShell over # launch-studio.ps1) and confirm it brings the backend up. This is the # only step that proves the shortcut launch is not silently broken. @@ -265,10 +265,10 @@ jobs: $owner = (Get-NetTCPConnection -LocalPort $foundPort -State Listen -ErrorAction Stop | Select-Object -First 1).OwningProcess if ($owner) { taskkill /PID $owner /T /F 2>$null | Out-Null } } catch {} - if (-not $foundPort) { throw "Studio did not become healthy when launched via the shortcut" } - Write-Host "Studio healthy on port $foundPort (launched via the shortcut)" + if (-not $foundPort) { throw "Unsloth did not become healthy when launched via the shortcut" } + Write-Host "Unsloth healthy on port $foundPort (launched via the shortcut)" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH # install.ps1 puts unsloth.exe at $StudioHome\bin\unsloth.exe # and adds that dir to the User PATH via the Windows registry. # Registry-level PATH updates don't propagate to a running @@ -284,7 +284,7 @@ jobs: fi # GITHUB_PATH wants Windows-style paths; convert via cygpath. cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH" - echo "Added Studio shim dir to PATH: $(cygpath -w "$SHIM_DIR")" + echo "Added Unsloth shim dir to PATH: $(cygpath -w "$SHIM_DIR")" - name: Install Playwright + Chromium # No --with-deps on Windows: that flag installs Linux apt @@ -294,7 +294,7 @@ jobs: python -m pip install 'playwright>=1.45' python -m playwright install chromium - - name: Reset auth + boot Studio + - name: Reset auth + boot Unsloth run: | unsloth studio reset-password mkdir -p logs @@ -339,13 +339,13 @@ jobs: mkdir -p logs/playwright python tests/studio/playwright_chat_ui.py - - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders) + - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders) if: always() run: | kill "${STUDIO_PID}" 2>/dev/null || true sleep 2 - - name: Reset auth + boot Studio for extra UI tests (port 18897) + - name: Reset auth + boot Unsloth for extra UI tests (port 18897) run: | unsloth studio reset-password mkdir -p logs @@ -372,7 +372,7 @@ jobs: echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV" echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV" - - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright + - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright env: BASE_URL: http://127.0.0.1:18897 STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }} @@ -386,7 +386,7 @@ jobs: mkdir -p logs/playwright_extra python tests/studio/playwright_extra_ui.py - - name: Stop second Studio + - name: Stop second Unsloth if: always() run: | kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml index 5b92f1a3e0..42d74d47d2 100644 --- a/.github/workflows/studio-windows-update-smoke.yml +++ b/.github/workflows/studio-windows-update-smoke.yml @@ -5,19 +5,19 @@ # studio-mac-update-smoke.yml. Verifies that on the FREE # windows-latest runner: # -# 1. install.ps1 --local --no-torch installs Studio AND auto-fetches +# 1. install.ps1 --local --no-torch installs Unsloth AND auto-fetches # the prebuilt llama.cpp Windows binary (app--windows-x64-cpu # from unslothai/llama.cpp). Hitting the source-build fallback is -# treated as an Unsloth bug -- Studio must always pick the +# treated as an Unsloth bug -- Unsloth must always pick the # prebuilt on Windows. # 2. unsloth studio update --local is idempotent. Two consecutive # runs both report "prebuilt up to date and validated", no # source-build fallback. The CLI's _find_setup_script picks # setup.ps1 on Windows automatically. -# 3. The installed Studio still boots and /api/health returns +# 3. The installed Unsloth still boots and /api/health returns # healthy after the update path. -name: Windows Studio Update CI +name: Windows Unsloth Update CI on: pull_request: @@ -45,7 +45,7 @@ permissions: jobs: update-idempotency: - name: Studio Updating Tests + name: Unsloth Updating Tests runs-on: windows-latest timeout-minutes: 30 defaults: @@ -53,7 +53,7 @@ jobs: shell: bash env: # Force UTF-8 for stdio (Windows defaults to cp1252; hf - # download / Studio CLI print "✓" checkmarks and crash + # download / Unsloth CLI print "✓" checkmarks and crash # otherwise). PYTHONIOENCODING: utf-8 PYTHONUTF8: '1' @@ -90,7 +90,7 @@ jobs: # reuses the existing Node with no download. # # (2) Defender. windows-latest's real-time scan opens / hashes - # every file Studio writes during install (Vite output = + # every file Unsloth writes during install (Vite output = # thousands of small chunks, uv pip = wheel-extraction = # thousands of small files). The latency dominates the # 200 s frontend build and the 90 s deps install. Adding @@ -109,7 +109,7 @@ jobs: # setup.ps1 line 1281-1296's mtime-based "is the frontend # stale?" check into "up to date, skip rebuild", because the # newly-created dist's mtime is younger than every source - # file. Studio then boots with an empty dist and 500s on + # file. Unsloth then boots with an empty dist and 500s on # GET / with FileNotFoundError: dist\index.html. See run # 25546676715 / job 74984469728. # Add-MpPreference accepts paths that do not yet exist; the @@ -129,7 +129,7 @@ jobs: } } - - name: Install Studio (--local, --no-torch) + - name: Install Unsloth (--local, --no-torch) shell: pwsh env: GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} @@ -168,7 +168,7 @@ jobs: echo "install.ps1 installed the Windows prebuilt llama.cpp:" cat "$INFO" - - name: Add Studio shim to GITHUB_PATH + - name: Add Unsloth shim to GITHUB_PATH run: | SHIM_DIR=~/.unsloth/studio/bin if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then @@ -212,7 +212,7 @@ jobs: grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log echo "second update was clean" - - name: Boot Studio briefly to confirm the install is still usable + - name: Boot Unsloth briefly to confirm the install is still usable run: | mkdir -p logs UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18891 \ @@ -239,13 +239,13 @@ jobs: sleep 1 done if [ -z "$HEALTHY" ]; then - echo "Studio failed to come up after \`update\`" + echo "Unsloth failed to come up after \`update\`" tail -200 logs/studio.log kill "$PID" 2>/dev/null || true exit 1 fi kill "$PID" 2>/dev/null || true - echo "post-update Studio /api/health OK" + echo "post-update Unsloth /api/health OK" - name: Uninstall and verify clean # Round-trip through scripts/uninstall.ps1 against the default diff --git a/.github/workflows/wheel-smoke.yml b/.github/workflows/wheel-smoke.yml index 3de3c33ca2..cdad617027 100644 --- a/.github/workflows/wheel-smoke.yml +++ b/.github/workflows/wheel-smoke.yml @@ -3,7 +3,7 @@ # Builds the PyPI wheel from the PR branch, then verifies the built wheel # actually contains what we expect to ship and does NOT contain the broken -# Studio bundle that 2026.5.1 published. This is the single workflow that +# Unsloth bundle that 2026.5.1 published. This is the single workflow that # would have blocked the 2026.5.1 release before twine upload. # # Verified locally end-to-end against this branch: @@ -12,7 +12,7 @@ # lockfile shipped, frontend dist shipped, # no node_modules in wheel, no bun.lock in wheel, # main bundle has unstable_Provider hits=1 (assistant-ui internals only). -# - Studio backend imports cleanly from the installed wheel with the +# - Unsloth backend imports cleanly from the installed wheel with the # lightweight dep set below. name: Wheel CI @@ -101,7 +101,7 @@ jobs: hits = data.count("unstable_Provider:") print(f"main bundle: {js[0]}") print(f"unstable_Provider hits: {hits} (>=4 indicates 2026.5.1 regression)") - checks["bundle has no Studio unstable_Provider call site"] = (hits < 4) + checks["bundle has no Unsloth unstable_Provider call site"] = (hits < 4) print() for k, v in checks.items(): @@ -109,7 +109,7 @@ jobs: sys.exit(0 if all(checks.values()) else 1) PY - - name: Studio backend import smoke + - name: Unsloth backend import smoke # Imports `studio.backend.main:app` from the freshly-installed wheel in # a clean venv. This catches the class of bug that 2026.5.1 shipped with: # frontend dist missing, package-lock.json missing, or the wheel's Python @@ -125,7 +125,7 @@ jobs: /tmp/v/bin/pip install --no-deps dist/unsloth-*.whl # Run from /tmp so Python imports the installed package, not the source tree. cd /tmp - /tmp/v/bin/python -c "from studio.backend.main import app; print('Studio backend OK:', app.title)" + /tmp/v/bin/python -c "from studio.backend.main import app; print('Unsloth backend OK:', app.title)" - name: Upload wheel on failure if: failure() diff --git a/README.md b/README.md index ef45b91430..085c7718e5 100644 --- a/README.md +++ b/README.md @@ -65,7 +65,7 @@ Unsloth Studio (Beta) works on **Windows, Linux, WSL** and **macOS**. * **CPU:** Supported for Chat and Data Recipes currently * **NVIDIA:** Training works on RTX 30/40/50, Blackwell, DGX Spark, Station and more * **macOS:** Training, MLX and GGUF inference are ALL supported. -* **AMD:** Chat + Data works. Train with [Unsloth Core](#unsloth-core-code-based). Studio support is out soon. +* **AMD:** Chat + Data works. Train with [Unsloth Core](#unsloth-core-code-based). Unsloth Studio support is out soon. * **Multi-GPU:** Available now, with a major upgrade on the way #### macOS, Linux, WSL: @@ -86,7 +86,7 @@ unsloth studio -p 8888 ``` For LAN or cloud access, add `-H 0.0.0.0` (raw port only; add `--cloudflare` for a public URL). By default, Unsloth is accessible only locally. -To reach Studio over HTTPS, use `unsloth studio --secure`. Studio stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Studio reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below). +To reach Unsloth over HTTPS, use `unsloth studio --secure`. Unsloth stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Unsloth reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below). #### Docker Use our [Docker image](https://hub.docker.com/r/unsloth/unsloth) ```unsloth/unsloth``` container. Run: @@ -208,7 +208,7 @@ unsloth studio -p 8888 #### Remote access: `--secure` (HTTPS tunnel) vs raw port By default `unsloth studio` binds to `127.0.0.1` (this machine only). To reach it from another device, pick one of: -- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Studio stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed. +- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Unsloth stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed. ```bash unsloth studio --secure -p 8888 ``` @@ -218,7 +218,7 @@ unsloth studio -H 0.0.0.0 -p 8888 ``` The Cloudflare tunnel is **off by default**: `-H 0.0.0.0` exposes the raw port only, not a public internet URL. Pair the wildcard bind with `--cloudflare` (`unsloth studio -H 0.0.0.0 --cloudflare`) to also publish a public `https://*.trycloudflare.com` link, or prefer `--secure` (above), which keeps the raw port private. `--cloudflare` has no effect on a loopback bind. -The first time Studio is published on a public URL (`--secure` or `--cloudflare`) with the auto-generated admin password still in place, it asks for a new admin password in the terminal (masked input with confirmation) before the public link goes up. Without an attached terminal it warns instead and keeps the bootstrap deadline: Studio shuts down after `UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT` (default 1 hour) unless the password is changed in the web UI. +The first time Unsloth is published on a public URL (`--secure` or `--cloudflare`) with the auto-generated admin password still in place, it asks for a new admin password in the terminal (masked input with confirmation) before the public link goes up. Without an attached terminal it warns instead and keeps the bootstrap deadline: Unsloth shuts down after `UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT` (default 1 hour) unless the password is changed in the web UI. For headless setups that cannot answer that prompt, set the initial admin password non-interactively with `--password` (only takes effect when no password is set yet; if one already exists it is a hard error, so rotate later with `unsloth studio reset-password`): @@ -230,7 +230,7 @@ printf '%s\n' 'your-strong-password' | unsloth studio --secure --password - # A literal `--password VALUE` is visible in the process list and shell history, so prefer the `UNSLOTH_STUDIO_PASSWORD` env var or `--password -` (stdin) for automation. This applies to any launch (public or a headless `-H 0.0.0.0` bind), and the password is set in the parent before the server binds, so it never reaches a re-executed child process. -Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Studio. +Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Unsloth. #### Advanced launch options Installer options can be passed as environment variables. On macOS, Linux and WSL place the variable after the pipe so the shell passes it to `sh`; on Windows set it with `$env:` before piping to `iex`. @@ -243,7 +243,7 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh $env:UNSLOTH_NO_TORCH=1; irm https://unsloth.ai/install.ps1 | iex ``` -Skip the post-install prompt that starts Studio (useful for automated installs): +Skip the post-install prompt that starts Unsloth (useful for automated installs): ```bash curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_SKIP_AUTOSTART=1 sh ``` @@ -279,9 +279,9 @@ UNSLOTH_NPM_REGISTRY=https://artifactory.example.com/api/npm/npm/ ./install.sh - ```powershell $env:UNSLOTH_NPM_REGISTRY='https://artifactory.example.com/api/npm/npm/'; .\install.ps1 --local ``` -It is threaded as `--registry` into the Studio frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force. +It is threaded as `--registry` into the Unsloth frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force. -Cap Studio's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`. +Cap Unsloth's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`. #### Uninstall The recommended way to fully remove Unsloth Studio is the matching uninstall script for your OS. It stops any running servers, removes the install dir, the launcher data dir, the desktop shortcut, and any platform-specific entries (macOS `.app` bundle + Launch Services on Mac; Start Menu, `HKCU\Software\Unsloth` registry key and user `PATH` entries on Windows): diff --git a/build.sh b/build.sh index dc272f0de1..2a836e19d9 100644 --- a/build.sh +++ b/build.sh @@ -4,9 +4,9 @@ set -euo pipefail -# PyPI/Studio release publishing must use `./build.sh publish` (or an -# equivalent stamp -> build -> verify-dist -> upload flow) so packaged Studio -# artifacts include the display-only Studio release version. +# PyPI/Unsloth release publishing must use `./build.sh publish` (or an +# equivalent stamp -> build -> verify-dist -> upload flow) so packaged Unsloth +# artifacts include the display-only Unsloth release version. # 1. Build frontend (Vite outputs to dist/) cd studio/frontend @@ -87,7 +87,7 @@ cd ../.. # 2. Clean old artifacts rm -rf build dist *.egg-info -# 3. Stamp display-only Studio release metadata for packaged builds. +# 3. Stamp display-only Unsloth release metadata for packaged builds. _STUDIO_BUILD_INFO="studio/backend/utils/_studio_release_build.py" _STUDIO_BUILD_INFO_BACKUP="$(mktemp)" cp "$_STUDIO_BUILD_INFO" "$_STUDIO_BUILD_INFO_BACKUP" diff --git a/install.ps1 b/install.ps1 index bbbaf3b416..3e674fe15b 100644 --- a/install.ps1 +++ b/install.ps1 @@ -176,7 +176,7 @@ function Install-UnslothStudio { $envOverride = $env:STUDIO_HOME.Trim() } - # Custom Studio roots are not supported with --tauri (desktop app still + # Custom Unsloth roots are not supported with --tauri (desktop app still # resolves %USERPROFILE%\.unsloth\studio). Pass through if override == legacy. if ($TauriMode -and $envOverride) { $_tauriOverride = $envOverride @@ -756,7 +756,7 @@ function Find-FreeLaunchPort { return `$null } -# If Studio is already healthy on any expected port, just open it and exit. +# If Unsloth is already healthy on any expected port, just open it and exit. `$existingPort = Find-HealthyStudioPort if (`$existingPort) { Start-Process "http://localhost:`$existingPort" @@ -772,7 +772,7 @@ try { `$haveMutex = `$true } if (-not `$haveMutex) { - # Another launcher is already running; wait for it to bring Studio up + # Another launcher is already running; wait for it to bring Unsloth up `$deadline = (Get-Date).AddSeconds(`$timeoutSec) while ((Get-Date) -lt `$deadline) { `$port = Find-HealthyStudioPort @@ -1438,7 +1438,7 @@ exit 0 if (Test-Path -LiteralPath $VenvPython) { # why: matching guard to the .venv branch below -- in env-mode # $StudioHome is a user-chosen workspace, so refuse to nuke an - # existing $StudioHome\unsloth_studio that lacks Studio sentinels. + # existing $StudioHome\unsloth_studio that lacks Unsloth sentinels. # -PathType Leaf rejects a directory at the sentinel path. Accept the # in-VENV ownership marker so partial-install retries are not blocked. if ( @@ -1449,7 +1449,7 @@ exit 0 ) { Write-Host "[ERROR] $VenvDir already exists but does not look like an Unsloth Studio install." -ForegroundColor Red Write-Host " Move it aside or choose an empty UNSLOTH_STUDIO_HOME." -ForegroundColor Yellow - throw "Refusing to delete non-Studio venv at $VenvDir" + throw "Refusing to delete non-Unsloth venv at $VenvDir" } # New layout already exists -- replace only after preserving rollback copy. substep "preserving existing environment for rollback..." @@ -1468,7 +1468,7 @@ exit 0 # workspace root (e.g. user's existing project Python venv). $OldVenv = Join-Path $StudioHome ".venv" $OldPy = Join-Path $OldVenv "Scripts\python.exe" - substep "found legacy Studio environment, validating..." + substep "found legacy Unsloth environment, validating..." $prevEAP2 = $ErrorActionPreference $ErrorActionPreference = "Continue" try { @@ -1498,7 +1498,7 @@ exit 0 # Skip in env-mode so we don't relocate the default-install venv into # the workspace root. $CwdVenv = Join-Path $env:USERPROFILE "unsloth_studio" - substep "found CWD-relative Studio environment, migrating to $VenvDir..." + substep "found CWD-relative Unsloth environment, migrating to $VenvDir..." Move-Item -LiteralPath $CwdVenv -Destination $VenvDir -Force substep "moved ~/unsloth_studio -> ~/.unsloth/studio/unsloth_studio" $_Migrated = $true @@ -1517,7 +1517,7 @@ exit 0 substep "$VenvDir" } - # Mark the freshly-created venv as Studio-owned so a partial install can be + # Mark the freshly-created venv as Unsloth-owned so a partial install can be # repaired by re-running install.ps1; the env-mode deletion guard above # accepts this marker as the primary sentinel. if (Test-Path -LiteralPath $VenvDir -PathType Container) { @@ -1526,7 +1526,7 @@ exit 0 # ── Helper: run amd-smi without triggering a UAC elevation prompt ── # amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing - # DiskPart UAC prompt mid-install (Studio backend amd.py hits the same). + # DiskPart UAC prompt mid-install (Unsloth backend amd.py hits the same). # __COMPAT_LAYER=RunAsInvoker forces it (and helpers it spawns) to run # un-elevated; on failure the WMI name -> gfx fallback still resolves the arch. function Invoke-AmdSmiNoElevate { @@ -1653,7 +1653,7 @@ exit 0 function Test-HipinfoIsVenvInternal { param([AllowNull()][string]$HipinfoPath) if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false } - # Also derive the venv from the setup python + default Studio home, so + # Also derive the venv from the setup python + default Unsloth home, so # the venv hipInfo is caught when VenvDir/VIRTUAL_ENV are unset. $venvRoots = @() if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV } @@ -1663,7 +1663,7 @@ exit 0 try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {} } if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") } - # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the + # A custom Unsloth home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the # venv off the default path; seed it too or its hipInfo escapes the filter. $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null } if ($studioHomeEnv) { @@ -1942,7 +1942,7 @@ exit 0 substep " Ensure the ROCm compute driver is installed alongside the display driver:" "Yellow" substep " https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow" } elseif ($ROCmGfxArch) { - # Known arch: Studio setup installs AMD's bundled-runtime ROCm PyTorch wheels + # Known arch: Unsloth setup installs AMD's bundled-runtime ROCm PyTorch wheels # (repo.amd.com), which ship their own runtime -- HIP SDK optional. step "gpu" "AMD ROCm ($ROCmGfxArch)" "Cyan" substep "Detected: $ROCmGpuLabel" "Cyan" @@ -2223,8 +2223,8 @@ exit 0 $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $torchSpec $visionSpec $audioSpec } if ($torchInstallExit -ne 0) { # Transient AMD-index failure: fall back to a CPU base so the install - # still completes; Studio setup retries ROCm afterwards. - substep "ROCm PyTorch install failed (exit $torchInstallExit); using a CPU base, Studio setup retries ROCm." "Yellow" + # still completes; Unsloth setup retries ROCm afterwards. + substep "ROCm PyTorch install failed (exit $torchInstallExit); using a CPU base, Unsloth setup retries ROCm." "Yellow" # --force-reinstall: a failed ROCm install can leave an unpinned ROCm # torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still satisfies the CPU # torch>= range, so without it uv would keep the ROCm build and only swap @@ -2426,7 +2426,7 @@ exit 0 Write-TauriLog "ERROR" "unsloth CLI was not installed correctly" Write-Host "[ERROR] unsloth CLI was not installed correctly." -ForegroundColor Red Write-Host " Expected: $UnslothExe" -ForegroundColor Yellow - Write-Host " This usually means an older unsloth version was installed that does not include the Studio CLI." -ForegroundColor Yellow + Write-Host " This usually means an older unsloth version was installed that does not include the Unsloth CLI." -ForegroundColor Yellow Write-Host " Try re-running the installer or see: https://github.com/unslothai/unsloth?tab=readme-ov-file#-quickstart" -ForegroundColor Yellow return (Exit-InstallFailure "unsloth CLI was not installed correctly") } @@ -2537,7 +2537,7 @@ exit 0 Write-Host " Move or remove it manually, then re-run the installer." -ForegroundColor Yellow throw "Cannot create unsloth launcher: $ShimExe is a directory." } - # try/catch: if unsloth.exe is locked (Studio running), keep the old shim. + # try/catch: if unsloth.exe is locked (Unsloth running), keep the old shim. $shimUpdated = $false try { if (Test-Path -LiteralPath $ShimExe) { Remove-Item -LiteralPath $ShimExe -Force -ErrorAction Stop } @@ -2555,7 +2555,7 @@ exit 0 if (Test-Path -LiteralPath $ShimExe) { Write-Host "[WARN] Could not refresh unsloth launcher at $ShimExe." -ForegroundColor Yellow Write-Host " This usually means a running 'unsloth studio' process still holds the file open." -ForegroundColor Yellow - Write-Host " Close Studio and re-run the installer to pick up the latest launcher." -ForegroundColor Yellow + Write-Host " Close Unsloth and re-run the installer to pick up the latest launcher." -ForegroundColor Yellow Write-Host " Continuing with the existing launcher." -ForegroundColor Yellow } else { Write-Host "[WARN] Could not create unsloth launcher at $ShimExe" -ForegroundColor Yellow @@ -2620,7 +2620,7 @@ exit 0 # Diagnostic only; never block install on a probe failure. } - # In interactive terminals, ask the user before starting Studio unless the + # In interactive terminals, ask the user before starting Unsloth unless the # caller explicitly disabled the post-install prompt. # In non-interactive environments (CI, Docker) just print instructions. $IsInteractive = (-not $SkipAutostart) -and [Environment]::UserInteractive -and (-not [Console]::IsInputRedirected) diff --git a/install.sh b/install.sh index 0b4732e86d..e119ec8049 100755 --- a/install.sh +++ b/install.sh @@ -97,7 +97,7 @@ if [ "$_VERBOSE" = true ]; then export UNSLOTH_VERBOSE=1 fi -# Custom Studio roots are not supported with --tauri (desktop app still +# Custom Unsloth roots are not supported with --tauri (desktop app still # resolves ~/.unsloth/studio). Pass through if the override == legacy default. if [ "$TAURI_MODE" = true ]; then _tauri_override_var="" @@ -472,11 +472,13 @@ _on_install_exit() { _restore_studio_venv_replacement fi [ -n "${_UV_OVERRIDE_TMPDIR:-}" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true + [ -n "${_UNSLOTH_TORCH_OVERRIDES:-}" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES" 2>/dev/null || true exit "$_status" } -# Empty so an inherited value can never reach the trap's rm; only a temp dir -# this script creates below (Apple Silicon, spaced path) is ever removed. +# Empty so an inherited value never reaches the trap's rm; only temp paths this +# script creates below (spaced-path dir, torch-trio overrides) are removed. _UV_OVERRIDE_TMPDIR="" +_UNSLOTH_TORCH_OVERRIDES="" trap _on_install_exit EXIT # ── Helper: download a URL to a file (supports curl and wget) ── @@ -663,7 +665,7 @@ POLL_INTERVAL_SEC=0.25 LOG_FILE="$DATA_DIR/studio.log" # why: in env-override mode multiple installs share an OS user; namespace the # lock and remember our own healthy port so we never attach to an unrelated -# Studio listening on the global 8888..8908 range. +# Unsloth listening on the global 8888..8908 range. LOCK_DIR="${XDG_RUNTIME_DIR:-/tmp}/unsloth-studio-launcher-$(id -u).lock" PORT_FILE="" # why: gate on the install-time mode (baked above) instead of the runtime env @@ -734,7 +736,7 @@ _candidate_ports() { _find_healthy_port() { if [ -n "$PORT_FILE" ] && [ -f "$PORT_FILE" ]; then # why: env-mode installs only attach to a port we previously launched - # ourselves; never to a sibling Studio that happens to be healthy. + # ourselves; never to a sibling Unsloth that happens to be healthy. _p=$(cat "$PORT_FILE" 2>/dev/null || true) case "$_p" in ''|*[!0-9]*) ;; @@ -901,7 +903,7 @@ _acquire_lock() { # Lock dir exists -- check if owner is still alive _old_pid=$(cat "$LOCK_DIR/pid" 2>/dev/null || true) if [ -n "$_old_pid" ] && kill -0 "$_old_pid" 2>/dev/null; then - # Another launcher is running; wait for it to bring Studio up + # Another launcher is running; wait for it to bring Unsloth up _deadline=$(($(date +%s) + TIMEOUT_SEC)) while [ "$(date +%s)" -lt "$_deadline" ]; do _port=$(_find_healthy_port) && { @@ -1371,7 +1373,7 @@ WSLPS1_EOF # shortcut wasn't created; tell the user how to launch / re-enable it. if [ "$_css_created" -ne 1 ]; then substep "Couldn't create the Windows shortcut (WSL interop may be disabled)." "$C_WARN" - substep " Launch Studio from Windows: wsl -d \"$_css_distro\" -- bash -lc 'unsloth studio'" "$C_WARN" + substep " Launch Unsloth from Windows: wsl -d \"$_css_distro\" -- bash -lc 'unsloth studio'" "$C_WARN" substep " (re-enable shortcuts: turn WSL interop back on, e.g. run 'wsl --shutdown' then reopen WSL.)" "$C_WARN" fi fi @@ -1439,7 +1441,7 @@ if [ "$MAC_INTEL" = true ]; then echo "" echo " NOTE: Intel Mac (x86_64) detected." echo " PyTorch is unavailable for this platform (dropped Jan 2024)." - echo " Studio will install in GGUF-only mode." + echo " Unsloth will install in GGUF-only mode." echo " Chat, inference via GGUF, and data recipes will work." echo " Training requires Apple Silicon or Linux with GPU." echo "" @@ -1671,7 +1673,7 @@ _maybe_reroute_strixhalo_to_2404() { _maybe_reroute_strixhalo_to_2404 || true # ── Check system dependencies ── -# cmake/git are only needed to *build* llama.cpp from source. Studio downloads a +# cmake/git are only needed to *build* llama.cpp from source. Unsloth downloads a # prebuilt by default, and setup.sh self-skips the source build when they're # absent -- so macOS doesn't block on cmake (requiring it would force a manual # Homebrew install). Linux keeps requiring them; its package manager has them. @@ -1821,11 +1823,13 @@ tauri_log "STEP" "Creating virtual environment" mkdir -p "$STUDIO_HOME" _MIGRATED=false +# Empty so an inherited value can never masquerade as a probed torch version. +_PREV_TORCH_VER="" if [ -x "$VENV_DIR/bin/python" ]; then # why: matching guard to the .venv branch below -- in env-mode # $STUDIO_HOME is a user-chosen workspace, so refuse to nuke an - # existing $STUDIO_HOME/unsloth_studio that lacks Studio sentinels. + # existing $STUDIO_HOME/unsloth_studio that lacks Unsloth sentinels. # Accept the in-VENV ownership marker so partial-install retries are # not blocked. Sentinels must be regular files: -f follows symlinks # to files (the legitimate ln -s shim shape) but rejects directories @@ -1838,6 +1842,12 @@ if [ -x "$VENV_DIR/bin/python" ]; then echo " Move it aside or choose an empty UNSLOTH_STUDIO_HOME." >&2 exit 1 fi + # Record the existing venv's torch BEFORE the replacement moves it aside: a re-run + # rebuilds the venv for clean state, but must keep the torch release the user + # already has (see _previous_torch_pin below). Last line only: sitecustomize or + # import-hook noise on stdout must not corrupt the version. + _PREV_TORCH_VER=$("$VENV_DIR/bin/python" -c \ + "import torch; print(torch.__version__)" 2>/dev/null | tail -n 1 || true) # New layout already exists — replace only after preserving rollback copy. substep "preserving existing environment for rollback..." _start_studio_venv_replacement "$VENV_DIR" @@ -1846,7 +1856,7 @@ elif [ "$_STUDIO_HOME_REDIRECT" != "env" ] && [ -x "$STUDIO_HOME/.venv/bin/pytho # Skip in env-mode so we don't rm -rf an unrelated .venv at the # workspace root (e.g. user's existing project Python venv). # In no-torch mode, a missing torch package is expected; validate Python only. - substep "found legacy Studio environment, validating..." + substep "found legacy Unsloth environment, validating..." _legacy_ok=false if [ "$SKIP_TORCH" = true ]; then if "$STUDIO_HOME/.venv/bin/python" -c "import sys; print(sys.executable)" >/dev/null 2>&1; then @@ -1903,7 +1913,7 @@ if [ ! -x "$VENV_DIR/bin/python" ]; then fi fi -# Mark the freshly-created venv as Studio-owned so a partial install can be +# Mark the freshly-created venv as Unsloth-owned so a partial install can be # repaired by re-running install.sh; the env-mode deletion guard above accepts # this marker as the primary sentinel. if [ -x "$VENV_DIR/bin/python" ]; then @@ -2202,6 +2212,68 @@ _torch_flavor_tag() { esac } +# Whether release base $1 (X.Y[.Z...]) falls inside constraint window $2 +# ("torch>=A.B[.C],="*",<"*) ;; + *) echo "no"; return ;; + esac + _trw_floor="${_trw_con#torch>=}"; _trw_floor="${_trw_floor%%,*}" + _trw_ceil="${_trw_con##*,<}" + _v_maj="${1%%.*}"; _v_rest="${1#*.}"; _v_min="${_v_rest%%.*}" + _f_maj="${_trw_floor%%.*}"; _f_rest="${_trw_floor#*.}"; _f_min="${_f_rest%%.*}" + _c_maj="${_trw_ceil%%.*}"; _c_rest="${_trw_ceil#*.}"; _c_min="${_c_rest%%.*}" + for _trw_n in "$_v_maj" "$_v_min" "$_f_maj" "$_f_min" "$_c_maj" "$_c_min"; do + case "$_trw_n" in ''|*[!0-9]*) echo "no"; return ;; esac + done + if [ "$_v_maj" -gt "$_f_maj" ] || { [ "$_v_maj" -eq "$_f_maj" ] && [ "$_v_min" -ge "$_f_min" ]; }; then + if [ "$_v_maj" -lt "$_c_maj" ] || { [ "$_v_maj" -eq "$_c_maj" ] && [ "$_v_min" -lt "$_c_min" ]; }; then + echo "yes" + return + fi + fi + echo "no" +} + +# Whether a re-run should keep the previous venv's torch: echo "torch==X.Y.Z" when the +# probed previous version ($1) has a flavor tag matching the freshly chosen cu*/cpu index +# leaf ($2) AND sits inside the active constraint window ($3), else "". Re-running +# `curl | sh` rebuilds the venv for clean state, but a healthy torch the user already +# validated must not be silently moved to a newer release (2.10 -> 2.11); a flavor +# change (cpu <-> cuda, cu126 -> cu130) still installs the correct new build, rocm +# leaves keep their floors (rocm7.2 must land 2.11 for the Strix _grouped_mm fix), and +# a release outside the window (2.3.x manual install, 2.12.x manual upgrade) is never +# kept: the installer's own bounds win. Opt out with UNSLOTH_TORCH_UPGRADE=1 to get +# the newest release. +_previous_torch_pin() { + _ptp_ver="$1" + _ptp_leaf="$2" + _ptp_con="$3" + [ -n "$_ptp_ver" ] || { echo ""; return; } + [ "${UNSLOTH_TORCH_UPGRADE:-0}" = "1" ] && { echo ""; return; } + case "$_ptp_leaf" in + cu[0-9]*|cpu) ;; + *) echo ""; return ;; + esac + _ptp_base="${_ptp_ver%%+*}" + # The base must look like a release (probe noise / garbage must never become a pin). + case "$_ptp_base" in + [0-9]*.[0-9]*) ;; + *) echo ""; return ;; + esac + [ "$(_torch_release_in_window "$_ptp_base" "$_ptp_con")" = "yes" ] || { echo ""; return; } + if [ "$(_torch_flavor_tag "$_ptp_ver")" = "$_ptp_leaf" ]; then + echo "torch==$_ptp_base" + else + echo "" + fi +} + # Expected tag from the index leaf ($1): cuXXX / cpu / rocm (rocmX.Y and gfx* -> # rocm). Empty on an unknown leaf (odd mirror) so the repair safely no-ops. _expected_torch_flavor_tag() { @@ -2350,7 +2422,7 @@ _pick_radeon_wheel() { # the installer -- always returns 0. Runs the idempotent helper (ROCm 7.2 + # librocdxg), then sources the env it persisted so detection finds the GPU. # Export the ROCm-on-WSL env into this process and persist it to /etc/profile.d -# so non-login Studio/llama launches inherit it. Idempotent (writes only when +# so non-login Unsloth/llama launches inherit it. Idempotent (writes only when # the drop-in is missing); no-op without librocdxg, so never fires off WSL. # /etc/profile.d is root-owned -- sudo-tee when not root, else ROCm vanishes # after this shell on a non-root reinstall. Best-effort either way. @@ -2395,7 +2467,7 @@ _maybe_bootstrap_rocm_wsl() { rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9]/ && !/generic/{found=1} END{exit !found}'; then # rocminfo may work only via the transient env _ensure_rocm_probe_env # just set, which dies with the installer. Persist the drop-in so login - # shells (Studio, llama.cpp) inherit it -- else a reinstall over an + # shells (Unsloth, llama.cpp) inherit it -- else a reinstall over an # existing /opt/rocm (uninstall keeps ROCm but drops it) loses the GPU. _persist_rocm_wsl_dropin return 0 @@ -2417,7 +2489,7 @@ _maybe_bootstrap_rocm_wsl() { # shellcheck disable=SC1091 . /etc/profile.d/unsloth-rocm-wsl.sh || true else - # librocdxg present but the env drop-in is gone (e.g. a Studio + # librocdxg present but the env drop-in is gone (e.g. an Unsloth # uninstall removed it while keeping shared ROCm). Restore the env. _persist_rocm_wsl_dropin fi @@ -2493,12 +2565,32 @@ case "$_torch_index_leaf" in *) export UNSLOTH_TORCH_BACKEND="cuda" ;; esac -# rocm7.2 ships torch 2.11.0 -- adjust the constraint to allow it. -# All other ROCm tags and CUDA stay within <2.11.0. -case "$TORCH_INDEX_URL" in - */rocm7.2) TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0" ;; +# rocm7.2 and the CUDA cu12x/cu13x indexes now ship torch 2.11.x, so widen the +# ceiling to <2.12.0 (matches the base image and _CUDA_TORCH_PKG_SPEC in +# studio/install_python_stack.py). Keep the >=2.4 floor so an older CUDA index +# (e.g. cu118) still resolves. Match on _torch_index_leaf, not the full URL, so +# a mirror whose base path contains cu*/rocm7.2 but resolves to a cpu/older-rocm +# leaf keeps the default <2.11.0. +case "$_torch_index_leaf" in + rocm7.2) TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0" ;; + cu[0-9]*) TORCH_CONSTRAINT="torch>=2.4,<2.12.0" ;; esac +# Re-run over an existing install: keep the previous venv's torch release instead of +# resolving the newest in range. The range stays in _PREV_FALLBACK_CONSTRAINT so the +# install can fall back when the exact release is not on the chosen index (custom +# mirrors may prune old wheels). Skipped for --no-torch (no previous probe runs). +_PREV_TORCH_PIN="" +_PREV_FALLBACK_CONSTRAINT="$TORCH_CONSTRAINT" +if [ "$SKIP_TORCH" = false ]; then + _prev_pin=$(_previous_torch_pin "$_PREV_TORCH_VER" "$_torch_index_leaf" "$TORCH_CONSTRAINT") + if [ -n "$_prev_pin" ]; then + _PREV_TORCH_PIN="$_prev_pin" + TORCH_CONSTRAINT="$_prev_pin" + substep "existing install has torch $_PREV_TORCH_VER -- keeping it (set UNSLOTH_TORCH_UPGRADE=1 to get the newest release)" + fi +fi + # Auto-detect GPU for AMD ROCm based # get_torch_index_url must have chosen */rocm* # (gfx in rocminfo or amd-smi list). Then require rocminfo "Marketing Name:.*Radeon". @@ -2720,6 +2812,43 @@ esac # ── Install unsloth directly into the venv (no activation needed) ── tauri_log "STEP" "Installing PyTorch" _VENV_PY="$VENV_DIR/bin/python" + +# A released unsloth wheel can pin an older torch (unsloth 2026.7.2 declares +# torch<2.11.0); a with-deps PyPI resolve then downgrades the whole trio, +# swapping the pinned +cuXXX/+rocm build for PyPI's default. The flavor guard +# below misses this (PyPI's torch 2.10 default is itself cu128-flavored), so +# freeze the trio via uv --overrides (overrides replace dependency requirements +# during resolution) while unsloth's other deps resolve normally. Sets +# _UNSLOTH_TORCH_OVERRIDES from the trio in the venv; every with-deps unsloth +# install (migrated and fresh) must call this before resolving and rm it after. +_build_unsloth_torch_overrides() { + _UNSLOTH_TORCH_OVERRIDES="" + [ "$SKIP_TORCH" = false ] || return 0 + _torch_trio_pins=$("$_VENV_PY" -c " +from importlib.metadata import version, PackageNotFoundError +for _p in ('torch', 'torchvision', 'torchaudio'): + try: + print(_p + '==' + version(_p)) + except PackageNotFoundError: + pass +" 2>/dev/null) || _torch_trio_pins="" + case "$_torch_trio_pins" in + torch==*) + _UNSLOTH_TORCH_OVERRIDES=$(mktemp) + printf '%s\n' "$_torch_trio_pins" > "$_UNSLOTH_TORCH_OVERRIDES" + # The CLI --overrides flag replaces any UV_OVERRIDE env file (same + # uv setting; macOS arm64 exports one here), so fold its pins in. + # awk, not cat: it drops inherited torch-trio lines (uv intersects + # duplicate overrides, so a conflicting pin would make resolution + # unsatisfiable) and newline-terminates the last line so an + # unterminated file cannot join two requirements into one. + for _ov_file in ${UV_OVERRIDE:-}; do + [ -f "$_ov_file" ] && awk '!/^[[:space:]]*torch(vision|audio)?([[:space:]<>=!~;@[]|$)/' "$_ov_file" >> "$_UNSLOTH_TORCH_OVERRIDES" + done + ;; + esac +} + if [ "$_MIGRATED" = true ]; then # Migrated env: force-reinstall unsloth+unsloth-zoo to ensure clean state # in the new venv location, while preserving existing torch/CUDA @@ -2744,9 +2873,13 @@ if [ "$_MIGRATED" = true ]; then else # Pin mlx-lm away from 0.31.3 here too: a curl-piped migration has no # overrides file, so UV_OVERRIDE is unset and this positional is the only cover. + _build_unsloth_torch_overrides run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \ + ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \ --reinstall-package unsloth --reinstall-package unsloth-zoo \ "unsloth>=2026.7.3" "unsloth-zoo>=2026.7.3" ${_MLX_LM_EXCLUDE_ARG:-} + [ -n "$_UNSLOTH_TORCH_OVERRIDES" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES" + _UNSLOTH_TORCH_OVERRIDES="" fi if [ "$STUDIO_LOCAL_INSTALL" = true ]; then substep "overlaying local repo (editable)..." @@ -2928,8 +3061,20 @@ elif [ -n "$TORCH_INDEX_URL" ]; then fi else substep "installing PyTorch ($TORCH_INDEX_URL)..." - run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \ - --default-index "$TORCH_INDEX_URL" + if [ -n "$_PREV_TORCH_PIN" ]; then + # Kept previous release: fall back to the supported range if the exact + # release is not resolvable from the chosen index (pruned mirror). + if ! run_install_cmd_retry "install PyTorch (kept release)" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \ + --default-index "$TORCH_INDEX_URL"; then + substep "[WARN] $_PREV_TORCH_PIN is not installable from $TORCH_INDEX_URL -- installing the newest supported release instead" "$C_WARN" + TORCH_CONSTRAINT="$_PREV_FALLBACK_CONSTRAINT" + run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \ + --default-index "$TORCH_INDEX_URL" + fi + else + run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \ + --default-index "$TORCH_INDEX_URL" + fi fi # AMD ROCm: install bitsandbytes (once, after torch, for all ROCm paths). # Gate on SKIP_TORCH=false so a user running with --no-torch on a ROCm @@ -2942,9 +3087,10 @@ elif [ -n "$TORCH_INDEX_URL" ]; then ;; esac fi - # Fresh: Step 2 - install unsloth, preserving pre-installed torch + # Fresh: Step 2 - install unsloth, preserving the torch Step 1 installed tauri_log "STEP" "Installing Unsloth" substep "installing unsloth (this may take a few minutes)..." + _build_unsloth_torch_overrides if [ "$SKIP_TORCH" = true ]; then # No-torch: install unsloth + unsloth-zoo with --no-deps, then # runtime deps (typer, safetensors, transformers, etc.) with --no-deps. @@ -2968,6 +3114,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then fi elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \ + ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \ --upgrade-package unsloth "unsloth>=2026.7.3" "unsloth-zoo>=2026.7.3" substep "overlaying local repo (editable)..." run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps @@ -2977,8 +3124,11 @@ elif [ -n "$TORCH_INDEX_URL" ]; then "$_ZOO_GIT_SPEC" else run_install_cmd_retry "install unsloth" uv pip install --python "$_VENV_PY" \ + ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \ --upgrade-package unsloth -- "$PACKAGE_NAME" ${_MLX_LM_EXCLUDE_ARG:-} fi + [ -n "$_UNSLOTH_TORCH_OVERRIDES" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES" + _UNSLOTH_TORCH_OVERRIDES="" # AMD ROCm: repair torch if the unsloth/unsloth-zoo install pulled in # CUDA torch from PyPI, overwriting the ROCm wheels installed in Step 1. if [ "$SKIP_TORCH" = false ]; then @@ -3048,7 +3198,7 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then fi # ── Run studio setup ── -tauri_log "STEP" "Running Studio setup" +tauri_log "STEP" "Running Unsloth setup" # When --local, use the repo's own setup.sh directly. # Otherwise, find it inside the installed package. SETUP_SH="" @@ -3242,7 +3392,7 @@ printf " ${C_TITLE}%s${C_RST}\n" "Unsloth Studio installed!" printf " ${C_DIM}%s${C_RST}\n" "$RULE" echo "" -# In interactive terminals, ask the user before starting Studio unless the +# In interactive terminals, ask the user before starting Unsloth unless the # caller explicitly disabled the post-install prompt. # In non-interactive environments (Docker, CI, cloud-init) just print instructions. if [ "$_SKIP_AUTOSTART" != true ] && [ -t 1 ]; then diff --git a/scripts/install_rocm_wsl_strixhalo.sh b/scripts/install_rocm_wsl_strixhalo.sh index aa560fc432..697aae933f 100644 --- a/scripts/install_rocm_wsl_strixhalo.sh +++ b/scripts/install_rocm_wsl_strixhalo.sh @@ -219,7 +219,7 @@ fi echo "${ROCM_DIR}/lib" | $SUDO tee /etc/ld.so.conf.d/rocm.conf >/dev/null $SUDO ldconfig -# ── Step 4: persist environment (system-wide so Studio's worker inherits it) ── +# ── Step 4: persist environment (system-wide so Unsloth's worker inherits it) ── say "Persisting ROCm-on-WSL environment" _envfile="/etc/profile.d/unsloth-rocm-wsl.sh" $SUDO tee "$_envfile" >/dev/null < str: return f'''# SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -"""Build-stamped Studio release metadata.""" +"""Build-stamped Unsloth release metadata.""" STUDIO_RELEASE_VERSION = {literal} ''' @@ -168,7 +168,7 @@ def stamp(require_release: bool) -> int: version, source = resolve_version() if version is not None and not is_valid_version(version): print( - f"Invalid Studio release version from {source}: {version!r}", + f"Invalid Unsloth release version from {source}: {version!r}", file = sys.stderr, ) return 2 @@ -196,9 +196,9 @@ def stamp(require_release: bool) -> int: if version is None: if require_release: print( - "No Studio release version available. Set " + "No Unsloth release version available. Set " "UNSLOTH_STUDIO_RELEASE_VERSION, build from a GitHub tag, " - "or run from an exact local Studio release tag.", + "or run from an exact local Unsloth release tag.", file = sys.stderr, ) return 2 @@ -207,7 +207,7 @@ def stamp(require_release: bool) -> int: return 0 _atomic_write_text(BUILD_INFO_PATH, build_info_source(version), encoding = "utf-8") - print(f"Stamping Studio release version {version} from {source}", file = sys.stderr) + print(f"Stamping Unsloth release version {version} from {source}", file = sys.stderr) print(version) return 0 @@ -233,7 +233,7 @@ def _read_sdist_member(path: Path) -> str | None: def verify_dist(expected: str, dist_dir: Path) -> int: if not is_valid_version(expected): - print(f"Invalid expected Studio release version: {expected!r}", file = sys.stderr) + print(f"Invalid expected Unsloth release version: {expected!r}", file = sys.stderr) return 2 artifacts = list(dist_dir.glob("*.whl")) + list(dist_dir.glob("*.tar.gz")) @@ -251,14 +251,14 @@ def verify_dist(expected: str, dist_dir: Path) -> int: if content is None: failures.append(f"{artifact.name}: missing {BUILD_INFO_SUFFIX}") elif expected_line not in content: - failures.append(f"{artifact.name}: Studio release version mismatch") + failures.append(f"{artifact.name}: Unsloth release version mismatch") if failures: for failure in failures: print(failure, file = sys.stderr) return 2 - print(f"Verified Studio release version {expected} in {len(artifacts)} artifact(s)") + print(f"Verified Unsloth release version {expected} in {len(artifacts)} artifact(s)") return 0 diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1 index 88defb9ea0..9b6e6ebb86 100644 --- a/scripts/uninstall.ps1 +++ b/scripts/uninstall.ps1 @@ -83,7 +83,7 @@ function Uninstall-UnslothStudio { } } - # A path is a Studio-owned root iff one of install.ps1's sentinels exists: + # A path is an Unsloth-owned root iff one of install.ps1's sentinels exists: # \share\studio.conf, \unsloth_studio\.unsloth-studio-owned, # or \bin\unsloth.exe. function _IsStudioRoot { @@ -164,7 +164,7 @@ function Uninstall-UnslothStudio { return $p } - # Discover non-default Studio roots from env vars + studio.conf files. + # Discover non-default Unsloth roots from env vars + studio.conf files. # Mirrors install.ps1's precedence: UNSLOTH_STUDIO_HOME wins, STUDIO_HOME # is ignored when both are set, so uninstalling install A doesn't also # delete install B if the user has a stale STUDIO_HOME pointing at B. @@ -207,7 +207,7 @@ function Uninstall-UnslothStudio { # Return $true iff the PID's image path lives under one of $KnownRoots. # Prevents killing an unrelated process that happens to listen on a stale - # Studio port. + # Unsloth port. function _PidUnderKnownRoot { param([int]$Pid_, [string[]]$KnownRoots) if (-not $KnownRoots -or $KnownRoots.Count -eq 0) { return $false } @@ -223,8 +223,8 @@ function Uninstall-UnslothStudio { return $false } - # Stop a Studio backend whose port is recorded in \studio.port. - # Only kills if the listening PID's exe path is under a known Studio root. + # Stop an Unsloth backend whose port is recorded in \studio.port. + # Only kills if the listening PID's exe path is under a known Unsloth root. function _StopByPortFile { param([string]$PortFile, [string[]]$KnownRoots) if (-not (Test-Path -LiteralPath $PortFile -PathType Leaf)) { return } @@ -372,7 +372,7 @@ function Uninstall-UnslothStudio { continue } if (-not (_IsStudioRoot $r)) { - _Substep "refusing to remove non-Studio path: $r" "Yellow" + _Substep "refusing to remove non-Unsloth path: $r" "Yellow" continue } _RemovePath $r @@ -436,7 +436,7 @@ function Uninstall-UnslothStudio { $entries = $rawPath -split ';' $kept = New-Object System.Collections.ArrayList $removedAny = $false - # Only remove PATH entries that live inside a Studio root we + # Only remove PATH entries that live inside an Unsloth root we # actually own (default or env-mode). A literal substring # match on `unsloth_studio` would clobber unrelated user # virtualenvs that happen to share the name. diff --git a/scripts/uninstall.sh b/scripts/uninstall.sh index 31e851fcbb..957d2b7af2 100755 --- a/scripts/uninstall.sh +++ b/scripts/uninstall.sh @@ -12,7 +12,7 @@ set -e -# Stop a Studio server via its PID file (written by install.sh's _spawn_terminal). +# Stop an Unsloth server via its PID file (written by install.sh's _spawn_terminal). _kill_pid_file() { _pid_file="$1" [ -f "$_pid_file" ] || return 0 @@ -47,7 +47,7 @@ _pkill_studio() { command -v pkill >/dev/null 2>&1 || return 0 # Scope fallback patterns to the install roots we are removing so a - # different Studio install (different UNSLOTH_STUDIO_HOME) is not touched. + # different Unsloth install (different UNSLOTH_STUDIO_HOME) is not touched. _kill_roots="$HOME/.unsloth/studio" _roots_from_conf=$(_custom_studio_roots 2>/dev/null || true) [ -n "$_roots_from_conf" ] && _kill_roots="$_kill_roots @@ -89,7 +89,7 @@ _remove_path() { fi } -# Accept as Studio root only if Studio sentinels exist (matches install.sh's +# Accept as Unsloth root only if Unsloth sentinels exist (matches install.sh's # env-mode ownership guard at install.sh:1358-1361). A bare unsloth_studio/ # directory is NOT enough -- require the install-time owner marker so a user # directory that happens to contain a folder named "unsloth_studio" is safe. @@ -175,8 +175,8 @@ _custom_studio_roots() { _from_conf "$HOME/.local/share/unsloth/studio.conf" } -# Remove $HOME/.local/bin/unsloth only if it's a Studio-managed symlink. -# Studio's install.sh writes this as a symlink into the studio venv +# Remove $HOME/.local/bin/unsloth only if it's an Unsloth-managed symlink. +# Unsloth's install.sh writes this as a symlink into the studio venv # (install.sh: `ln -sfn "$VENV_DIR/bin/unsloth" "$_shim_path"`). A # pip-installed `unsloth` CLI is a regular file — leave it alone to avoid # wiping an unrelated install. @@ -206,7 +206,7 @@ _custom_studio_roots | while IFS= read -r _custom_root; do continue fi if ! _is_studio_root "$_custom_root"; then - echo " refusing to remove non-Studio path: $_custom_root" >&2 + echo " refusing to remove non-Unsloth path: $_custom_root" >&2 continue fi _remove_path "$_custom_root" @@ -234,7 +234,7 @@ _remove_path "$HOME/.unsloth/rocm-smoketest" # Drop ~/.unsloth only if now empty (rmdir refuses non-empty, so user content is kept). rmdir "$HOME/.unsloth" 2>/dev/null || true _remove_path "$HOME/.local/share/unsloth" -# CLI shim: only the symlink Studio created, never a pip-installed file. +# CLI shim: only the symlink Unsloth created, never a pip-installed file. _remove_cli_shim echo "Removing desktop shortcut and launcher lock..." diff --git a/studio/MCP.md b/studio/MCP.md index 91b39fcc69..127a85a116 100644 --- a/studio/MCP.md +++ b/studio/MCP.md @@ -1,10 +1,10 @@ # Unsloth Studio MCP server -Studio can expose a local MCP server so an MCP client can inspect models and +Unsloth can expose a local MCP server so an MCP client can inspect models and GPU state, validate recipes, start or stop training, inspect recipe output, and export a loaded model. -The server is disabled by default. Enable it for a local Studio process with: +The server is disabled by default. Enable it for a local Unsloth process with: ```bash UNSLOTH_STUDIO_ENABLE_MCP=1 \ @@ -12,8 +12,8 @@ UNSLOTH_STUDIO_MCP_TOKEN='use-a-local-secret' \ unsloth studio ``` -The endpoint is `http://127.0.0.1:8888/mcp/` when Studio uses its default port -(a request to `/mcp` redirects to the canonical `/mcp/`). Use the actual Studio +The endpoint is `http://127.0.0.1:8888/mcp/` when Unsloth uses its default port +(a request to `/mcp` redirects to the canonical `/mcp/`). Use the actual Unsloth port when it is configured differently. The high-impact tools are: @@ -23,9 +23,9 @@ The high-impact tools are: - `validate_recipe`, `get_recipe_job_status`, and `get_recipe_job_dataset` - `load_checkpoint` and `export_gguf` -`start_training` accepts the same fields as the Studio `TrainingStartRequest`. +`start_training` accepts the same fields as the Unsloth `TrainingStartRequest`. The request is validated by the existing Pydantic model before a subprocess is -started. Export paths use the existing Studio validation as well. +started. Export paths use the existing Unsloth validation as well. The endpoint always requires `UNSLOTH_STUDIO_MCP_TOKEN` and checks an exact Bearer token for both HTTP and WebSocket connections. Keep it on localhost diff --git a/studio/Unsloth_Studio_Colab.ipynb b/studio/Unsloth_Studio_Colab.ipynb index 619395bd6d..44282b2255 100644 --- a/studio/Unsloth_Studio_Colab.ipynb +++ b/studio/Unsloth_Studio_Colab.ipynb @@ -33,7 +33,7 @@ "\n", "We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n", "\n", - "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Studio Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)" + "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Unsloth Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)" ] }, { diff --git a/studio/backend/assets/chat_templates/gemma-4-edge.jinja b/studio/backend/assets/chat_templates/gemma-4-edge.jinja index 0266127233..74fa73ddd3 100644 --- a/studio/backend/assets/chat_templates/gemma-4-edge.jinja +++ b/studio/backend/assets/chat_templates/gemma-4-edge.jinja @@ -3,7 +3,7 @@ Source: google/gemma-4-31B-it HF discussion/PR #118 (adds the preserve_thinking flag plus null-rendering, string-arguments validation, balanced turn tags, empty messages handling, and OpenAI image_url/input_audio aliases). - Studio-local changes vs PR #118: + Unsloth-local changes vs PR #118: 1. preserve_thinking defaults to false (see SETUP block below). 2. The empty "<|channel>thought\n" block on enable_thinking=false is NOT emitted. Google ships a distinct template for E2B/E4B (google/gemma-4-E2B-it, diff --git a/studio/backend/assets/chat_templates/gemma-4.jinja b/studio/backend/assets/chat_templates/gemma-4.jinja index 65ab39df57..cc5f98065f 100644 --- a/studio/backend/assets/chat_templates/gemma-4.jinja +++ b/studio/backend/assets/chat_templates/gemma-4.jinja @@ -3,7 +3,7 @@ Source: google/gemma-4-31B-it HF discussion/PR #118 (adds the preserve_thinking flag plus null-rendering, string-arguments validation, balanced turn tags, empty messages handling, and OpenAI image_url/input_audio aliases). - Studio-local change: preserve_thinking defaults to false (see SETUP block below). + Unsloth-local change: preserve_thinking defaults to false (see SETUP block below). Applied to unsloth/gemma-4-*-GGUF models so the embedded GGUF template does not need re-downloading. Keep in sync with upstream if PR #118 changes. -#} diff --git a/studio/backend/auth/authentication.py b/studio/backend/auth/authentication.py index b13cd1c851..dfb8fc513e 100644 --- a/studio/backend/auth/authentication.py +++ b/studio/backend/auth/authentication.py @@ -148,7 +148,7 @@ async def authenticated_via_api_key( ) -> bool: """True when the caller used an sk-unsloth API key, not a UI session JWT. - Lets routes treat programmatic API callers differently from the Studio UI + Lets routes treat programmatic API callers differently from the Unsloth UI (e.g. refuse a teardown the UI would allow). """ return bool(credentials and credentials.credentials.startswith(API_KEY_PREFIX)) diff --git a/studio/backend/auth/bootstrap_timeout.py b/studio/backend/auth/bootstrap_timeout.py index 728433dc54..97a8086f04 100644 --- a/studio/backend/auth/bootstrap_timeout.py +++ b/studio/backend/auth/bootstrap_timeout.py @@ -1,13 +1,13 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -"""Auto-shutdown for an exposed first-run Studio whose admin password is unchanged. +"""Auto-shutdown for an exposed first-run Unsloth whose admin password is unchanged. On a fresh install the seeded bootstrap admin password stays a valid login credential until first login changes it. When the web UI is put on the network (``--secure`` / ``0.0.0.0``) and nobody completes that first-login change within -a deadline, tear Studio down so a fresh, unconfigured instance does not stay -publicly reachable indefinitely. If the password was changed, Studio keeps +a deadline, tear Unsloth down so a fresh, unconfigured instance does not stay +publicly reachable indefinitely. If the password was changed, Unsloth keeps running. Scope: web UI launches only (never ``--api-only``, which authenticates by API @@ -98,7 +98,7 @@ def enforce_bootstrap_password_deadline( ) -> bool: """Deadline handler: shut down iff the seeded admin password is still unchanged. - Returns True if it shut Studio down, False if it left it running (the + Returns True if it shut Unsloth down, False if it left it running (the password was changed in time). """ try: @@ -106,7 +106,7 @@ def enforce_bootstrap_password_deadline( except Exception: return False if not still_default: - return False # password changed in time -> leave Studio running + return False # password changed in time -> leave Unsloth running message = ( "\nUnsloth Studio was exposed on the network but its default admin " diff --git a/studio/backend/auth/storage.py b/studio/backend/auth/storage.py index 9bb3ab5735..39fa691304 100644 --- a/studio/backend/auth/storage.py +++ b/studio/backend/auth/storage.py @@ -146,7 +146,7 @@ def get_connection() -> sqlite3.Connection: pass conn.row_factory = sqlite3.Row # WAL lets token reads run concurrently with refresh-token writes; - # busy_timeout bounds lock waits. Matches the other Studio SQLite stores. + # busy_timeout bounds lock waits. Matches the other Unsloth SQLite stores. # Set busy_timeout first: switching journal_mode needs a lock, so if a # refresh-token write already holds one, journal_mode=WAL raises SQLITE_BUSY; # with busy_timeout already in effect it waits instead of failing and leaving @@ -305,8 +305,8 @@ def get_or_create_identity_secret() -> bytes: def compute_identity_proof(nonce: bytes, host: str, port: int) -> str: """HMAC-SHA256 proof that the caller holds this install's identity secret, bound to the loopback address and port the connection landed on. A proof - relayed from a Studio on a different address/port (a squatter proxying to the - real one, e.g. localhost resolving to ::1 while Studio is on 127.0.0.1) was + relayed from an Unsloth on a different address/port (a squatter proxying to the + real one, e.g. localhost resolving to ::1 while Unsloth is on 127.0.0.1) was computed for that other endpoint and won't match the one the client dialed.""" try: host = ipaddress.ip_address(host).compressed # normalise 127.0.0.1 / ::1 forms diff --git a/studio/backend/auth/terminal_prompt.py b/studio/backend/auth/terminal_prompt.py index 8491019ae9..e855f4078b 100644 --- a/studio/backend/auth/terminal_prompt.py +++ b/studio/backend/auth/terminal_prompt.py @@ -2,14 +2,14 @@ # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 """Interactive terminal prompt that forces a bootstrap password change before -Studio is exposed on a public Cloudflare URL (``--secure`` / ``--cloudflare``). +Unsloth is exposed on a public Cloudflare URL (``--secure`` / ``--cloudflare``). Masked input echoes one ``*`` per keystroke (unlike ``getpass``). Works on Windows (``msvcrt``) and Linux/macOS (``termios``). All output goes to stderr so redirected stdout never swallows the prompt. Mirrored for the CLI at ``unsloth_cli/commands/_password_prompt.py`` (the CLI -cannot import the Studio backend package); keep the two in sync. +cannot import the Unsloth backend package); keep the two in sync. """ from __future__ import annotations @@ -252,7 +252,7 @@ def prompt_for_password_change( out.flush() return True except (KeyboardInterrupt, EOFError): - out.write("Password change aborted; not exposing Studio.\n") + out.write("Password change aborted; not exposing Unsloth.\n") out.flush() return False diff --git a/studio/backend/cloudflare_tunnel.py b/studio/backend/cloudflare_tunnel.py index ef7bacba67..b1ddc74c32 100644 --- a/studio/backend/cloudflare_tunnel.py +++ b/studio/backend/cloudflare_tunnel.py @@ -1,13 +1,13 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -"""Free Cloudflare quick tunnel for Studio's 0.0.0.0 launches. +"""Free Cloudflare quick tunnel for Unsloth's 0.0.0.0 launches. The raw http://: is often unreachable (https-vs-http, blocked ports, closed security groups); a cloudflared quick tunnel gives a free https://*.trycloudflare.com URL that works anywhere, with no account or domain. -Best-effort throughout: any failure collapses to "no URL" and Studio keeps +Best-effort throughout: any failure collapses to "no URL" and Unsloth keeps running. Stdlib only (back-end imports are lazy) so it is safe to import early. """ @@ -95,7 +95,7 @@ def _cache_path() -> Optional[Path]: def find_cloudflared() -> Optional[str]: - """Locate an existing cloudflared: PATH first, then the Studio bin cache.""" + """Locate an existing cloudflared: PATH first, then the Unsloth bin cache.""" on_path = shutil.which("cloudflared") if on_path: return on_path @@ -309,7 +309,7 @@ class CloudflareTunnel: pass -# Single serving process per Studio launch, so one module-level tunnel handle is +# Single serving process per Unsloth launch, so one module-level tunnel handle is # enough; the lock guards the start/stop/shutdown races. _active_tunnel: Optional[CloudflareTunnel] = None _active_lock = threading.Lock() diff --git a/studio/backend/colab.py b/studio/backend/colab.py index e04543b3aa..1762469bcf 100644 --- a/studio/backend/colab.py +++ b/studio/backend/colab.py @@ -129,7 +129,7 @@ def start_cloudflare_tunnel(port: int) -> "str | None": logger.warning( "Cloudflare link not started: the admin account still has its temporary " "bootstrap password, which is exposed to anyone who can load the page. " - "Open Studio in this tab, log in and change the admin password, then re-run " + "Open Unsloth in this tab, log in and change the admin password, then re-run " "start(cloudflare=True) to get the shareable link." ) return None @@ -203,7 +203,7 @@ def _shareable_link_html(cloudflare_url: str) -> str: display: flex; align-items: center; gap: 12px;"> - Shareable Studio Link is Ready! + Shareable Unsloth Link is Ready! ParsedUpdate | None: source = "github", status = "rate_limited", retry_after_sec = seconds, - message = ("Waiting for GitHub rate limit. Studio will resume automatically."), + message = ("Waiting for GitHub rate limit. Unsloth will resume automatically."), ), ) @@ -147,7 +147,7 @@ def parse_log_message(msg: str) -> ParsedUpdate | None: status = "rate_limited", retry_after_sec = seconds, message = ( - "Waiting for GitHub secondary rate limit. Studio will resume automatically." + "Waiting for GitHub secondary rate limit. Unsloth will resume automatically." ), ), ) @@ -161,7 +161,7 @@ def parse_log_message(msg: str) -> ParsedUpdate | None: source = "github", status = "rate_limited", retry_after_sec = seconds, - message = ("Waiting for GitHub rate limit. Studio will resume automatically."), + message = ("Waiting for GitHub rate limit. Unsloth will resume automatically."), ), ) diff --git a/studio/backend/core/data_recipe/local_callable_validators.py b/studio/backend/core/data_recipe/local_callable_validators.py index ebb1d39dfb..ffc81669ae 100644 --- a/studio/backend/core/data_recipe/local_callable_validators.py +++ b/studio/backend/core/data_recipe/local_callable_validators.py @@ -238,7 +238,7 @@ def _run_oxc_batch( if not node_executable: return _fallback_results( len(code_values), - "Node.js not found (install Node >= 20.19, or re-run Studio setup to provision it).", + "Node.js not found (install Node >= 20.19, or re-run Unsloth setup to provision it).", ) try: tmp_dir = ensure_dir(oxc_validator_tmp_root()) diff --git a/studio/backend/core/data_recipe/service.py b/studio/backend/core/data_recipe/service.py index 4647dc098d..9770e88b7f 100644 --- a/studio/backend/core/data_recipe/service.py +++ b/studio/backend/core/data_recipe/service.py @@ -280,8 +280,8 @@ def create_data_designer(recipe: dict[str, Any], *, artifact_path: str | None = from data_designer.interface.data_designer import DataDesigner # pyright: ignore[reportMissingImports] if artifact_path is None: - # DataDesigner defaults to cwd/artifacts; packaged Studio can run with - # cwd=/, so keep default callers on Studio's writable recipe artifact root. + # DataDesigner defaults to cwd/artifacts; packaged Unsloth can run with + # cwd=/, so keep default callers on Unsloth's writable recipe artifact root. artifact_path = str(recipe_datasets_root()) recipe = _strip_frontend_model_config_metadata(recipe) diff --git a/studio/backend/core/inference/__init__.py b/studio/backend/core/inference/__init__.py index ad78157418..1491dfa749 100644 --- a/studio/backend/core/inference/__init__.py +++ b/studio/backend/core/inference/__init__.py @@ -11,7 +11,7 @@ subprocess and can be imported directly from .inference when needed. Public names are resolved lazily (PEP 562): importing this package -- or a dependency-light leaf like ``core.inference.chat_eos`` -- must NOT eagerly pull the orchestrator / llama_cpp import chain (httpx, subprocess plumbing, the ML -backend and its Studio dependencies). Those load only when a public name is +backend and its Unsloth dependencies). Those load only when a public name is actually accessed, so standalone helpers stay unit-testable without the full inference stack. """ diff --git a/studio/backend/core/inference/anthropic_compat.py b/studio/backend/core/inference/anthropic_compat.py index 3c7a4cb182..34445cc58e 100644 --- a/studio/backend/core/inference/anthropic_compat.py +++ b/studio/backend/core/inference/anthropic_compat.py @@ -539,7 +539,7 @@ class AnthropicPassthroughEmitter: Only calls naming a tool in ``allowed_tools`` (the client's declared tools) are promoted; everything else streams as text exactly as before. - Never enabled for Studio's own tool loop. + Never enabled for Unsloth's own tool loop. """ from core.inference.passthrough_healing import StreamToolCallHealer diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py index 5113eebb36..528c059fbc 100644 --- a/studio/backend/core/inference/chat_template_helpers.py +++ b/studio/backend/core/inference/chat_template_helpers.py @@ -150,7 +150,7 @@ def _split_partial_marker(text: str, marker: str) -> tuple[str, str]: class ReasoningChannelNormalizer: """Incrementally convert one native reasoning channel to ````. - The parser follows mlx-vlm's streaming boundary behavior but emits Studio's + The parser follows mlx-vlm's streaming boundary behavior but emits Unsloth's established canonical text contract. Only the configured opening and closing markers are consumed; tool-call and other control markers remain available to downstream parsers. diff --git a/studio/backend/core/inference/chat_templates.py b/studio/backend/core/inference/chat_templates.py index 58f63ff61b..04c0db6aae 100644 --- a/studio/backend/core/inference/chat_templates.py +++ b/studio/backend/core/inference/chat_templates.py @@ -4,13 +4,13 @@ """Bundled chat-template selection for GGUF inference. Some shipped GGUF quants embed an older chat template. Rather than re-cutting and -asking users to re-download every quant, Studio can override the embedded template +asking users to re-download every quant, Unsloth can override the embedded template at llama-server launch time with a bundled, up-to-date Jinja template for known model families. The override is wired through the existing ``chat_template_override`` -> ``--chat-template-file`` path in ``LlamaCppBackend.load_model``. Currently this covers ``unsloth/gemma-4-*-GGUF``, which gains the upstream PR #118 -``preserve_thinking`` flag (defaulted OFF here) so the Studio "Preserve thinking" +``preserve_thinking`` flag (defaulted OFF here) so the Unsloth "Preserve thinking" toggle appears while staying disabled by default. """ diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py index 20312e067c..2debf946e9 100644 --- a/studio/backend/core/inference/external_provider.py +++ b/studio/backend/core/inference/external_provider.py @@ -473,7 +473,7 @@ def _apply_mistral_reasoning_controls( # handles every provider without storing credentials. def _create_shared_http_client() -> httpx.AsyncClient: # Unsupported env proxy schemes (socks:// etc) raise at construction and - # would crash Studio startup (#6090); retry ignoring env proxies instead. + # would crash Unsloth startup (#6090); retry ignoring env proxies instead. try: return httpx.AsyncClient() except (ImportError, ValueError) as exc: @@ -858,7 +858,7 @@ class ExternalProviderClient: if not self._is_openai_compatible(): # Gemini speaks its own native REST shape (contents/parts); # `_stream_gemini` translates request/response into the OpenAI - # Chat Completions chunk format the rest of Studio expects. + # Chat Completions chunk format the rest of Unsloth expects. # API ref: https://ai.google.dev/gemini-api/docs if self.provider_type == "gemini": async for line in self._stream_gemini( @@ -1706,7 +1706,7 @@ class ExternalProviderClient: # Translate OpenAI multimodal parts -> Anthropic native shapes. # - `image_url` -> `{type:"image", source:...}` # - `input_document` -> `{type:"document", source:...}` - # (Studio extension; mirrors Anthropic's document block, + # (Unsloth extension; mirrors Anthropic's document block, # which supports PDFs as base64 or URL per # https://platform.claude.com/docs/en/build-with-claude/vision) anthropic_parts: list[dict[str, Any]] = [] @@ -1749,7 +1749,7 @@ class ExternalProviderClient: } ) elif part.get("type") == "input_document": - # Studio's normalised PDF/doc type (file_data data-URI or + # Unsloth's normalised PDF/doc type (file_data data-URI or # file_url) -> Anthropic's native `document` block. url = part.get("file_url") or "" data_uri = part.get("file_data") or "" @@ -4704,7 +4704,7 @@ class ExternalProviderClient: {"type": "image_generation_call", "id": call_id} ) elif part_type == "input_document": - # Map Studio's `input_document` onto Responses' `input_file`. + # Map Unsloth's `input_document` onto Responses' `input_file`. # https://developers.openai.com/api/docs/guides/images-vision file_url = part.get("file_url") file_data = part.get("file_data") @@ -6010,7 +6010,7 @@ class ExternalProviderClient: if not models and self.provider_type == "ollama": models = await self._list_ollama_native_models() # Gemini's native /v1beta/models uses a different shape; repackage - # into the OpenAI-compatible one Studio expects. + # into the OpenAI-compatible one Unsloth expects. if not models and self.provider_type == "gemini": models = self._parse_gemini_models(data) return models @@ -6213,7 +6213,7 @@ def _friendly_provider_error_text( *, model: str | None = None, ) -> str: - """Rewrite common provider errors into actionable Studio copy.""" + """Rewrite common provider errors into actionable Unsloth copy.""" if status_code == 404 and model: lowered = raw_message.lower() if "not found" in lowered or "not_found" in lowered: diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py index fec5214cf9..d7c7eed518 100644 --- a/studio/backend/core/inference/llama_cpp.py +++ b/studio/backend/core/inference/llama_cpp.py @@ -11,6 +11,7 @@ import atexit import contextlib import functools import json +import math import os import re import struct @@ -23,11 +24,22 @@ import sys import threading import time from pathlib import Path -from typing import Callable, Collection, Generator, Iterable, List, Mapping, Optional, Union +from typing import ( + Callable, + Collection, + Generator, + Iterable, + List, + Literal, + Mapping, + Optional, + Union, +) import httpx from core.inference.llama_server_args import ( + _LAYER_OFFLOAD_FLAGS, _effective_tensor_parallel, _tensor_parallel_matches_loaded, extra_args_disable_mmproj, @@ -116,7 +128,7 @@ LLAMA_SERVER_NOT_FOUND_DETAIL = ( # llama-server can serve HTTP 200 while running a model entirely on CPU when a # GPU backend fails to init (#5807 / #5106 / #5830). Classify the startup log so -# Studio can warn. Priority: explicit "offloaded N/M layers to GPU" counts +# Unsloth can warn. Priority: explicit "offloaded N/M layers to GPU" counts # (authoritative), then GPU "model buffer size" lines (host-pinned _Host # excluded), then the "device_info:" device table (disconfirm only). _GPU_OFFLOAD_MARKERS = ( @@ -234,8 +246,7 @@ def _wsl_system_rocm_lib_dirs() -> "list[str]": return out -# Plan-without-action re-prompt state (intent signal, caps, message) now lives -# in tool_call_parser, imported above under its old aliases. +# Plan-without-action re-prompt state now lives in tool_call_parser (imported above). # Default max_tokens to the effective context when known. The floor is high # enough for reasoning-heavy GGUFs and max_tokens-omitting API clients. @@ -1363,7 +1374,7 @@ def _kv_bytes_per_elem(cache_type: Optional[str]) -> float: def _env_main_cache_type_for_budget(env: Optional[Mapping[str, str]] = None) -> Optional[str]: """Heavier of the inherited LLAMA_ARG_CACHE_TYPE_K/_V env types when it - exceeds the f16 default, else None. Studio emits --cache-type only for the + exceeds the f16 default, else None. Unsloth emits --cache-type only for the param/extras path, so a heavier env (f32) would otherwise reach the child unbudgeted; quantized env types stay over-reserved by f16 (-> None).""" e = os.environ if env is None else env @@ -1431,7 +1442,10 @@ def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool: return _extra_args_set_any_flag(extra_args, {"--spec-type", "--spec-default"}) -_GPU_OFFLOAD_OVERRIDE_FLAGS = frozenset({"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"}) +# Layer-offload override detection. Single-sourced from llama_server_args, which +# also strips these (plus the MoE flags) from inherited extras; sharing the layer +# set keeps detection and stripping from drifting. +_GPU_OFFLOAD_OVERRIDE_FLAGS = _LAYER_OFFLOAD_FLAGS _THREAD_OVERRIDE_FLAGS = frozenset({"-t", "--threads"}) @@ -1682,7 +1696,7 @@ def _build_ngram_mod_flags( return [] -# Canonical Speculative Decoding modes exposed by the Studio chat UI. +# Canonical Speculative Decoding modes exposed by the Unsloth chat UI. # Dropdown renders five (auto, mtp, ngram, mtp+ngram, off); the load API # also accepts legacy values the original Switch and external callers emit # (default, draft-mtp, ngram-mod, ngram-simple). @@ -1731,7 +1745,7 @@ def _backfill_usage_from_timings(usage, timings): """Synthesize ``usage`` from llama-server's ``timings`` when the OpenAI-style usage block is missing or reports zero tokens. - The Studio chat UI computes generation t/s from + The Unsloth chat UI computes generation t/s from ``meta.usage.completion_tokens / totalStreamTime``. llama-server always populates ``timings.predicted_n`` (true decoded count) and ``timings.prompt_n``, but the final SSE chunk's ``usage`` can be absent @@ -1804,7 +1818,7 @@ def _llama_lib_dir(binary: str) -> Path: def _is_external_link(path: Path) -> bool: """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink or a Windows directory junction / reparse point. Such a link resolves into - the user's own llama.cpp checkout, which Studio does not own.""" + the user's own llama.cpp checkout, which Unsloth does not own.""" try: if os.path.islink(path): return True @@ -1895,6 +1909,17 @@ class LlamaCppBackend: self._cache_type_kv: Optional[str] = None # Whether --split-mode tensor was applied on the active load. self._tensor_parallel: bool = False + # GPU memory strategy applied on the active load ("auto"/"manual"). + self._gpu_memory_mode: str = "auto" + # Manual-mode load options (echoed back so the UI round-trips them). + self._gpu_layers: int = -1 + # MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none. + self._n_cpu_moe: int = 0 + # Relative model share per GPU (--tensor-split), in GPU order; None = + # default (llama.cpp splits by free VRAM). + self._tensor_split: Optional[List[float]] = None + # User-picked physical GPU indices (None = automatic selection). + self._gpu_ids: Optional[List[int]] = None # 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 @@ -1909,6 +1934,11 @@ class LlamaCppBackend: self._spec_draft_n_max: Optional[int] = None # KV-cache estimation fields (populated by _read_gguf_metadata) self._n_layers: Optional[int] = None + # MoE metadata (populated by _read_gguf_metadata): expert count (>0 = + # MoE) and leading dense-layer count (offsets --n-cpu-moe, which counts + # from layer 0). See the n_moe_layers property. + self._n_experts: Optional[int] = None + self._leading_dense_block_count: Optional[int] = None self._n_kv_heads: Optional[int] = None self._n_kv_heads_by_layer: Optional[list[int]] = None self._n_heads: Optional[int] = None @@ -1960,7 +1990,7 @@ class LlamaCppBackend: # observes it (direct proxy endpoints, or nothing in flight). self._mtp_watchdog_thread: Optional[threading.Thread] = None self._mtp_watchdog_stop = threading.Event() - # True when the launch actually runs MTP+tensor (Studio- or user/env-driven); + # True when the launch actually runs MTP+tensor (Unsloth- or user/env-driven); # gates the probe, watchdog, and recovery so pass-through MTP is covered. self._mtp_runtime_fallback_active = False self._stdout_lines: list[str] = [] @@ -2329,6 +2359,79 @@ class LlamaCppBackend: """Whether --split-mode tensor is active on the loaded server.""" return self._tensor_parallel + @property + def gpu_memory_mode(self) -> str: + """Active GPU memory strategy: 'auto' or 'manual' (gpu_layers < 0 = Auto/--fit, >= 0 = pinned).""" + return self._gpu_memory_mode + + @property + def gpu_layers(self) -> int: + """Requested --gpu-layers for manual mode (-1 when not manual).""" + return self._gpu_layers + + @property + def n_cpu_moe(self) -> int: + """MoE expert layers manual mode kept on CPU (--n-cpu-moe); 0 = none.""" + return self._n_cpu_moe + + @property + def tensor_split(self) -> Optional[List[float]]: + """Manual-mode relative model share per GPU (--tensor-split); None = + default (split by free VRAM).""" + return self._tensor_split + + @property + def gpu_ids(self) -> Optional[List[int]]: + """User-picked physical GPU indices, or None for automatic selection.""" + return self._gpu_ids + + @property + def n_layers(self) -> Optional[int]: + """Model layer count (GGUF block_count), or None if unknown.""" + return self._n_layers + + @property + def n_moe_layers(self) -> int: + """Number of MoE expert layers (the --n-cpu-moe ceiling), 0 if not MoE. + + block_count minus the leading dense layers (which carry no experts): + --n-cpu-moe counts from layer 0, so those dense layers are no-ops. + """ + if not self._n_experts or not self._n_layers: + return 0 + return max(0, self._n_layers - (self._leading_dense_block_count or 0)) + + @staticmethod + def _resolve_cpu_moe_flag( + n_cpu_moe: int, n_moe_layers: int, leading_dense: int + ) -> Optional[int]: + """The --n-cpu-moe value (absolute first-N layers), or None to omit it. + + Clamps the requested count to the model's MoE layers, then offsets past + the leading dense layers (--n-cpu-moe counts from layer 0). Returns None + for nothing-to-offload (0 requested) or a non-MoE model. + """ + if n_cpu_moe <= 0 or n_moe_layers <= 0: + return None + return leading_dense + min(n_cpu_moe, n_moe_layers) + + @staticmethod + def _sanitize_tensor_split(tensor_split: Optional[List[float]]) -> List[float]: + """Per-GPU shares with negative and non-finite entries clamped to 0. + + A direct caller's negative entry would launch a placement different + from the ratio the UI showed, and inf would pass a plain ``> 0`` total + gate and emit ``--tensor-split inf,...``. Returns [] for input that + can't be read as floats (the length gate at the call site then drops + the split). + """ + try: + return [ + x if math.isfinite(x) and x > 0.0 else 0.0 for x in (float(v) for v in tensor_split) + ] + except (TypeError, ValueError, OverflowError): + return [] + @property def layer_preserves_tensor_intent(self) -> bool: """True when a downgraded tensor request kept this layer load multi-GPU.""" @@ -2353,7 +2456,7 @@ class LlamaCppBackend: @staticmethod def _resolved_studio_root_and_is_legacy() -> "tuple[Optional[Path], bool]": - """Resolve the Studio install root and classify it as the legacy + """Resolve the Unsloth install root and classify it as the legacy ~/.unsloth/studio root vs. a custom (env/venv-inferred) root. Returns (resolved_root, is_legacy). On any import/resolution failure the @@ -2530,6 +2633,7 @@ class LlamaCppBackend: "spec_draft_n_max_flag": None, "supports_kv_unified": False, "supports_fit_ctx": False, + "supports_fit_target": False, "supports_cache_ram": False, "supports_ctx_checkpoints": False, "supports_no_cache_prompt": False, @@ -2549,6 +2653,7 @@ class LlamaCppBackend: spec_draft_n_max_flag: Optional[str] = None supports_kv_unified = False supports_fit_ctx = False + supports_fit_target = False supports_cache_ram = False supports_ctx_checkpoints = False supports_no_cache_prompt = False @@ -2646,6 +2751,7 @@ class LlamaCppBackend: supports_kv_unified = _is_real("--kv-unified") supports_fit_ctx = _is_real("--fit-ctx") + supports_fit_target = _is_real("--fit-target") supports_cache_ram = _is_real("--cache-ram") supports_ctx_checkpoints = _is_real("--ctx-checkpoints") supports_no_cache_prompt = _is_real("--no-cache-prompt") @@ -2662,6 +2768,7 @@ class LlamaCppBackend: "spec_draft_n_max_flag": spec_draft_n_max_flag, "supports_kv_unified": supports_kv_unified, "supports_fit_ctx": supports_fit_ctx, + "supports_fit_target": supports_fit_target, "supports_cache_ram": supports_cache_ram, "supports_ctx_checkpoints": supports_ctx_checkpoints, "supports_no_cache_prompt": supports_no_cache_prompt, @@ -2746,6 +2853,57 @@ class LlamaCppBackend: except ValueError: return None + @staticmethod + def _emit_child_gpu_visibility(env: dict, pinned: str) -> None: + """Write the child's GPU visibility mask (CUDA, plus the HIP mirror on + ROCm, where narrowing only CUDA_VISIBLE_DEVICES leaves an AMD child + seeing the full set). Do NOT also set ROCR_VISIBLE_DEVICES: ROCR and HIP + mask at different layers, so the same indices apply twice -- ROCR reduces + and re-indexes from 0, then a non-zero HIP pin points out of range, HIP + enumerates 0 devices, and llama.cpp falls back to CPU. The HIP mask alone + narrows correctly; clear any inherited ROCR mask so it can't double up.""" + env["CUDA_VISIBLE_DEVICES"] = pinned + try: + import torch as _torch + if getattr(_torch.version, "hip", None) is not None: + env["HIP_VISIBLE_DEVICES"] = pinned + env.pop("ROCR_VISIBLE_DEVICES", None) + except Exception as e: + logger.debug("Failed to set ROCm visibility env vars for child: %s", e) + + @staticmethod + def _pin_visible_gpu_order_for_split(env: dict) -> None: + """Pin the child's GPU enumeration to the picker's order for a manual + ``--tensor-split`` across the whole visible set. CUDA's default + FASTEST_FIRST enumeration applies the shares to the wrong cards on + heterogeneous hosts (#5025), and CUDA_DEVICE_ORDER only fixes the + numbering base: an inherited numeric visibility mask ALSO defines + enumeration order, so a reordered parent mask (CUDA_VISIBLE_DEVICES=3,1) + would still hand the shares to the wrong cards. The UI built the split + positionally over get_backend_visible_gpu_info's device list (ascending + physical via nvidia-smi, inherited mask order on the torch fallback), so + re-emit the same set in that report order -- not an assumed ascending + sort. The visible set itself never changes. No mask, an empty mask, or a + UUID/MIG mask (which resolves to None) is left alone -- the multi-GPU + controls are hidden for the latter.""" + env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" + inherited = LlamaCppBackend._resolve_visible_physical_ids() + if not inherited: + return + order = None + try: + from utils.hardware import get_backend_visible_gpu_info + info = get_backend_visible_gpu_info() + if info.get("available") and info.get("index_kind") == "physical": + reported = [d["index"] for d in info.get("devices", [])] + if sorted(reported) == sorted(inherited): + order = reported + except Exception as e: + logger.debug("Could not read reported GPU order for split pin: %s", e) + if order is None: + order = sorted(inherited) + LlamaCppBackend._emit_child_gpu_visibility(env, ",".join(str(i) for i in order)) + @staticmethod def _amd_apu_wants_unified_memory(gpu_indices = None) -> bool: """True only for AMD unified-memory APUs (gfx1150/gfx1151), where @@ -3241,7 +3399,7 @@ class LlamaCppBackend: return prev = curr - # Free-VRAM fraction at which Studio pins the GPU directly instead of + # Free-VRAM fraction at which Unsloth pins the GPU directly instead of # deferring to ``--fit on``. 3% headroom: the compute buffer is now modelled in # the fit, so this only guards fragmentation + multi-GPU per-device CUDA context # (~2-3%); kept >= 3% as a floor (0.90 dropped 91-94% fits to CPU offload, #5106). @@ -3262,6 +3420,20 @@ class LlamaCppBackend: # aborts a --split-mode tensor load, so it's dropped for the tensor attempt. _TENSOR_PARALLEL_KV_TYPES = frozenset({"f16", "bf16", "f32"}) + # Main-model placement settings that Manual mode owns. They must not leak + # from Studio's parent environment into llama-server and silently override + # the command assembled from the current request. Draft-model placement is + # intentionally separate and remains available to speculative decoding. + _MANUAL_PLACEMENT_ENV_VARS = ( + "LLAMA_ARG_CPU_MOE", + "LLAMA_ARG_N_CPU_MOE", + "LLAMA_ARG_N_GPU_LAYERS", + "LLAMA_ARG_TENSOR_SPLIT", + "LLAMA_ARG_FIT", + "LLAMA_ARG_FIT_TARGET", + "LLAMA_ARG_FIT_CTX", + ) + # (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. @@ -3426,6 +3598,12 @@ class LlamaCppBackend: return env + @classmethod + def _clear_manual_placement_env(cls, env: dict[str, str]) -> None: + """Remove inherited main-model placement owned by Manual mode.""" + for name in cls._MANUAL_PLACEMENT_ENV_VARS: + env.pop(name, None) + @staticmethod def _select_gpus( model_size_bytes: int, @@ -3800,7 +3978,7 @@ class LlamaCppBackend: return total if total > 0 else None return draft_kv + weights + target_ctx_copy - _DEFAULT_N_UBATCH = 512 # llama.cpp --ubatch default; Studio does not override it + _DEFAULT_N_UBATCH = 512 # llama.cpp --ubatch default; Unsloth does not override it _COMPUTE_BUFFER_SAFETY = 1.15 # upper-bound margin on the compute-buffer estimate # Soft VRAM the modeled terms omit; charged to the fit budget on tight tiers (#6682). _CUDA_CONTEXT_RESERVE_BYTES = 320 * 1024 * 1024 # CUDA ctx + cuBLAS workspace (~330 MiB) @@ -3940,7 +4118,7 @@ class LlamaCppBackend: n_ubatch: Optional[int] = None, ) -> tuple[Optional[list[int]], bool, int]: """Largest serving-slot count in [1, n_parallel) whose fully-on-GPU footprint fits, - so Studio keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers + so Unsloth keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers to host and collapses decode ~3x (oobabooga #6718). ``base_footprint_bytes`` is the slot-independent footprint (weights + soft overhead + MTP + context-linear compute, minus the folded compute buffer); each candidate re-adds the slot-sized compute buffer @@ -4246,6 +4424,8 @@ class LlamaCppBackend: self._supports_preserve_thinking = False self._supports_tools = False self._n_layers = None + self._n_experts = None + self._leading_dense_block_count = None self._n_kv_heads = None self._n_kv_heads_by_layer = None self._n_heads = None @@ -4335,6 +4515,8 @@ class LlamaCppBackend: arch_keys = { f"{arch}.context_length": "context_length", f"{arch}.block_count": "n_layers", + f"{arch}.expert_count": "n_experts", + f"{arch}.leading_dense_block_count": "leading_dense_block_count", f"{arch}.attention.head_count_kv": "n_kv_heads", f"{arch}.attention.head_count": "n_heads", f"{arch}.embedding_length": "embedding_length", @@ -4416,7 +4598,7 @@ class LlamaCppBackend: ] # Otherwise hand off to the resolver (cache / bootstrap / transformers / HF). Diffusion models - # skip it: they do not use Studio's SWA pattern and the resolver can raise for them. + # skip it: they do not use Unsloth's SWA pattern and the resolver can raise for them. if ( self._sliding_window_pattern is None and self._sliding_window @@ -4523,6 +4705,28 @@ class LlamaCppBackend: return None + @staticmethod + def _diffusion_gpu_arg(gpu_ids: Optional[List[int]], *, cpu_only: bool = False) -> str: + """Device token passed to the diffusion visual-server child. + + The visual engine replaces its child's CUDA visibility mask with this + token, so an unpinned load must carry forward the first token from the + parent's mask rather than turning a parent-relative ordinal into a new + physical selection. + """ + if gpu_ids: + return str(sorted(gpu_ids)[0]) + if cpu_only: + return "" + if "DG_GPU" in os.environ: + return os.environ["DG_GPU"] + parent_mask = os.environ.get("CUDA_VISIBLE_DEVICES") + if parent_mask: + first = next((token.strip() for token in parent_mask.split(",") if token.strip()), "") + if first and first != "-1": + return first + return "0" + def _start_diffusion_server( self, *, @@ -4533,10 +4737,11 @@ class LlamaCppBackend: model_identifier: str, n_ctx: int, extra_args: Optional[List[str]], + gpu_ids: Optional[List[int]] = None, ) -> bool: """Launch the OpenAI-compat diffusion shim (which drives the on-device visual decoder) and wait for health. Presents the same /v1 + /health - interface as llama-server, so the rest of Studio is unchanged. + interface as llama-server, so the rest of Unsloth is unchanged. """ assets = self._find_diffusion_assets() if assets is None: @@ -4558,7 +4763,11 @@ class LlamaCppBackend: # CUDA_VISIBLE_DEVICES="" to force CPU serving. Keep the visual-server child # CPU-masked (empty --gpu) so the shim does not re-expose GPU 0 via its default. cpu_only = self._effective_gpu_count() == 0 - gpu = "" if cpu_only else os.environ.get("DG_GPU", "0") + # Honor the GPU picker first: the diffusion runner takes a single device, + # so use the lowest selected GPU (matches the sorted set recorded below, so + # the device used == the echoed gpu_ids[0]). With no pick, fall back to the + # CPU-only mask, else DG_GPU / 0. + gpu = self._diffusion_gpu_arg(gpu_ids, cpu_only = cpu_only) cmd = list(shim_cmd) + [ "--gguf", @@ -4586,6 +4795,11 @@ class LlamaCppBackend: env.setdefault("UNSLOTH_ALLOW_CPU", "1") env["DG_VISUAL_BIN"] = visual_bin env["DG_GPU"] = gpu + if gpu_ids: + # The visual server remasks via CUDA_VISIBLE_DEVICES=; pin PCI + # order (as the llama-server path does) so the picked physical id maps + # to the GPU the picker showed, not CUDA's default fastest-first order. + env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # The file-override shim imports its sibling visual_engine; put its dir on PYTHONPATH. # (The zoo-package shim is an installed module and needs no PYTHONPATH change.) if extra_pythonpath: @@ -4608,7 +4822,7 @@ class LlamaCppBackend: logger.debug(f"Could not open diffusion runner log file: {e}") # The shim (and its visual server) die with this backend process, so a - # Studio crash/restart never orphans a GPU process. + # Unsloth crash/restart never orphans a GPU process. self._process = subprocess.Popen( cmd, stdout = subprocess.PIPE, @@ -4631,6 +4845,23 @@ class LlamaCppBackend: self._model_identifier = model_identifier self._cache_type_kv = None self._gpu_offload_active = True + # Diffusion doesn't use the llama.cpp GPU-memory knobs; reset them to + # defaults (the picked device is still recorded below) so /load, /status + # and reload dedup don't report a previous GGUF's manual settings. + self._gpu_memory_mode = "auto" + self._gpu_layers = -1 + self._n_cpu_moe = 0 + self._tensor_split = None + # Diffusion is never tensor-parallel; clear any state left by a prior TP + # chat load (load_model phase 1 only kills the process, it doesn't run + # the unload reset) so /status doesn't misreport TP and an identical + # re-Apply doesn't reload against stale tensor-parallel state. + self._tensor_parallel = False + # Record only the single device the runner actually uses (the lowest + # selected GPU, chosen above) -- not the whole pick. The diffusion runner + # is single-device, so echoing a multi-GPU list would misreport placement + # in /status and let a re-Apply dedup against GPUs the runner never used. + self._gpu_ids = [sorted(gpu_ids)[0]] if gpu_ids else None if hf_variant: self._hf_variant = hf_variant elif gguf_path: @@ -5242,7 +5473,7 @@ class LlamaCppBackend: return ( f"'{arch}' is a diffusion (image-generation) GGUF, which " "llama-server cannot run as a chat/completion model. Use " - "Studio's Images page to generate with local diffusion " + "Unsloth's Images page to generate with local diffusion " "GGUFs such as FLUX and Qwen-Image." ) if is_ollama: @@ -5721,6 +5952,11 @@ class LlamaCppBackend: speculative_type: Optional[str] = None, spec_draft_n_max: Optional[int] = None, tensor_parallel: bool = False, + gpu_memory_mode: Literal["auto", "manual"] = "auto", + gpu_layers: int = -1, + n_cpu_moe: int = 0, + tensor_split: Optional[List[float]] = None, + gpu_ids: Optional[List[int]] = None, n_threads: Optional[int] = None, n_gpu_layers: Optional[int] = None, # caller compat, unused n_parallel: int = 1, @@ -5753,6 +5989,14 @@ class LlamaCppBackend: "speculative_type": speculative_type, "spec_draft_n_max": spec_draft_n_max, "tensor_parallel": tensor_parallel, + # GPU-memory placement: replayed on respawn so a server SIGKILL'd by + # GPU/RAM pressure reloads onto the same devices with the same + # offload, not the auto defaults. + "gpu_memory_mode": gpu_memory_mode, + "gpu_layers": gpu_layers, + "n_cpu_moe": n_cpu_moe, + "tensor_split": list(tensor_split) if tensor_split is not None else None, + "gpu_ids": list(gpu_ids) if gpu_ids is not None else None, "n_threads": n_threads, "n_gpu_layers": n_gpu_layers, "n_parallel": n_parallel, @@ -5779,6 +6023,11 @@ class LlamaCppBackend: speculative_type = speculative_type, spec_draft_n_max = spec_draft_n_max, tensor_parallel = tensor_parallel, + gpu_memory_mode = gpu_memory_mode, + gpu_layers = gpu_layers, + n_cpu_moe = n_cpu_moe, + tensor_split = tensor_split, + gpu_ids = gpu_ids, chat_template_override = chat_template_override, extra_args = extra_args, is_vision = is_vision, @@ -5899,6 +6148,7 @@ class LlamaCppBackend: model_identifier = model_identifier, n_ctx = n_ctx, extra_args = extra_args, + gpu_ids = gpu_ids, ) if not binary: @@ -5960,6 +6210,59 @@ class LlamaCppBackend: # use the same helper so a healthy env-driven tensor server matches. split_mode_override = parse_split_mode_override(extra_args) tensor_parallel = _effective_tensor_parallel(extra_args, tensor_parallel) + # gpu_layers=0 leaves nothing to split, yet --split-mode tensor or + # a per-GPU ratio still launches tensor mode -- and under the + # CPU-only mask below (no visible devices) that aborts the server + # instead of loading on CPU. Drop both here (nothing to split). + if gpu_memory_mode == "manual" and gpu_layers == 0: + if tensor_parallel or tensor_split: + logger.info( + "Manual gpu_layers=0: dropping tensor split/parallel " + "flags (nothing to split on the GPU)" + ) + tensor_parallel = False + tensor_split = None + # Record the requested strategy for /status and the load + # response. 'manual' has no fallback, so the request value is the + # value actually applied. + self._gpu_memory_mode = gpu_memory_mode + # The layer/MoE/split knobs apply only with an explicit offload + # (manual + gpu_layers >= 0); else record defaults so /status and + # /load don't report knobs the server never applied. + if gpu_memory_mode == "manual" and gpu_layers >= 0: + self._gpu_layers = gpu_layers + self._n_cpu_moe = n_cpu_moe + self._tensor_split = tensor_split + else: + self._gpu_layers = -1 + self._n_cpu_moe = 0 + self._tensor_split = None + self._gpu_ids = sorted(gpu_ids) if gpu_ids else None + # Manual offload skips the TP planner but still emits --split-mode + # tensor at launch; drop it when fewer than 2 GPUs are in use -- + # tensor split is a no-op there and aborts on some architectures. + # Done before the cache-drop below so a quantized KV survives. + if ( + tensor_parallel + and gpu_memory_mode == "manual" + and gpu_layers >= 0 + and self._effective_gpu_count(sorted(gpu_ids) if gpu_ids else None) < 2 + ): + logger.info( + "Tensor parallelism requested in manual mode but fewer " + "than 2 GPUs are in use; ignoring (needs >= 2)." + ) + tensor_parallel = False + # Drop TP for manual + Auto layers before the cache-drop below (like + # the <2-GPU guard above), so a requested quantized KV survives into + # the --fit load rather than being stripped for a tensor attempt. + if tensor_parallel and gpu_memory_mode == "manual" and gpu_layers < 0: + logger.info( + "Manual mode with Auto layers hands memory management to " + "llama.cpp --fit, which is incompatible with tensor " + "parallelism; ignoring the tensor split." + ) + tensor_parallel = False # Tensor mode aborts on a quantized KV cache, so drop it for the # tensor attempt (and strip any inherited/explicit --cache-type # that would re-impose it when appended last). Layer split does @@ -6040,6 +6343,12 @@ class LlamaCppBackend: "Vision-capable GGUF loaded without a usable mmproj; " "image input will be disabled for this session" ) + # Seed before the try: the except (GPU-selection failure -> + # --fit on) falls through to the launch which reads this, and the + # probe that assigns it may throw first. Captured before manual + # empty `gpus` so the speculative defaults stay GPU-aware and the + # CPU-fallback check still knows GPUs were present. + _detected_gpus: list[tuple[int, int]] = [] 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). @@ -6057,6 +6366,18 @@ class LlamaCppBackend: _gpu_mem = self._get_gpu_memory(binary) gpus = [(idx, free) for idx, free, _t in _gpu_mem] total_by_idx = {idx: total for idx, _f, total in _gpu_mem} + # GPU picker: restrict every mode to the chosen devices, so + # auto selection only considers them and manual mask to + # them (the env block below pins CUDA/HIP_VISIBLE_DEVICES). + if gpu_ids: + _picked = set(gpu_ids) + gpus = [g for g in gpus if g[0] in _picked] + + # GPUs the model will run on -- captured before manual + # empty `gpus` to bypass the planner. bool() drives the + # GPU-aware speculative defaults; the list feeds the + # CPU-fallback check. + _detected_gpus = list(gpus) def _gpu_usable(g, frac = _CTX_FIT_VRAM_FRACTION): # Per-GPU usable budget for ranking: free - (1-frac)*total. @@ -6088,6 +6409,44 @@ class LlamaCppBackend: # GPU/VRAM-fit logic below may shrink it on limited HW. max_available_ctx = self._context_length or effective_ctx + # Manual + Auto layers (the Manual default): hand memory + # management to llama.cpp's --fit. Emptying the probed GPU set + # no-ops the selection/TP planning below, leaving gpu_indices + # None (an explicit gpu_ids pick still pins below) and use_fit + # True. An explicit context is honored (--fit optimizes around + # it); 0 lets --fit size it. + if gpu_memory_mode == "manual" and gpu_layers < 0: + # Tensor parallelism was already dropped above (before the + # cache-drop), so a quantized KV survives into this --fit load. + gpus = [] + effective_ctx = requested_ctx if requested_ctx > 0 else 0 + original_ctx = effective_ctx + # --fit aborts under --split-mode tensor; a raw extras + # --split-mode/--tensor-split (appended last) would + # otherwise reach llama-server. Strip it like the TP + # downgrade does. + extra_args = strip_split_mode_only(extra_args) + elif gpu_memory_mode == "manual": + # Manual offload (--gpu-layers + --fit off): no automatic + # device masking (a gpu_ids pick still pins below) or + # context cap -- the user owns both. tensor_parallel is + # honored but skips the memory-based planner (gpus = []); + # the toggle just emits --split-mode tensor (split by free + # VRAM, or by the Split ratio if set). + gpus = [] + effective_ctx = ( + requested_ctx if requested_ctx > 0 else (self._context_length or 0) + ) + original_ctx = effective_ctx + # Strip the user --split-mode when the toggle owns the split + # (TP engaged -> Studio emits --split-mode tensor) or when the + # user asked for tensor (which aborts on a single GPU even if + # the manual <2-GPU guard downgraded TP). Otherwise keep their + # non-tensor mode (row/none/layer) -- the toggle can't express + # those. + if tensor_parallel or split_mode_override == "tensor": + extra_args = strip_split_mode_only(extra_args) + # Will MTP engage? If so, auto-fit reserves draft-model VRAM. # Mirrors _build_speculative_flags: forced mtp/mtp+ngram always # engage; auto only on an MTP model >= 3B; ngram/off never. A @@ -6103,7 +6462,7 @@ class LlamaCppBackend: and not bool(mtp_draft_path) ) # LLAMA_ARG_SPEC_TYPE only reaches the child when neither extras - # nor Studio emit a spec flag (mode "off", no user --spec-type), + # nor Unsloth emit a spec flag (mode "off", no user --spec-type), # since _build_speculative_flags emits one for every other mode. # Consult the env for the reserve only then, else a stale MTP env # would over-reserve. @@ -6112,7 +6471,7 @@ class LlamaCppBackend: if (not _extra_args_set_spec_type(extra_args) and _mtp_canonical == "off") else {} ) - # Extras can run MTP even when Studio suppresses its own emission. + # Extras can run MTP even when Unsloth suppresses its own emission. _user_mtp_via_extras = _extra_args_requests_mtp(extra_args, env = _spec_env) # A non-MTP model-based draft mode (draft-simple/draft-eagle3) in # extras also loads a separate draft model that needs reserving; @@ -6175,10 +6534,13 @@ class LlamaCppBackend: _extra_n_max = _extra_args_spec_draft_n_max(extra_args) _mtp_eff_n_max = _extra_n_max if _extra_n_max is not None else spec_draft_n_max if _mtp_eff_n_max is None: - _mtp_eff_n_max = 2 if gpus else 3 + # _detected_gpus (not gpus) so manual -- which empty + # gpus to bypass the planner -- keep the GPU draft depth the + # launch flags also use, instead of the CPU default. + _mtp_eff_n_max = 2 if _detected_gpus else 3 # Separate-drafter weights live on GPU (an embedded head is # already in model_size). Size the drafter the launch loads, by - # precedence: extras --model-draft (last-wins), else Studio's + # precedence: extras --model-draft (last-wins), else Unsloth's # emitted mtp_draft_path, else the env drafter. Sizing the wrong # one would under-reserve and OOM. _cli_draft_for_budget = _extra_args_mtp_draft_path(extra_args, env = {}) @@ -6313,7 +6675,8 @@ class LlamaCppBackend: # honor it, cap only if it fits no combination. Auto (native): # prefer fewer GPUs with reduced context (multi-GPU is slower). gpu_indices, use_fit = None, True - # Per-GPU weight proportions for tensor mode (None = even). + # Per-GPU weight proportions for tensor mode (None lets + # llama.cpp split by free VRAM). tp_tensor_split: Optional[list[int]] = None explicit_ctx = requested_ctx > 0 # Flat MTP reserve fraction: used only as the fallback when the @@ -6388,7 +6751,12 @@ class LlamaCppBackend: # GPUs below that reserve from the set up front (gpu_indices # becomes the CUDA_VISIBLE_DEVICES mask, fully excluding them). tp_gpus = gpus - if tensor_parallel: + # Manual mode owns the layer count and context, so it skips + # the memory-based planner; its toggle still emits + # --split-mode tensor below (split by free VRAM, or by the + # Split ratio if set). auto plans here. + plan_tp = tensor_parallel and gpu_memory_mode != "manual" + if plan_tp: # Deterministic per-device compute buffer (replicated on # every device in tensor mode); flat fallback when dims # are unavailable. _plan_tensor_parallel uses the same. @@ -6407,7 +6775,7 @@ class LlamaCppBackend: # free yet have no budget left. tp_gpus = [g for g in gpus if _gpu_usable(g) >= reserve_mib] - if tensor_parallel and len(tp_gpus) < 2: + if plan_tp and len(tp_gpus) < 2: # Tensor parallelism needs >= 2 usable GPUs. On a single # GPU --split-mode tensor is a no-op; with 0 GPUs (CPU-only # or probe failed) it must not reach llama-server; and a @@ -6823,6 +7191,12 @@ class LlamaCppBackend: tp_tensor_split = None effective_ctx = requested_ctx # fall back to original + # GPU picker: when no narrower subset was chosen (manual, or + # a failed/file-size selection), pin the whole picked set so the + # model can't spill onto an unpicked GPU. + if gpu_ids and gpu_indices is None: + gpu_indices = sorted(gpu_ids) + # Unified-memory APUs load weights into system RAM (under WSL the VM # cap, not the ROCm-reported VRAM, is the real ceiling); refuse an # oversize load the OS would otherwise kill mid-flight. Base model @@ -6859,8 +7233,6 @@ class LlamaCppBackend: model_path, "--port", str(self._port), - "-c", - str(effective_ctx) if effective_ctx > 0 else "0", "--parallel", str(n_parallel), "--flash-attn", @@ -6868,6 +7240,17 @@ class LlamaCppBackend: # Error out at n_ctx instead of silently rotating the KV cache; frontend catches it and points the user at "Context Length". "--no-context-shift", ] + # A positive context is always passed (in auto-fit, --fit then + # optimizes the gpu-layer offload around it). When auto-fit has + # no explicit context, omit -c so --fit sizes it to fit VRAM: + # "-c 0" would instead pin the FULL native context (llama.cpp's + # -c handler sets fit_params_min_ctx = UINT32_MAX on value 0, + # disabling --fit's reduction). See gpu_memory_mode. + auto_fit = gpu_memory_mode == "manual" and gpu_layers < 0 + if effective_ctx > 0: + cmd.extend(["-c", str(effective_ctx)]) + elif not auto_fit: + cmd.extend(["-c", "0"]) # 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 @@ -6879,7 +7262,63 @@ class LlamaCppBackend: cmd.extend(["--alias", _alias]) fully_gpu_offloaded = False - if use_fit: + # Set when a positional --tensor-split is emitted, so the env block + # can pin CUDA to PCI order even without a GPU subset (see below). + manual_tensor_split_emitted = False + if gpu_memory_mode == "manual" and gpu_layers >= 0: + # Pin the user's layer count and disable auto-fit. --fit off + # also means _ctx_integrity_flags must not add --fit-ctx. + use_fit = False + cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"]) + # Keep the first n_cpu_moe MoE layers' experts on CPU. + moe_flag = self._resolve_cpu_moe_flag( + n_cpu_moe, + self.n_moe_layers, + self._leading_dense_block_count or 0, + ) + if moe_flag is not None: + cmd.extend(["--n-cpu-moe", str(moe_flag)]) + elif n_cpu_moe: + # Requested on a dense model: nothing was emitted, so + # don't report a count llama-server never received. + self._n_cpu_moe = 0 + # Distribute the model across GPUs by the user's per-GPU shares + # (--tensor-split). Works with layer split and tensor + # parallelism; --fit off means no fit/tensor abort. Only emit + # when >1 GPU is in use AND the list length matches that count: + # the field is hidden (not cleared) when the picker narrows to + # one, and a direct caller can send a stale ratio for a different + # GPU set. Studio drops any mismatch to the free-VRAM default + # (llama.cpp would silently zero-pad a short list, or abort past + # its 16-device cap). + _split_gpus = self._effective_gpu_count(gpu_indices) + if tensor_split and _split_gpus > 1: + # An all-zero/non-positive sanitized split assigns nothing + # anywhere, so fall through to the free-VRAM default in + # that case. + _sanitized_split = self._sanitize_tensor_split(tensor_split) + _split_total = sum(_sanitized_split) + if len(_sanitized_split) == _split_gpus and _split_total > 0: + cmd.extend( + ["--tensor-split", ",".join(f"{x:g}" for x in _sanitized_split)] + ) + self._tensor_split = _sanitized_split + manual_tensor_split_emitted = True + else: + logger.warning( + "Dropping manual --tensor-split (%d entries for " + "%d GPUs, sanitized total %s); llama.cpp's " + "free-VRAM split applies instead", + len(tensor_split), + _split_gpus, + _split_total, + ) + self._tensor_split = None + elif tensor_split: + # Single effective GPU: the split is never emitted, so + # don't report it as active via /status and /load. + self._tensor_split = None + elif use_fit: cmd.extend(["--fit", "on"]) elif gpu_indices is not None: # Fits on selected GPU(s) -- force all layers on GPU. --fit off is @@ -6897,6 +7336,7 @@ class LlamaCppBackend: self._ctx_integrity_flags( n_parallel, use_fit, + auto_fit, requested_ctx, effective_ctx, server_caps, @@ -6960,9 +7400,11 @@ class LlamaCppBackend: self._cache_type_kv = None # Tensor parallelism: split the model across GPUs by tensor - # rather than by layer. Multi-GPU only -- a no-op on a single - # GPU. Default (layer split) is left implicit by omitting the - # flag. See llama.cpp --split-mode. + # rather than by layer. The UI only offers it on multi-GPU; a + # direct single-GPU caller is redundant (supported archs no-op, + # unsupported ones abort and the /load path retries layer split). + # Default (layer split) is left implicit by omitting the flag. + # See llama.cpp --split-mode. if tensor_parallel: cmd.extend(["--split-mode", "tensor"]) if tp_tensor_split and len(tp_tensor_split) > 1: @@ -6994,7 +7436,7 @@ class LlamaCppBackend: extra_args = extra_args, model_identifier = model_identifier, model_path = model_path, - gpus = bool(gpus), + gpus = bool(_detected_gpus), binary = binary, mtp_draft_path = launch_mtp_draft_path, ) @@ -7097,12 +7539,12 @@ class LlamaCppBackend: # Vulkan pins via --device (a cmd arg, unlike the env-based # CUDA/ROCm pin below), emitted BEFORE user extras so llama.cpp's - # last-wins parsing lets a user --device override Studio's pick. + # last-wins parsing lets a user --device override Unsloth's pick. if is_vulkan_backend and gpu_indices is not None: cmd += LlamaCppBackend._vulkan_pin_args(gpu_indices) # User pass-through args go last so llama.cpp's last-wins parsing - # lets the user override Studio's auto-set flags. Already + # lets the user override Unsloth's auto-set flags. Already # validated by the route via validate_extra_args(). if extra_args: cmd.extend(str(a) for a in extra_args) @@ -7112,15 +7554,17 @@ class LlamaCppBackend: # Library paths so llama-server finds its shared libs and CUDA DLLs. env = self._llama_server_env_for_binary(binary) + if gpu_memory_mode == "manual": + self._clear_manual_placement_env(env) # Omitting --threads relies on llama.cpp's physical-core default, so # drop an inherited LLAMA_ARG_THREADS that would otherwise feed the # arg handler and silently force hardware_concurrency(). #5692 if "--threads" not in cmd: env.pop("LLAMA_ARG_THREADS", None) - # Reconcile the inherited LLAMA_ARG_* env with Studio's final + # Reconcile the inherited LLAMA_ARG_* env with Unsloth's final # decision: stripping CLI extras on a tensor->layer downgrade - # can't remove env vars, so the child could run a mode/KV Studio + # can't remove env vars, so the child could run a mode/KV Unsloth # didn't budget. if not tensor_parallel: # Layer split: clear a non-layer inherited split mode (and any @@ -7130,7 +7574,7 @@ class LlamaCppBackend: env.pop("LLAMA_ARG_SPLIT_MODE", None) env.pop("LLAMA_ARG_TENSOR_SPLIT", None) else: - # Studio owns the tensor split: it emits --tensor-split when it + # Unsloth owns the tensor split: it emits --tensor-split when it # picks an uneven one (CLI wins) and nothing when an even split # is safe. Clear any inherited LLAMA_ARG_TENSOR_SPLIT so the even # case can't be overridden by a stale env (the layer branch above @@ -7170,28 +7614,39 @@ class LlamaCppBackend: # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full set, so # set HIP_VISIBLE_DEVICES too. Vulkan is pinned via --device # (above), not here. - if gpu_indices is not None and not is_vulkan_backend: - pinned = ",".join(str(i) for i in gpu_indices) - env["CUDA_VISIBLE_DEVICES"] = pinned - try: - import torch as _torch - if getattr(_torch.version, "hip", None) is not None: - env["HIP_VISIBLE_DEVICES"] = pinned - # Do NOT also set ROCR_VISIBLE_DEVICES to the same - # value. ROCR_VISIBLE_DEVICES filters at the HSA/ROCr - # layer and HIP_VISIBLE_DEVICES at the HIP layer, so - # setting both with the same physical indices applies - # the mask twice: ROCR reduces the visible set and - # re-indexes it from 0, then HIP indexes into the - # already-reduced set. A single non-zero pin (e.g. - # "1") then points out of range at the HIP layer, HIP - # enumerates 0 devices, and llama.cpp falls back to - # CPU ("ggml_cuda_init: no ROCm-capable device is - # detected"). The HIP mask alone narrows correctly; - # clear any inherited ROCR mask so it can't double up. - env.pop("ROCR_VISIBLE_DEVICES", None) - except Exception as e: - logger.debug("Failed to set ROCm visibility env vars for child: %s", e) + # A deliberate zero-offload load with no GPU companions runs + # entirely on CPU, yet a visible CUDA device still costs the child + # ~0.5 GB (context + compute scratch) that the CPU-only + # classification below reports as free. Hide the GPUs so the load + # is exactly what it claims: zero VRAM (verified: GPU stays at idle + # baseline and generation runs). Companion loads keep the normal + # masking, and a user device pin (in extras or an inherited + # LLAMA_ARG_DEVICE) keeps control of its own devices -- the child + # aborts on a pin it can't see. The draft-device forms count too: + # llama-server parses them even with no drafter loaded. + _cpu_only_zero_offload = ( + gpu_memory_mode == "manual" + and gpu_layers == 0 + and not is_vulkan_backend + and not self._zero_offload_keeps_gpu_visible(cmd, env) + ) + if _cpu_only_zero_offload: + self._emit_child_gpu_visibility(env, "-1") + elif gpu_indices is not None and not is_vulkan_backend: + # When the user picked GPUs by index, align CUDA's ordering + # with the PCI-bus order the picker enumerated (nvidia-smi), + # so "GPU 1" in the UI is GPU 1 to llama.cpp -- not CUDA's + # default FASTEST_FIRST order (#5025). + if gpu_ids: + env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" + self._emit_child_gpu_visibility(env, ",".join(str(i) for i in gpu_indices)) + elif manual_tensor_split_emitted and not is_vulkan_backend: + # A manual per-GPU ratio across ALL GPUs (no explicit pick, so + # no CUDA_VISIBLE_DEVICES mask above): the UI built the + # --tensor-split list in ascending physical/PCI index order, + # so pin the child's enumeration to that order too. The whole + # visible set stays in use; only its ordering is fixed. + self._pin_visible_gpu_order_for_split(env) # Captured before any text-only fallback strips it from cmd. launched_with_mmproj = "--mmproj" in cmd @@ -7201,7 +7656,7 @@ class LlamaCppBackend: # 'on') even when -ngl is explicit. That step has aborted on # some ROCm hosts (ggml-cuda.cu ROCm error during worst-case # estimation, e.g. MTP + mmproj models on gfx1151). When - # Studio's own VRAM math already placed the model + # Unsloth's own VRAM math already placed the model # (use_fit=False), the step is redundant second-guessing -- # retry once with --fit off before declaring the load failed. # Never retry when fit was requested (use_fit) or the caller @@ -7284,7 +7739,7 @@ class LlamaCppBackend: and _startup_crashed and not _split_axis_crash ): - # We forced --fit off because Studio's (conservative) VRAM + # We forced --fit off because Unsloth's (conservative) VRAM # math placed the model fully on GPU. A startup crash here # means that estimate was optimistic, so fall back to --fit # on and let llama.cpp offload rather than fail the load. @@ -7296,7 +7751,7 @@ class LlamaCppBackend: self._process.returncode, self._llama_log_path, ) - # Flip Studio's own --fit off (added first, before any + # Flip Unsloth's own --fit off (added first, before any # user extra args) to on; a user's later --fit still wins # by last-arg. Defensive: if absent, the default is already # --fit on, so leave it. @@ -7313,7 +7768,7 @@ class LlamaCppBackend: ): logger.warning( "llama-server crashed during startup (exit code %s) " - "with the default memory-fit step enabled; Studio " + "with the default memory-fit step enabled; Unsloth " "already verified the model fits, retrying once " "with --fit off. Crash log: %s", self._process.returncode, @@ -7350,7 +7805,6 @@ class LlamaCppBackend: self._effective_context_length = ( effective_ctx if effective_ctx > 0 else self._context_length ) - self._reconcile_effective_ctx_with_server() self._max_context_length = ( max_available_ctx if max_available_ctx > 0 else self._effective_context_length ) @@ -7393,7 +7847,7 @@ class LlamaCppBackend: cmd = _fa_cmd healthy = _spawn_and_wait(_fa_cmd, label = "-noflash") - # MTP from Studio's spec flags or the user's (extra_args + # MTP from Unsloth's spec flags or the user's (extra_args # --spec-type / LLAMA_ARG_SPEC_TYPE). The env reaches the child # only when neither emits a spec flag, so consult it only then. _launch_spec_env: Mapping[str, str] = ( @@ -7535,6 +7989,10 @@ class LlamaCppBackend: "session; run 'unsloth studio update' to enable vision." ) cmd = self._strip_mmproj_args(_last_spawn_cmd) + # This retry bypasses _spawn_and_wait, so refresh the + # launched-argv snapshot itself -- the zero-offload + # classification below must not see the stripped --mmproj. + _last_spawn_cmd = list(cmd) self._is_vision = False self._mmproj_has_audio = False self._start_llama_process(cmd, env) @@ -7566,6 +8024,13 @@ class LlamaCppBackend: self._healthy = True self._commit_effective_parallel_slots(n_parallel) + # Server is up: adopt the real per-request context it allocated + # -- the length --fit chose, or a --parallel slot split -- so the + # reported context_length matches reality. (Querying /props + # before the spawn above always failed; the seeded value was the + # requested/native length.) + self._reconcile_effective_ctx_with_server() + # Commit caller intent only after _healthy=True so a failed start # can't poison the next inheritance check. None keeps prior, [] # clears, list sets. Source records hf_variant for the route's @@ -7580,18 +8045,31 @@ class LlamaCppBackend: self._mtp_runtime_fallback_active = _mtp_active_for_launched_server self._start_mtp_crash_watchdog() - # Catch silent CPU fallback when GPU was intended (#5106). - self._gpu_offload_active = self._classify_gpu_offload( - gpu_indices is not None or use_fit, gpus or [] - ) - if self._gpu_offload_active is False: + # Catch silent CPU fallback when GPU was intended (#5106). Manual + # offload (no picker) leaves gpu_indices None and use_fit False, so + # include its GPU-layer intent; use the preserved probe since + # auto-layers/manual empty `gpus`. A deliberate zero-offload load + # classifies by its launched argv instead: the main model is + # CPU-only by construction and must read False (not None), or + # training needlessly unloads a server holding no VRAM. + _deliberate_cpu_only = gpu_memory_mode == "manual" and gpu_layers == 0 + if _deliberate_cpu_only: + self._gpu_offload_active = self._zero_offload_gpu_flag( + _last_spawn_cmd, _detected_gpus, env + ) + else: + self._gpu_offload_active = self._classify_gpu_offload( + gpu_indices is not None or use_fit or gpu_memory_mode == "manual", + _detected_gpus, + ) + if self._gpu_offload_active is False and not _deliberate_cpu_only: logger.warning( "llama-server appears to have loaded the model entirely " - "on CPU even though Studio detected at least one GPU. " + "on CPU even though Unsloth detected at least one GPU. " "This usually means the prebuilt binary's GPU backend " "failed to load -- on Windows, cudart64_X.dll / " "cublas64_X.dll could not be resolved. Reinstall the " - "Studio llama.cpp prebuilt or install a matching CUDA " + "Unsloth llama.cpp prebuilt or install a matching CUDA " "toolkit (issue unslothai/unsloth#5106).", ) @@ -7888,7 +8366,7 @@ class LlamaCppBackend: logger.info( "Auto: MLA embedded-MTP model detected; llama.cpp's MLA/DSA " "MTP path is slower than no speculation, so using ngram-mod " - "instead. Override via the Studio Speculative Decoding " + "instead. Override via the Unsloth Speculative Decoding " "dropdown or UNSLOTH_MLA_MTP_ENABLED=1." ) _emit_ngram_mod() @@ -7916,7 +8394,7 @@ class LlamaCppBackend: f"MTP GGUF detected but model size {_mtp_size_b:.1f}B " "is below the 3B speedup threshold; using ngram-mod " "only (zero-VRAM, no draft head). Override via " - "--spec-type or the Studio Speculative Decoding " + "--spec-type or the Unsloth Speculative Decoding " "dropdown." ) _emit_ngram_mod() @@ -7947,6 +8425,11 @@ class LlamaCppBackend: gguf_path: Optional[str] = None, spec_draft_n_max: Optional[int] = None, tensor_parallel: bool = False, + gpu_memory_mode: Literal["auto", "manual"] = "auto", + gpu_layers: int = -1, + n_cpu_moe: int = 0, + tensor_split: Optional[List[float]] = None, + gpu_ids: Optional[List[int]] = None, mtp_draft_path: Optional[str] = None, preserve_multi_gpu_on_layer: bool = False, ) -> bool: @@ -8003,6 +8486,38 @@ class LlamaCppBackend: ): return False + # The diffusion runner is mode-agnostic (always "auto", ignores the + # layer/MoE/split knobs), so a standing manual preference in the + # request must not force a needless reload -- only the GPU pick matters. + if not self._is_diffusion: + # A GPU-memory-mode flip (Unsloth / manual) must always reload. + if self._gpu_memory_mode != gpu_memory_mode: + return False + # Manual: a layer-count change always reloads (covers Auto(-1) <-> a + # pinned count); MoE/split only matter with an explicit offload. + if gpu_memory_mode == "manual" and ( + self._gpu_layers != gpu_layers + or ( + gpu_layers >= 0 + and ( + self._n_cpu_moe != n_cpu_moe + or (self._tensor_split or None) != (tensor_split or None) + ) + ) + ): + return False + # A changed GPU pick must reload (compare order-insensitively; None/[] + # both mean automatic). The diffusion runner collapses a multi-GPU pick + # to its single lowest device, so self._gpu_ids holds just that device; + # normalize the request the same way, or a multi-GPU pick that resolves + # to the same device needlessly reloads. + if self._is_diffusion: + requested_gpu_pick = [sorted(gpu_ids)[0]] if gpu_ids else None + else: + requested_gpu_pick = sorted(gpu_ids) if gpu_ids else None + if (self._gpu_ids or None) != requested_gpu_pick: + return False + # Compare on the canonical requested mode. With --spec-type in # extra_args the backend stores None; mirror that here. if _extra_args_set_spec_type(extra_args): @@ -8071,6 +8586,78 @@ class LlamaCppBackend: return None return classify_gpu_offload_lines(self._stdout_lines) + @staticmethod + def _cmd_has_gpu_companion(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool: + """True when the argv/env carries a GPU companion: any --mmproj form, or + a drafter (Studio's --model-draft, the extras aliases, or the + LLAMA_ARG_SPEC_DRAFT_* env) -- these offload to the GPU regardless of + the main ``--gpu-layers``. A drafter explicitly forced to CPU + (--spec-draft-ngl 0 / --spec-draft-device cpu) doesn't count.""" + if any(str(a).startswith("--mmproj") for a in cmd): + return True + if _extra_args_mtp_draft_path(cmd, env) is None: + return False + return not _extra_args_draft_offloaded_to_cpu(cmd, env) + + @staticmethod + def _zero_offload_keeps_gpu_visible(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool: + """Whether a zero-layer launch still has a reason to use visible GPUs. + + Keep this shared by child masking and post-launch residency bookkeeping: + a device pin, surviving tensor mode, mmproj, or GPU drafter prevents the + launch from being a confirmed zero-VRAM server. + """ + return ( + LlamaCppBackend._cmd_has_gpu_device_pin(cmd, env) + or _effective_tensor_parallel(cmd, False, env) + or LlamaCppBackend._cmd_has_gpu_companion(cmd, env) + ) + + @staticmethod + def _cmd_has_gpu_device_pin(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool: + """True when the effective main or draft ``--device`` pin names a GPU.""" + main_flags = {"--device", "-dev"} + draft_flags = {"--spec-draft-device", "-devd", "--device-draft"} + last_main: Optional[str] = None + last_draft: Optional[str] = None + args = [str(arg) for arg in cmd] + for index, raw in enumerate(args): + flag, equals, inline = raw.partition("=") + if flag not in main_flags and flag not in draft_flags: + continue + value = inline if equals else (args[index + 1] if index + 1 < len(args) else "") + if flag in main_flags: + last_main = value + else: + last_draft = value + if last_main is None: + last_main = (env or {}).get("LLAMA_ARG_DEVICE") + + def _names_gpu(value: Optional[str]) -> bool: + if value is None: + return False + devices = [item.strip().lower() for item in value.split(",") if item.strip()] + return not devices or any(item not in ("cpu", "none") for item in devices) + + return _names_gpu(last_main) or _names_gpu(last_draft) + + @staticmethod + def _zero_offload_gpu_flag( + spawn_cmd: list, + detected_gpus: list, + env: Optional[Mapping[str, str]] = None, + ) -> Optional[bool]: + """GPU-residency flag for a deliberate manual zero-offload load. The + main model is CPU-only by construction, but device pins, tensor mode, + mmproj, and GPU drafters can still make the server hold VRAM. The counted + offload classifier cannot see those allocations. This uses the same + predicate as the launch-time zero-VRAM mask; None means no GPU signal.""" + if not detected_gpus: + return None + if LlamaCppBackend._is_vulkan_backend(): + return True + return LlamaCppBackend._zero_offload_keeps_gpu_visible(spawn_cmd, env) + def load_cancelled(self) -> bool: """True if a load was cancelled (e.g. via unload/_cancel_event) and not yet consumed by the next load_model. Lets the tensor->layer fallback @@ -8114,11 +8701,18 @@ class LlamaCppBackend: self._supports_tools = False self._cache_type_kv = None self._tensor_parallel = False + self._gpu_memory_mode = "auto" + self._gpu_layers = -1 + self._n_cpu_moe = 0 + self._tensor_split = None + self._gpu_ids = None self._layer_preserves_tensor_intent = False self._speculative_type = None self._requested_spec_mode = None self._spec_draft_n_max = None self._n_layers = None + self._n_experts = None + self._leading_dense_block_count = None self._n_kv_heads = None self._n_kv_heads_by_layer = None self._n_heads = None @@ -8181,6 +8775,10 @@ class LlamaCppBackend: # Clear healthy so a /load during the replacement's warm-up can't # short-circuit against the previous server's health (#5401). self._healthy = False + # Reset to unknown so the training guard treats the next (still + # loading) server as VRAM-resident rather than reading the killed + # server's stale zero-offload flag until the health probe reclassifies. + self._gpu_offload_active = None # Drives _wait_for_vram_settle in the next load_model; set in finally # so both in-process and frontend Apply paths record the kill. self._last_kill_monotonic = time.monotonic() @@ -8294,7 +8892,7 @@ class LlamaCppBackend: def _pid_parent_is_alive(pid: int) -> bool: """True if the recorded server's parent is still running, i.e. the server is NOT orphaned. Lets the cross-session reap kill only a true orphan (parent - gone) and never a live server owned by a running Studio, regardless of which + gone) and never a live server owned by a running Unsloth, regardless of which process performs the sweep. Biased toward "alive" on uncertainty so a live server is never mistakenly reaped.""" try: @@ -8334,9 +8932,9 @@ class LlamaCppBackend: @classmethod def _reap_recorded_pid(cls) -> int: """Kill the exact llama-server PID recorded at spawn, but only when it is a - genuine orphan -- its parent (the Studio that spawned it) is gone. This is + genuine orphan -- its parent (the Unsloth that spawned it) is gone. This is the cross-session backstop the parent-death reaper (Job Object / - PR_SET_PDEATHSIG) cannot cover: an orphan left by an already-dead Studio + PR_SET_PDEATHSIG) cannot cover: an orphan left by an already-dead Unsloth (macOS, a best-effort failure, or a pre-existing orphan). Path-independent, so it also catches an orphan the install-root match would miss. @@ -8393,7 +8991,7 @@ class LlamaCppBackend: """Kill orphaned llama-server processes started by studio. Only kills processes whose resolved binary lives under a known - Studio install dir (or matches an exact env-var override), to avoid + Unsloth install dir (or matches an exact env-var override), to avoid terminating unrelated llama-server instances. Mirrors every location _find_llama_server_binary() can return, so orphans from any supported install path are cleaned up. @@ -8413,7 +9011,7 @@ class LlamaCppBackend: try: # -- Build the ownership allowlist -------------------------------- # exact_binaries -- env var overrides (exact path match). - # install_roots -- Studio-owned dir trees (binary must be under one). + # install_roots -- Unsloth-owned dir trees (binary must be under one). install_roots: list[Path] = [] # Env-mode custom root (mirrors _find_llama_server_binary). @@ -8423,7 +9021,7 @@ class LlamaCppBackend: install_roots.append(_resolved_sr / "llama.cpp") # Primary install dir (default mode only). Env-mode skips this so a - # custom-root Studio can't kill a default-install Studio's server. + # custom-root Unsloth can't kill a default-install Unsloth's server. if not _is_custom_root: install_roots.append(Path.home() / ".unsloth" / "llama.cpp") @@ -8497,7 +9095,7 @@ class LlamaCppBackend: if not is_ours: continue - # A live parent means a running Studio (or the user's + # A live parent means a running Unsloth (or the user's # shell) still owns it -- not an orphan. if LlamaCppBackend._pid_parent_is_alive(proc.info["pid"]): continue @@ -8577,7 +9175,7 @@ class LlamaCppBackend: def _fit_off_retry_eligible(cmd: "list[str]", use_fit: bool) -> bool: """Whether a llama-server startup crash may be retried with --fit off. - Only when Studio's own VRAM math placed the model (use_fit=False) + Only when Unsloth's own VRAM math placed the model (use_fit=False) and nothing on the command line set the fit mode explicitly (-fit / --fit, space- or equals-form). --fit-ctx / --fit-target / -fitc / -fitt tune the fit step but do not select the mode, so @@ -8785,7 +9383,12 @@ class LlamaCppBackend: @staticmethod def _ctx_integrity_flags( - n_parallel: int, use_fit: bool, requested_ctx: int, effective_ctx: int, caps: dict + n_parallel: int, + use_fit: bool, + auto_fit: bool, + requested_ctx: int, + effective_ctx: int, + caps: dict, ) -> list[str]: """Flags that keep the per-request window equal to the advertised ctx. @@ -8793,14 +9396,28 @@ class LlamaCppBackend: ``--kv-unified`` default, silently splitting ``-c`` into per-slot windows of ``-c / N``; restore the shared pool so one request can use the full context. With ``--fit on``, ``--fit-ctx`` floors the fit step - at an explicitly requested ctx (default floor is 4096) so it offloads - or fails instead of silently shrinking the window. + at an explicitly requested ctx so it offloads or fails instead of + silently shrinking the window. The 8192 auto-floor and the tighter + ``--fit-target`` margin apply only under Manual + Auto (``auto_fit``), + which omits ``-c``: on the legacy auto path ``-c 0`` already pins the + native window and ``--fit-ctx 8192`` would override it down to 8192. """ flags: list[str] = [] if n_parallel > 1 and caps.get("supports_kv_unified"): flags.append("--kv-unified") - if use_fit and requested_ctx > 0 and effective_ctx > 0 and caps.get("supports_fit_ctx"): - flags.extend(["--fit-ctx", str(effective_ctx)]) + if use_fit and caps.get("supports_fit_ctx"): + if requested_ctx > 0 and effective_ctx > 0: + # Floor the fit step at the explicitly requested ctx. + flags.extend(["--fit-ctx", str(effective_ctx)]) + elif auto_fit: + # Manual + Auto omits -c, so floor at 8192 so --fit doesn't + # shrink the window below a usable size. + flags.extend(["--fit-ctx", "8192"]) + if use_fit and auto_fit and caps.get("supports_fit_target"): + # llama.cpp's --fit leaves 1 GiB free per device by default; + # tighten that to 512 MiB so it packs more of the model onto + # the GPU before spilling to system RAM. + flags.extend(["--fit-target", "512"]) return flags def _query_server_n_ctx(self) -> Optional[int]: @@ -8821,7 +9438,7 @@ class LlamaCppBackend: return None def _reconcile_effective_ctx_with_server(self) -> None: - """Adopt the server's real ``n_ctx`` when it is below Studio's value. + """Adopt the server's real ``n_ctx`` when it is below Unsloth's value. Keeps ``context_length`` (load response, status route, passthrough ``max_tokens`` ceiling) honest; clients sized to the requested value diff --git a/studio/backend/core/inference/llama_keepwarm.py b/studio/backend/core/inference/llama_keepwarm.py index 4ce663c3ce..86a8c8a404 100644 --- a/studio/backend/core/inference/llama_keepwarm.py +++ b/studio/backend/core/inference/llama_keepwarm.py @@ -59,7 +59,7 @@ _INFERENCE_SUFFIXES = ( "/messages/count_tokens", # counts via the loaded tokenizer; protect like /messages "/embeddings", "/responses", - "/generate/stream", # Studio's own streaming route on the same llama-server + "/generate/stream", # Unsloth's own streaming route on the same llama-server "/audio/generate", # direct GGUF TTS; can outlive the idle TTL ) diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py index f400d2ae40..e72e10e071 100644 --- a/studio/backend/core/inference/llama_server_args.py +++ b/studio/backend/core/inference/llama_server_args.py @@ -3,10 +3,10 @@ """Boundary validator for user-supplied llama-server pass-through args. -Reject only flags Studio manages (model identity, auth, network, parallel +Reject only flags Unsloth manages (model identity, auth, network, parallel slots). Everything else (sampling, ``-c``, ``-ngl``, ``--flash-attn``, ``--cache-type-*``, ``--spec-*``, ``--jinja``, ...) is appended after -Studio's auto-set flags so llama.cpp's last-wins parser lets the user override. +Unsloth's auto-set flags so llama.cpp's last-wins parser lets the user override. Ref: https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md """ @@ -22,12 +22,12 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = ( # Parallel slots: owned by typer --parallel; a pass-through would desync # app.state.llama_parallel_slots from llama-server. frozenset({"-np", "--parallel", "--n-parallel"}), - # Model identity: Studio resolves it from LoadRequest; a second -m would - # load a different model than Studio thinks it loaded. + # Model identity: Unsloth resolves it from LoadRequest; a second -m would + # load a different model than Unsloth thinks it loaded. frozenset({"-m", "--model"}), - # Public model id: Studio sets a sanitized --alias so the OpenAI API never + # Public model id: Unsloth 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 + # Unsloth'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"}), @@ -39,14 +39,14 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = ( frozenset({"-hft", "--hf-token"}), frozenset({"-mm", "--mmproj"}), frozenset({"-mmu", "--mmproj-url"}), - # Networking: Studio binds + proxies; retargeting orphans the proxy. + # Networking: Unsloth binds + proxies; retargeting orphans the proxy. frozenset({"--host"}), frozenset({"--port"}), frozenset({"--path"}), frozenset({"--api-prefix"}), frozenset({"--reuse-port"}), - # Auth / TLS: Studio terminates auth; upstream --api-key / TLS shadows - # Studio's key and breaks the proxy hop. + # Auth / TLS: Unsloth terminates auth; upstream --api-key / TLS shadows + # Unsloth's key and breaks the proxy hop. frozenset({"--api-key"}), frozenset({"--api-key-file"}), frozenset({"--ssl-key-file"}), @@ -64,11 +64,11 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = ( frozenset({"--models-max"}), frozenset({"--models-autoload", "--no-models-autoload"}), # Server-mode flips: --embedding / --rerank restrict llama-server to - # those endpoints, breaking Studio's /v1/chat/completions hop. + # those endpoints, breaking Unsloth's /v1/chat/completions hop. frozenset({"--embedding", "--embeddings"}), frozenset({"--rerank", "--reranking"}), # llama-server's own built-in tools flag would silently stack on top of - # Studio's --enable-tools / --disable-tools policy resolver. + # Unsloth's --enable-tools / --disable-tools policy resolver. frozenset({"--tools"}), ) @@ -120,7 +120,7 @@ def validate_extra_args(args: Optional[Iterable[str]]) -> list[str]: def is_managed_flag(flag: str) -> bool: - """True if ``flag`` is Studio-managed. Normalises via ``_flag_name`` so + """True if ``flag`` is Unsloth-managed. Normalises via ``_flag_name`` so `-np8` / `--parallel=8` classify like the canonical tokens.""" normalised = _flag_name(flag) return normalised is not None and normalised in _DENYLIST @@ -142,7 +142,7 @@ _SPEC_FLAGS: frozenset[str] = frozenset( "--draft-min", "--draft-max", # MTP path (llama.cpp #22673). The drafter selectors (local --model-draft - # and HF --spec-draft-hf aliases) are Studio-managed since the separate- + # and HF --spec-draft-hf aliases) are Unsloth-managed since the separate- # drafter support (Gemma 4): an inherited copy must not last-wins-override # the auto-detected drafter. Explicit extras for the current load are never # stripped. The per-drafter tuning knobs (--spec-draft-type-*, -ngld, @@ -179,25 +179,38 @@ _TEMPLATE_FLAGS: frozenset[str] = frozenset( # (--split-mode tensor). Pass-through stays allowed so users keep the # row/none/layer modes the toggle doesn't expose, but it's stripped on # inherit and reconciled into the round-tripped tensor_parallel state. -# --tensor-split is coupled to the split mode and is stripped with it: Studio +# --tensor-split is coupled to the split mode and is stripped with it: Unsloth # owns the tensor-mode split ratios, so an inherited/stale --tensor-split must -# not last-wins-override Studio's computed asymmetric split. +# not last-wins-override Unsloth's computed asymmetric split. _SPLIT_MODE_FLAGS: frozenset[str] = frozenset({"-sm", "--split-mode"}) _TENSOR_SPLIT_FLAGS: frozenset[str] = frozenset({"-ts", "--tensor-split"}) _SPLIT_SHADOWING_FLAGS: frozenset[str] = _SPLIT_MODE_FLAGS | _TENSOR_SPLIT_FLAGS +# GPU-offload flags. Stripped only when the GPU Memory mode owns offload +# (manual emits --fit / --gpu-layers / --n-cpu-moe); in auto, a user's +# inherited -ngl is respected (the offload_overridden path), so this group is +# opt-in, not default. Layer flags are shared with llama_cpp's override +# detection; the MoE flags are strip-only (manual's --n-cpu-moe slider owns them). +_LAYER_OFFLOAD_FLAGS: frozenset[str] = frozenset( + {"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"} +) +_MOE_OFFLOAD_FLAGS: frozenset[str] = frozenset({"-ncmoe", "--n-cpu-moe", "-cmoe", "--cpu-moe"}) +_OFFLOAD_SHADOWING_FLAGS: frozenset[str] = _LAYER_OFFLOAD_FLAGS | _MOE_OFFLOAD_FLAGS + _SHADOWING_FLAGS: frozenset[str] = ( _CONTEXT_FLAGS | _CACHE_FLAGS | _SPEC_FLAGS | _TEMPLATE_FLAGS | _SPLIT_SHADOWING_FLAGS ) # Shadowing flags that take no value -- strip the flag only, not the next token. -_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset({"--spec-default", "--jinja", "--no-jinja"}) +_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset( + {"--spec-default", "--jinja", "--no-jinja", "-cmoe", "--cpu-moe"} +) def parse_ctx_override(args: Optional[Iterable[str]]) -> Optional[int]: """Return the last user-supplied ``-c`` / ``--ctx-size`` value. - Mirrors llama.cpp's last-wins parsing for the one numeric knob Studio's + Mirrors llama.cpp's last-wins parsing for the one numeric knob Unsloth's load-time fit logic needs. """ if not args: @@ -286,7 +299,7 @@ def parse_cache_override(args: Optional[Iterable[str]]) -> Optional[str]: Mirrors parse_ctx_override but for cache type. Recognises both -ctk (key) and -ctv (value). When both flags appear, returns the last-wins value, treating key and value cache flags as the same setting because - Studio's KV estimate has a single cache_type_kv knob. + Unsloth's KV estimate has a single cache_type_kv knob. """ return _last_flag_value(args, _CACHE_FLAGS) @@ -341,7 +354,7 @@ def resolve_tensor_parallel(args: Optional[Iterable[str]], fallback_tensor_paral def _env_split_mode_is_tensor(env: Optional[Mapping[str, str]] = None) -> bool: - """True when the inherited LLAMA_ARG_SPLIT_MODE env selects tensor. Studio + """True when the inherited LLAMA_ARG_SPLIT_MODE env selects tensor. Unsloth emits --split-mode only on its tensor branch, so a tensor env on the layer path would run the child tensor-parallel unbudgeted; this flips the budget to tensor. Only tensor is heavier, so other modes are ignored.""" @@ -424,14 +437,22 @@ def strip_shadowing_flags( strip_spec: bool = True, strip_template: bool = True, strip_split_mode: bool = True, + strip_tensor_split: bool = False, + strip_offload: bool = False, ) -> list[str]: - """Strip flags that shadow first-class Studio settings. + """Strip flags that shadow first-class Unsloth settings. Used when inheriting a previous load's ``llama_extra_args`` so an inherited `-c 4096` can't override the current `max_seq_length` (same for cache / spec / template / split-mode). Each ``strip_*`` toggle controls one group; the route only strips groups whose first-class field the caller actually supplied. + + ``strip_split_mode`` removes both ``--split-mode`` and the coupled + ``--tensor-split`` (the Tensor Parallelism toggle owns the whole split). + ``strip_tensor_split`` removes ``--tensor-split`` *alone*, so manual mode can + replace an inherited per-GPU ratio while leaving the user's ``--split-mode`` + row/none/layer choice intact. """ shadowing: set[str] = set() if strip_context: @@ -444,6 +465,10 @@ def strip_shadowing_flags( shadowing |= _TEMPLATE_FLAGS if strip_split_mode: shadowing |= _SPLIT_SHADOWING_FLAGS + if strip_tensor_split: + shadowing |= _TENSOR_SPLIT_FLAGS + if strip_offload: + shadowing |= _OFFLOAD_SHADOWING_FLAGS tokens = [str(a) for a in (args or [])] out: list[str] = [] diff --git a/studio/backend/core/inference/llama_stats.py b/studio/backend/core/inference/llama_stats.py index 6047aedbc0..ab0d287e8c 100644 --- a/studio/backend/core/inference/llama_stats.py +++ b/studio/backend/core/inference/llama_stats.py @@ -5,7 +5,7 @@ engine-stats log line (generation/prompt throughput, requests in flight). llama-server already computes these (it needs `--metrics`); this lifts them -into Studio's structured log so the terminal shows serving health, not just +into Unsloth's structured log so the terminal shows serving health, not just per-request access lines. Emitted only while there is activity. """ diff --git a/studio/backend/core/inference/local_model_resolver.py b/studio/backend/core/inference/local_model_resolver.py index 002cafe2c8..64ab38ec75 100644 --- a/studio/backend/core/inference/local_model_resolver.py +++ b/studio/backend/core/inference/local_model_resolver.py @@ -130,7 +130,7 @@ def info_has_local_gguf(info) -> bool: def _build_index() -> dict[str, _LocalGgufEntry]: """Map normalized id/model_id/display_name -> local GGUF entry. - Scans the same roots Studio's model picker lists (./models, the active plus + Scans the same roots Unsloth's model picker lists (./models, the active plus legacy/default HF caches, LM Studio dirs, and user scan folders) so a named local model is never missed and silently served as the loaded one. Ollama's scanner is skipped: it creates symlinks as a side effect and this runs on the @@ -199,9 +199,13 @@ def _build_index() -> dict[str, _LocalGgufEntry]: raw_id = getattr(info, "id", None) if not raw_id: continue - # Skip what Studio hides from its pickers (validation probe, RAG embed + # Skip what Unsloth hides from its pickers (validation probe, RAG embed # weights): not chat models, so never an auto-switch target. - if _is_hidden_model(raw_id, getattr(info, "path", None)): + if _is_hidden_model( + raw_id, + getattr(info, "model_id", None), + getattr(info, "path", None), + ): continue # Advertise a client-facing alias, not an absolute filesystem path. loader_id = _advertised_loader_id(info) diff --git a/studio/backend/core/inference/mcp_client.py b/studio/backend/core/inference/mcp_client.py index 6b5ce02216..0256df944e 100644 --- a/studio/backend/core/inference/mcp_client.py +++ b/studio/backend/core/inference/mcp_client.py @@ -906,7 +906,7 @@ def _call_stdio_tool( def _remaining() -> Optional[float]: return None if deadline is None else max(0.0, deadline - time.monotonic()) - # Callers without a Studio session id must retain the former one-shot + # Callers without an Unsloth session id must retain the former one-shot # behavior: no browser/cookie/tool state can leak into another request. # Use an ephemeral key (and close it below) rather than the shared empty # scope that the persistent-session cache used previously. diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py index e7a90b4307..e78c93b6f3 100644 --- a/studio/backend/core/inference/mlx_inference.py +++ b/studio/backend/core/inference/mlx_inference.py @@ -8,6 +8,7 @@ instead of torch/transformers for model loading and generation. import json import os import threading +from contextlib import contextmanager from typing import Optional, Generator from core.inference.message_content import content_to_text from core.inference.runtime_context import runtime_context_length @@ -20,6 +21,63 @@ from loggers import get_logger logger = get_logger(__name__) +def _mlx_adapter_modules(model): + """Return bypassable adapter entries and unsupported wrapper paths.""" + adapters = [] + unsupported = [] + for path, module in model.named_modules(): + if not path or not (hasattr(module, "lora_a") and hasattr(module, "lora_b")): + continue + base = getattr(module, "linear", None) + if base is None: + base = getattr(module, "embedding", None) + if base is None: + unsupported.append(path) + else: + adapters.append((path, module, base)) + return adapters, unsupported + + +@contextmanager +def _temporary_mlx_adapter_state(model, use_adapter): + """Select base or adapter modules for one request, then restore the tree.""" + if use_adapter is None: + yield + return + if isinstance(use_adapter, str): + raise NotImplementedError( + "Unsloth MLX: named adapter selection is not supported; use True for " + "the loaded adapter or False for the base model." + ) + if use_adapter is not True and use_adapter is not False: + raise TypeError("Unsloth MLX: use_adapter must be None, True, False, or a string.") + + adapters, unsupported = _mlx_adapter_modules(model) + if use_adapter is True: + if not adapters and not unsupported: + logger.warning("MLX adapter requested, but the active model has no adapter layers") + yield + return + if unsupported: + raise RuntimeError( + "Unsloth MLX: cannot disable adapter layers without their base modules: " + + ", ".join(unsupported[:5]) + ) + if not adapters: + yield + return + + from mlx.utils import tree_unflatten + + base_modules = tree_unflatten([(path, base) for path, _, base in adapters]) + adapter_modules = tree_unflatten([(path, wrapper) for path, wrapper, _ in adapters]) + try: + model.update_modules(base_modules) + yield + finally: + model.update_modules(adapter_modules) + + def _mlx_vlm_model_config(model): """Return the loaded MLX model config and its type, preferring whichever of config / _config actually carries a model_type.""" @@ -508,6 +566,7 @@ class MLXInferenceBackend: reasoning_effort = None, preserve_thinking = None, presence_penalty = 0.0, + _adapter_state = None, ) -> Generator[str, None, None]: if self._model is None: raise RuntimeError("No model loaded") @@ -552,6 +611,7 @@ class MLXInferenceBackend: reasoning_effort = reasoning_effort, preserve_thinking = preserve_thinking, presence_penalty = presence_penalty, + _adapter_state = _adapter_state, ) else: stream = self._generate_text( @@ -568,6 +628,7 @@ class MLXInferenceBackend: reasoning_effort = reasoning_effort, preserve_thinking = preserve_thinking, presence_penalty = presence_penalty, + _adapter_state = _adapter_state, ) yield from stream @@ -587,6 +648,7 @@ class MLXInferenceBackend: reasoning_effort = None, preserve_thinking = None, presence_penalty = 0.0, + _adapter_state = None, ): from mlx_lm import stream_generate from mlx_lm.sample_utils import make_sampler, make_logits_processors @@ -635,10 +697,6 @@ class MLXInferenceBackend: think_prefix = detect_think_prefill( prompt, getattr(self._tokenizer, "all_special_tokens", None) ) - # Emit it before the first token so the block renders during prefill. - if think_prefix: - yield think_prefix - sampler = make_sampler( temp = temperature, top_p = top_p, @@ -680,9 +738,12 @@ class MLXInferenceBackend: type(self._model).__name__, type(self._tokenizer).__name__, ) - with self._generation_lock: + with self._generation_lock, _temporary_mlx_adapter_state(self._model, _adapter_state): final_response = None try: + # Enter request-scoped model state before yielding any response. + if think_prefix: + yield think_prefix gen_kwargs = dict( prompt = prompt, max_tokens = max_new_tokens, @@ -749,6 +810,7 @@ class MLXInferenceBackend: reasoning_effort = None, preserve_thinking = None, presence_penalty = 0.0, + _adapter_state = None, ): from mlx_vlm import stream_generate as vlm_stream @@ -852,9 +914,6 @@ class MLXInferenceBackend: # Re-emit an open prefill from the prompt (see _generate_text). cumulative = detect_think_prefill(prompt, getattr(chat_target, "all_special_tokens", None)) - # Emit it before the first token so the block renders during prefill. - if cumulative: - yield cumulative logger.info( "VLM generating: prompt_len=%d, has_image=%s", len(prompt), @@ -891,9 +950,18 @@ class MLXInferenceBackend: def _stream_vlm_snapshots(): nonlocal cumulative - with self._generation_lock: + # Hold the generation lock AND the request-scoped adapter state for the + # whole stream so Base-vs-LoRA compare mode honors use_adapter and the + # wrapper tree is restored on completion, cancellation, or close. + with self._generation_lock, _temporary_mlx_adapter_state(self._model, _adapter_state): final_response = None try: + # Emit any prefilled block before the first token so the + # UI renders it during prefill, matching _generate_text. Done + # inside the adapter context so an unsupported request raises + # before any output escapes. + if cumulative: + yield cumulative for response in vlm_stream( self._model, self._processor, @@ -927,8 +995,11 @@ class MLXInferenceBackend: cancel_event = None, **gen_kwargs, ) -> Generator[str, None, None]: - # MLX LoRA adapter toggling not yet supported; generate normally - yield from self.generate_chat_response(cancel_event = cancel_event, **gen_kwargs) + yield from self.generate_chat_response( + cancel_event = cancel_event, + _adapter_state = use_adapter, + **gen_kwargs, + ) def reset_generation_state(self): import mlx.core as mx diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py index 3afda74411..eaa474d9b8 100644 --- a/studio/backend/core/inference/orchestrator.py +++ b/studio/backend/core/inference/orchestrator.py @@ -1502,14 +1502,27 @@ class InferenceOrchestrator: Uses the dispatcher path (no _gen_lock) so compare-mode requests don't block each other; the subprocess serializes them via its - sequential command loop. + sequential command loop. Backend failures raise instead of becoming + assistant text. """ - yield from self._generate_dispatched( + stream = self._generate_dispatched( use_adapter = use_adapter, cancel_event = cancel_event, stats_holder = stats_holder, **gen_kwargs, ) + try: + for chunk in stream: + if isinstance(chunk, GenStreamError): + # Preserve the public/operational flag so the route can surface + # the real message (e.g. "model is being unloaded") instead of a + # generic error. Mirrors the safetensors tool loop's _single_turn. + raise GenStreamErrorRaised(str(chunk), public = chunk.public) + yield chunk + finally: + close = getattr(stream, "close", None) + if callable(close): + close() def _generate_inner( self, diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py index ed7c7ecfcf..e6da0a22b0 100644 --- a/studio/backend/core/inference/passthrough_healing.py +++ b/studio/backend/core/inference/passthrough_healing.py @@ -5,7 +5,7 @@ With server-side tools disabled (``unsloth run --disable-tools``, every ``unsloth start`` coding agent), requests carrying the client's own ``tools`` -bypass Studio's tool loop and are relayed to/from llama-server verbatim. Small +bypass Unsloth's tool loop and are relayed to/from llama-server verbatim. Small GGUF models often emit their tool calls as TEXT (``{...}``, Gemma ``<|tool_call>...``, ```` XML) instead of structured ``tool_calls`` -- on the passthrough that text reaches the agent as prose and @@ -18,7 +18,7 @@ promotes calls whose function name exactly matches a declared tool. Promotion removes EXACTLY the promoted calls' markup spans (the parser reports them): undeclared calls, unparseable blocks, and suppressed alternate formats keep every byte and relay as text, so healing can never silently delete model -output. Responses without a tool signal, requests without tools, and Studio's +output. Responses without a tool signal, requests without tools, and Unsloth's own enable-tools loop are untouched. Per-request opt-out: ``auto_heal_tool_calls: false``. Process kill-switch: ``UNSLOTH_DISABLE_TOOL_CALL_HEALING=1``. diff --git a/studio/backend/core/inference/pricing.py b/studio/backend/core/inference/pricing.py index 3b611d3596..30fec47723 100644 --- a/studio/backend/core/inference/pricing.py +++ b/studio/backend/core/inference/pricing.py @@ -122,12 +122,12 @@ def calculate_cost(provider: str, model: str, usage: dict[str, Any]) -> dict[str "priced": bool(prices), } - # Accept raw (input_tokens/output_tokens) and Studio chat-style + # Accept raw (input_tokens/output_tokens) and Unsloth chat-style # (prompt_tokens/completion_tokens) envelopes. Cache buckets differ: # raw Anthropic: input_tokens EXCLUDES cache buckets # raw OpenAI: input_tokens INCLUDES cache_read - # Studio Anthropic: prompt_tokens INCLUDES cache_creation + cache_read - # Studio OpenAI: prompt_tokens == raw input_tokens + # Unsloth Anthropic: prompt_tokens INCLUDES cache_creation + cache_read + # Unsloth OpenAI: prompt_tokens == raw input_tokens # Clamp >=0 so corrupted payloads can't produce a negative bill. cache_creation = max(0, int(usage.get("cache_creation_input_tokens") or 0)) cache_read_native_present = ( @@ -160,7 +160,7 @@ def calculate_cost(provider: str, model: str, usage: dict[str, Any]) -> dict[str output_tokens = max(0, int(usage.get("completion_tokens") or 0)) if provider == "openai": # Cached tokens land on input_tokens_details (raw Responses) or - # prompt_tokens_details (Studio chat-style). + # prompt_tokens_details (Unsloth chat-style). for key in ("input_tokens_details", "prompt_tokens_details"): details = usage.get(key) or {} if isinstance(details, dict): diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py index 5b72373c03..d3bffc2f3d 100644 --- a/studio/backend/core/inference/providers.py +++ b/studio/backend/core/inference/providers.py @@ -276,8 +276,9 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = { "auth_header": "Authorization", "auth_prefix": "Bearer ", "notes": ( - "Local Ollama server. OpenAI-compatible /v1/chat/completions; " - "no API key. Surfaced via CUSTOM_PROVIDER_PRESETS in the frontend." + "Ollama server (local or cloud). OpenAI-compatible " + "/v1/chat/completions; API key optional (required by Ollama " + "cloud). Surfaced via CUSTOM_PROVIDER_PRESETS in the frontend." ), "hidden": True, }, diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py index 9110315815..40731de57b 100644 --- a/studio/backend/core/inference/safetensors_agentic.py +++ b/studio/backend/core/inference/safetensors_agentic.py @@ -995,7 +995,7 @@ def run_safetensors_tool_loop( if not safety_tc: # Re-prompt once on plan-without-action, before any tool runs # (GGUF loop parity). The retry is gated on nudge_tool_calls so - # Studio callers (which send True) always nudge, while API callers + # Unsloth callers (which send True) always nudge, while API callers # who omit the flag keep today's no-reprompt behavior (opt-in). intent_text = _reprompt_intent_text( content_accum, diff --git a/studio/backend/core/inference/sandbox_site/sitecustomize.py b/studio/backend/core/inference/sandbox_site/sitecustomize.py index d655e8e35a..244fa95145 100644 --- a/studio/backend/core/inference/sandbox_site/sitecustomize.py +++ b/studio/backend/core/inference/sandbox_site/sitecustomize.py @@ -4,7 +4,7 @@ """Sandbox-side compatibility shim for ChatGPT code-interpreter paths. Models habitually write to /mnt/data (or /mnt/outputs, /home/sandbox, -/workspace), none of which exist in the Studio sandbox. This module sits on the +/workspace), none of which exist in the Unsloth sandbox. This module sits on the sandbox subprocess PYTHONPATH (see ``tools._build_safe_env``), so it loads at interpreter startup in every sandboxed ``python`` run and any Python the ``terminal`` tool launches. diff --git a/studio/backend/core/inference/tool_loop_controller.py b/studio/backend/core/inference/tool_loop_controller.py index f7ed450d11..61643b5795 100644 --- a/studio/backend/core/inference/tool_loop_controller.py +++ b/studio/backend/core/inference/tool_loop_controller.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 -"""Shared controller state for Studio local agentic tool loops. +"""Shared controller state for Unsloth local agentic tool loops. This module is intentionally dependency-light: it owns only per-response ledger state and value objects used by the GGUF and safetensors loops. diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py index 5fd57e1b2c..bc9ffe85c2 100644 --- a/studio/backend/core/inference/tools.py +++ b/studio/backend/core/inference/tools.py @@ -2502,7 +2502,7 @@ def _build_safe_env(workdir: str) -> dict[str, str]: shim directory. """ # Start from the running interpreter's dir so 'python'/'pip' resolve to the - # same environment the Studio server runs in. + # same environment the Unsloth server runs in. exe_dir = os.path.dirname(sys.executable) path_entries = [exe_dir] if exe_dir else [] @@ -2792,7 +2792,7 @@ def _bypass_preexec(): """Minimal pre-exec for bypass exec: os.setsid() only. Required, not a restriction: _kill_process_tree does killpg(getpgid(child)), - so without a new session a timeout/cancel would kill the Studio server too. + so without a new session a timeout/cancel would kill the Unsloth server too. """ try: os.setsid() @@ -2800,13 +2800,13 @@ def _bypass_preexec(): pass -# Hardening the Studio parent is done once (PR_SET_DUMPABLE is process-global +# Hardening the Unsloth parent is done once (PR_SET_DUMPABLE is process-global # and sticky); guarded so repeated bypass calls do not re-issue the prctl. _parent_proc_hardened = False def _harden_parent_against_proc_env_leak() -> bool: - """Make the Studio process's /proc//environ unreadable to its children. + """Make the Unsloth process's /proc//environ unreadable to its children. Stripping the child env is not enough on Linux: a bypassed same-UID child can read /proc//environ to recover the parent's unfiltered @@ -5482,7 +5482,7 @@ def _truncate(text: str, limit: int = _MAX_OUTPUT_CHARS) -> str: # ChatGPT code-interpreter path conventions models write out of habit; none -# exist in the Studio sandbox, so a failure on one earns the retry hint. +# exist in the Unsloth sandbox, so a failure on one earns the retry hint. _MISSING_PATH_PREFIXES = ( "/mnt/data", "/mnt/outputs", @@ -5688,7 +5688,7 @@ def _python_exec( # Close the /proc//environ secret-recovery path first; if it # cannot be applied, fail closed rather than leak the parent environ. return ( - "Execution error: could not harden the Studio process against " + "Execution error: could not harden the Unsloth process against " "/proc environment reads; refusing bypass execution." ) @@ -5833,7 +5833,7 @@ def _bash_exec( # Close the /proc//environ secret-recovery path first; if it # cannot be applied, fail closed rather than leak the parent environ. return ( - "Execution error: could not harden the Studio process against " + "Execution error: could not harden the Unsloth process against " "/proc environment reads; refusing bypass execution." ) diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py index e4628dcea8..9f301ba37e 100644 --- a/studio/backend/core/inference/worker.py +++ b/studio/backend/core/inference/worker.py @@ -513,20 +513,25 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None: logger.info("Starting text generation for request_id=%s", request_id) - for cumulative_text in generator: - # cancel_event is an mp.Event — checked instantly, no queue polling. - if cancel_event.is_set(): - logger.info("Generation cancelled for request %s", request_id) - break + try: + for cumulative_text in generator: + # cancel_event is an mp.Event — checked instantly, no queue polling. + if cancel_event.is_set(): + logger.info("Generation cancelled for request %s", request_id) + break - _send_response( - resp_queue, - { - "type": "token", - "request_id": request_id, - "text": cumulative_text, - }, - ) + _send_response( + resp_queue, + { + "type": "token", + "request_id": request_id, + "text": cumulative_text, + }, + ) + finally: + close = getattr(generator, "close", None) + if callable(close): + close() _send_response( resp_queue, diff --git a/studio/backend/core/rag/captioner.py b/studio/backend/core/rag/captioner.py index 6d1512a770..8398506f21 100644 --- a/studio/backend/core/rag/captioner.py +++ b/studio/backend/core/rag/captioner.py @@ -6,7 +6,7 @@ Both turn pixels into indexable text and are a no-op (never raise) without a loaded vision model. They reuse the chat model's vision endpoint, so it must be served with ``--ubatch-size`` >= one image's tokens (some encoders, e.g. Gemma, attend -non-causally and abort otherwise); Studio's vision chat already requires this.""" +non-causally and abort otherwise); Unsloth's vision chat already requires this.""" from __future__ import annotations diff --git a/studio/backend/core/rag/config.py b/studio/backend/core/rag/config.py index 2de32a68e4..f54d795731 100644 --- a/studio/backend/core/rag/config.py +++ b/studio/backend/core/rag/config.py @@ -87,6 +87,22 @@ def _names_gguf(model: str) -> bool: return "gguf" in re.split(r"[^a-z0-9]+", model.lower()) +def gguf_repo_for_embedding_model(model: str) -> str: + """GGUF repo for ``model``, honoring an explicit companion override.""" + if "RAG_EMBED_GGUF_REPO" in os.environ: + return EMBED_GGUF_REPO + if model == DEFAULT_EMBEDDING_MODEL: + return EMBED_GGUF_REPO + if _names_gguf(model): + return model + return f"{model}-GGUF" + + +def default_gguf_repo() -> str: + """GGUF companion for the env/default embedding model.""" + return gguf_repo_for_embedding_model(EMBEDDING_MODEL) + + def effective_gguf_repo() -> str: """GGUF repo for the llama-server backend, tracking the effective model. @@ -95,14 +111,7 @@ def effective_gguf_repo() -> str: ``-GGUF`` companion repo (the unsloth convention the default pair follows), or is used as-is when it already names a GGUF repo. """ - if "RAG_EMBED_GGUF_REPO" in os.environ: - return EMBED_GGUF_REPO - model = effective_embedding_model() - if model == DEFAULT_EMBEDDING_MODEL: - return EMBED_GGUF_REPO - if _names_gguf(model): - return model - return f"{model}-GGUF" + return gguf_repo_for_embedding_model(effective_embedding_model()) # llama-server backend only. F16 over Q8_0: faster (no per-block dequant for this diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py index 46a282c939..b141e59422 100644 --- a/studio/backend/core/rag/embed_llama_server.py +++ b/studio/backend/core/rag/embed_llama_server.py @@ -10,7 +10,7 @@ Opt-in (``RAG_EMBED_BACKEND=llama-server``). Runs a dedicated Device is ``auto`` (GPU when present, else CPU, falling back to CPU if a GPU start fails); ``RAG_EMBED_DEVICE`` forces it. We call only llama_cpp's *static* helpers (no torch), copying the instance-coupled bits locally, since constructing a -``LlamaCppBackend`` runs an ``__init__`` reaper that kills any Studio llama-server +``LlamaCppBackend`` runs an ``__init__`` reaper that kills any Unsloth llama-server -- so each request re-spawns ours if it died (self-heal). """ diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py index b0ecedd593..15be7f1249 100644 --- a/studio/backend/core/rag/embeddings.py +++ b/studio/backend/core/rag/embeddings.py @@ -39,7 +39,7 @@ _model = None _name: str | None = None -# Studio device -> torch device string. Apple has no torch device -> CPU. +# Unsloth device -> torch device string. Apple has no torch device -> CPU. _TORCH_DEVICE = {DeviceType.CUDA: "cuda", DeviceType.XPU: "xpu"} diff --git a/studio/backend/core/training/resume.py b/studio/backend/core/training/resume.py index 2a4a198610..bbd9a895ab 100644 --- a/studio/backend/core/training/resume.py +++ b/studio/backend/core/training/resume.py @@ -53,7 +53,7 @@ def get_resume_checkpoint_path(path_value: str) -> Optional[str]: def normalize_resume_output_dir(path_value: str) -> str: path = resolve_output_dir(path_value) if not _is_under_outputs(path): - raise ValueError("Resume checkpoint must be inside Studio outputs.") + raise ValueError("Resume checkpoint must be inside Unsloth outputs.") return str(path) diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py index 883a535a89..26720865f4 100644 --- a/studio/backend/core/training/trainer.py +++ b/studio/backend/core/training/trainer.py @@ -797,7 +797,7 @@ class UnslothTrainer: ) logger.info("Loaded text model") - raise_if_offloaded(self.model, device_map, "Studio training") + raise_if_offloaded(self.model, device_map, "Unsloth training") if self.should_stop: return False diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py index 38f6b92f6d..b407ba39a5 100644 --- a/studio/backend/core/training/training.py +++ b/studio/backend/core/training/training.py @@ -140,7 +140,7 @@ def should_use_mlx_training_backend(*, device: Optional[Any] = None) -> bool: def _build_training_worker_config(values: dict[str, Any]) -> dict[str, Any]: - """Build the normalized worker config shared by Studio and the CLI adapter.""" + """Build the normalized worker config shared by Unsloth and the CLI adapter.""" config = { "model_name": values["model_name"], "project_name": values.get("project_name"), @@ -307,7 +307,7 @@ PLOT_HEIGHT = 3.5 @dataclass class TrainingProgress: - """Shared training progress payload for Studio and backend-aware trainers.""" + """Shared training progress payload for Unsloth and backend-aware trainers.""" epoch: float = 0 step: int = 0 @@ -328,7 +328,7 @@ class TrainingProgress: class _MLXTrainerAdapter: - """Adapts the legacy UnslothTrainer API to the shared Studio MLX worker path.""" + """Adapts the legacy UnslothTrainer API to the shared Unsloth MLX worker path.""" def __init__(self): self.model = None diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py index c52adbe8fa..111f4fdd0f 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -1100,7 +1100,7 @@ _MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE = {} def _mlx_vlm_resized_image_layout(processor = None) -> str | None: - """Return the numpy image layout expected after Studio-side VLM resizing.""" + """Return the numpy image layout expected after Unsloth-side VLM resizing.""" image_processor = getattr(processor, "image_processor", None) if image_processor is None: return None @@ -1257,7 +1257,7 @@ _MLX_STUDIO_LR_SCHEDULERS = {"linear", "cosine", "constant"} # Fallback alias map mirroring unsloth_zoo._normalize_mlx_optimizer_name, used -# only when mlx (Apple Silicon) is not importable so Studio config validation +# only when mlx (Apple Silicon) is not importable so Unsloth config validation # still works on non-MLX hosts. The zoo function stays the source of truth. _MLX_STUDIO_ADAMW_ALIASES = frozenset( ( @@ -1309,7 +1309,7 @@ def _normalize_mlx_studio_scheduler(value): def _resolve_mlx_local_dataset_files(file_paths: list) -> list[str]: - """Resolve CLI paths and Studio local dataset uploads without importing the GPU trainer.""" + """Resolve CLI paths and Unsloth local dataset uploads without importing the GPU trainer.""" from utils.paths import resolve_dataset_path all_files: list[str] = [] @@ -1912,7 +1912,7 @@ def _run_mlx_training(event_queue, stop_queue, config): if "max_grad_leaf_norm" in _supported_fields: mlx_config_kwargs["max_grad_leaf_norm"] = max_grad_leaf_norm if "append_eos" in _supported_fields: - # Studio SFT formatting owns rendered examples; raw/CPT text still + # Unsloth SFT formatting owns rendered examples; raw/CPT text still # needs MLX to append EOS like the CUDA raw-text path. mlx_config_kwargs["append_eos"] = bool(raw_text_mode) @@ -2121,7 +2121,7 @@ def run_mlx_training_process( config: dict, transformers_activated: bool = False, ) -> None: - """MLX worker entrypoint shared by Studio subprocesses and the CLI adapter.""" + """MLX worker entrypoint shared by Unsloth subprocesses and the CLI adapter.""" model_name = config["model_name"] backend_path = str(Path(__file__).resolve().parent.parent.parent) @@ -2780,7 +2780,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) -> ) # Unified Windows APUs: the WDDM budget is user-raisable, but # nothing on the box says so -- users see "48 GB VRAM" on a - # 96 GB machine and assume a Studio bug. Say where the limit + # 96 GB machine and assume an Unsloth bug. Say where the limit # comes from and how to raise it. if _is_unified and sys.platform == "win32": try: diff --git a/studio/backend/hub/services/download_lifecycle.py b/studio/backend/hub/services/download_lifecycle.py index e5d48872c1..23f8c7c911 100644 --- a/studio/backend/hub/services/download_lifecycle.py +++ b/studio/backend/hub/services/download_lifecycle.py @@ -76,7 +76,7 @@ def spawn_worker( env["HF_HUB_DISABLE_PROGRESS_BARS"] = "1" env["HF_HUB_DISABLE_TELEMETRY"] = "1" env["HF_HUB_DISABLE_XET"] = "0" if use_xet else "1" - # No token in Studio settings: fall back to the backend's own HF_TOKEN so + # No token in Unsloth settings: fall back to the backend's own HF_TOKEN so # private repos stay downloadable (needed while inkling repos are private). if not hf_token: hf_token = os.environ.get("HF_TOKEN") or None diff --git a/studio/backend/hub/services/models/cache_inventory.py b/studio/backend/hub/services/models/cache_inventory.py index 1f38af9381..54a25482f2 100644 --- a/studio/backend/hub/services/models/cache_inventory.py +++ b/studio/backend/hub/services/models/cache_inventory.py @@ -37,6 +37,13 @@ from hub.services.models.common import ( _runtime_for_format, ) +# Imported at module scope (not inside the per-repo scan loop) so a broken +# import surfaces at startup instead of silently emptying the inventory: the +# scan loop swallows per-repo exceptions and would drop every repo. Lives under +# ``utils`` (not ``utils.models``) to avoid the eager model-config/checkpoint +# imports in ``utils/models/__init__.py``. +from utils.hidden_models import is_hidden_model + logger = get_logger(__name__) _repo_size_cache: "OrderedDict[tuple[str, str, str], tuple[int, frozenset[str], float]]" = ( @@ -243,6 +250,13 @@ def invalidate_hf_cache_scans() -> None: hf_cache_scan.invalidate_hf_cache_scans() +def _is_hidden_infra_repo(*values: str | None) -> bool: + """True for infra-only repos (the RAG embedder and the llama.cpp install + validation probe) that are cached as a side effect of Studio itself and are + not usable chat models.""" + return is_hidden_model(*values) + + def _scan_cached_gguf() -> list[dict]: """Synchronous HF-cache disk walk for GGUF repos; runs in a worker thread.""" cache_scans = all_hf_cache_scans() @@ -254,13 +268,24 @@ def _scan_cached_gguf() -> list[dict]: if str(repo_info.repo_type) != "model": continue repo_id = repo_info.repo_id + repo_path = Path(repo_info.repo_path) + snapshot_path = _cached_model_snapshot_path(repo_path) total_size = _repo_gguf_size_bytes(repo_info) has_variant_state, variant_state_size = _gguf_variant_state_summary(repo_id) + is_hidden_infra = _is_hidden_infra_repo( + repo_id, + str(repo_path), + str(snapshot_path) if snapshot_path is not None else None, + ) + # Hide infra repos unless the user downloaded a variant via + # the Hub; variant state only exists for user downloads. + if is_hidden_infra and not has_variant_state: + continue if total_size == 0 and not has_variant_state: continue partial = hf_cache_scan.is_gguf_repo_partial( repo_id, - Path(repo_info.repo_path), + repo_path, ) if total_size == 0 and not partial: continue @@ -283,6 +308,9 @@ def _scan_cached_gguf() -> list[dict]: requires_variant = True, ) ) + # Visible infra variants remain management-only. + if is_hidden_infra: + row["capabilities"]["can_chat"] = False if _prefer_cache_row(row, existing): seen_lower[key] = row except Exception as e: @@ -475,6 +503,15 @@ def _scan_cached_models() -> list[dict]: if str(repo_info.repo_type) != "model": continue repo_id = repo_info.repo_id + repo_path = Path(repo_info.repo_path) + snapshot_path = _cached_model_snapshot_path(repo_path) + # The non-GGUF embedder has no variant downloads; always hide. + if _is_hidden_infra_repo( + repo_id, + str(repo_path), + str(snapshot_path) if snapshot_path is not None else None, + ): + continue has_main_gguf = _repo_has_gguf_files(repo_info) payload = _repo_non_gguf_model_payload(repo_info) if payload.size_bytes == 0: @@ -486,7 +523,6 @@ def _scan_cached_models() -> list[dict]: continue key = repo_id.lower() existing = seen_lower.get(key) - repo_path = Path(repo_info.repo_path) snapshot_partial = hf_cache_scan.is_snapshot_partial( "model", repo_id, diff --git a/studio/backend/hub/services/models/folder_browser.py b/studio/backend/hub/services/models/folder_browser.py index d56b62c318..7d9c3ac665 100644 --- a/studio/backend/hub/services/models/folder_browser.py +++ b/studio/backend/hub/services/models/folder_browser.py @@ -165,7 +165,7 @@ def _looks_like_model_dir(directory: Path) -> bool: def _build_browse_allowlist( media_roots: Optional[list[Path]] = None, drive_roots: Optional[list[Path]] = None ) -> list[Path]: - """Root directories the browser may walk (also seeds the suggestion chips): HOME, resolved HF cache dirs, Studio outputs/exports/root, registered scan folders, and well-known local-LLM dirs. Each is added only if it resolves to a real directory so the sandbox has no dead boundary. + """Root directories the browser may walk (also seeds the suggestion chips): HOME, resolved HF cache dirs, Unsloth outputs/exports/root, registered scan folders, and well-known local-LLM dirs. Each is added only if it resolves to a real directory so the sandbox has no dead boundary. *media_roots* / *drive_roots* let the caller pass already-probed removable-media and Windows drive roots so they aren't scanned again (a diff --git a/studio/backend/hub/services/models/local_inventory.py b/studio/backend/hub/services/models/local_inventory.py index a3782efead..b34532fa35 100644 --- a/studio/backend/hub/services/models/local_inventory.py +++ b/studio/backend/hub/services/models/local_inventory.py @@ -36,6 +36,7 @@ from hub.utils.paths import ( ) from hub.services.models import common as model_common from hub.services.models.ollama import scan_ollama_dir +from utils.hidden_models import is_hidden_model logger = get_logger(__name__) _MAX_MODELS_PER_CUSTOM_FOLDER = 200 @@ -623,6 +624,20 @@ def _dedupe_local_models(local_models: List[LocalModelInfo]) -> list[LocalModelI ) +def _filter_hidden_models(local_models: List[LocalModelInfo]) -> list[LocalModelInfo]: + """Remove infrastructure-only models from the shared local inventory.""" + visible: list[LocalModelInfo] = [] + for model in local_models: + resolved_cache_path = ( + hf_cache_scan.resolve_hf_cache_realpath(Path(model.path)) + if model.source == "hf_cache" + else None + ) + if not is_hidden_model(model.id, model.model_id, model.path, resolved_cache_path): + visible.append(model) + return visible + + async def list_local_models_response(models_dir: str = "./models") -> LocalModelListResponse: """List local model candidates from every supported on-device source.""" hf_cache_dir = _resolve_hf_cache_dir() @@ -653,7 +668,7 @@ async def list_local_models_response(models_dir: str = "./models") -> LocalModel ollama_dirs, ) local_models += await _collect_models_from_custom_folders() - models = _dedupe_local_models(local_models) + models = _dedupe_local_models(_filter_hidden_models(local_models)) return LocalModelListResponse( models_dir = str(models_root), diff --git a/studio/backend/hub/services/models/ollama.py b/studio/backend/hub/services/models/ollama.py index 96a4114620..2ccdbb44f1 100644 --- a/studio/backend/hub/services/models/ollama.py +++ b/studio/backend/hub/services/models/ollama.py @@ -85,7 +85,7 @@ def _contained_link_path(link_dir: Path, link_name: str) -> Optional[Path]: def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]: - """Writable directory for Ollama ``.gguf`` symlinks. Prefers ``/.studio_links/`` next to the blobs; falls back to Studio's cache (read-only system installs), then the temp dir (sandboxed installs).""" + """Writable directory for Ollama ``.gguf`` symlinks. Prefers ``/.studio_links/`` next to the blobs; falls back to Unsloth's cache (read-only system installs), then the temp dir (sandboxed installs).""" def _ensure_writable_dir(path: Path) -> Optional[Path]: try: diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py index 2c33e09b2b..693d945ee1 100644 --- a/studio/backend/hub/tests/test_model_services.py +++ b/studio/backend/hub/tests/test_model_services.py @@ -439,6 +439,287 @@ def test_cached_gguf_scan_includes_variant_state_without_completed_gguf(monkeypa assert row["capabilities"]["requires_variant"] is True +def test_cached_gguf_scan_hides_infra_repos_without_user_downloads(monkeypatch, tmp_path): + probe = _repo( + "ggml-org/models", + [_file("tinyllamas/stories260K.gguf", 1_200_000)], + tmp_path / "probe", + ) + embedder = _repo( + "unsloth/bge-small-en-v1.5-GGUF", + [_file("bge-small-en-v1.5-f16.gguf", 60_000_000)], + tmp_path / "embedder", + ) + chat = _repo("Org/Chat-GGUF", [_file("Q4_K_M.gguf", 100)], tmp_path / "chat") + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [probe, embedder, chat])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_gguf_repo_partial", + lambda _repo_id, _path: False, + ) + + result = {"cached": cache_inventory._scan_cached_gguf()} + + assert [row["repo_id"] for row in result["cached"]] == ["Org/Chat-GGUF"] + + +def test_cached_gguf_scan_keeps_infra_repo_with_user_downloaded_variant(monkeypatch, tmp_path): + monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state") + embedder = _repo( + "unsloth/bge-small-en-v1.5-GGUF", + [ + _file("bge-small-en-v1.5-f16.gguf", 60_000_000), + _file("bge-small-en-v1.5-Q8_0.gguf", 35_000_000), + ], + tmp_path / "embedder", + ) + # Variant manifests only exist for user Hub downloads, not auto-downloads. + assert download_manifest.write_manifest( + "model", + "unsloth/bge-small-en-v1.5-GGUF", + "Q8_0", + [download_manifest.ExpectedFile(path = "bge-small-en-v1.5-Q8_0.gguf", size = 35_000_000)], + "http", + ) + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [embedder])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_gguf_repo_partial", + lambda _repo_id, _path: False, + ) + + result = {"cached": cache_inventory._scan_cached_gguf()} + + assert [row["repo_id"] for row in result["cached"]] == ["unsloth/bge-small-en-v1.5-GGUF"] + assert result["cached"][0]["capabilities"]["can_chat"] is False + + +def test_cached_models_scan_hides_non_gguf_embedder(monkeypatch, tmp_path): + embedder_path = tmp_path / "hub" / "models--unsloth--bge-small-en-v1.5" + embedder_path.mkdir(parents = True) + embedder = _repo( + "unsloth/bge-small-en-v1.5", + [_file("config.json", 12), _file("model.safetensors", 130_000_000)], + embedder_path, + ) + chat_path = tmp_path / "hub" / "models--Org--Chat" + chat_path.mkdir(parents = True) + chat = _repo( + "Org/Chat", + [_file("config.json", 12), _file("model.safetensors", 100)], + chat_path, + ) + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [embedder, chat])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_snapshot_partial", + lambda _kind, _repo_id, _path: False, + ) + + result = {"cached": cache_inventory._scan_cached_models()} + + assert [row["repo_id"] for row in result["cached"]] == ["Org/Chat"] + + +def test_cached_scans_hide_embedders_configured_by_cache_path(monkeypatch, tmp_path): + from core.rag import config as rag_config + + gguf_path = tmp_path / "hub" / "models--Org--PathEmbedder-GGUF" + gguf_path.mkdir(parents = True) + gguf = _repo( + "Org/PathEmbedder-GGUF", + [_file("model-F16.gguf", 60_000_000)], + gguf_path, + ) + model_path = tmp_path / "hub" / "models--Org--PathEmbedder" + model_path.mkdir(parents = True) + model = _repo( + "Org/PathEmbedder", + [_file("config.json", 12), _file("model.safetensors", 130_000_000)], + model_path, + ) + monkeypatch.setattr( + rag_config, + "effective_embedding_model", + lambda: str(model_path), + ) + monkeypatch.setattr( + rag_config, + "effective_gguf_repo", + lambda: str(gguf_path), + ) + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [gguf, model])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_gguf_repo_partial", + lambda _repo_id, _path: False, + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_snapshot_partial", + lambda _kind, _repo_id, _path: False, + ) + + assert cache_inventory._scan_cached_gguf() == [] + assert cache_inventory._scan_cached_models() == [] + + +def test_cached_scans_hide_embedders_configured_by_snapshot_path(monkeypatch, tmp_path): + from core.rag import config as rag_config + + gguf_path = tmp_path / "hub" / "models--Org--SnapshotEmbedder-GGUF" + gguf_snapshot = gguf_path / "snapshots" / "gguf-revision" + gguf_snapshot.mkdir(parents = True) + gguf = _repo( + "Org/SnapshotEmbedder-GGUF", + [_file("model-F16.gguf", 60_000_000)], + gguf_path, + ) + model_path = tmp_path / "hub" / "models--Org--SnapshotEmbedder" + model_snapshot = model_path / "snapshots" / "model-revision" + model_snapshot.mkdir(parents = True) + model = _repo( + "Org/SnapshotEmbedder", + [_file("config.json", 12), _file("model.safetensors", 130_000_000)], + model_path, + ) + monkeypatch.setattr( + rag_config, + "effective_embedding_model", + lambda: str(model_snapshot), + ) + monkeypatch.setattr( + rag_config, + "effective_gguf_repo", + lambda: str(gguf_snapshot), + ) + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [gguf, model])], + ) + + def _resolve_snapshot(repo_path): + return str( + { + gguf_path: gguf_snapshot, + model_path: model_snapshot, + }.get(Path(repo_path), Path(repo_path)) + ) + + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "resolve_hf_cache_realpath", + _resolve_snapshot, + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_gguf_repo_partial", + lambda _repo_id, _path: False, + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_snapshot_partial", + lambda _kind, _repo_id, _path: False, + ) + + assert cache_inventory._scan_cached_gguf() == [] + assert cache_inventory._scan_cached_models() == [] + + +def test_cached_models_scan_keeps_unrelated_repo_with_custom_generic_embedder( + monkeypatch, tmp_path +): + # A custom embedder with a generic basename ("org/model") must be hidden by + # EXACT repo-id match only. An unrelated cached chat model whose id merely + # contains "model" (e.g. "user/model-chat") must stay on device: substring + # basename matching used to drop real chat models from the inventory. + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF") + + def _model_repo(repo_id: str): + path = tmp_path / "hub" / f"models--{repo_id.replace('/', '--')}" + path.mkdir(parents = True) + return _repo( + repo_id, + [_file("config.json", 12), _file("model.safetensors", 100)], + path, + ) + + embedder = _model_repo("org/model") + chat = _model_repo("user/model-chat") + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [embedder, chat])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_snapshot_partial", + lambda _kind, _repo_id, _path: False, + ) + + result = {"cached": cache_inventory._scan_cached_models()} + + assert [row["repo_id"] for row in result["cached"]] == ["user/model-chat"] + + +def test_cached_scans_hide_stale_default_embedder_after_custom_setting(monkeypatch, tmp_path): + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF") + + gguf = _repo( + "unsloth/bge-small-en-v1.5-GGUF", + [_file("bge-small-en-v1.5-f16.gguf", 60_000_000)], + tmp_path / "default-gguf", + ) + weights_path = tmp_path / "hub" / "models--unsloth--bge-small-en-v1.5" + weights_path.mkdir(parents = True) + weights = _repo( + "unsloth/bge-small-en-v1.5", + [_file("config.json", 12), _file("model.safetensors", 130_000_000)], + weights_path, + ) + monkeypatch.setattr( + cache_inventory, + "all_hf_cache_scans", + lambda: [SimpleNamespace(repos = [gguf, weights])], + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_gguf_repo_partial", + lambda _repo_id, _path: False, + ) + monkeypatch.setattr( + cache_inventory.hf_cache_scan, + "is_snapshot_partial", + lambda _kind, _repo_id, _path: False, + ) + + assert cache_inventory._scan_cached_gguf() == [] + assert cache_inventory._scan_cached_models() == [] + + def test_gguf_variant_requirements_include_split_files_and_preferred_mmproj(): requirements = gguf_variants._build_gguf_variant_requirements( [ @@ -1610,6 +1891,63 @@ def test_hf_cache_scan_uses_gguf_partial_row_for_variant_state(monkeypatch, tmp_ assert rows[0].capabilities.requires_variant is True +def test_local_inventory_filters_custom_embedder_hf_cache_row(monkeypatch, tmp_path): + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/embedder") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF") + + def _row(repo_id: str): + repo_path = tmp_path / f"models--{repo_id.replace('/', '--')}" + return model_common._local_model_info( + scan_path = repo_path, + load_path = repo_path, + source = "hf_cache", + model_format = "safetensors", + model_id = repo_id, + ) + + rows = local_inventory._filter_hidden_models([_row("org/embedder"), _row("org/chat-model")]) + + assert [row.model_id for row in rows] == ["org/chat-model"] + + +def test_local_inventory_filters_embedder_configured_by_snapshot_path(monkeypatch, tmp_path): + from core.rag import config as rag_config + + embedder_path = tmp_path / "hub" / "models--org--embedder" + embedder_snapshot = embedder_path / "snapshots" / "revision" + embedder_snapshot.mkdir(parents = True) + chat_path = tmp_path / "hub" / "models--org--chat-model" + chat_path.mkdir(parents = True) + monkeypatch.setattr( + rag_config, + "effective_embedding_model", + lambda: str(embedder_snapshot), + ) + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF") + monkeypatch.setattr( + local_inventory.hf_cache_scan, + "resolve_hf_cache_realpath", + lambda path: str(embedder_snapshot) if Path(path) == embedder_path else str(path), + ) + + def _row(repo_id: str, repo_path: Path): + return model_common._local_model_info( + scan_path = repo_path, + load_path = repo_path, + source = "hf_cache", + model_format = "safetensors", + model_id = repo_id, + ) + + rows = local_inventory._filter_hidden_models( + [_row("org/embedder", embedder_path), _row("org/chat-model", chat_path)] + ) + + assert [row.model_id for row in rows] == ["org/chat-model"] + + def test_model_download_job_helpers_preserve_idle_shape(): key = downloads._download_job_key("Org/Model", None) status = downloads._job_status(key) diff --git a/studio/backend/hub/utils/state_dir.py b/studio/backend/hub/utils/state_dir.py index 183e934724..898c03c87d 100644 --- a/studio/backend/hub/utils/state_dir.py +++ b/studio/backend/hub/utils/state_dir.py @@ -3,7 +3,7 @@ """Filesystem layout for Hub download state. -State directory sits beside HF's cache (under Studio's own cache root) +State directory sits beside HF's cache (under Unsloth's own cache root) so it survives ``huggingface-cli delete-cache`` and any other HF-side cache lifecycle. Two subdirectories: diff --git a/studio/backend/main.py b/studio/backend/main.py index 6e16dc00ca..81d4c16e52 100644 --- a/studio/backend/main.py +++ b/studio/backend/main.py @@ -19,7 +19,7 @@ os.environ["PYTHONWARNINGS"] = "ignore" # Pin GPU index ordering to PCI bus id before any torch import creates a CUDA # context. Without this, torch/CUDA default to FASTEST_FIRST while nvidia-smi -# (and Studio's VRAM probes) use PCI-bus order, so a GPU index chosen from +# (and Unsloth's VRAM probes) use PCI-bus order, so a GPU index chosen from # nvidia-smi data can resolve to a different physical card via # CUDA_VISIBLE_DEVICES. setdefault so an explicit user override wins. See # utils/hardware/hardware.py for the full rationale; set here too so the entry @@ -93,7 +93,7 @@ if sys.platform == "win32": # ── Windows AMD ROCm: make hipInfo.exe resolvable for subprocess probes ── # bitsandbytes' get_rocm_gpu_arch() runs `hipinfo.exe` via PATH at import # time; the AMD torch wheel ships it in the venv Scripts dir, which is on - # PATH only when the venv is activated -- Studio launches python directly. + # PATH only when the venv is activated -- Unsloth launches python directly. # Without this, every bitsandbytes import logs a scary (but harmless) # "Could not detect ROCm GPU architecture: [WinError 2]" ERROR + WARNING. # Gated on the file existing: only AMD ROCm wheels ship hipInfo.exe, so @@ -252,7 +252,7 @@ def _read_studio_install_id() -> str: Returns "" when absent or not a 64-char lowercase-hex token; then /api/health emits "" and the launcher accepts any healthy backend. - Carries no install-path info (matters when Studio runs -H 0.0.0.0).""" + Carries no install-path info (matters when Unsloth runs -H 0.0.0.0).""" try: token = (_STUDIO_ROOT_RESOLVED / "share" / "studio_install_id").read_text().strip() except (OSError, ValueError): @@ -573,7 +573,7 @@ async def lifespan(app: FastAPI): print("DEFAULT ADMIN ACCOUNT CREATED") print(f" username: {storage.DEFAULT_ADMIN_USERNAME}") print(f" password saved to: {bootstrap_path}") - print(" Open the Studio UI to sign in and change it.") + print(" Open the Unsloth UI to sign in and change it.") print("=" * 60 + "\n") else: app.state.bootstrap_password = ( @@ -613,7 +613,7 @@ app = FastAPI( ) # The MCP surface is opt-in because it can start GPU jobs and write model -# artifacts. Mount it only when explicitly enabled by the Studio process. +# artifacts. Mount it only when explicitly enabled by the Unsloth process. if os.environ.get("UNSLOTH_STUDIO_ENABLE_MCP") == "1": from fastmcp.utilities.lifespan import combine_lifespans @@ -973,7 +973,7 @@ app.include_router(training_router, prefix = "/api/train", tags = ["training"]) app.include_router(models_router, prefix = "/api/models", tags = ["models"]) app.include_router(chat_history_router, prefix = "/api/chat", tags = ["chat"]) app.include_router(inference_router, prefix = "/api/inference", tags = ["inference"]) -# Studio-only inference endpoints (cancel, etc.) are NOT exposed on the /v1 +# Unsloth-only inference endpoints (cancel, etc.) are NOT exposed on the /v1 # OpenAI-compat prefix below. app.include_router(inference_studio_router, prefix = "/api/inference", tags = ["inference"]) @@ -1080,7 +1080,7 @@ def studio_install_source(_current_subject: str = Depends(get_current_subject)): @app.get("/api/studio/update-status") def studio_update_status(_current_subject: str = Depends(get_current_subject)): - """Return source-aware manual update status for browser-served Studio.""" + """Return source-aware manual update status for browser-served Unsloth.""" return get_studio_update_status(UNSLOTH_VERSION) @@ -1156,9 +1156,23 @@ def _get_cached_system_gpu_info(logger) -> dict[str, Any]: enriched_dev["vram_utilization_pct"] = util.get("vram_utilization_pct") enriched_devices.append(enriched_dev) + # Whether GGUF loads accept an explicit gpu_ids pick: /load and + # /validate 400 picks on XPU hosts (no visibility mask speaks torch-xpu + # ordinals) and on Vulkan-only builds (--device pins ggml's own + # ordinals), so the picker must not offer them. + try: + from core.inference.llama_cpp import LlamaCppBackend + from utils.hardware import DeviceType, get_device + gpu_ids_supported = ( + get_device() != DeviceType.XPU and not LlamaCppBackend._is_vulkan_backend() + ) + except Exception as e: + logger.debug(f"Could not resolve gpu_ids support: {e}") + gpu_ids_supported = True gpu_info = { "available": visibility_info.get("available", False), "devices": enriched_devices, + "gguf_gpu_ids_supported": gpu_ids_supported, } _system_gpu_cache = (time.monotonic(), gpu_info) return gpu_info diff --git a/studio/backend/mcp_server.py b/studio/backend/mcp_server.py index f837f46425..e93490411d 100644 --- a/studio/backend/mcp_server.py +++ b/studio/backend/mcp_server.py @@ -3,7 +3,7 @@ """Curated MCP tools for driving an Unsloth Studio instance. -The MCP surface deliberately wraps the existing Studio services instead of +The MCP surface deliberately wraps the existing Unsloth services instead of duplicating training or export logic. It is opt-in because several tools can start GPU work or write model artifacts. """ @@ -17,14 +17,14 @@ from fastmcp import FastMCP class BearerTokenMiddleware: - """Require an exact bearer token when Studio MCP is exposed remotely.""" + """Require an exact bearer token when Unsloth MCP is exposed remotely.""" def __init__(self, app: Any, token: str) -> None: if not token or not token.strip(): - raise ValueError("Studio MCP bearer token must be a non-empty value") + raise ValueError("Unsloth MCP bearer token must be a non-empty value") if not token.isascii(): # A non-ASCII token cannot be sent in an HTTP header; reject it here. - raise ValueError("Studio MCP bearer token must contain ASCII characters only") + raise ValueError("Unsloth MCP bearer token must contain ASCII characters only") self.app = app # Compare on raw header bytes: str hmac.compare_digest raises on non-ASCII # input, which would surface as a 500 instead of a clean 401. @@ -76,18 +76,18 @@ def _dump(value: Any) -> Any: def _clamp(value: int, low: int, high: int) -> int: """Clamp an MCP-supplied integer into an inclusive range. - MCP tools call the Studio route functions directly, which skips FastAPI's + MCP tools call the Unsloth route functions directly, which skips FastAPI's Query(ge=, le=) validation, so we re-apply the same bounds here. """ return max(low, min(value, high)) def create_studio_mcp() -> FastMCP: - """Create the Studio MCP server and register the high-value tools.""" + """Create the Unsloth MCP server and register the high-value tools.""" mcp = FastMCP( "Unsloth Studio", instructions = ( - "Use read tools to inspect the local Studio state before starting GPU work. " + "Use read tools to inspect the local Unsloth state before starting GPU work. " "Training and export tools can consume substantial VRAM and write files. " "Never expose tokens or local paths from tool results unless the user asks." ), @@ -116,7 +116,7 @@ def create_studio_mcp() -> FastMCP: @mcp.tool async def list_local_models(models_dir: str = "./models") -> dict[str, Any]: - """List local and cached models available to Studio.""" + """List local and cached models available to Unsloth.""" from routes.models import list_local_models as list_models return _dump(await list_models(models_dir = models_dir, current_subject = "mcp")) @@ -128,9 +128,9 @@ def create_studio_mcp() -> FastMCP: @mcp.tool async def start_training(config: dict[str, Any]) -> dict[str, Any]: - """Start a validated Studio training job from a TrainingStartRequest-shaped object. + """Start a validated Unsloth training job from a TrainingStartRequest-shaped object. - The config is validated by the same Pydantic model used by the Studio UI. + The config is validated by the same Pydantic model used by the Unsloth UI. Call get_training_status first and do not start work while another job runs. """ from models import TrainingStartRequest @@ -138,7 +138,7 @@ def create_studio_mcp() -> FastMCP: request = TrainingStartRequest.model_validate(config) # Pass via_api_key explicitly (a direct call leaves it a Depends object). - # MCP drives Studio like the UI session, so it coexists and frees VRAM. + # MCP drives Unsloth like the UI session, so it coexists and frees VRAM. return _dump(await start(request, current_subject = "mcp", via_api_key = False)) @mcp.tool @@ -159,7 +159,7 @@ def create_studio_mcp() -> FastMCP: @mcp.tool def validate_recipe(recipe: dict[str, Any]) -> dict[str, Any]: - """Validate a Data Recipe with the same validator used by Studio.""" + """Validate a Data Recipe with the same validator used by Unsloth.""" from models.data_recipe import RecipePayload from routes.data_recipe.validate import validate @@ -225,7 +225,7 @@ def create_studio_mcp() -> FastMCP: imatrix: bool = False, imatrix_path: str | None = None, ) -> dict[str, Any]: - """Export the loaded model to GGUF using Studio's existing path validation. + """Export the loaded model to GGUF using Unsloth's existing path validation. quantization_method may be a single method or a list to produce several GGUFs from one load. Pass hf_token when push_to_hub is set (the backend diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py index 3ae974448e..d51d35189b 100644 --- a/studio/backend/models/inference.py +++ b/studio/backend/models/inference.py @@ -64,7 +64,7 @@ class LoadRequest(BaseModel): ) gpu_ids: Optional[List[int]] = Field( None, - description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. Not supported for GGUF models.", + description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. For GGUF models the picked devices are pinned via CUDA/HIP_VISIBLE_DEVICES.", ) speculative_type: Optional[str] = Field( None, @@ -100,12 +100,72 @@ class LoadRequest(BaseModel): "No effect on a single GPU. Ignored for non-GGUF models." ), ) + gpu_memory_mode: Literal["auto", "manual"] = Field( + "auto", + description = ( + "GPU memory strategy for GGUF models. 'auto' (default): Unsloth " + "selects GPUs and caps context to fit VRAM. 'manual': you own the " + "offload. Leave gpu_layers at -1 (Auto) to hand memory management to " + "llama.cpp's --fit (no device masking, no context auto-reduce, no " + "gpu-layer/tensor-split planning); set gpu_layers >= 0 to pin layers " + "and n_cpu_moe yourself (--fit off), with tensor_parallel still " + "applying (split by free VRAM unless tensor_split is set, no planner). " + "Ignored for non-GGUF." + ), + ) + gpu_layers: int = Field( + -1, + ge = -1, + description = ( + "Manual mode only: number of layers to offload to the GPU " + "(--gpu-layers, with --fit off). A value >= the model's layer count " + "offloads all of them. -1 = Auto: hand layer + context sizing to " + "llama.cpp's --fit. Ignored unless gpu_memory_mode is 'manual'." + ), + ) + n_cpu_moe: int = Field( + 0, + ge = 0, + description = ( + "Manual mode only: keep the first N MoE expert layers on the CPU " + "(--n-cpu-moe) to save VRAM on MoE models. 0 = none, N = number of " + "MoE layers offloaded (the backend offsets past any leading dense " + "layers). Ignored unless gpu_memory_mode is 'manual' with gpu_layers >= 0." + ), + ) + tensor_split: Optional[List[float]] = Field( + None, + description = ( + "Manual mode only: relative share of the model per GPU (--tensor-split), " + "in the order of the GPUs in use, e.g. [2, 1] for 2:1. Omit it to let " + "llama.cpp use its default, which splits by free VRAM. Any list given is " + "passed through as-is, so send [1, 1] to force an even split. Ignored " + "unless gpu_memory_mode is 'manual' with gpu_layers >= 0." + ), + ) + + @field_validator("tensor_split") + @classmethod + def _reject_degenerate_tensor_split(cls, value: Optional[List[float]]) -> Optional[List[float]]: + # A negative / non-finite / all-zero split is silently dropped at launch + # (stored as None) yet still compared raw in the reload dedupe, so an + # identical Apply reloads forever. Reject it up front; [] = no split. + if not value: + return value + import math + + if any((not math.isfinite(v)) or v < 0 for v in value): + raise ValueError("tensor_split entries must be finite and non-negative") + if sum(value) <= 0: + raise ValueError("tensor_split must have a positive total") + return value + llama_extra_args: Optional[List[str]] = Field( None, description = ( "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, " + "Unsloth-managed flags (model identity, port, context length, GPU placement, " "auth, UI/server mode) are rejected. Ignored for non-GGUF models." ), ) @@ -133,6 +193,14 @@ class ValidateModelRequest(BaseModel): max_seq_length: int = Field(0, ge = 0, le = 1048576) load_in_4bit: bool = Field(True) gpu_ids: Optional[List[int]] = Field(None) + gpu_memory_mode: Literal["auto", "manual"] = Field( + "auto", + description = ( + "GGUF GPU-memory strategy intended for the follow-up load. Manual " + "placement bypasses the training coexistence estimate: Auto layers " + "delegate fitting to llama.cpp, while explicit layers are user-owned." + ), + ) include_context_length: bool = Field( False, description = "Also read the native context length from the local GGUF header. " @@ -151,13 +219,13 @@ class TransformersUpgradeInfo(BaseModel): ) supported_in_pypi: bool = Field( False, - description = "True if the latest PyPI release ships this model_type; Studio can " + description = "True if the latest PyPI release ships this model_type; Unsloth can " "install it into a persistent sidecar after user consent.", ) supported_in_main: bool = Field( False, description = "True if transformers GitHub main ships this model_type (dev-only; " - "not installable through Studio yet).", + "not installable through Unsloth yet).", ) @@ -188,6 +256,16 @@ class ValidateModelResponse(BaseModel): description = "Native training context length, read from the GGUF header when the file " "is already downloaded locally; None for non-GGUF, gated, or not-yet-downloaded models.", ) + layer_count: Optional[int] = Field( + None, + description = "Total layer count (GGUF block_count), the manual gpu-layers ceiling, read " + "from the header alongside context_length; None when not read.", + ) + moe_layer_count: Optional[int] = Field( + None, + description = "MoE expert-layer count (the manual --n-cpu-moe ceiling), read from the GGUF " + "header alongside context_length; 0 for dense models, None when not read.", + ) # Additive fields; the consuming consent dialog ships in a follow-up frontend PR. requires_transformers_upgrade: bool = Field( False, @@ -333,6 +411,34 @@ class LoadResponse(BaseModel): False, description = "Whether tensor-parallel split (--split-mode tensor) is active.", ) + gpu_memory_mode: Literal["auto", "manual"] = Field( + "auto", + description = "Active GPU memory strategy ('auto' or 'manual').", + ) + gpu_layers: int = Field( + -1, + description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).", + ) + n_cpu_moe: int = Field( + 0, + description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.", + ) + tensor_split: Optional[List[float]] = Field( + None, + description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).", + ) + n_layers: Optional[int] = Field( + None, + description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.", + ) + n_moe_layers: int = Field( + 0, + description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.", + ) + gpu_ids: Optional[List[int]] = Field( + None, + description = "Physical GPU indices the model is pinned to, or None for automatic selection.", + ) class UnloadResponse(BaseModel): @@ -461,6 +567,42 @@ class InferenceStatusResponse(BaseModel): False, description = "Whether tensor-parallel split (--split-mode tensor) is active.", ) + gpu_memory_mode: Literal["auto", "manual"] = Field( + "auto", + description = "Active GPU memory strategy ('auto' or 'manual').", + ) + gpu_layers: int = Field( + -1, + description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).", + ) + n_cpu_moe: int = Field( + 0, + description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.", + ) + tensor_split: Optional[List[float]] = Field( + None, + description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).", + ) + requested_context_length: Optional[int] = Field( + None, + description = ( + "The n_ctx the active GGUF load was invoked with (0 = Auto). Lets the " + "UI re-seed a Manual + Auto-layers context pin on hydration, where " + "context_length only exposes the resolved value. None for non-GGUF." + ), + ) + n_layers: Optional[int] = Field( + None, + description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.", + ) + n_moe_layers: int = Field( + 0, + description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.", + ) + gpu_ids: Optional[List[int]] = Field( + None, + description = "Physical GPU indices the model is pinned to, or None for automatic selection.", + ) llama_cpp_supports_mtp: bool = Field( True, description = ( @@ -533,7 +675,7 @@ class ImageContentPart(BaseModel): class InputDocumentContentPart(BaseModel): """Document (PDF / file) content part in a multimodal message. - Studio-normalised shape (file_data or file_url, plus optional filename/media_type). + Unsloth-normalised shape (file_data or file_url, plus optional filename/media_type). Mapped onto Anthropic ``document`` / OpenAI ``input_file`` for vision providers; dropped for non-vision providers. """ @@ -689,7 +831,7 @@ class ThinkingConfig(BaseModel): """Anthropic-compatible thinking/reasoning configuration. Use type='disabled' to turn off thinking, or type='enabled' to turn it on. Only type is read; extra fields (e.g. budget_tokens) are ignored, since - Studio sets provider thinking budgets itself. + Unsloth sets provider thinking budgets itself. """ type: Literal["disabled", "enabled"] = "disabled" @@ -748,7 +890,7 @@ class ChatCompletionRequest(BaseModel): None, description = ( "OpenAI function-tool definitions. When provided without `enable_tools=true`, " - "Studio forwards the tools to the backend so the model returns structured " + "Unsloth forwards the tools to the backend so the model returns structured " "tool_calls for the client to execute (standard OpenAI function calling)." ), ) @@ -1160,7 +1302,7 @@ class ChatCompletionRequest(BaseModel): and (self.enable_tools is True or bool(self.mcp_enabled)) ): # "Ask" gates every call, so a direct API caller that omits the legacy - # confirm flag must still hit the confirmation gate for Studio's own + # confirm flag must still hit the confirmation gate for Unsloth's own # tool loop. An explicit confirm_tool_calls=False wins over the mode # (mirrors _permission_mode_confirm and the Anthropic pre-switch guard), # so only self-enable when the flag is unset. Only self-enable when that @@ -1168,7 +1310,7 @@ class ChatCompletionRequest(BaseModel): # (enable_tools / mcp_enabled) -- the router enters the loop on those # signals, not on enabled_tools alone (which merely filters which tools # run). A plain client-tool passthrough (client-supplied `tools` that - # Studio does not execute) must route verbatim, and external-provider + # Unsloth does not execute) must route verbatim, and external-provider # routing rejects confirm_tool_calls with tools, so skip the fold there. # # "auto" is deliberately NOT folded: it only prompts for a call the diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py index ff815a2fa9..0b50f63b95 100644 --- a/studio/backend/models/training.py +++ b/studio/backend/models/training.py @@ -446,7 +446,7 @@ class TrainingStartRequest(BaseModel): random_seed: int = Field( 3407, description = ( - "Random seed; matches the Studio backend / MLX worker default " + "Random seed; matches the Unsloth backend / MLX worker default " "and unsloth's historical recommended value." ), ) diff --git a/studio/backend/plugins/data-designer-github-repo-seed/README.md b/studio/backend/plugins/data-designer-github-repo-seed/README.md index 346d94b305..44519496f5 100644 --- a/studio/backend/plugins/data-designer-github-repo-seed/README.md +++ b/studio/backend/plugins/data-designer-github-repo-seed/README.md @@ -4,7 +4,7 @@ A Data Designer seed-reader plugin for **Unsloth Studio** that scrapes real GitHub data (issues, pull requests, commits) from one or more repositories and hands it to the recipe pipeline as a seed dataset. -Designed to ship with Studio as a default seed source so any user with a +Designed to ship with Unsloth as a default seed source so any user with a GitHub token can build training datasets straight from live repos. ## What it does @@ -64,7 +64,7 @@ sleeps until reset when the budget drops below a safety threshold. ## Install -Shipped as a default Studio plugin. For development: +Shipped as a default Unsloth plugin. For development: ```bash pip install -e . diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py index 62ecb2e280..d4d46da370 100644 --- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py +++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py @@ -3,4 +3,4 @@ # Intentionally empty. Data-designer loads submodules lazily via qualified names # in plugin.py, so importing this package must not touch data_designer.engine.* -# during Studio bootstrap (circular import). +# during Unsloth bootstrap (circular import). diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py index 637193e8b3..1af8133cc5 100644 --- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py +++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.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 -"""Multi-repo GitHub scraper for the Studio seed plugin. +"""Multi-repo GitHub scraper for the Unsloth seed plugin. Drives the GraphQL scraper in `scraper_impl/` per repo, capped via trial_limits to stop at `limit` items per resource. Then reads the per-resource JSONL shards diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt index 5830a47789..3361af50dd 100644 --- a/studio/backend/requirements/extras-no-deps.txt +++ b/studio/backend/requirements/extras-no-deps.txt @@ -5,7 +5,7 @@ julius torchcodec==0.10.0 snac -# peft 0.19.0 causes export subprocess shutdown issues in Studio; +# peft 0.19.0 causes export subprocess shutdown issues in Unsloth; # installing with --no-deps to avoid pulling in torch>=0.11.0 peft==0.18.1 diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt index de321f80ed..378fb33a60 100644 --- a/studio/backend/requirements/no-torch-runtime.txt +++ b/studio/backend/requirements/no-torch-runtime.txt @@ -70,7 +70,7 @@ cut_cross_entropy pillow # RAG store + document parsing, mirroring studio.txt. Pinned here because -# this file installs --no-deps; without them Studio runs with RAG disabled. +# this file installs --no-deps; without them Unsloth runs with RAG disabled. sqlite-vec==0.1.9 pymupdf==1.27.2.3 # 0.3.x keeps pymupdf-layout (which pulls onnxruntime) an optional extra; the diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt index 0ed2bf8b26..0a5619924a 100644 --- a/studio/backend/requirements/single-env/constraints.txt +++ b/studio/backend/requirements/single-env/constraints.txt @@ -4,7 +4,7 @@ transformers==4.57.6 trl==0.23.1 huggingface-hub==0.36.2 -# Studio stack +# Unsloth stack datasets==4.3.0 pyarrow==23.0.1 diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt index 6f4a5c3292..0c7503a5ca 100644 --- a/studio/backend/requirements/studio.txt +++ b/studio/backend/requirements/studio.txt @@ -1,4 +1,4 @@ -# Studio UI backend dependencies +# Unsloth UI backend dependencies typer fastapi uvicorn @@ -9,7 +9,7 @@ pandas nest_asyncio datasets==4.3.0 pyjwt -# gradio>=4.0.0 # 148 MB - Studio uses React + FastAPI, not Gradio +# gradio>=4.0.0 # 148 MB - Unsloth uses React + FastAPI, not Gradio huggingface-hub==0.36.2 structlog>=24.1.0 diceware diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py index c61c1a16e4..d779c8784e 100644 --- a/studio/backend/routes/auth.py +++ b/studio/backend/routes/auth.py @@ -338,11 +338,11 @@ def _clear_login_bucket(key: tuple[str, str]) -> None: # so FastAPI runs it in the threadpool rather than blocking the event loop. @router.get("/identity") def identity(nonce: str, request: Request) -> dict: - """Challenge-response proof this is the real local Studio: caller sends a nonce, + """Challenge-response proof this is the real local Unsloth: caller sends a nonce, gets HMAC(install identity secret, nonce, connection address + port). Unauthenticated and side-effect free; a process that can't read the same-user secret can't forge a proof, and binding to the address/port the connection - landed on stops a squatter relaying a proof from the real Studio elsewhere.""" + landed on stops a squatter relaying a proof from the real Unsloth elsewhere.""" try: raw = base64.urlsafe_b64decode(nonce) except Exception: diff --git a/studio/backend/routes/data_recipe/jobs.py b/studio/backend/routes/data_recipe/jobs.py index 59714380da..e870e8855e 100644 --- a/studio/backend/routes/data_recipe/jobs.py +++ b/studio/backend/routes/data_recipe/jobs.py @@ -37,7 +37,7 @@ def _resolve_local_v1_endpoint(request: Request) -> str: Resolution order: 1. ``app.state.server_port`` (run.py, post-bind) - survives proxies/tunnels. - 2. ``request.scope["server"]`` - when Studio starts outside ``run_server``. + 2. ``request.scope["server"]`` - when Unsloth starts outside ``run_server``. 3. parsed ``request.base_url`` - last resort for test fixtures. """ port: Any = getattr(request.app.state, "server_port", None) diff --git a/studio/backend/routes/datasets.py b/studio/backend/routes/datasets.py index 46319ca2ba..5456080f34 100644 --- a/studio/backend/routes/datasets.py +++ b/studio/backend/routes/datasets.py @@ -485,7 +485,7 @@ async def upload_dataset( # Stream to disk in chunks to avoid holding the whole file in memory. The # route-level cap gives a clear training-dataset error and avoids leaving - # oversized partial files in the Studio uploads directory. + # oversized partial files in the Unsloth uploads directory. upload_limit_bytes = get_upload_limit_bytes() total_bytes = 0 upload_complete = False diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py index 52ea1f86a3..136e4f7645 100644 --- a/studio/backend/routes/inference.py +++ b/studio/backend/routes/inference.py @@ -13,7 +13,7 @@ from pathlib import Path from fastapi import APIRouter, Depends, HTTPException, Request, status from fastapi.responses import StreamingResponse, JSONResponse, Response from starlette.requests import ClientDisconnect -from typing import Any, Callable, List, Optional, Union +from typing import Any, Callable, List, Literal, Optional, Union import json import httpx from loggers import get_logger @@ -92,7 +92,7 @@ def _mlx_distributed_launch_detected() -> bool: def _install_httpcore_asyncgen_silencer() -> None: """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13. - When Studio proxies a llama-server stream via httpx, the innermost + When Unsloth proxies a llama-server stream via httpx, the innermost ``HTTP11ConnectionByteStream.__aiter__`` async generator is finalised by the asyncgen GC hook on a task different from the one that opened it. Its ``aclose`` calls ``anyio.Lock.acquire`` → ``cancel_shielded_checkpoint``, @@ -229,14 +229,14 @@ def _friendly_upstream_error(text: str) -> str: parse grammar" / "failed to initialize samplers"). This surfaces to coding agents as a hard 400 on every tool-bearing turn. It is a llama-server limitation with some model/quant + tool-schema combinations, and recent llama.cpp builds handle the common - coding-agent tools, so point the user at updating Studio rather than the raw body. + coding-agent tools, so point the user at updating Unsloth rather than the raw body. """ lowered = text.lower() if "failed to parse grammar" in lowered or "failed to initialize samplers" in lowered: return ( "The model couldn't compile a tool-calling grammar for this request. This is a " "llama-server limitation with some model/quant and tool-schema combinations. " - "Update Studio (it installs the latest llama.cpp, which handles the common " + "Update Unsloth (it installs the latest llama.cpp, which handles the common " "coding-agent tools) or try a different GGUF model." ) return f"llama-server error: {text}" @@ -731,7 +731,7 @@ def _openai_passthrough_sse_line_terminal_state(raw_line: str) -> Optional[str]: Some llama-server builds can emit the logical final chunk (``finish_reason``) and optional usage chunk, then keep the HTTP stream open without sending the - OpenAI ``data: [DONE]`` sentinel. Classifying those chunks lets Studio close + OpenAI ``data: [DONE]`` sentinel. Classifying those chunks lets Unsloth close the client stream promptly while preserving an optional trailing usage chunk. """ if not raw_line.startswith("data:"): @@ -1786,7 +1786,7 @@ import numpy as np from datetime import date as _date router = APIRouter() -# Studio-only router (not mounted on /v1 OpenAI-compat). +# Unsloth-only router (not mounted on /v1 OpenAI-compat). studio_router = APIRouter() @@ -2108,9 +2108,9 @@ def _effective_enable_tools(payload) -> Optional[bool]: def _explicit_studio_tool_loop_requested(payload) -> bool: - """True when the request itself asks Studio to execute local tools. + """True when the request itself asks Unsloth to execute local tools. - Process-wide CLI policy can default Studio's tool loop on for ordinary chat, + Process-wide CLI policy can default Unsloth's tool loop on for ordinary chat, but it must not steal OpenAI-compatible client tools or response_format requests from the llama-server passthrough path. A policy of ``False`` (--disable-tools) vetoes even an explicit ``enable_tools: true`` ask. @@ -2122,7 +2122,7 @@ def _explicit_studio_tool_loop_requested(payload) -> bool: def _permission_mode_confirm(payload) -> bool: - """Effective confirm-gate intent for Studio's own local tool loop. + """Effective confirm-gate intent for Unsloth's own local tool loop. Honors the documented default that an unset permission_mode behaves as "ask". An explicit confirm_tool_calls (True or False) wins; explicit @@ -2144,7 +2144,7 @@ def _permission_mode_confirm(payload) -> bool: def _confirm_gate_needs_stream(payload) -> bool: - """Whether Studio's local tool-loop confirm gate still requires stream=true. + """Whether Unsloth's local tool-loop confirm gate still requires stream=true. The gate can only prompt while streaming, so a non-streaming request that will prompt must 400 up front. auto ("Approve for me") only prompts for a call the @@ -3115,13 +3115,16 @@ def _normalise_settings_str(value: Optional[str]) -> Optional[str]: def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[str]]) -> bool: - """Whether an inherited --split-mode should be stripped on reload. + """Whether an inherited --split-mode (and its coupled --tensor-split) should + be stripped on reload. The binary Tensor Parallelism toggle can't carry --split-mode's row/none/ layer modes, so only strip when the toggle overrides it: tensor being turned on, or the inherited mode is tensor (toggle turning it off). Non-tensor modes - survive. Shared by the inheritance strip and the already-loaded stale check - so they agree on what reload would do. + survive. A manual per-GPU ratio is handled by _should_strip_tensor_split, + which strips only --tensor-split so the inherited mode is kept. Shared by the + inheritance strip and the already-loaded stale check so they agree on what + reload would do. """ fields_set = getattr(request, "model_fields_set", set()) return "tensor_parallel" in fields_set and ( @@ -3129,6 +3132,25 @@ def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[ ) +def _should_strip_tensor_split(request: LoadRequest) -> bool: + """Whether an inherited --tensor-split alone should be stripped on reload. + + Manual explicit offload (gpu_layers >= 0) owns the per-GPU split: with a ratio + it emits its own --tensor-split (an inherited one, appended last, would + override it), and with the ratio cleared it wants llama.cpp's default + free-VRAM split. Either way an inherited --tensor-split must go, else the + cleared case silently keeps the stale ratio while status reports None. + Unlike _should_strip_split_mode this leaves --split-mode untouched, so a + user's row/none/layer mode survives a Studio split-ratio edit. When the + Tensor Parallelism toggle IS overriding the mode, _should_strip_split_mode + (called alongside this at every site) strips --split-mode anyway. + """ + return ( + getattr(request, "gpu_memory_mode", "auto") == "manual" + and getattr(request, "gpu_layers", -1) >= 0 + ) + + def _carry_preserved_tensor_intent( *, preserved: bool, same_model: bool, explicit_drop: bool ) -> bool: @@ -3143,7 +3165,7 @@ 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 + A bare tensor_parallel field is NOT a drop: the Unsloth 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 @@ -3187,12 +3209,44 @@ def _request_matches_loaded_settings( else strip_shadowing_flags( backend_extra, strip_split_mode = _should_strip_split_mode(request, backend_extra), + strip_tensor_split = _should_strip_tensor_split(request), + strip_offload = request.gpu_memory_mode == "manual", ) ) if not _tensor_parallel_matches_loaded( effective_extra, request.tensor_parallel, llama_backend.tensor_parallel ): return False + # The diffusion runner is mode-agnostic (it always reports "auto" and ignores + # the layer/MoE/split knobs), so a standing manual preference in the request + # must not force a needless reload -- only the GPU pick matters. + if not llama_backend.is_diffusion: + if request.gpu_memory_mode != llama_backend.gpu_memory_mode: + return False + # Manual: a layer-count change always reloads; MoE/split only matter with + # an explicit offload (gpu_layers >= 0), so a leftover value under Auto + # must not force one. Mirrors LlamaCppBackend._already_in_target_state. + if request.gpu_memory_mode == "manual" and ( + request.gpu_layers != llama_backend.gpu_layers + or ( + request.gpu_layers >= 0 + and ( + request.n_cpu_moe != llama_backend.n_cpu_moe + or (request.tensor_split or None) != (llama_backend.tensor_split or None) + ) + ) + ): + return False + # A changed GPU pick must reload. The diffusion runner collapses a multi-GPU + # request to its single lowest device (it drives one device only), so the + # backend records just that device; compare the request the same way, or a + # multi-GPU pick that resolves to the same device needlessly reloads. + if llama_backend.is_diffusion: + _req_gpu_ids = [sorted(request.gpu_ids)[0]] if request.gpu_ids else None + else: + _req_gpu_ids = sorted(request.gpu_ids) if request.gpu_ids else None + if _req_gpu_ids != llama_backend.gpu_ids: + 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 @@ -3235,14 +3289,17 @@ def _request_matches_loaded_settings( # contain any shadow flag, so the reload path strips them rather than # leaving a stale override in effect. (backend_extra computed above.) if request.llama_extra_args is None: - # Mirror the reload's conditional split-mode strip, so a preserved - # non-tensor mode (row/none/layer) isn't seen as stale and doesn't - # trigger a needless reload of a healthy server. + # Mirror the reload's conditional strips, so a preserved non-tensor mode + # (row/none/layer) isn't seen as stale and doesn't trigger a needless + # reload of a healthy server, while an inherited offload/ratio flag that + # the reload *would* strip is correctly seen as stale. if ( backend_extra and strip_shadowing_flags( backend_extra, strip_split_mode = _should_strip_split_mode(request, backend_extra), + strip_tensor_split = _should_strip_tensor_split(request), + strip_offload = request.gpu_memory_mode == "manual", ) != backend_extra ): @@ -3861,6 +3918,46 @@ def _estimate_gguf_required_gb( return None +def _classify_diffusion_gguf(config: ModelConfig) -> Optional[bool]: + """Classify a GGUF as diffusion, normal, or unknown before it is loaded. + + ``None`` is important here: a remote GGUF whose header is not cached can + still be routed to the single-GPU diffusion runner after download. Treating + that case as normal would let Manual mode skip the training guard even + though the runner ignores Manual's llama-server placement controls. + """ + identity = " ".join( + str(getattr(config, attr, "") or "") for attr in ("identifier", "gguf_hf_repo", "gguf_file") + ).lower() + if "diffusion" in identity: + return True + + try: + main = getattr(config, "gguf_file", None) + if not (main and Path(main).is_file()): + repo = getattr(config, "gguf_hf_repo", None) + variant = getattr(config, "gguf_variant", None) + if repo and variant: + from hub.utils.gguf import resolve_local_gguf_path + main = resolve_local_gguf_path(repo, variant) + if not main or not Path(main).is_file(): + return None + + probe = LlamaCppBackend() + probe._read_gguf_metadata(str(main)) + if probe.is_diffusion: + return True + # A successfully decoded architecture proves that this is a normal + # llama-server GGUF. No architecture means the lightweight probe could + # not establish the routing decision, so preserve the unknown state. + if getattr(probe, "_architecture", None): + return False + return None + except Exception as e: + logger.debug("Could not identify diffusion GGUF for training guard: %s", e) + return None + + def _guard_chat_load_against_training( config: ModelConfig, *, @@ -3871,11 +3968,19 @@ def _guard_chat_load_against_training( requested_gpu_ids: Optional[List[int]], llama_extra_args: Optional[list[str]] = None, n_parallel: int = 1, + gpu_memory_mode: Literal["auto", "manual"] = "auto", ) -> None: - """Refuse loading a local chat model that would OOM an active training run. + """Protect active training from automatically placed chat-model loads. + No-op when training is inactive or unknown. `load_in_4bit` must be the - effective quantization (see _effective_load_in_4bit). Raises HTTP 409 when the - model would not fit alongside training.""" + effective quantization (see _effective_load_in_4bit). Manual chat-GGUF + placement is an explicit override: Auto layers delegate fitting to + llama.cpp's ``--fit`` and pinned layers are owned by the user, so neither is + estimated here. Diffusion is still guarded because its mode-agnostic runner + ignores those controls and uses one GPU. An unclassified GGUF is guarded as + potentially diffusion until its local header proves otherwise. Other loads + raise HTTP 409 when they would not fit beside training. + """ from core.training import get_training_backend from routes.training_vram import can_load_chat_during_training @@ -3887,6 +3992,19 @@ def _guard_chat_load_against_training( return is_gguf = bool(getattr(config, "is_gguf", False)) + diffusion_kind = _classify_diffusion_gguf(config) if is_gguf else False + if is_gguf and gpu_memory_mode == "manual" and diffusion_kind is False: + return + + diffusion_gpu = None + if is_gguf and diffusion_kind is not False: + # Use the same token selection as the runner: an explicit pick wins, + # followed by DG_GPU, the first parent-visible token, then GPU 0. + diffusion_gpu = LlamaCppBackend._diffusion_gpu_arg( + requested_gpu_ids, + cpu_only = LlamaCppBackend._effective_gpu_count() == 0, + ) + required_override_gb = ( _estimate_gguf_required_gb( config, @@ -3907,6 +4025,7 @@ def _guard_chat_load_against_training( requested_gpu_ids = requested_gpu_ids, is_gguf = is_gguf, required_override_gb = required_override_gb, + single_device_gpu = diffusion_gpu, ) if ok: return @@ -3934,6 +4053,98 @@ def _guard_chat_load_against_training( raise HTTPException(status_code = 409, detail = detail) +def _resolve_inherited_extra_args( + request, + config: ModelConfig, + model_identifier: str, + extra_llama_args: Optional[list[str]], + effective_chat_template_override: Optional[str] = None, +) -> Optional[list[str]]: + """Effective pass-through extras for a GGUF request that omitted the field: + the previous same-model load's extras, shadow-stripped, so a settings-Apply + reload (which does not round-trip the extras field) keeps them (#5401).""" + if getattr(request, "llama_extra_args", None) is not None: + return extra_llama_args + if not getattr(config, "is_gguf", False): + return extra_llama_args + llama_backend = get_llama_cpp_backend() + if not llama_backend.extra_args: + return extra_llama_args + # Inherit the previous load's extras (the chat-settings Apply path doesn't + # round-trip them; an explicit [] still clears). Gated on (model_identifier, + # hf_variant) to refuse cross-model pickup, and shadowing flags are + # stripped so an inherited override can't win the last-wins CLI + # parse against a freshly-supplied first-class field. + source = llama_backend.extra_args_source + # Compare against the resolved variant, not the request field: callers + # commonly omit gguf_variant for local ``.gguf`` paths and HF auto-pick + # flows. ``config.gguf_variant`` is the variant load_model was actually + # invoked with, so both sides of the comparison key off the same string. + resolved_variant = (config.gguf_variant or "").lower() + request_variant = (request.gguf_variant or "").lower() + stored_variant = (source[1] or "").lower() if source else "" + same_model = bool(source and source[0] and source[0].lower() == model_identifier.lower()) + if request.gguf_variant: + variant_mismatch = request_variant != stored_variant + else: + variant_mismatch = bool(stored_variant and resolved_variant != stored_variant) + same_source = same_model and not variant_mismatch + if not same_source: + logger.info( + "Not inheriting llama_extra_args: stored args came from %s, loading %s", + source, + (model_identifier, resolved_variant), + ) + # Cross-model: clear explicitly so the backend doesn't + # inherit via "no opinion" semantics. + extra_llama_args = [] + else: + # Strip only the groups whose first-class field was set by the caller, so + # an inherited --chat-template-file survives an Apply that omits + # chat_template_override. A bundled family template (e.g. gemma-4) counts as + # a first-class template even when the request omits chat_template_override, + # so strip the inherited --chat-template-file then too -- else the stale arg + # (appended last) shadows the bundled template while Studio reports its caps. + fields_set = getattr(request, "model_fields_set", set()) + stripped = strip_shadowing_flags( + llama_backend.extra_args, + strip_context = "max_seq_length" in fields_set, + strip_cache = "cache_type_kv" in fields_set, + strip_spec = ("speculative_type" in fields_set or "spec_draft_n_max" in fields_set), + strip_template = ( + "chat_template_override" in fields_set + or effective_chat_template_override is not None + ), + strip_split_mode = _should_strip_split_mode(request, llama_backend.extra_args), + # manual + per-GPU ratio emits its own --tensor-split; drop + # an inherited one (appended last would override it) while + # keeping the user's --split-mode row/none/layer choice. + strip_tensor_split = _should_strip_tensor_split(request), + # manual emits its own --fit/--gpu-layers, so an inherited offload flag + # must not last-wins-override it. auto leaves a user's inherited -ngl + # alone. getattr: a validate request reuses this resolver, no offload fields. + strip_offload = getattr(request, "gpu_memory_mode", "auto") == "manual", + ) + try: + extra_llama_args = validate_extra_args(stripped) + except ValueError: + # Shouldn't happen on already-validated args; degrade to + # no-extras rather than 400 if managed flags changed. + logger.warning( + "Stored llama_extra_args failed revalidation; loading without them: %s", + stripped, + ) + extra_llama_args = [] + else: + if extra_llama_args: + logger.info( + "Inheriting llama_extra_args from previous " + "load (same model, shadow-stripped): %s", + extra_llama_args, + ) + return extra_llama_args + + def _model_json_response(model, status_code: int = 200) -> Response: """Serialize a pydantic response once via pydantic-core. @@ -4040,6 +4251,35 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre None if request.llama_extra_args is None else extra_llama_args ) + # Manual mode owns the offload flags: strip them from EXPLICIT extras + # too (the inherited path already does), or a last-wins --gpu-layers / + # --fit in extras re-enables GPU offload on a load status reports as + # CPU-only. Manual + per-GPU ratio owns --tensor-split the same way. + if request.gpu_memory_mode == "manual" and extra_llama_args: + _stripped_explicit = strip_shadowing_flags( + extra_llama_args, + strip_context = False, + strip_cache = False, + strip_spec = False, + strip_template = False, + strip_split_mode = False, + strip_tensor_split = _should_strip_tensor_split(request), + strip_offload = True, + ) + if _stripped_explicit != extra_llama_args: + logger.info( + "Manual GPU memory owns the offload flags; stripping them " + "from explicit llama_extra_args: %s -> %s", + extra_llama_args, + _stripped_explicit, + ) + extra_llama_args = _stripped_explicit + + # Keep every downstream consumer on the normalized explicit list. In + # particular, the already-loaded comparator must not compare the raw + # request's managed offload flags against the stripped launch state. + request = request.model_copy(update = {"llama_extra_args": extra_llama_args}) + model_identifier, model_log_label, native_grant_backed = ( _resolve_model_identifier_for_request(request, operation = "load-model") ) @@ -4121,6 +4361,13 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre speculative_type = llama_backend.requested_spec_mode, spec_draft_n_max = llama_backend.spec_draft_n_max, tensor_parallel = llama_backend.tensor_parallel, + gpu_memory_mode = llama_backend.gpu_memory_mode, + gpu_layers = llama_backend.gpu_layers, + n_cpu_moe = llama_backend.n_cpu_moe, + tensor_split = llama_backend.tensor_split, + n_layers = llama_backend.n_layers, + n_moe_layers = llama_backend.n_moe_layers, + gpu_ids = llama_backend.gpu_ids, ) else: if ( @@ -4187,18 +4434,47 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre # Normalize gpu_ids: empty list means auto-selection, same as None effective_gpu_ids = request.gpu_ids if request.gpu_ids else None - # Reject GGUF + gpu_ids first so the guard can't mask it with a VRAM 409. + # GGUF supports gpu_ids: validate the pick up front (before the training + # guard) so a bad pick is a clean 400, not masked by a VRAM 409. Rejects + # negative / out-of-range / duplicate ids and UUID/MIG parents. XPU hosts + # are rejected outright: the picker's indices are torch-xpu ordinals neither + # applicator speaks (CUDA/HIP masks don't apply, the Vulkan --device pin + # uses ggml's own Vulkan ordinals), so a pick could land on the wrong device. if config.is_gguf and effective_gpu_ids is not None: - raise HTTPException( - status_code = 400, - detail = "gpu_ids is not supported for GGUF models yet.", - ) + from utils.hardware import DeviceType, get_device + from utils.hardware.hardware import resolve_requested_gpu_ids + + if get_device() == DeviceType.XPU: + raise HTTPException( + status_code = 400, + detail = ( + "GPU selection (gpu_ids) is not supported on Intel XPU. " + "Omit gpu_ids to use all devices." + ), + ) + # Same reasoning for a Vulkan-only build: --device pins ggml's own + # Vulkan ordinals, so a physical pick can land on the wrong card on + # masked or non-contiguous hosts. + if LlamaCppBackend._is_vulkan_backend(): + raise HTTPException( + status_code = 400, + detail = ( + "GPU selection (gpu_ids) is not supported with a Vulkan " + "llama.cpp build: physical GPU ids have no defined " + "mapping to Vulkan device ordinals. Omit gpu_ids to use " + "all devices." + ), + ) + try: + resolve_requested_gpu_ids(effective_gpu_ids) + except ValueError as exc: + raise HTTPException(status_code = 400, detail = str(exc)) from exc if not config.is_gguf and _mlx_distributed_launch_detected(): raise HTTPException( status_code = 400, detail = ( - "Studio does not support distributed MLX inference under " - "mlx.launch. Use `mlx.launch ... unsloth chat` or run Studio " + "Unsloth does not support distributed MLX inference under " + "mlx.launch. Use `mlx.launch ... unsloth chat` or run Unsloth " "without the distributed launcher." ), ) @@ -4222,8 +4498,20 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre "architectures)" ) - # Refuse a load that would OOM active training, before the unload step below - # frees the resident model. Off-loop: guard does sync nvidia-smi / HF work. + # Inherit the previous same-model load's pass-through extras when this + # request omits the field (a settings-Apply reload doesn't round-trip + # them); shadow-stripped so an inherited flag can't override a + # first-class field the caller did set (#5401). + extra_llama_args = _resolve_inherited_extra_args( + request, + config, + model_identifier, + extra_llama_args, + effective_chat_template_override, + ) + + # Apply the training coexistence policy before the unload step below + # frees the resident model. Off-loop: the default-mode guard does sync work. await asyncio.to_thread( _guard_chat_load_against_training, config, @@ -4234,6 +4522,7 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre requested_gpu_ids = effective_gpu_ids, llama_extra_args = extra_llama_args, n_parallel = getattr(fastapi_request.app.state, "llama_parallel_slots", 1), + gpu_memory_mode = request.gpu_memory_mode, ) # ── GGUF path: load via llama-server ────────────────────── @@ -4245,84 +4534,6 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre from core.inference.llama_cpp import gguf_load_in_flight gguf_load_stack.enter_context(gguf_load_in_flight(config.gguf_hf_repo)) - # Inherit llama_extra_args from the previous load when the request - # omits the field (the chat-settings Apply path doesn't round-trip - # them; explicit [] still clears). Gated on (model_identifier, - # hf_variant) to refuse cross-model pickup, and shadowing flags are - # stripped so an inherited override can't win the last-wins CLI - # parse against a freshly-supplied first-class field. - if request.llama_extra_args is None and llama_backend.extra_args: - source = llama_backend.extra_args_source - # Compare against the resolved variant, not the request - # field: callers commonly omit gguf_variant for local - # ``.gguf`` paths and HF auto-pick flows. ``config.gguf_ - # variant`` is the variant load_model was actually - # invoked with (see the HF / local branches below), so - # both sides of the comparison key off the same string. - resolved_variant = (config.gguf_variant or "").lower() - request_variant = (request.gguf_variant or "").lower() - stored_variant = (source[1] or "").lower() if source else "" - same_model = bool( - source and source[0] and source[0].lower() == model_identifier.lower() - ) - if request.gguf_variant: - variant_mismatch = request_variant != stored_variant - else: - variant_mismatch = bool(stored_variant and resolved_variant != stored_variant) - same_source = same_model and not variant_mismatch - if not same_source: - logger.info( - "Not inheriting llama_extra_args: stored args came from %s, loading %s", - source, - (model_identifier, resolved_variant), - ) - # Cross-model: clear explicitly so the backend doesn't - # inherit via "no opinion" semantics. - extra_llama_args = [] - else: - # Strip only the groups whose first-class field was set by - # the caller, so an inherited --chat-template-file survives - # an Apply that omits chat_template_override. A bundled family - # template (e.g. the gemma-4 override) is an effective - # first-class template setting even when the raw request - # omits chat_template_override, so strip the inherited - # --chat-template-file in that case too -- otherwise the stale - # extra arg (appended last) shadows the bundled template while - # Studio reports the bundled template's capabilities. - fields_set = getattr(request, "model_fields_set", set()) - stripped = strip_shadowing_flags( - llama_backend.extra_args, - strip_context = "max_seq_length" in fields_set, - strip_cache = "cache_type_kv" in fields_set, - strip_spec = ( - "speculative_type" in fields_set or "spec_draft_n_max" in fields_set - ), - strip_template = ( - "chat_template_override" in fields_set - or effective_chat_template_override is not None - ), - strip_split_mode = _should_strip_split_mode( - request, llama_backend.extra_args - ), - ) - try: - extra_llama_args = validate_extra_args(stripped) - except ValueError: - # Shouldn't happen on already-validated args; degrade to - # no-extras rather than 400 if managed flags changed. - logger.warning( - "Stored llama_extra_args failed revalidation; loading without them: %s", - stripped, - ) - extra_llama_args = [] - else: - if extra_llama_args: - logger.info( - "Inheriting llama_extra_args from previous " - "load (same model, shadow-stripped): %s", - extra_llama_args, - ) - # Block cache writes that would race the download manager. This runs # after pass-through argument inheritance so a carried --no-mmproj # changes the companion requirement exactly as it does for the load. @@ -4370,6 +4581,11 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre cache_type_kv = request.cache_type_kv, speculative_type = request.speculative_type, spec_draft_n_max = request.spec_draft_n_max, + gpu_memory_mode = request.gpu_memory_mode, + gpu_layers = request.gpu_layers, + n_cpu_moe = request.n_cpu_moe, + tensor_split = request.tensor_split, + gpu_ids = effective_gpu_ids, n_parallel = _n_parallel, ) if config.gguf_hf_repo: @@ -4537,6 +4753,13 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre speculative_type = llama_backend.requested_spec_mode, spec_draft_n_max = llama_backend.spec_draft_n_max, tensor_parallel = llama_backend.tensor_parallel, + gpu_memory_mode = llama_backend.gpu_memory_mode, + gpu_layers = llama_backend.gpu_layers, + n_cpu_moe = llama_backend.n_cpu_moe, + tensor_split = llama_backend.tensor_split, + n_layers = llama_backend.n_layers, + n_moe_layers = llama_backend.n_moe_layers, + gpu_ids = llama_backend.gpu_ids, ) # ── Standard path: load via Unsloth/transformers ────────── @@ -4727,7 +4950,7 @@ def _requires_trust_remote_code_for_model( model_identifier: str, hf_token: Optional[str] = None ) -> bool: """Whether loading this model would execute custom repo code, so the consent - dialog must run first. True if the Studio YAML default enables + dialog must run first. True if the Unsloth YAML default enables ``trust_remote_code`` OR the raw config declares an ``auto_map`` (Hub/local, config.json or tokenizer_config.json). Reads raw JSON only; never imports model code.""" @@ -4795,7 +5018,9 @@ def _requires_security_review_for_model( @router.post("/validate", response_model = ValidateModelResponse) async def validate_model( - request: ValidateModelRequest, current_subject: str = Depends(get_current_subject) + request: ValidateModelRequest, + fastapi_request: Request = None, + current_subject: str = Depends(get_current_subject), ): """ Lightweight validation endpoint for model identifiers. @@ -4823,15 +5048,39 @@ async def validate_model( detail = f"Invalid model identifier: {model_log_label}", ) - # Refuse early (before the frontend unloads to load this) if it can't fit - # alongside training, using the same settings /load uses so they agree. + # Apply the same training coexistence policy as /load before the frontend + # unloads the current model. effective_gpu_ids = request.gpu_ids if request.gpu_ids else None - # Mirror /load: reject GGUF + gpu_ids before the guard so both return 400. + # Mirror /load: GGUF supports gpu_ids, so validate the pick (a bad one is + # a clean 400) before the guard sizes the model against training VRAM. + # XPU-host picks are rejected like /load (no defined mapping from the + # picker's torch-xpu ordinals to the launcher's device spaces). if config.is_gguf and effective_gpu_ids is not None: - raise HTTPException( - status_code = 400, - detail = "gpu_ids is not supported for GGUF models yet.", - ) + from utils.hardware import DeviceType, get_device + from utils.hardware.hardware import resolve_requested_gpu_ids + + if get_device() == DeviceType.XPU: + raise HTTPException( + status_code = 400, + detail = ( + "GPU selection (gpu_ids) is not supported on Intel XPU. " + "Omit gpu_ids to use all devices." + ), + ) + if LlamaCppBackend._is_vulkan_backend(): + raise HTTPException( + status_code = 400, + detail = ( + "GPU selection (gpu_ids) is not supported with a Vulkan " + "llama.cpp build: physical GPU ids have no defined " + "mapping to Vulkan device ordinals. Omit gpu_ids to use " + "all devices." + ), + ) + try: + resolve_requested_gpu_ids(effective_gpu_ids) + except ValueError as exc: + raise HTTPException(status_code = 400, detail = str(exc)) from exc effective_load_in_4bit = _effective_load_in_4bit(config, request.load_in_4bit) # Both checks cover the [adapter, base] set (matching the scan route and workers): @@ -4895,16 +5144,32 @@ async def validate_model( latest_tier_active_for, config.identifier, request.hf_token ): effective_load_in_4bit = False - # Off-loop: guard does sync nvidia-smi / HF work. - await asyncio.to_thread( - _guard_chat_load_against_training, - config, - model_identifier = model_identifier, - hf_token = request.hf_token, - load_in_4bit = effective_load_in_4bit, - max_seq_length = request.max_seq_length, - requested_gpu_ids = effective_gpu_ids, - ) + # A metadata-only probe just reads the GGUF header and allocates no VRAM, + # so it must not be refused by the training guard. Real loads validate + # without include_context_length and /load applies the guard again. + if not request.include_context_length: + # Match /load's inherited llama.cpp extras and parallel slot count so + # validation cannot pass a smaller estimate than the subsequent load. + effective_extra_args = _resolve_inherited_extra_args( + request, config, model_identifier, None + ) + # Off-loop: guard does sync nvidia-smi / HF work. + await asyncio.to_thread( + _guard_chat_load_against_training, + config, + model_identifier = model_identifier, + hf_token = request.hf_token, + load_in_4bit = effective_load_in_4bit, + max_seq_length = request.max_seq_length, + requested_gpu_ids = effective_gpu_ids, + llama_extra_args = effective_extra_args, + n_parallel = ( + getattr(fastapi_request.app.state, "llama_parallel_slots", 1) + if fastapi_request is not None + else 1 + ), + gpu_memory_mode = request.gpu_memory_mode, + ) # A selected GGUF loads via llama.cpp: auto_map Python and root pickle weights in a # mixed repo are inert for this load, so gating on them is a false positive. Only @@ -4918,10 +5183,15 @@ async def validate_model( # Native context length, read from the local GGUF header when present. # Lets the staged ("Load on selection" off) flow populate the context # slider before the GPU load; None until the file is downloaded. + # Staged header dims (one read): native context, total layer count, and + # MoE expert-layer count -- let the staged flow size the context, GPU- + # layers and manual --n-cpu-moe sliders before the load. context_length: Optional[int] = None + layer_count: Optional[int] = None + moe_layer_count: Optional[int] = None if request.include_context_length and is_gguf: from hub.utils.gguf import resolve_local_gguf_path - from utils.models.gguf_metadata import read_gguf_context_length + from utils.models.gguf_metadata import read_gguf_staged_dims # Best-effort: a header-read failure must never fail validation of an # otherwise-valid model (the outer except turns it into a 400). @@ -4937,9 +5207,15 @@ async def validate_model( model_identifier, request.gguf_variant ) if local_gguf: - context_length = read_gguf_context_length(local_gguf) + # Header walk reads tokenizer arrays for dense models (tens of + # ms); keep it off the event loop. + dims = await asyncio.to_thread(read_gguf_staged_dims, local_gguf) + if dims: + context_length = dims["context_length"] + layer_count = dims["layer_count"] + moe_layer_count = dims["moe_layer_count"] except Exception as e: - logger.debug("Context-length probe failed for %s: %s", model_log_label, e) + logger.debug("Header probe failed for %s: %s", model_log_label, e) return ValidateModelResponse( valid = True, @@ -4954,6 +5230,8 @@ async def validate_model( requires_trust_remote_code = requires_trust_remote_code, requires_security_review = requires_security_review, context_length = context_length, + layer_count = layer_count, + moe_layer_count = moe_layer_count, requires_transformers_upgrade = transformers_upgrade is not None, transformers_upgrade = transformers_upgrade, ) @@ -5360,7 +5638,7 @@ async def confirm_tool_call( @studio_router.get("/monitor") async def get_api_monitor(current_subject: str = Depends(get_current_subject)): - """Return recent OpenAI-compatible API activity for Studio.""" + """Return recent OpenAI-compatible API activity for Unsloth.""" active_model = _monitor_active_model() active_requests = api_monitor.active_count(subject = current_subject) if active_requests: @@ -5548,7 +5826,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)): _display_model_id = os.path.basename(_model_id) _inference_cfg = load_inference_config(_model_id) if _model_id else None _audio_type = getattr(llama_backend, "_audio_type", None) - # Don't surface Studio's auto-applied bundled family template (e.g. the + # Don't surface Unsloth's auto-applied bundled family template (e.g. the # gemma-4 override) as a user-authored override: the frontend adopts # status.chat_template_override as editable state and would otherwise # re-send it as an explicit override for a later, unrelated model. Only @@ -5593,6 +5871,14 @@ async def get_status(current_subject: str = Depends(get_current_subject)): speculative_type = llama_backend.requested_spec_mode, spec_draft_n_max = llama_backend.spec_draft_n_max, tensor_parallel = llama_backend.tensor_parallel, + gpu_memory_mode = llama_backend.gpu_memory_mode, + gpu_layers = llama_backend.gpu_layers, + n_cpu_moe = llama_backend.n_cpu_moe, + tensor_split = llama_backend.tensor_split, + requested_context_length = llama_backend.requested_n_ctx, + n_layers = llama_backend.n_layers, + n_moe_layers = llama_backend.n_moe_layers, + gpu_ids = llama_backend.gpu_ids, llama_cpp_supports_mtp = _supports_mtp, spec_fallback_reason = llama_backend.spec_fallback_reason, llama_cpp_prebuilt_stale = _stale, @@ -6173,7 +6459,7 @@ def _build_external_messages( metadata; strip it for providers that can't parse the unknown key. 2. Marked server-side builtin cards (`_server_tool: true` on a canonical builtin name, or a Gemini `native_part` payload) are - Studio-internal tool cards from a prior native Gemini turn; + Unsloth-internal tool cards from a prior native Gemini turn; forwarding them to OpenAI / Anthropic / custom OAI-compat gateways sends an orphan `tool_calls` entry (no matching tool declaration, often no matching `role="tool"` reply) that can be rejected. We @@ -6859,7 +7145,7 @@ async def openai_chat_completions( # is invalid and must not evict the resident model first. # # Enter the local-loop arm exactly when the passthrough router below would - # run Studio's own tool loop. That gate is `_tools_on or _mcp_allowed` + # run Unsloth's own tool loop. That gate is `_tools_on or _mcp_allowed` # (see the use_tools block): _effective_enable_tools (which lets a # process-wide --enable-tools policy force the loop on) plus mcp_enabled # honoring --disable-tools, and tool_choice="none" disabling it unless the @@ -6884,7 +7170,7 @@ async def openai_chat_completions( or bool(payload.openai_code_exec_container_id) or bool(payload.anthropic_code_exec_container_id) # A JSON-schema response_format is guided-decoding structured output the - # router forwards to the llama-server passthrough, not Studio's tool + # router forwards to the llama-server passthrough, not Unsloth's tool # loop, so a --enable-tools policy must not 400 it as a local-confirm # request under ask/auto. or bool(_extract_response_format(payload)) @@ -6961,7 +7247,7 @@ async def openai_chat_completions( using_gguf = llama_backend.is_loaded # OpenAI-SDK clients send ``chat_template_kwargs`` via ``extra_body``, which - # the SDK spreads into the request body at the top level. Studio's + # the SDK spreads into the request body at the top level. Unsloth's # ChatCompletionRequest has ``extra="allow"`` so pydantic stashes them in # ``model_extra``, but downstream generators consume the typed # ``payload.enable_thinking``. Lift ``enable_thinking`` from the extra-body @@ -7215,7 +7501,7 @@ async def openai_chat_completions( # ── Standard OpenAI function-calling pass-through (GGUF only) ──── # When a client (opencode / Claude Code via OpenAI compat / Cursor / - # Continue / ...) sends standard OpenAI `tools` without Studio's + # Continue / ...) sends standard OpenAI `tools` without Unsloth's # `enable_tools` shorthand, forward the request to llama-server # verbatim so structured `tool_calls` flow back to the client. This # branch runs BEFORE `_extract_content_parts` because that helper is @@ -7238,7 +7524,7 @@ async def openai_chat_completions( _has_tool_catalog = bool(payload.tools and len(payload.tools) > 0) _has_active_tool_catalog = _has_tool_catalog and payload.tool_choice != "none" _has_client_tool_contract = _has_active_tool_catalog or _has_tool_messages - # The Studio tool loop needs a tool-capable backend, so a request that asks + # The Unsloth tool loop needs a tool-capable backend, so a request that asks # for it on a backend that can't run it (DiffusionGemma forces supports_tools # off) must not steal client tools from the passthrough (#6851). _studio_tool_loop_requested = ( @@ -7434,7 +7720,7 @@ async def openai_chat_completions( use_tools = False if use_tools: - # permission_mode ask/auto require the confirm gate for Studio's own + # permission_mode ask/auto require the confirm gate for Unsloth's own # tool loop. The request validator self-enables confirm only for # request-level tool signals (enable_tools/enabled_tools/mcp_enabled); # when a CLI policy (--enable-tools) forces the loop on without those, @@ -8711,7 +8997,7 @@ async def openai_chat_completions( _sf_model_info = backend.models.get(backend.active_model_name, {}) _sf_tpl = (_sf_model_info.get("chat_template_info") or {}).get("template") # Named templates may expose native reasoning only in their ``tool_use`` - # branch. Use a truthy placeholder for Studio-managed tools, whose concrete + # branch. Use a truthy placeholder for Unsloth-managed tools, whose concrete # schemas are selected below, and the request schemas for client passthrough. _sf_server_tool_intent = bool( _effective_enable_tools(payload) or _explicit_studio_tool_loop_requested(payload) @@ -8790,7 +9076,7 @@ async def openai_chat_completions( _sf_use_tools = False if _sf_use_tools: - # permission_mode ask/auto require the confirm gate for Studio's own tool + # permission_mode ask/auto require the confirm gate for Unsloth's own tool # loop; when a CLI policy (--enable-tools) forces the loop on without a # request-level tool signal, derive confirm here so the mode still gates # the call (matching the GGUF path). off/full never prompt. @@ -9443,6 +9729,13 @@ async def openai_chat_completions( backend.reset_generation_state() api_monitor.finish(monitor_id, "cancelled") raise + except GenStreamErrorRaised as exc: + # Adapter-controlled (compare-mode) backend failure. Honor the + # public flag so operational errors surface their real message. + backend.reset_generation_state() + _msg = _friendly_gen_stream_error(exc) + api_monitor.fail(monitor_id, _msg) + yield _openai_stream_error_sse({"error": {"message": _msg, "type": "server_error"}}) except Exception as e: backend.reset_generation_state() logger.error(f"Error during OpenAI streaming: {e}", exc_info = True) @@ -9591,6 +9884,13 @@ async def openai_chat_completions( except HTTPException: raise + except GenStreamErrorRaised as exc: + # Adapter-controlled (compare-mode) backend failure. Honor the public + # flag so operational errors surface their real message. + backend.reset_generation_state() + _msg = _friendly_gen_stream_error(exc) + api_monitor.fail(monitor_id, _msg) + raise HTTPException(status_code = 500, detail = _msg) except Exception as e: backend.reset_generation_state() logger.error(f"Error during OpenAI completion: {e}", exc_info = True) @@ -12036,7 +12336,7 @@ def _anthropic_requested_studio_tools(tools: Optional[list]) -> set[str]: def _select_anthropic_server_tools( all_tools: list[dict], requested_studio_tools: set[str], enabled_tools: Optional[list[str]] ) -> list[dict]: - """Select Studio tools requested through Anthropic tools and extensions.""" + """Select Unsloth tools requested through Anthropic tools and extensions.""" if not requested_studio_tools and enabled_tools is None: return all_tools @@ -12275,7 +12575,7 @@ async def anthropic_messages( ), ) - # Reject an unsupported confirm-gated permission mode for Studio's own + # Reject an unsupported confirm-gated permission mode for Unsloth's own # ("server") Anthropic tools before the switch, mirroring the malformed- and # mixed-tool checks above. ask always wants a per-call pause this passthrough # cannot offer, so it 400s whenever server tools are selected. auto only needs @@ -12764,11 +13064,11 @@ async def _anthropic_tool_stream( ends_on_tool_use = True elif etype == "tool_end": tool_blocks_emitted += 1 - # A tool_end means Studio executed the tool server-side, so + # A tool_end means Unsloth executed the tool server-side, so # the response no longer ends on a pending client action. # Without this, a server tool that produces no trailing text # would be mislabeled stop_reason "tool_use", telling the - # client to run a tool Studio already ran. + # client to run a tool Unsloth already ran. ends_on_tool_use = False elif etype == "content" and event.get("text"): ends_on_tool_use = False @@ -13684,7 +13984,7 @@ def _openai_messages_for_passthrough(payload) -> list[dict]: structured ``tool_calls``. Content-parts images already in the list are left untouched. - When a client uses Studio's legacy ``image_base64`` top-level field, the + When a client uses Unsloth's legacy ``image_base64`` top-level field, the image is re-encoded to PNG (llama-server's stb_image has limited format support) and spliced into the last user message as an OpenAI ``image_url`` content part so vision + function-calling requests work transparently. @@ -13818,7 +14118,7 @@ def _build_openai_passthrough_body( ) -> dict: """Assemble the llama-server request body from a ChatCompletionRequest. - Only known OpenAI / llama-server fields are forwarded, so Studio-specific + Only known OpenAI / llama-server fields are forwarded, so Unsloth-specific extensions (``enable_tools``, ``enabled_tools``, ``session_id``, ...) never leak to the backend. """ @@ -14068,7 +14368,7 @@ async def _openai_passthrough_stream_admitted( admission_lease: LlamaAdmissionLease, tracker, ): - """Streaming client-side pass-through after Studio granted an upstream slot. + """Streaming client-side pass-through after Unsloth granted an upstream slot. Forwards the client's OpenAI function-calling request to llama-server and relays the SSE stream back with minimal normalization (reasoning-only diff --git a/studio/backend/routes/mcp_servers.py b/studio/backend/routes/mcp_servers.py index 71f0fd2874..dc018d163a 100644 --- a/studio/backend/routes/mcp_servers.py +++ b/studio/backend/routes/mcp_servers.py @@ -82,7 +82,7 @@ def _validate_url(url: str) -> str: if _looks_like_command(trimmed): detail = ( "Local commands aren't enabled on this server. To allow them, " - "set UNSLOTH_STUDIO_ALLOW_STDIO_MCP=1 and restart Studio, or use " + "set UNSLOTH_STUDIO_ALLOW_STDIO_MCP=1 and restart Unsloth, or use " "an http:// or https:// URL instead." ) else: diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py index b8526c75e7..0806c2f513 100644 --- a/studio/backend/routes/models.py +++ b/studio/backend/routes/models.py @@ -59,59 +59,12 @@ def _safe_is_dir(path) -> bool: return False -# Hub repo id shape ("owner/name", no leading separator); anything else is -# treated as a local filesystem path. -_HF_REPO_ID_RE = re.compile(r"^[A-Za-z0-9][\w.\-]*/[\w.\-]+$") - - -def _is_hidden_model(*values: str | None) -> bool: - """True if any id/path is the RAG embedding model (EMBEDDING_MODEL or - EMBED_GGUF_REPO basename) or the llama.cpp install validation probe - (ggml-org/models / stories260K), so pickers hide them (GGUF and non-GGUF). - None are usable chat models; the probe can be cached as a side effect of - installing the prebuilt llama-server and otherwise sorts smallest, so it - would be auto-selected. A local-path embedder is matched by exact resolved - path only: a generic basename like "model" must not substring-hide - unrelated chat models.""" - from core.rag import config as rag_config - - needles = [ - # The validation probe's repo (matches the cached repo id) and its exact - # filename (matches the on-disk path). The filename carries the .gguf so - # it does not hide unrelated repos like ``user/stories260K-finetune-GGUF``. - "ggml-org/models", - "stories260k.gguf", - ] - exact_paths: list[str] = [] - for model in ( - rag_config.effective_embedding_model(), - rag_config.effective_gguf_repo(), - ): - if _HF_REPO_ID_RE.match(model): - needles.append(model.split("/")[-1].lower()) - else: - resolved = _safe_resolve(Path(model).expanduser()) - if resolved: - exact_paths.append(resolved.lower()) - for v in values: - if not v: - continue - low = v.lower() - if any(n in low for n in needles): - return True - if exact_paths: - resolved = _safe_resolve(Path(v).expanduser()) - if resolved and resolved.lower() in exact_paths: - return True - return False - - -def _safe_resolve(path: Path) -> Optional[str]: - """resolve() to a string, or None when the path is inaccessible.""" - try: - return str(path.resolve()) - except OSError: - return None +# Shared with the hub inventory scans; keep the private aliases so existing +# importers (core.inference.local_model_resolver, tests) stay valid. +from utils.hidden_models import ( + _safe_resolve, + is_hidden_model as _is_hidden_model, +) backend_path = Path(__file__).parent.parent.parent @@ -544,7 +497,7 @@ def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]: """Return a writable directory for Ollama ``.gguf`` symlinks. Prefers ``/.studio_links/`` so links sit next to their - blobs; falls back to a per-ollama-dir namespace under Studio's cache + blobs; falls back to a per-ollama-dir namespace under Unsloth's cache when the models dir is read-only (common for system installs). """ from utils.paths.storage_roots import cache_root @@ -555,7 +508,7 @@ def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]: return primary except OSError as e: logger.debug( - "Ollama dir %s not writable for .studio_links (%s); falling back to Studio cache", + "Ollama dir %s not writable for .studio_links (%s); falling back to Unsloth cache", ollama_dir, e, ) @@ -594,7 +547,7 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca model, keyed by a short hash of the manifest path, so ``detect_mmproj_file`` only sees that model's projector). Links are symlinks when possible, else hardlinks; the link dir is - ``.studio_links/`` when writable, else Studio's cache. + ``.studio_links/`` when writable, else Unsloth's cache. """ manifests_root = ollama_dir / "manifests" if not manifests_root.is_dir(): @@ -853,7 +806,7 @@ def collect_local_models(models_root: Path) -> List[LocalModelInfo]: 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)] + return [m for m in models if not _is_hidden_model(m.id, m.model_id, m.path)] @router.get("/local", response_model = LocalModelListResponse) @@ -1194,7 +1147,7 @@ def _build_browse_allowlist( """Return the root directories the folder browser may walk. The same list seeds the sidebar suggestion chips, so chip targets are - always reachable. Roots: HOME, resolved HF cache dirs, Studio's + always reachable. Roots: HOME, resolved HF cache dirs, Unsloth's outputs/exports/studio root, registered scan folders, and well-known local-LLM dirs (LM Studio, Ollama, ``~/models``); each added only if it resolves to a real directory. @@ -1486,7 +1439,7 @@ def browse_folders( "Directory to list. If omitted, defaults to the current user's " "home directory. Tilde (`~`) and relative paths are expanded. " "Must resolve inside the allowlist of browseable roots (HOME, " - "HF cache, Studio dirs, registered scan folders, well-known " + "HF cache, Unsloth dirs, registered scan folders, well-known " "model dirs)." ), ), @@ -2251,15 +2204,15 @@ async def delete_finetuned_model( gguf_variant: Optional[str] = Body(None), current_subject: str = Depends(get_current_subject), ): - """Delete a Studio-trained or exported model from disk. + """Delete an Unsloth-trained or exported model from disk. - Only paths under Studio's outputs/exports roots are accepted. + Only paths under Unsloth's outputs/exports roots are accepted. Exported GGUF entries can delete one quant variant at a time. """ if source not in {"training", "exported"}: raise HTTPException( status_code = 400, - detail = "Only trained or exported Studio models can be deleted", + detail = "Only trained or exported Unsloth models can be deleted", ) if not model_path or not model_path.strip(): @@ -2291,14 +2244,14 @@ async def delete_finetuned_model( if not _is_path_under_lexically(delete_path, allowed_root): raise HTTPException( status_code = 400, - detail = "Model path is outside Studio storage", + detail = "Model path is outside Unsloth storage", ) if export_type == "gguf" and gguf_variant: target_path = delete_path.resolve() if not _is_path_under(target_path, allowed_root): raise HTTPException( status_code = 400, - detail = "Model path is outside Studio storage", + detail = "Model path is outside Unsloth storage", ) else: target_path = delete_path @@ -2311,7 +2264,7 @@ async def delete_finetuned_model( if should_check_resolved_path and not _is_path_under(target_path, allowed_root): raise HTTPException( status_code = 400, - detail = "Model path is outside Studio storage", + detail = "Model path is outside Unsloth storage", ) if target_path == allowed_root: raise HTTPException( @@ -2778,7 +2731,11 @@ async def get_gguf_variants( ], has_vision = response.has_vision, default_variant = response.default_variant, - context_length = _read_native_context_length(repo_id, is_local = local), + # The header walk reads tokenizer arrays on dense models (tens of + # ms per uncached file); keep it off the event loop. + context_length = await asyncio.to_thread( + _read_native_context_length, repo_id, is_local = local + ), ) except HTTPException: raise @@ -3456,7 +3413,7 @@ _EXPORT_SIZE_CACHE: dict[str, tuple[int, int, str]] = {} def _is_sizable_local_path(model: str) -> bool: - """True only for local paths under a Studio data root. + """True only for local paths under an Unsloth data root. Containment is decided lexically (no filesystem access) before the path is touched, then the path is symlink-resolved and re-checked so a symlink diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py index 1ddfc0eacb..ab0fd2fd99 100644 --- a/studio/backend/routes/settings.py +++ b/studio/backend/routes/settings.py @@ -10,6 +10,7 @@ from pydantic import BaseModel, ConfigDict, Field, field_validator from auth.authentication import get_current_subject from auth.storage import rotate_preview_link_secret +from core.rag.config import default_gguf_repo, effective_gguf_repo from loggers import get_logger from utils.utils import safe_error_detail, log_and_http_error from utils.personalization_settings import ( @@ -263,14 +264,18 @@ class EmbeddingModelPayload(BaseModel): class EmbeddingModelResponse(BaseModel): embedding_model: str + embedding_gguf_repo: str default_embedding_model: str + default_embedding_gguf_repo: str is_custom: bool def _embedding_model_response() -> EmbeddingModelResponse: return EmbeddingModelResponse( embedding_model = get_rag_embedding_model(), + embedding_gguf_repo = effective_gguf_repo(), default_embedding_model = default_embedding_model(), + default_embedding_gguf_repo = default_gguf_repo(), is_custom = get_stored_embedding_model() is not None, ) diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py index d53e8f2bbc..53b1c4d991 100644 --- a/studio/backend/routes/training.py +++ b/studio/backend/routes/training.py @@ -127,9 +127,9 @@ async def start_training( try: logger.info(f"Starting training job with model: {request.model_name}") - # When Studio is driven as an inference API (API-key auth), refuse to start + # When Unsloth is driven as an inference API (API-key auth), refuse to start # training while a request is in flight: training frees VRAM by unloading - # the chat model, which would kill the stream. The Studio UI (session auth) + # the chat model, which would kill the stream. The Unsloth UI (session auth) # still starts training and coexists/frees VRAM as before. (A mixed UI+API # session is not yet special-cased.) if via_api_key is True: @@ -139,7 +139,7 @@ async def start_training( status_code = 409, detail = ( "Cannot start training over the API while an inference request is in " - "progress. Wait for it to finish, or start training from the Studio UI." + "progress. Wait for it to finish, or start training from the Unsloth UI." ), ) diff --git a/studio/backend/routes/training_vram.py b/studio/backend/routes/training_vram.py index fb361d3359..fd96fe2175 100644 --- a/studio/backend/routes/training_vram.py +++ b/studio/backend/routes/training_vram.py @@ -197,15 +197,18 @@ def can_load_chat_during_training( requested_gpu_ids: Optional[List[int]], is_gguf: bool = False, required_override_gb: Optional[float] = None, + single_device_gpu: Optional[str] = None, ) -> Tuple[bool, Dict[str, Any]]: """Decide if a NEW chat model can load without OOMing active training (inverse of can_keep_chat_during_training: training is already resident, so size the chat model against the free VRAM that remains). Sizes/places it the same way the loader will: HF auto reuses auto_select_gpu_ids; HF explicit requires an even-share per-GPU floor for device_map="balanced"; GGUF sizes from - required_override_gb over the visible pool. `load_in_4bit` must be effective - (LoRA can flip 4-bit -> 16-bit). Non-CUDA allows the load; default-deny on any - CUDA case it can't size, so a load never OOMs training.""" + required_override_gb over the visible pool. ``single_device_gpu`` is the + exact physical device token selected by a single-device runner. + `load_in_4bit` must be effective (LoRA can flip 4-bit -> 16-bit). Non-CUDA + allows the load; default-deny on any CUDA case it can't size, so a load never + OOMs training.""" try: from utils.hardware import ( DeviceType, @@ -251,26 +254,49 @@ def can_load_chat_during_training( } # Explicit GPUs, or GGUF: size directly and check live free VRAM. + if single_device_gpu is not None: + mode = "single_device" + elif is_gguf: + mode = "gguf" + else: + mode = "explicit" required_gb = required_override_gb if required_gb is None: required_gb, _meta = estimate_required_model_memory_gb(model_name, **est_kwargs) if required_gb is None: - mode = "explicit" if requested_gpu_ids else "gguf" return False, {"mode": mode, "reason": "estimate_unavailable"} free_by_index = _free_vram_by_index(get_visible_gpu_utilization().get("devices", [])) - if requested_gpu_ids: + if single_device_gpu is not None: + token = str(single_device_gpu).strip() + if not token: + # Empty token = a CPU-only single-device runner (e.g. a CPU + # diffusion GGUF): it uses no GPU VRAM, so it never threatens + # active training and can always load. + return True, {"mode": "single_device", "reason": "cpu_only"} + try: + selected_gpu = int(token) + if selected_gpu < 0: + raise ValueError + except (TypeError, ValueError): + # A non-numeric device token (e.g. a CUDA UUID / MIG handle) + # can't be mapped to a free-VRAM index, but the runner still + # drives ONE device. Size against the worst-case visible device + # (min free), never the aggregate pool, so a single-device load + # is never OK'd on capacity it can't use and OOMs training. + free_vals = [min(free_by_index.values())] if free_by_index else [] + else: + free_vals = [free_by_index.get(selected_gpu, 0.0)] + elif requested_gpu_ids: # Invalid ids -> load_model 400s first, so don't block; missing id = 0. try: resolved = resolve_requested_gpu_ids(requested_gpu_ids) except ValueError: - return True, {"mode": "explicit", "reason": "invalid_gpu_ids"} + return True, {"mode": mode, "reason": "invalid_gpu_ids"} free_vals = [free_by_index.get(i, 0.0) for i in resolved] - mode = "explicit" else: # GGUF: llama.cpp picks the GPU(s); any visible GPU is a candidate. free_vals = list(free_by_index.values()) - mode = "gguf" if not free_vals: return False, {"mode": mode, "reason": "no_visible_gpus"} diff --git a/studio/backend/run.py b/studio/backend/run.py index 4f105b53b1..398943cc2c 100644 --- a/studio/backend/run.py +++ b/studio/backend/run.py @@ -232,7 +232,7 @@ def _working_local_url(port: int) -> "str | None": def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None": """Return the IPv4 loopback URL when localhost won't reach 127.0.0.1. - Local Studio binds to 127.0.0.1. Where localhost resolves to IPv6 only (::1), + Local Unsloth binds to 127.0.0.1. Where localhost resolves to IPv6 only (::1), http://localhost: fails (or hits a different process on ::1) even though http://127.0.0.1: works. Return the IPv4 URL for the caller to surface. """ @@ -243,7 +243,7 @@ def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None": ipv4_url = f"http://127.0.0.1:{port}" - # Only warn once Studio is confirmed answering on IPv4 loopback. + # Only warn once Unsloth is confirmed answering on IPv4 loopback. if _working_local_url(port) != ipv4_url: return None @@ -265,7 +265,7 @@ def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None": if host == "::1": has_ipv6_loopback = True - # A connection to ::1 is NOT evidence Studio is reachable there: Studio binds + # A connection to ::1 is NOT evidence Unsloth is reachable there: Unsloth binds # 127.0.0.1 only, so anything on ::1 is a different process. Dual-stack # localhost is fine (browsers fall back to 127.0.0.1), so only the IPv6-only # case strands the user. @@ -287,7 +287,7 @@ def _stdout_color_ok() -> bool: def _print_localhost_ipv6_mismatch_warning(local_url: str, port: int) -> None: - """Warn that localhost points at ::1 while Studio is bound to 127.0.0.1.""" + """Warn that localhost points at ::1 while Unsloth is bound to 127.0.0.1.""" use_color = _stdout_color_ok() warn_c = "\033[38;5;215;1m" if use_color else "" reset = "\033[0m" if use_color else "" @@ -303,7 +303,7 @@ def _print_localhost_ipv6_mismatch_warning(local_url: str, port: int) -> None: def _verify_global_reachability(display_host: str, port: int) -> None: """Probe check-host.net to confirm display_host:port is reachable from the public internet. Synchronous so output lands between the banner URLs and the - stop hint. Bounded at ~15s; failures swallowed (verifier failing != Studio + stop hint. Bounded at ~15s; failures swallowed (verifier failing != Unsloth failing). Only meaningful for a wildcard bind.""" global _public_reachable # Reset to "unknown" each run; set True/False only when the probe decides. @@ -563,15 +563,15 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 " Cloudflare tunnel: ON. This Cloudflare URL is PUBLIC, and the " "raw port is also publicly reachable. --no-cloudflare disables " f"only the Cloudflare URL; bind {loopback_host} or close firewall " - "access to keep Studio private.", + "access to keep Unsloth private.", warn, ) else: _emit( " Cloudflare tunnel: ON. This is a PUBLIC internet URL: anyone " - "who has it can reach this Studio. Relaunch with --no-cloudflare " + "who has it can reach this Unsloth. Relaunch with --no-cloudflare " f"to disable the Cloudflare URL; bind {loopback_host} or close " - "firewall access to keep Studio private.", + "firewall access to keep Unsloth private.", warn, ) return @@ -580,12 +580,12 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 _emit( " Cloudflare tunnel: requested but failed to start. The raw port is " "still reachable from the public internet (see the reachability check " - "above): anyone who can reach it can access this Studio.", + "above): anyone who can reach it can access this Unsloth.", warn, ) elif _public_reachable is False: _emit( - " Cloudflare tunnel: requested but failed to start. Studio is reachable " + " Cloudflare tunnel: requested but failed to start. Unsloth is reachable " "on your local network only (no public link).", warn, ) @@ -593,7 +593,7 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 _emit( " Cloudflare tunnel: requested but failed to start. There is no " "Cloudflare public link. Raw port reachability was not verified; " - f"bind {loopback_host} or close firewall access to keep Studio private.", + f"bind {loopback_host} or close firewall access to keep Unsloth private.", warn, ) elif _cloudflare_flag: @@ -601,19 +601,19 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 _emit( " Cloudflare tunnel: OFF for this mode. The raw port is still " "reachable from the public internet (see the reachability check above): " - "anyone who can reach it can access this Studio.", + "anyone who can reach it can access this Unsloth.", warn, ) elif _public_reachable is False: _emit( - " Cloudflare tunnel: OFF for this mode. Studio is reachable on your " + " Cloudflare tunnel: OFF for this mode. Unsloth is reachable on your " "local network only (no public link)." ) else: _emit( " Cloudflare tunnel: OFF for this mode. There is no Cloudflare public " "link. Raw port reachability was not verified; " - f"bind {loopback_host} or close firewall access to keep Studio private.", + f"bind {loopback_host} or close firewall access to keep Unsloth private.", warn, ) elif _cloudflare_flag is False or _cloudflare_flag is None: @@ -624,12 +624,12 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 f" Cloudflare tunnel: OFF ({_reason}). The raw port is still " "reachable from the public internet (see the reachability check above): " "pass --cloudflare to also expose a public Cloudflare HTTPS link, or " - f"bind {loopback_host} to keep Studio private.", + f"bind {loopback_host} to keep Unsloth private.", warn, ) elif _public_reachable is False: _emit( - f" Cloudflare tunnel: OFF ({_reason}). Studio is reachable on your " + f" Cloudflare tunnel: OFF ({_reason}). Unsloth is reachable on your " "local network only. Pass --cloudflare to expose a public " "Cloudflare HTTPS link." ) @@ -638,7 +638,7 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1 f" Cloudflare tunnel: OFF ({_reason}). There is no Cloudflare " "public link. Raw port reachability was not verified; pass --cloudflare " "to expose a public Cloudflare HTTPS link, or " - f"bind {loopback_host} or close firewall access to keep Studio private.", + f"bind {loopback_host} or close firewall access to keep Unsloth private.", warn, ) @@ -674,7 +674,7 @@ def _is_port_free(host: str, port: int) -> bool: For a ``0.0.0.0`` wildcard host, also check whether anything is listening on ``127.0.0.1`` (and ``::1`` when IPv6 exists): an SSH tunnel may hold loopback - while the wildcard bind succeeds, making Studio unreachable via ``localhost``. + while the wildcard bind succeeds, making Unsloth unreachable via ``localhost``. """ import socket @@ -1087,7 +1087,7 @@ def _terminal_password_gate( ) -> Tuple[bool, bool]: """Force a terminal password change before the public tunnel goes up. - When the tunnel is about to publish Studio and the seeded admin password was + When the tunnel is about to publish Unsloth and the seeded admin password was never changed, ask for a new one (masked, confirmed) before any public URL exists. The CLI normally does this before re-exec'ing the backend; this is the backstop for direct `python run.py` launches and older-CLI installs. @@ -1147,7 +1147,7 @@ def _terminal_password_gate( ) if not deadline_arms: print( - "Refusing to publish Studio on a public Cloudflare URL: the " + "Refusing to publish Unsloth on a public Cloudflare URL: the " "default admin password was never changed, no terminal is " "attached to change it here, and the bootstrap shutdown " "deadline does not apply to this launch (api-only, or " @@ -1163,11 +1163,11 @@ def _terminal_password_gate( # terminal-attached run / reset-password instead of reading it from disk. print( " WARNING: the default admin password is still active while " - "Studio is about to be published on a public Cloudflare URL, and " + "Unsloth is about to be published on a public Cloudflare URL, and " "no terminal is attached to change it here. The public page will " "NOT auto-fill the bootstrap credential. Set a new password by " "running `unsloth studio` locally with a terminal attached, or " - "`unsloth studio reset-password`. Studio shuts down after the " + "`unsloth studio reset-password`. Unsloth shuts down after the " "bootstrap deadline (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT, default 1h) " "unless the password is changed.", file = sys.stderr, @@ -1222,7 +1222,7 @@ def _apply_supplied_password(password_value: "Optional[str]") -> None: _auth_storage.ensure_default_admin() if not _auth_storage.requires_password_change(_admin): print( - "Error: a Studio admin password is already set; --password only sets " + "Error: an Unsloth admin password is already set; --password only sets " "the initial password. Run `unsloth studio reset-password` first.", file = sys.stderr, flush = True, @@ -1337,7 +1337,7 @@ def run_server( pass # Persist a session log + native-crash stacks BEFORE importing main, so - # even import-time failures leave evidence on disk. Field report: Studio + # even import-time failures leave evidence on disk. Field report: Unsloth # "terminates without a warning" -- a native crash in the GPU runtime # kills the process with no Python traceback, and a desktop-shortcut # console closes before anything can be read. Console-only logging made @@ -1406,7 +1406,7 @@ def run_server( ensure_studio_directories() logger.info( - "Ensured Studio directories in %.1fms", + "Ensured Unsloth directories in %.1fms", (time.perf_counter() - boot_started) * 1000, ) @@ -1455,7 +1455,7 @@ def run_server( installer_bin = home / "unsloth_studio" / "bin" / "unsloth" tried_lines = "\n".join(f" - {p}" for p in attempted) or " (none)" raise SystemExit( - "[ERROR] Studio frontend build not found.\n" + "[ERROR] Unsloth frontend build not found.\n" f"Tried:\n{tried_lines}\n" "\n" "Likely cause: another 'unsloth' on PATH is shadowing the " @@ -1557,7 +1557,7 @@ def run_server( ) if not _pw_proceed: print( - "Not starting Studio; set a new admin password first, or launch " + "Not starting Unsloth; set a new admin password first, or launch " "without --secure/--cloudflare.", file = sys.stderr, flush = True, @@ -1695,7 +1695,7 @@ def run_server( logger = logger, ) logger.info( - "Studio will shut down in %ds unless the default admin password is changed.", + "Unsloth will shut down in %ds unless the default admin password is changed.", _bootstrap_timeout, ) except Exception as e: # best-effort: never block startup on the timeout @@ -1753,11 +1753,11 @@ def _build_arg_parser(): "--cloudflare", action = argparse.BooleanOptionalAction, default = None, - help = "Expose Studio on a PUBLIC internet URL via a free Cloudflare HTTPS " + help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS " "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; " "pass --cloudflare to enable it (--secure implies it), --no-cloudflare to " "force it off. It does not change a raw wildcard bind. If the admin " - "password was never changed, Studio asks for a new one in the terminal " + "password was never changed, Unsloth asks for a new one in the terminal " "before publishing the URL.", ) parser.add_argument( @@ -1767,7 +1767,7 @@ def _build_arg_parser(): help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed " "if the tunnel can't start. Without it, --no-secure also serves the raw " "0.0.0.0 port, which is reachable from anywhere on the network. If the " - "admin password was never changed, Studio asks for a new one in the " + "admin password was never changed, Unsloth asks for a new one in the " "terminal before publishing the URL.", ) # Back-compat: accept --not-secure as a hidden alias for --no-secure. diff --git a/studio/backend/startup_banner.py b/studio/backend/startup_banner.py index ea951a4325..9ec7a4f91c 100644 --- a/studio/backend/startup_banner.py +++ b/studio/backend/startup_banner.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 -"""Terminal banner for Studio startup. +"""Terminal banner for Unsloth startup. Stdlib only -- safe to import without the rest of the backend. """ @@ -172,7 +172,7 @@ def print_studio_access_banner( secondary, ), style( - " Only on trusted networks -- anyone who reaches this machine can use Studio.", + " Only on trusted networks -- anyone who reaches this machine can use Unsloth.", secondary, ), ] diff --git a/studio/backend/tests/conftest.py b/studio/backend/tests/conftest.py index b0b9ee309c..c2216104a3 100644 --- a/studio/backend/tests/conftest.py +++ b/studio/backend/tests/conftest.py @@ -101,13 +101,13 @@ def studio_server(request): @pytest.fixture def base_url(studio_server): - """Base URL for the e2e Studio server (from ``studio_server``).""" + """Base URL for the e2e Unsloth server (from ``studio_server``).""" return studio_server[0] @pytest.fixture def api_key(studio_server): - """API key for the e2e Studio server (from ``studio_server``).""" + """API key for the e2e Unsloth server (from ``studio_server``).""" return studio_server[1] diff --git a/studio/backend/tests/test_amd_apu_unified_memory.py b/studio/backend/tests/test_amd_apu_unified_memory.py index 4df9e85b30..9fd8260bf2 100644 --- a/studio/backend/tests/test_amd_apu_unified_memory.py +++ b/studio/backend/tests/test_amd_apu_unified_memory.py @@ -91,7 +91,7 @@ class TestApuRamShortfall: """On a unified-memory APU the weights load into system RAM, so a model larger than available RAM (the field case: a 64.6 GB GGUF on a WSL VM capped well below the ROCm-reported APU budget) must be refused before spawning, - not left to OOM-kill the Studio process.""" + not left to OOM-kill the Unsloth process.""" def test_field_case_wsl_cap_refuses(self): # 64.6 GB weights, ~46 GB available (WSL VM): refuse with guidance. diff --git a/studio/backend/tests/test_anthropic_compaction.py b/studio/backend/tests/test_anthropic_compaction.py index 1528eebe8b..acc0acc2e0 100644 --- a/studio/backend/tests/test_anthropic_compaction.py +++ b/studio/backend/tests/test_anthropic_compaction.py @@ -4,7 +4,7 @@ """Unit tests for Anthropic server-side context compaction wiring. Compaction is a beta (header ``compact-2026-01-12``) gated to Opus 4.6/4.7, -Sonnet 4.6, and Mythos preview. When enabled, Studio attaches +Sonnet 4.6, and Mythos preview. When enabled, Unsloth attaches ``context_management.edits[{type:"compact_20260112", trigger:{type:"input_tokens", value:N}}]``; the 50k-token minimum is clamped up so the request doesn't 400. diff --git a/studio/backend/tests/test_anthropic_fast_mode_edge.py b/studio/backend/tests/test_anthropic_fast_mode_edge.py index dd69d77590..03f5d1c0eb 100644 --- a/studio/backend/tests/test_anthropic_fast_mode_edge.py +++ b/studio/backend/tests/test_anthropic_fast_mode_edge.py @@ -330,7 +330,7 @@ def test_refusal_chunk_is_proper_openai_delta_shape(monkeypatch): def test_refusal_tool_event_chunk_shape(monkeypatch): - """Drop signal rides a Studio `_toolEvent` envelope (delta={}, + """Drop signal rides an Unsloth `_toolEvent` envelope (delta={}, finish_reason=null); the frontend latches on `_toolEvent.type == "anthropic_refusal"`.""" _, lines = _capture(monkeypatch, sse = _refusal_sse(), model = "claude-opus-4-7") @@ -409,7 +409,7 @@ def _fast_speed_sse(model: str = "claude-opus-4-7", speed: str = "fast") -> byte def test_usage_speed_propagates_to_final_usage_chunk_fast(monkeypatch): - """``usage.speed == "fast"`` from upstream must reach the Studio usage chunk.""" + """``usage.speed == "fast"`` from upstream must reach the Unsloth usage chunk.""" _, lines = _capture(monkeypatch, sse = _fast_speed_sse(speed = "fast")) usage_lines = [l for l in lines if l.startswith("data: ") and '"usage"' in l] assert usage_lines, lines @@ -428,7 +428,7 @@ def test_usage_speed_propagates_to_final_usage_chunk_standard(monkeypatch): def test_usage_speed_absent_when_anthropic_does_not_report(monkeypatch): - """Studio must not invent ``usage.speed`` when upstream omits it.""" + """Unsloth must not invent ``usage.speed`` when upstream omits it.""" _, lines = _capture(monkeypatch) parsed = [ json.loads(l[len("data: ") :]) for l in lines if l.startswith("data: ") and '"usage"' in l diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py index 3b0ea37372..9ccc3f44dd 100644 --- a/studio/backend/tests/test_anthropic_messages.py +++ b/studio/backend/tests/test_anthropic_messages.py @@ -1418,7 +1418,7 @@ class TestNormalizeAnthropicOpenAIImages: # ===================================================================== -# Studio-tool alias detection (/v1/messages tool routing) +# Unsloth-tool alias detection (/v1/messages tool routing) # ===================================================================== @@ -1436,7 +1436,7 @@ class TestAnthropicRequestedStudioTools: def test_client_tool_named_python_is_not_misclassified(self): # input_schema is the client-tool discriminator; its presence must - # prevent the name from being treated as a Studio alias. + # prevent the name from being treated as an Unsloth alias. tools = [ { "name": "python", @@ -1747,9 +1747,9 @@ class TestAnthropicMessagesToolRouting: assert "name" in exc.value.detail def test_alias_named_client_tool_without_schema_rejected_with_400(self, monkeypatch): - # Regression: a typo'd client tool whose name collides with a Studio + # Regression: a typo'd client tool whose name collides with an Unsloth # alias (e.g. a custom "python" tool missing input_schema) must - # surface a 400, not silently switch into Studio's built-in python + # surface a 400, not silently switch into Unsloth's built-in python # execution. _mock_backend(monkeypatch) payload = _basic_payload(tools = [{"name": "python"}]) @@ -1770,7 +1770,7 @@ class TestAnthropicMessagesToolRouting: def test_disable_tools_policy_overrides_server_tool_alias(self, monkeypatch): # CLI `unsloth run --disable-tools` sets policy=False. A request with - # a Studio server-tool alias must NOT enter the agentic loop then. + # an Unsloth server-tool alias must NOT enter the agentic loop then. backend = _mock_backend(monkeypatch) set_tool_policy(False) payload = _basic_payload( diff --git a/studio/backend/tests/test_cached_gguf_routes.py b/studio/backend/tests/test_cached_gguf_routes.py index d4a7cae208..b3e6255d55 100644 --- a/studio/backend/tests/test_cached_gguf_routes.py +++ b/studio/backend/tests/test_cached_gguf_routes.py @@ -120,12 +120,151 @@ def test_is_hidden_model_hides_validation_probe_everywhere(): assert models_route._is_hidden_model( None, "/hf/models--ggml-org--models/snapshots/abc/tinyllamas/stories260K.gguf" ) + # A Windows-style snapshot path must match too, even on a POSIX interpreter + # (the filename check splits on both separators). + assert models_route._is_hidden_model( + r"C:\Users\u\.cache\huggingface\hub\models--ggml-org--models\snapshots\abc\tinyllamas\stories260K.gguf" + ) assert not models_route._is_hidden_model("unsloth/gemma-3-270m-it-GGUF") # The exact-filename needle must not hide a real repo that merely # references stories260K in its name. assert not models_route._is_hidden_model("user/stories260K-finetune-GGUF") +def test_is_hidden_model_matches_repo_ids_exactly(monkeypatch): + """A custom embedder with a generic basename is hidden by EXACT repo-id + match only, so unrelated cached repos that merely contain the basename stay + visible. Regression: substring basename matching hid real chat models like + ``user/model-chat`` from the On Device inventory.""" + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF") + + # The exact embedder repo and its GGUF companion are hidden. + assert models_route._is_hidden_model("org/model") + assert models_route._is_hidden_model("org/model-GGUF") + # Unrelated repos that merely contain "model" must NOT be hidden. + assert not models_route._is_hidden_model("user/model-chat") + assert not models_route._is_hidden_model("org/model-instruct") + assert not models_route._is_hidden_model("acme/remodelled-chat") + # The validation probe stays hidden regardless of embedder config. + assert models_route._is_hidden_model("ggml-org/models") + + +def test_is_hidden_model_matches_repo_derived_local_paths(monkeypatch): + """Match exact repo-derived cache and LM Studio paths.""" + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF") + + assert models_route._is_hidden_model( + "/cache/models--org--model/snapshots/abc/model.safetensors" + ) + assert models_route._is_hidden_model( + r"C:\Users\u\.cache\huggingface\hub\models--org--model-GGUF\snapshots\abc" + ) + assert models_route._is_hidden_model("/lm-studio/org/model-GGUF/model-Q8_0.gguf") + assert not models_route._is_hidden_model("/lm-studio/user/model-chat/model-Q8_0.gguf") + assert not models_route._is_hidden_model("/cache/models--org--model-instruct") + + +def test_is_hidden_model_prefers_existing_relative_path(monkeypatch, tmp_path): + """Prefer an existing relative path over repo-id syntax.""" + from core.rag import config as rag_config + + embedder = tmp_path / "models" / "embedder" + embedder.mkdir(parents = True) + monkeypatch.chdir(tmp_path) + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "models/embedder") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF") + + assert models_route._is_hidden_model(str(embedder)) + + +def test_is_hidden_model_keeps_stale_default_embedder_hidden(monkeypatch): + """Keep default embedders hidden after a settings change.""" + from core.rag import config as rag_config + + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF") + + assert models_route._is_hidden_model("unsloth/bge-small-en-v1.5") + assert models_route._is_hidden_model("unsloth/bge-small-en-v1.5-GGUF") + assert models_route._is_hidden_model("/models/bge-small-en-v1.5") + assert models_route._is_hidden_model("/models/bge-small-en-v1.5-F16.gguf") + assert models_route._is_hidden_model(r"C:\models\bge-small-en-v1.5-Q8_0.gguf") + # Repo IDs still use exact matching, and similar local basenames must have + # a real separator after the static default name. + assert not models_route._is_hidden_model("user/bge-small-en-v1.5-chat") + assert not models_route._is_hidden_model("/models/bge-small-en-v1.50") + + +def test_is_hidden_model_keeps_env_default_hidden_after_override(monkeypatch): + """A persisted override must not expose the deployment's env default.""" + from core.rag import config as rag_config + + monkeypatch.delenv("RAG_EMBED_GGUF_REPO", raising = False) + monkeypatch.setattr(rag_config, "EMBEDDING_MODEL", "org/env-default") + monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom") + monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF") + + assert models_route._is_hidden_model("org/env-default") + assert models_route._is_hidden_model("org/env-default-GGUF") + assert models_route._is_hidden_model("org/custom") + assert models_route._is_hidden_model("org/custom-GGUF") + assert not models_route._is_hidden_model("org/env-default-chat") + + +def test_hidden_models_importable_without_heavy_model_stack(): + """The hub cache scanner imports ``is_hidden_model`` at module scope, so it + must not drag in ``utils/models/__init__`` (the model-config + checkpoint + stack). Verify in a clean interpreter that importing the helper touches + neither ``utils.models`` nor those heavy submodules, and still classifies + the probe.""" + import os + import subprocess + import textwrap + + backend = Path(__file__).resolve().parents[1] + code = textwrap.dedent( + """ + import sys + + class _Blocker: + _blocked = ( + "utils.models", + "utils.models.model_config", + "utils.models.checkpoints", + ) + + def find_spec(self, name, path=None, target=None): + if name in self._blocked: + raise ImportError("blocked heavy import: " + name) + return None + + sys.meta_path.insert(0, _Blocker()) + from utils.hidden_models import is_hidden_model + + loaded = sorted(m for m in sys.modules if m.startswith("utils.models")) + assert not loaded, loaded + assert is_hidden_model("ggml-org/models") is True + assert is_hidden_model("unsloth/gemma-3-270m-it-GGUF") is False + print("HIDDEN_MODELS_IMPORT_OK") + """ + ) + env = dict(os.environ, PYTHONPATH = str(backend)) + proc = subprocess.run( + [sys.executable, "-c", code], + capture_output = True, + text = True, + env = env, + ) + assert proc.returncode == 0, proc.stderr + assert "HIDDEN_MODELS_IMPORT_OK" in proc.stdout + + def test_list_cached_gguf_hides_llama_validation_probe(monkeypatch, tmp_path): """The ggml-org/models / stories260K install validation probe can land in the HF cache as a side effect of installing the prebuilt llama-server. diff --git a/studio/backend/tests/test_chat_load_during_training.py b/studio/backend/tests/test_chat_load_during_training.py index 63dba8579c..7daa4224aa 100644 --- a/studio/backend/tests/test_chat_load_during_training.py +++ b/studio/backend/tests/test_chat_load_during_training.py @@ -168,11 +168,14 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase): devices, required_override = None, estimate = None, + single_device_gpu = None, + gpu_ids = None, ): with ( patch("utils.hardware.get_device", return_value = DeviceType.CUDA), patch("utils.hardware.estimate_required_model_memory_gb", return_value = (estimate, {})), patch("utils.hardware.get_visible_gpu_utilization", return_value = {"devices": devices}), + patch("utils.hardware.resolve_requested_gpu_ids", return_value = gpu_ids), patch("utils.hardware.auto_select_gpu_ids") as auto_mock, ): ok, info = tv.can_load_chat_during_training( @@ -180,9 +183,10 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase): hf_token = None, load_in_4bit = True, max_seq_length = 0, - requested_gpu_ids = None, + requested_gpu_ids = gpu_ids, is_gguf = True, required_override_gb = required_override, + single_device_gpu = single_device_gpu, ) return ok, info, auto_mock @@ -198,6 +202,88 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase): ok, _, _ = self._run(devices = _devices((0, 80, 35), (1, 80, 70)), required_override = 20.0) self.assertTrue(ok) + def test_no_per_gpu_floor_for_gguf_with_explicit_gpu_ids(self): + # gpu_ids narrows llama.cpp's candidate pool but does not turn its + # self-placement into HF device_map="balanced". The uneven selected + # pair therefore keeps the aggregate GGUF check without an even-share + # floor on the nearly-full card. + ok, info, _ = self._run( + devices = _devices((0, 80, 35), (1, 80, 70), (2, 80, 0)), + required_override = 20.0, + gpu_ids = [0, 1], + ) + self.assertTrue(ok) + self.assertEqual(info["mode"], "gguf") + + def test_single_device_uses_selected_gpu(self): + # The model needs 27 GB with headroom. GPU 0 has 45 GB free, while an + # unrelated training-heavy GPU 1 has only 10 GB free. + ok, info, _ = self._run( + devices = _devices((0, 80, 35), (1, 80, 70)), + required_override = 20.0, + single_device_gpu = "0", + ) + self.assertTrue(ok) + self.assertEqual(info["usable_gb"], 45.0) + + blocked, blocked_info, _ = self._run( + devices = _devices((0, 80, 35), (1, 80, 70)), + required_override = 20.0, + single_device_gpu = "1", + ) + self.assertFalse(blocked) + self.assertEqual(blocked_info["usable_gb"], 10.0) + + def test_single_device_unresolved_token_sizes_against_worst_device(self): + # A non-numeric device token (a CUDA UUID / MIG handle) can't map to a + # free-VRAM index. The runner still drives ONE device, so size against the + # worst-case visible device (min free), not the aggregate pool: one GPU + # with 80 GB free vs a 20 GB model -> allow. + ok, info, _ = self._run( + devices = _devices((0, 80, 0)), + required_override = 20.0, + single_device_gpu = "GPU-uuid", + ) + self.assertTrue(ok) + self.assertEqual(info["mode"], "single_device") + self.assertNotIn("reason", info) + + def test_single_device_unresolved_token_refuses_when_worst_device_full(self): + # Same UUID fallback, worst-case device nearly full (2 GB for a 20 GB + # model) -> refuse (default-deny), not on an unresolved-token technicality. + ok, info, _ = self._run( + devices = _devices((0, 80, 78)), + required_override = 20.0, + single_device_gpu = "GPU-uuid", + ) + self.assertFalse(ok) + self.assertNotEqual(info.get("reason"), "unresolved_gpu_id") + + def test_single_device_unresolved_token_uses_min_free_not_aggregate(self): + # The single-device runner uses ONE device but we can't tell which from a + # UUID token. Sizing against the aggregate pool would let a 20 GB model + # "fit" 160 GB of pooled free VRAM while landing on a 2 GB card and OOMing + # training. Min-free (2 GB) is the safe worst case -> refuse. + ok, info, _ = self._run( + devices = _devices((0, 80, 78), (1, 80, 0), (2, 80, 0)), + required_override = 20.0, + single_device_gpu = "GPU-uuid", + ) + self.assertFalse(ok) + self.assertEqual(info["mode"], "single_device") + + def test_single_device_cpu_token_allows(self): + # An empty device token = a CPU-only single-device runner (CPU diffusion + # GGUF): it uses no GPU VRAM, so it never threatens training -> allow + # regardless of how full the GPUs are. + ok, info, _ = self._run( + devices = _devices((0, 80, 78)), + required_override = 20.0, + single_device_gpu = "", + ) + self.assertTrue(ok) + self.assertEqual(info["reason"], "cpu_only") + def test_estimate_unavailable_refuses(self): # No override and the estimator can't size it -> default-deny. ok, info, _ = self._run(devices = _devices((0, 80, 0)), required_override = None, estimate = None) @@ -309,6 +395,8 @@ class TestChatLoadGuardRoute(unittest.TestCase): captured = None, training_active, decision, + gpu_memory_mode = "auto", + requested_gpu_ids = None, ): config = config or SimpleNamespace(is_gguf = False, is_lora = False, path = None) with _stub_guard_deps( @@ -320,7 +408,8 @@ class TestChatLoadGuardRoute(unittest.TestCase): hf_token = None, load_in_4bit = True, max_seq_length = 0, - requested_gpu_ids = None, + requested_gpu_ids = requested_gpu_ids, + gpu_memory_mode = gpu_memory_mode, ) def test_noop_when_training_inactive(self): @@ -332,6 +421,141 @@ class TestChatLoadGuardRoute(unittest.TestCase): def test_allows_when_fits(self): self._guard(training_active = True, decision = (True, {"mode": "auto"})) + def test_diffusion_detection_uses_name_before_download(self): + config = SimpleNamespace( + identifier = "unsloth/DiffusionGemma-GGUF", + gguf_hf_repo = "unsloth/DiffusionGemma-GGUF", + gguf_file = None, + ) + self.assertTrue(self.route._classify_diffusion_gguf(config)) + + def test_uncached_gguf_classification_remains_unknown(self): + config = SimpleNamespace( + identifier = "owner/renamed-model", + gguf_hf_repo = "owner/renamed-model", + gguf_variant = "Q4_K_M", + gguf_file = None, + ) + self.assertIsNone(self.route._classify_diffusion_gguf(config)) + + def test_diffusion_detection_reuses_loader_metadata_probe(self): + import tempfile + + seen = [] + + class _Probe: + is_diffusion = False + _architecture = None + + def _read_gguf_metadata(self, path): + seen.append(path) + self.is_diffusion = True + + with tempfile.TemporaryDirectory() as d: + model = Path(d) / "renamed.gguf" + model.write_bytes(b"GGUF") + config = SimpleNamespace(identifier = "local", gguf_file = str(model)) + with patch.object(self.route, "LlamaCppBackend", _Probe): + self.assertTrue(self.route._classify_diffusion_gguf(config)) + self.assertEqual(seen, [str(model)]) + + def test_local_chat_gguf_classification_is_definitive(self): + import tempfile + class _Probe: + is_diffusion = False + _architecture = "llama" + + def _read_gguf_metadata(self, _path): + pass + + with tempfile.TemporaryDirectory() as d: + model = Path(d) / "renamed.gguf" + model.write_bytes(b"GGUF") + config = SimpleNamespace(identifier = "local", gguf_file = str(model)) + with patch.object(self.route, "LlamaCppBackend", _Probe): + self.assertFalse(self.route._classify_diffusion_gguf(config)) + + def test_manual_known_normal_gguf_bypasses_training_estimate(self): + captured = [] + config = SimpleNamespace(is_gguf = True) + with patch.object(self.route, "_classify_diffusion_gguf", return_value = False): + self._guard( + config = config, + captured = captured, + training_active = True, + decision = (False, {"reason": "must not run"}), + gpu_memory_mode = "manual", + ) + self.assertEqual(captured, []) + + def test_manual_unknown_gguf_keeps_single_device_training_guard(self): + captured = [] + config = SimpleNamespace(is_gguf = True) + with ( + patch.object(self.route, "_classify_diffusion_gguf", return_value = None), + patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5), + patch.object( + self.route.LlamaCppBackend, + "_diffusion_gpu_arg", + return_value = "2", + ), + ): + self._guard( + config = config, + captured = captured, + training_active = True, + decision = (True, {"mode": "single_device"}), + gpu_memory_mode = "manual", + ) + self.assertEqual(len(captured), 1) + self.assertEqual(captured[0]["single_device_gpu"], "2") + + def test_manual_diffusion_uses_single_device_guard(self): + captured = [] + config = SimpleNamespace(is_gguf = True) + with ( + patch.object(self.route, "_classify_diffusion_gguf", return_value = True), + patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5), + ): + self._guard( + config = config, + captured = captured, + training_active = True, + decision = (True, {"mode": "gguf"}), + gpu_memory_mode = "manual", + requested_gpu_ids = [3, 1], + ) + self.assertEqual(len(captured), 1) + self.assertEqual(captured[0]["single_device_gpu"], "1") + self.assertEqual(captured[0]["requested_gpu_ids"], [3, 1]) + + def test_unpinned_diffusion_uses_runner_default_gpu(self): + captured = [] + config = SimpleNamespace(is_gguf = True) + with ( + patch.object(self.route, "_classify_diffusion_gguf", return_value = True), + patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5), + patch.object( + self.route.LlamaCppBackend, + "_effective_gpu_count", + return_value = 2, + ), + patch.object( + self.route.LlamaCppBackend, + "_diffusion_gpu_arg", + return_value = "3", + ) as gpu_arg, + ): + self._guard( + config = config, + captured = captured, + training_active = True, + decision = (True, {"mode": "single_device"}), + gpu_memory_mode = "manual", + ) + gpu_arg.assert_called_once_with(None, cpu_only = False) + self.assertEqual(captured[0]["single_device_gpu"], "3") + def test_refuses_with_headroom_number(self): info = {"required_gb": 30.0, "usable_gb": 6.0, "needed_gb": 39.0, "mode": "auto"} with self.assertRaises(HTTPException) as exc: @@ -467,36 +691,115 @@ class TestValidateRefusesDuringTraining(unittest.TestCase): self.assertEqual(captured[0]["load_in_4bit"], False) self.assertEqual(captured[0]["max_seq_length"], 4096) - def test_rejects_gguf_with_gpu_ids_before_guard(self): - # /validate must mirror /load's GGUF + gpu_ids 400, before the VRAM guard. + def test_validate_forwards_manual_gpu_memory_mode_to_guard(self): from models.inference import ValidateModelRequest - request = ValidateModelRequest(model_path = "x.gguf", gpu_ids = [0]) + request = ValidateModelRequest( + model_path = "unsloth/model-GGUF", + gguf_variant = "Q4_K_M", + gpu_memory_mode = "manual", + ) cfg = SimpleNamespace( - identifier = "x.gguf", - display_name = "x", + identifier = "unsloth/model-GGUF", + display_name = "model-GGUF", is_gguf = True, is_lora = False, is_vision = False, path = None, base_model = None, ) - captured = [] + captured = {} with ( patch.object( self.route, "_resolve_model_identifier_for_request", - return_value = ("x.gguf", "x.gguf", False), + return_value = ("unsloth/model-GGUF", "unsloth/model-GGUF", False), ), patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg), patch.object(self.route, "load_inference_config", return_value = {}), - _stub_guard_deps(training_active = True, decision = (True, {}), captured = captured), + patch.object( + self.route, + "_guard_chat_load_against_training", + lambda config, **kw: captured.update(kw), + ), ): - with self.assertRaises(HTTPException) as exc: - asyncio.run(self.route.validate_model(request, current_subject = "u")) - self.assertEqual(exc.exception.status_code, 400) - self.assertIn("gpu_ids is not supported for GGUF", exc.exception.detail) - self.assertEqual(captured, []) # guard never reached + asyncio.run(self.route.validate_model(request, current_subject = "u")) + self.assertEqual(captured.get("gpu_memory_mode"), "manual") + + def test_validate_forwards_inherited_extras_and_parallel_to_guard(self): + # Regression: /load resolves inherited same-model extras and passes the + # real slot count to the guard; validate must do the same, else it sizes + # a smaller estimate (no inherited -c/--model-draft, n_parallel=1) and + # /load then 409s after the frontend has already unloaded. + from models.inference import ValidateModelRequest + + request = ValidateModelRequest(model_path = "unsloth/Qwen3-1.7B", max_seq_length = 4096) + cfg = SimpleNamespace( + identifier = "unsloth/Qwen3-1.7B", + display_name = "Qwen3-1.7B", + is_gguf = False, + is_lora = False, + is_vision = False, + path = None, + base_model = None, + ) + captured = {} + with ( + patch.object( + self.route, + "_resolve_model_identifier_for_request", + return_value = ("unsloth/Qwen3-1.7B", "unsloth/Qwen3-1.7B", False), + ), + patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg), + patch.object(self.route, "load_inference_config", return_value = {}), + patch.object(self.route, "_resolve_inherited_extra_args", return_value = ["-c", "32768"]), + patch.object( + self.route, + "_guard_chat_load_against_training", + lambda config, **kw: captured.update(kw), + ), + ): + asyncio.run(self.route.validate_model(request, current_subject = "u")) + self.assertEqual(captured.get("llama_extra_args"), ["-c", "32768"]) + self.assertIn("n_parallel", captured) + + def test_metadata_probe_skips_training_guard(self): + # A header-only probe (include_context_length) allocates no VRAM, so the + # training guard must not run -- else the staging GPU-layers / MoE sliders + # it feeds are hidden exactly when a during-training user needs them. + from models.inference import ValidateModelRequest + + request = ValidateModelRequest( + model_path = "unsloth/Qwen3-1.7B", + max_seq_length = 4096, + include_context_length = True, + ) + cfg = SimpleNamespace( + identifier = "unsloth/Qwen3-1.7B", + display_name = "Qwen3-1.7B", + is_gguf = False, + is_lora = False, + is_vision = False, + path = None, + base_model = None, + ) + guard_called = [] + with ( + patch.object( + self.route, + "_resolve_model_identifier_for_request", + return_value = ("unsloth/Qwen3-1.7B", "unsloth/Qwen3-1.7B", False), + ), + patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg), + patch.object(self.route, "load_inference_config", return_value = {}), + patch.object( + self.route, + "_guard_chat_load_against_training", + lambda *a, **kw: guard_called.append(True), + ), + ): + asyncio.run(self.route.validate_model(request, current_subject = "u")) + self.assertEqual(guard_called, []) # ── _estimate_gguf_required_gb (sizes the same weights the loader loads) ────── diff --git a/studio/backend/tests/test_compute_buffer.py b/studio/backend/tests/test_compute_buffer.py index 8408f8203d..3e95acc98d 100644 --- a/studio/backend/tests/test_compute_buffer.py +++ b/studio/backend/tests/test_compute_buffer.py @@ -152,7 +152,7 @@ class TestFallback: class TestParallel1Default: - """At Studio's default --parallel 1 the buffer is negligible in pipeline.""" + """At Unsloth's default --parallel 1 the buffer is negligible in pipeline.""" def test_default_n_parallel(self): est = _backend()._estimate_compute_buffer_bytes() / MIB diff --git a/studio/backend/tests/test_cpu_threads.py b/studio/backend/tests/test_cpu_threads.py index 2930c9f081..9d8795b6c0 100644 --- a/studio/backend/tests/test_cpu_threads.py +++ b/studio/backend/tests/test_cpu_threads.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 -"""Tests for Studio's early CPU thread-pool configuration.""" +"""Tests for Unsloth's early CPU thread-pool configuration.""" import ast import os @@ -30,7 +30,7 @@ def test_cpu_thread_cap_seeds_native_pool_limits(): } -# Explicit per-library values win over the Studio knob via setdefault. +# Explicit per-library values win over the Unsloth knob via setdefault. def test_cpu_thread_cap_preserves_runtime_specific_override(): env = {"UNSLOTH_CPU_THREADS": "4", "OMP_NUM_THREADS": "2"} diff --git a/studio/backend/tests/test_cuda_torch_spec.py b/studio/backend/tests/test_cuda_torch_spec.py new file mode 100644 index 0000000000..928cef787e --- /dev/null +++ b/studio/backend/tests/test_cuda_torch_spec.py @@ -0,0 +1,73 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +"""Tests for _CUDA_TORCH_PKG_SPEC in install_python_stack.py. + +The CUDA repair path installs the torch trio from an exclusive --index-url (no +PyPI fallback), so these pinned ranges decide which torch the venv gets. The +upper bound is locked to the 2.11.x family to match the base image and rocm7.2 +spec and to keep the companions off a torch-2.12 wheel that would ABI-mismatch. +""" + +from __future__ import annotations + +import sys +from pathlib import Path + +import pytest +from packaging.requirements import Requirement + +# install_python_stack.py lives at repo_root/studio/install_python_stack.py +_INSTALL_SCRIPT = Path(__file__).resolve().parents[2] / "install_python_stack.py" + + +def _load_module(monkeypatch): + """(Re-)import and return install_python_stack (mirrors test_torchao_select).""" + sys.modules.pop("install_python_stack", None) + monkeypatch.syspath_prepend(str(_INSTALL_SCRIPT.parent)) + import install_python_stack + + return install_python_stack + + +def _spec_of(pkg_spec: str): + """Parse 'torch>=2.4,<2.12.0' into a packaging SpecifierSet.""" + return Requirement(pkg_spec).specifier + + +@pytest.mark.parametrize( + "index, allowed, rejected", + [ + # torch: 2.11.x allowed (matches base image); 2.12.x excluded. + (0, ["2.11.0", "2.11.2", "2.10.0", "2.4.0"], ["2.12.0", "2.3.0", "1.13.1"]), + # torchvision: 0.26.x (torch 2.11 companion) allowed; 0.27.x (torch 2.12) out. + (1, ["0.26.0", "0.26.1", "0.19.0"], ["0.27.0", "0.18.0"]), + # torchaudio: same 2.11.x window as torch. + (2, ["2.11.0", "2.10.0", "2.4.0"], ["2.12.0", "2.3.0"]), + ], +) +def test_cuda_spec_bounds(monkeypatch, index, allowed, rejected): + mod = _load_module(monkeypatch) + spec = _spec_of(mod._CUDA_TORCH_PKG_SPEC[index]) + for v in allowed: + assert spec.contains(v, prereleases = True), f"{v} should satisfy {spec}" + for v in rejected: + assert not spec.contains(v, prereleases = True), f"{v} should not satisfy {spec}" + + +def test_cuda_spec_matches_rocm72_upper_bound(monkeypatch): + """CUDA and rocm7.2 target the same torch 2.11.x family, so their upper + bounds must stay in lockstep (bump both together at 2.12.x).""" + mod = _load_module(monkeypatch) + rocm72 = mod._ROCM_TORCH_PKG_SPECS["rocm7.2"] + + def _upper(pkg_spec: str) -> str: + for clause in _spec_of(pkg_spec): + if clause.operator == "<": + return clause.version + raise AssertionError(f"no upper bound in {pkg_spec!r}") + + for cuda_pkg, rocm_pkg in zip(mod._CUDA_TORCH_PKG_SPEC, rocm72, strict = True): + assert _upper(cuda_pkg) == _upper( + rocm_pkg + ), f"CUDA {cuda_pkg!r} upper bound must match rocm7.2 {rocm_pkg!r}" diff --git a/studio/backend/tests/test_embedding_model_security_gate.py b/studio/backend/tests/test_embedding_model_security_gate.py index 940b35d7ba..b3fa98b604 100644 --- a/studio/backend/tests/test_embedding_model_security_gate.py +++ b/studio/backend/tests/test_embedding_model_security_gate.py @@ -52,6 +52,16 @@ def client(monkeypatch): monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False) monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", "")) monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model")) + monkeypatch.setattr( + settings, + "effective_gguf_repo", + lambda: f"{saved.get('model', 'unsloth/default-embed')}-GGUF", + ) + monkeypatch.setattr( + settings, + "default_gguf_repo", + lambda: "unsloth/default-embed-GGUF", + ) app = FastAPI() app.include_router(settings.router) @@ -257,6 +267,13 @@ def test_clean_repo_saves_under_force(client, monkeypatch): r = c.put("/embedding-model", json = {"embedding_model": "acme/clean-embed", "force": True}) assert r.status_code == 200 assert saved.get("model") == "acme/clean-embed" + assert r.json() == { + "embedding_model": "acme/clean-embed", + "embedding_gguf_repo": "acme/clean-embed-GGUF", + "default_embedding_model": "unsloth/default-embed", + "default_embedding_gguf_repo": "unsloth/default-embed-GGUF", + "is_custom": True, + } def test_load_sink_refuses_flagged_model(monkeypatch): diff --git a/studio/backend/tests/test_embedding_model_settings.py b/studio/backend/tests/test_embedding_model_settings.py index 3be4af0e32..bcf3ded71c 100644 --- a/studio/backend/tests/test_embedding_model_settings.py +++ b/studio/backend/tests/test_embedding_model_settings.py @@ -53,3 +53,10 @@ def test_custom_model_overrides_default_and_derives_gguf(settings_store, monkeyp assert ems.reset_rag_embedding_model() == rag_config.EMBEDDING_MODEL assert ems.get_stored_embedding_model() is None + + +def test_env_default_derives_its_gguf_companion(monkeypatch): + monkeypatch.delenv("RAG_EMBED_GGUF_REPO", raising = False) + monkeypatch.setattr(rag_config, "EMBEDDING_MODEL", "org/env-default-embedder") + + assert rag_config.default_gguf_repo() == "org/env-default-embedder-GGUF" diff --git a/studio/backend/tests/test_frontend_resolution.py b/studio/backend/tests/test_frontend_resolution.py index c3e0524a30..7ac2717aae 100644 --- a/studio/backend/tests/test_frontend_resolution.py +++ b/studio/backend/tests/test_frontend_resolution.py @@ -218,7 +218,7 @@ def test_systemexit_message_contains_actionable_fixes(tmp_path, monkeypatch): installer_bin = home / "unsloth_studio" / "bin" / "unsloth" tried_lines = "\n".join(f" - {p}" for p in attempted) message = ( - "[ERROR] Studio frontend build not found.\n" + "[ERROR] Unsloth frontend build not found.\n" f"Tried:\n{tried_lines}\n" "\n" "Likely cause: another 'unsloth' on PATH is shadowing the " diff --git a/studio/backend/tests/test_gemini_provider.py b/studio/backend/tests/test_gemini_provider.py index 85ceb04d27..c6ffa798d0 100644 --- a/studio/backend/tests/test_gemini_provider.py +++ b/studio/backend/tests/test_gemini_provider.py @@ -768,7 +768,7 @@ def test_cached_content_pass_through(monkeypatch): def test_boolean_caching_does_not_set_cached_content(monkeypatch): - """Studio's existing True/False signals shouldn't fabricate a cache id.""" + """Unsloth's existing True/False signals shouldn't fabricate a cache id.""" captured = _capture_body(monkeypatch, enable_prompt_caching = True) assert "cachedContent" not in captured["body"] @@ -2613,7 +2613,7 @@ def test_gemini_native_skips_orphan_function_response_for_native_part_replay(mon def test_gemini_native_part_falls_back_to_args_google(monkeypatch): """Round 27: a direct OpenAI-compat API caller (or imported third-party - thread) cannot use Studio's non-standard `tool_calls[].extra_content` + thread) cannot use Unsloth's non-standard `tool_calls[].extra_content` field, so the native_part payload round-trips through `function.arguments` as `{"google": {"native_part": {...}}}`. The synthetic-builtin detector recognizes that location, but the replay branch was only reading from diff --git a/studio/backend/tests/test_gemma4_chat_template_override.py b/studio/backend/tests/test_gemma4_chat_template_override.py index f726741aa5..9fb24a4cf6 100644 --- a/studio/backend/tests/test_gemma4_chat_template_override.py +++ b/studio/backend/tests/test_gemma4_chat_template_override.py @@ -3,7 +3,7 @@ """Auto-override of the chat template for ``unsloth/gemma-4-*-GGUF``. -Studio ships a bundled ``gemma-4.jinja`` (PR #118 based, ``preserve_thinking`` +Unsloth ships a bundled ``gemma-4.jinja`` (PR #118 based, ``preserve_thinking`` defaulted off) and applies it to gemma-4 GGUF loads via the existing ``chat_template_override`` -> ``--chat-template-file`` path, so users do not need to re-download quants. Pins the family matcher, the resolver precedence, the diff --git a/studio/backend/tests/test_gguf_metadata.py b/studio/backend/tests/test_gguf_metadata.py index a5be07f8e3..ec0330ce05 100644 --- a/studio/backend/tests/test_gguf_metadata.py +++ b/studio/backend/tests/test_gguf_metadata.py @@ -15,6 +15,7 @@ from utils.models.gguf_metadata import ( pairing_score, read_gguf_context_length, read_gguf_general_metadata, + read_gguf_staged_dims, read_mmproj_audio_capability, ) @@ -153,6 +154,78 @@ def test_context_length_ignores_foreign_arch_key(tmp_path: Path): assert read_gguf_context_length(str(p)) is None +# --- read_gguf_staged_dims (one pass: context + layer + moe counts) ---- + + +def test_staged_dims_none_for_missing_or_non_gguf(tmp_path: Path): + assert read_gguf_staged_dims(str(tmp_path / "nope.gguf")) is None + p = tmp_path / "garbage.gguf" + p.write_bytes(b"not a gguf at all") + assert read_gguf_staged_dims(str(p)) is None + + +def test_staged_dims_moe_with_leading_dense(tmp_path: Path): + # GLM-4.7-Flash shape: context + total layers + MoE layers in one read. + p = _write_synthetic_gguf( + tmp_path / "glm.gguf", + {"general.architecture": "deepseek2"}, + extra_uint32 = { + "deepseek2.context_length": 202752, + "deepseek2.block_count": 47, + "deepseek2.expert_count": 64, + "deepseek2.leading_dense_block_count": 1, + }, + ) + assert read_gguf_staged_dims(str(p)) == { + "context_length": 202752, + "layer_count": 47, + "moe_layer_count": 46, + } + + +def test_staged_dims_dense_model(tmp_path: Path): + # Dense: layer_count present, moe_layer_count 0 (slider hidden). + p = _write_synthetic_gguf( + tmp_path / "dense.gguf", + {"general.architecture": "qwen3"}, + extra_uint32 = {"qwen3.context_length": 40960, "qwen3.block_count": 36}, + ) + assert read_gguf_staged_dims(str(p)) == { + "context_length": 40960, + "layer_count": 36, + "moe_layer_count": 0, + } + + +def test_staged_dims_all_moe_no_leading_dense(tmp_path: Path): + # Experts present, no leading_dense key -> every block is a MoE layer. + p = _write_synthetic_gguf( + tmp_path / "moe.gguf", + {"general.architecture": "qwen35moe"}, + extra_uint32 = {"qwen35moe.block_count": 40, "qwen35moe.expert_count": 256}, + ) + assert read_gguf_staged_dims(str(p)) == { + "context_length": None, + "layer_count": 40, + "moe_layer_count": 40, + } + + +def test_staged_dims_uint64_block_count(tmp_path: Path): + # block_count stored as uint64 (vtype 10) still parses; moe == block_count. + p = _write_synthetic_gguf( + tmp_path / "moe64.gguf", + {"general.architecture": "gpt-oss"}, + extra_uint32 = {"gpt-oss.expert_count": 32}, + extra_uint64 = {"gpt-oss.block_count": 24}, + ) + assert read_gguf_staged_dims(str(p)) == { + "context_length": None, + "layer_count": 24, + "moe_layer_count": 24, + } + + def test_context_length_read_from_uint64(tmp_path: Path): # Some models store context_length as a uint64 (vtype 10). p = _write_synthetic_gguf( diff --git a/studio/backend/tests/test_gpu_memory_mode.py b/studio/backend/tests/test_gpu_memory_mode.py new file mode 100644 index 0000000000..b17274197f --- /dev/null +++ b/studio/backend/tests/test_gpu_memory_mode.py @@ -0,0 +1,879 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +"""Backend contract for the GPU Memory mode dropdown. + +The dropdown threads a single ``gpu_memory_mode`` ("auto" | "manual") from the +chat UI through the load request. "manual" lets the user own the offload: with +``gpu_layers < 0`` (Auto, the default) it hands all memory management to +llama.cpp's ``--fit on`` (no CUDA/HIP device masking, no context auto-reduce, no +gpu-layer or tensor-split planning); with ``gpu_layers >= 0`` it pins the layers +and MoE offload itself (``--fit off``). These tests pin: + + * the pydantic request/response/status contract (snake_case key, default + "auto", unknown values rejected), + * the backend ``gpu_memory_mode`` property and its reset on unload, + * the ``_already_in_target_state`` reload-detection branch, and + * that the manual + Auto-layers branch in ``load_model`` empties the probed + GPU set and drops tensor parallelism so the selection below no-ops, while + the explicit-offload branch emits ``--gpu-layers`` / ``--fit off``. +""" + +from __future__ import annotations + +import inspect +import sys +import types as _types +from pathlib import Path + +import pytest + +_BACKEND_DIR = str(Path(__file__).resolve().parent.parent) +if _BACKEND_DIR not in sys.path: + sys.path.insert(0, _BACKEND_DIR) + +# Same external-dep stubs as the other llama_cpp unit tests so importing +# the backend doesn't drag in structlog / httpx / loggers. +_loggers_stub = _types.ModuleType("loggers") +_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name) +sys.modules.setdefault("loggers", _loggers_stub) + +_structlog_stub = _types.ModuleType("structlog") +_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub") +sys.modules.setdefault("structlog", _structlog_stub) + +# httpx is a real, installed backend dependency: import it so the genuine module +# is in sys.modules. A hand-rolled stub here is inevitably incomplete and, since +# setdefault installs it before real httpx loads, would poison a combined pytest +# run -- routes/inference references httpx.Response (and other attrs) at def time. +import httpx # noqa: F401 + +from core.inference import llama_cpp as llama_cpp_module +from core.inference.llama_cpp import LlamaCppBackend +from models.inference import ( + InferenceStatusResponse, + LoadRequest, + LoadResponse, +) + + +# ── Pydantic contract (snake_case key, default "auto") ─────────────── + + +def test_load_request_defaults_gpu_memory_mode_auto(): + assert LoadRequest(model_path = "owner/repo").gpu_memory_mode == "auto" + + +def test_load_request_round_trips_json_key(): + req = LoadRequest.model_validate({"model_path": "owner/repo", "gpu_memory_mode": "manual"}) + assert req.gpu_memory_mode == "manual" + assert req.model_dump()["gpu_memory_mode"] == "manual" + + +def test_load_request_rejects_unknown_mode(): + with pytest.raises(ValueError): + LoadRequest(model_path = "owner/repo", gpu_memory_mode = "bogus") + + +@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse]) +def test_response_models_emit_gpu_memory_mode(model_cls): + if model_cls is LoadResponse: + default = model_cls( + status = "loaded", + model = "owner/repo", + display_name = "repo", + inference = {}, + ) + manual = model_cls( + status = "loaded", + model = "owner/repo", + display_name = "repo", + inference = {}, + gpu_memory_mode = "manual", + ) + else: + default = model_cls() + manual = model_cls(gpu_memory_mode = "manual") + assert default.model_dump()["gpu_memory_mode"] == "auto" + assert manual.model_dump()["gpu_memory_mode"] == "manual" + + +# ── Backend property + reset ───────────────────────────────────────── + + +class _FakeProcess: + """Stand-in for subprocess.Popen so _kill_process is a no-op.""" + + def terminate(self): + pass + + def wait(self, timeout = None): + return 0 + + def kill(self): + pass + + def poll(self): + return 0 + + +def test_gpu_memory_mode_property_defaults_auto(): + assert LlamaCppBackend().gpu_memory_mode == "auto" + + +def test_gpu_memory_mode_property_reflects_field(): + backend = LlamaCppBackend() + backend._gpu_memory_mode = "manual" + assert backend.gpu_memory_mode == "manual" + + +def test_unload_resets_gpu_memory_mode(): + backend = LlamaCppBackend() + backend._process = _FakeProcess() + backend._gpu_memory_mode = "manual" + backend.unload_model() + assert backend.gpu_memory_mode == "auto" + + +# ── _already_in_target_state reload-detection branch ───────────────── + + +def _loaded_backend(gpu_memory_mode: str) -> LlamaCppBackend: + backend = LlamaCppBackend() + backend._process = _FakeProcess() # is_loaded only checks "is not None" + backend._healthy = True + backend._model_identifier = "owner/repo" + backend._hf_variant = "Q4_K_M" + backend._requested_n_ctx = 8192 + backend._cache_type_kv = None + backend._requested_spec_mode = "auto" + backend._chat_template_override = None + backend._is_vision = False + backend._extra_args = None + backend._gguf_path = None + backend._gpu_memory_mode = gpu_memory_mode + return backend + + +def _target_state(backend: LlamaCppBackend, gpu_memory_mode: str) -> bool: + return backend._already_in_target_state( + gguf_path = None, + model_identifier = "owner/repo", + hf_variant = "Q4_K_M", + n_ctx = 8192, + cache_type_kv = None, + speculative_type = "auto", + chat_template_override = None, + extra_args = None, + is_vision = False, + gpu_memory_mode = gpu_memory_mode, + ) + + +@pytest.mark.parametrize("mode", ["auto", "manual"]) +def test_already_in_target_state_matches_same_mode(mode): + assert _target_state(_loaded_backend(mode), mode) is True + + +@pytest.mark.parametrize("loaded,requested", [("auto", "manual"), ("manual", "auto")]) +def test_already_in_target_state_reloads_on_mode_change(loaded, requested): + # Flipping the dropdown either direction must force a reload so the command + # is rebuilt with/without the Unsloth GPU masking. + assert _target_state(_loaded_backend(loaded), requested) is False + + +def test_already_in_target_state_ignores_mode_for_diffusion(): + # The diffusion runner is mode-agnostic (always "auto"), so a standing manual + # preference must not force a needless reload. + backend = _loaded_backend("auto") + backend._is_diffusion = True + assert _target_state(backend, "manual") is True + + +# ── load_model: manual + Auto layers bypasses Unsloth GPU management ── + + +def _load_model_source() -> str: + return inspect.getsource(llama_cpp_module.LlamaCppBackend.load_model) + + +def test_auto_layers_branch_empties_gpus_and_drops_tensor_parallel(): + # Emptying the probed set makes the selection / TP planning below no-op, so + # gpu_indices stays None and use_fit True (--fit on). + src = _load_model_source() + gate = src.find('if gpu_memory_mode == "manual" and gpu_layers < 0:') + assert gate != -1, "load_model must branch on manual + Auto layers (gpu_layers < 0)" + block = src[gate : gate + 1400] + assert "gpus = []" in block, "Auto-layers branch must empty the probed GPU set" + # --fit aborts under --split-mode tensor, so a raw-extras split-mode is stripped. + assert "strip_split_mode_only(extra_args)" in block + assert "requested_ctx if requested_ctx > 0 else 0" in block + # The branch sits before GPU selection assigns gpu_indices; --fit on is its emission. + assert gate < src.find("gpu_indices, use_fit = None, True") + assert 'cmd.extend(["--fit", "on"])' in src + # TP drops for this path, but at a guard BEFORE the quantized-KV cache-drop, so + # a requested quantized cache survives into the --fit load. + tp_drop = src.find('if tensor_parallel and gpu_memory_mode == "manual" and gpu_layers < 0:') + assert tp_drop != -1, "manual + Auto layers must drop tensor_parallel" + assert "tensor_parallel = False" in src[tp_drop : tp_drop + 400] + cache_drop = src.find("Tensor parallelism requires a non-quantized KV cache") + assert cache_drop != -1 + assert ( + tp_drop < cache_drop + ), "TP must drop before the cache-drop so a quantized KV survives --fit" + + +def test_auto_layers_never_sends_ctx_size_zero(): + # Sending "-c 0" sets fit_params_min_ctx = UINT32_MAX in llama.cpp, pinning + # the full native context and disabling --fit's reduction. So the base cmd + # must never carry -c, "-c 0" is emitted only outside the Auto-layers (--fit) + # case, and a positive context is passed through (which --fit optimizes + # layers around). + src = _load_model_source() + base_start = src.find("cmd = [") + base_end = src.find("\n ]", base_start) + base_block = src[base_start:base_end] + assert '"-c"' not in base_block, "-c must be conditional, not in the base cmd list" + assert 'cmd.extend(["-c", str(effective_ctx)])' in src, "positive ctx must pass -c" + assert 'auto_fit = gpu_memory_mode == "manual" and gpu_layers < 0' in src + zero = src.find('cmd.extend(["-c", "0"])') + assert zero != -1, '"-c 0" emission must exist outside the Auto-layers case' + guard = src.rfind("elif not auto_fit:", 0, zero) + assert guard != -1 and zero - guard < 120, '"-c 0" must sit under the not-auto_fit guard' + + +def test_manual_mode_clears_inherited_main_model_placement_env(): + env = {name: "inherited" for name in LlamaCppBackend._MANUAL_PLACEMENT_ENV_VARS} + env["LLAMA_ARG_N_GPU_LAYERS_DRAFT"] = "7" + env["UNRELATED"] = "kept" + + LlamaCppBackend._clear_manual_placement_env(env) + + assert not (set(env) & set(LlamaCppBackend._MANUAL_PLACEMENT_ENV_VARS)) + assert env["LLAMA_ARG_N_GPU_LAYERS_DRAFT"] == "7" + assert env["UNRELATED"] == "kept" + + +def test_load_model_sanitizes_manual_env_after_building_child_env(): + src = _load_model_source() + env_build = src.find("env = self._llama_server_env_for_binary(binary)") + env_clear = src.find("self._clear_manual_placement_env(env)", env_build) + launch = src.find("subprocess.Popen", env_build) + assert env_build != -1 + assert env_build < env_clear < launch + + +# ── Manual offload (--gpu-layers + --fit off + --n-cpu-moe) ─────────── + + +def test_load_request_accepts_manual(): + req = LoadRequest( + model_path = "owner/repo", + gpu_memory_mode = "manual", + gpu_layers = 20, + n_cpu_moe = 8, + tensor_split = [2, 1], + ) + assert req.gpu_memory_mode == "manual" + assert req.gpu_layers == 20 + assert req.n_cpu_moe == 8 + assert req.tensor_split == [2, 1] + + +def test_load_request_manual_defaults(): + req = LoadRequest(model_path = "owner/repo") + assert req.gpu_layers == -1 + assert req.n_cpu_moe == 0 + assert req.tensor_split is None + + +@pytest.mark.parametrize("bad", [[0, 0], [-1, 2], [float("inf"), 1], [float("nan"), 1]]) +def test_load_request_rejects_degenerate_tensor_split(bad): + # A negative/non-finite/all-zero split is dropped at launch but compared raw + # in the reload dedupe, so it would reload forever -- reject it up front. + with pytest.raises(ValueError): + LoadRequest(model_path = "owner/repo", tensor_split = bad) + + +@pytest.mark.parametrize("good", [[2, 1], [1, 1], [], None]) +def test_load_request_accepts_valid_tensor_split(good): + assert LoadRequest(model_path = "owner/repo", tensor_split = good).tensor_split == good + + +def test_route_normalizes_explicit_extras_before_reload_dedupe(): + route_src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8") + load_impl = route_src[route_src.index("async def _load_model_impl") :] + strip = load_impl.index("_stripped_explicit = strip_shadowing_flags") + normalize = load_impl.index( + 'request = request.model_copy(update = {"llama_extra_args": extra_llama_args})' + ) + dedupe = load_impl.index("and _request_matches_loaded_settings(") + assert strip < normalize < dedupe + + +@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse]) +def test_response_models_emit_manual_fields(model_cls): + if model_cls is LoadResponse: + obj = model_cls( + status = "loaded", + model = "owner/repo", + display_name = "repo", + inference = {}, + gpu_memory_mode = "manual", + gpu_layers = 20, + n_cpu_moe = 8, + tensor_split = [2, 1], + n_layers = 32, + n_moe_layers = 32, + ) + else: + obj = model_cls( + gpu_memory_mode = "manual", + gpu_layers = 20, + n_cpu_moe = 8, + tensor_split = [2, 1], + n_layers = 32, + n_moe_layers = 32, + ) + dumped = obj.model_dump() + assert dumped["gpu_memory_mode"] == "manual" + assert dumped["gpu_layers"] == 20 + assert dumped["n_cpu_moe"] == 8 + assert dumped["tensor_split"] == [2, 1] + assert dumped["n_layers"] == 32 + assert dumped["n_moe_layers"] == 32 + + +def test_manual_properties_default_and_reflect_and_reset(): + backend = LlamaCppBackend() + assert backend.gpu_layers == -1 and backend.n_cpu_moe == 0 + assert backend.tensor_split is None + backend._gpu_layers = 20 + backend._n_cpu_moe = 8 + backend._tensor_split = [2, 1] + assert backend.gpu_layers == 20 and backend.n_cpu_moe == 8 + assert backend.tensor_split == [2, 1] + backend._process = _FakeProcess() + backend.unload_model() + assert backend.gpu_layers == -1 and backend.n_cpu_moe == 0 + assert backend.tensor_split is None + + +def test_n_moe_layers_property(): + # 0 for a dense model (hides the slider); block_count for all-MoE; + # block_count - leading_dense otherwise (GLM-4.7-Flash: 47 - 1 -> 46). + b = LlamaCppBackend() + b._n_layers = 36 + b._n_experts = None + assert b.n_moe_layers == 0 + b._n_experts = 128 + b._leading_dense_block_count = None + assert b.n_moe_layers == 36 + b._n_layers = 47 + b._leading_dense_block_count = 1 + assert b.n_moe_layers == 46 + + +def _target_state_manual( + backend, + *, + gpu_layers, + n_cpu_moe, + tensor_split = None, +): + return backend._already_in_target_state( + gguf_path = None, + model_identifier = "owner/repo", + hf_variant = "Q4_K_M", + n_ctx = 8192, + cache_type_kv = None, + speculative_type = "auto", + chat_template_override = None, + extra_args = None, + is_vision = False, + gpu_memory_mode = "manual", + gpu_layers = gpu_layers, + n_cpu_moe = n_cpu_moe, + tensor_split = tensor_split, + ) + + +def test_manual_reloads_on_gpu_layers_or_n_cpu_moe_or_split_change(): + backend = _loaded_backend("manual") + backend._gpu_layers = 20 + backend._n_cpu_moe = 0 + backend._tensor_split = None + # Same knobs -> no reload. + assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0) is True + # Changed layer count -> reload. + assert _target_state_manual(backend, gpu_layers = 16, n_cpu_moe = 0) is False + # Changed MoE offload -> reload. + assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 8) is False + # Added a GPU split -> reload. + assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0, tensor_split = [2, 1]) is False + # Same GPU split -> no reload. + backend._tensor_split = [2, 1] + assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0, tensor_split = [2, 1]) is True + + +def test_auto_layers_reload_tracks_only_gpu_layers(): + # Under Auto (gpu_layers < 0) the MoE/split knobs don't apply, so a leftover + # request value must not reload -- only a gpu_layers change (Auto -> pinned) does. + backend = _loaded_backend("manual") + backend._gpu_layers = -1 + backend._n_cpu_moe = 0 + backend._tensor_split = None + # Same Auto, leftover MoE/split in the request -> still no reload. + assert _target_state_manual(backend, gpu_layers = -1, n_cpu_moe = 8, tensor_split = [2, 1]) is True + # Auto -> explicit offload reloads. + assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0) is False + + +def test_manual_offload_emits_gpu_layers_fit_off_and_n_cpu_moe(): + src = _load_model_source() + gate = src.find('elif gpu_memory_mode == "manual":') + assert gate != -1, "load_model must have an explicit-offload manual branch" + block = src[gate : gate + 700] + # Empties the probed set (skips the planner) but keeps the user's TP choice + # (only the Auto-layers branch above drops TP). + assert "gpus = []" in block + assert "tensor_parallel = False" not in block + # The cmd emits the layer count with fit disabled, gated on gpu_layers >= 0. + assert 'if gpu_memory_mode == "manual" and gpu_layers >= 0:' in src + assert 'cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"])' in src + # MoE offload uses --n-cpu-moe via _resolve_cpu_moe_flag (tested behaviorally below). + assert "_resolve_cpu_moe_flag(" in src + assert 'cmd.extend(["--n-cpu-moe", str(moe_flag)])' in src + # A count requested on a dense model is never emitted, so it must also be + # dropped from the recorded state -- else /status and /load report a count + # llama-server never received (same rule as the tensor-split drop below). + moe_emit = src.find('cmd.extend(["--n-cpu-moe", str(moe_flag)])') + assert "elif n_cpu_moe:" in src[moe_emit : moe_emit + 300] + assert "self._n_cpu_moe = 0" in src[moe_emit : moe_emit + 300] + # The offload path forces use_fit False so --fit-ctx is never added under --fit off. + emit = src.find('cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"])') + assert "use_fit = False" in src[src.rfind("\n", 0, emit) - 200 : emit + 80] + + +def test_status_reports_requested_context_length(): + # The hydration path re-seeds a Manual+Auto context pin from the REQUESTED + # n_ctx (0 = Auto); context_length only exposes the resolved value. + assert "requested_context_length" in InferenceStatusResponse.model_fields + s = InferenceStatusResponse(requested_context_length = 8192) + assert s.model_dump()["requested_context_length"] == 8192 + assert InferenceStatusResponse().model_dump()["requested_context_length"] is None + # The /status route must actually wire it from the backend (a declared-but- + # never-populated field would leave hydration silently reverting the pin). + from pathlib import Path as _P + + route_src = (_P(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8") + assert "requested_context_length = llama_backend.requested_n_ctx" in route_src + + +def test_manual_offload_emits_tensor_split(): + # The offload path emits --tensor-split from the per-GPU shares, only when + # provided, with >1 GPU in use, AND matching that count (a stale ratio on a + # narrowed picker or a mismatched direct-API list must not emit -- llama- + # server aborts on a split/GPU-count mismatch). + src = _load_model_source() + assert "if tensor_split and _split_gpus > 1:" in src + # Emit only on a length match AND a positive sanitized total: a mismatched + # or all-zero split aborts llama-server / assigns nothing, so it's dropped. + # The emitted list is the sanitized one (clamping tested behaviorally below). + assert "_sanitized_split = self._sanitize_tensor_split(tensor_split)" in src + assert "if len(_sanitized_split) == _split_gpus and _split_total > 0:" in src + assert '"--tensor-split"' in src + # Joined as a comma list (e.g. "2,1") within the explicit-offload cmd branch. + gate = src.find('if gpu_memory_mode == "manual" and gpu_layers >= 0:') + nxt = src.find("elif use_fit:", gate) + assert '","' in src[gate:nxt] and "tensor_split" in src[gate:nxt] + # A split with a single effective GPU is never emitted, so it must also be + # dropped from the recorded state -- else /status and /load report a ratio + # llama-server never received and the dedupe baseline preserves it. + assert "elif tensor_split:" in src[gate:nxt] + drop = src.find("elif tensor_split:", gate, nxt) + assert "self._tensor_split = None" in src[drop : drop + 250] + + +def test_sanitize_tensor_split_clamps_negative_and_non_finite(): + # Negative entries would launch a placement different from the ratio the + # UI showed; inf passes a plain > 0 total gate and would emit + # "--tensor-split inf,..." (llama.cpp normalizes shares by the running + # total, so an inf poisons the shares from that entry on). Both clamp to 0. + sanitize = LlamaCppBackend._sanitize_tensor_split + assert sanitize([2, 1]) == [2.0, 1.0] + assert sanitize([-1, 2]) == [0.0, 2.0] + assert sanitize([float("inf"), 1]) == [0.0, 1.0] + assert sanitize([float("nan"), 1]) == [0.0, 1.0] + # All-zero survives sanitization; the call site's total gate drops it. + assert sanitize([0, 0]) == [0.0, 0.0] + # Unreadable input -> []; the call site's length gate drops it. + assert sanitize(["x", 1]) == [] + assert sanitize([10**400, 1]) == [] + + +def test_zero_offload_mask_honors_device_pin_spellings(): + # A user device pin must keep the GPUs visible: llama-server aborts on a + # pin it can't see ('error: invalid device'). The pin can arrive as + # --device or its -dev alias, as the draft forms (parsed even with no + # drafter loaded), or as an inherited LLAMA_ARG_DEVICE env var. + load_src = _load_model_source() + assert "self._zero_offload_keeps_gpu_visible(cmd, env)" in load_src + block = inspect.getsource(LlamaCppBackend._cmd_has_gpu_device_pin) + for flag in ( + '"--device"', + '"-dev"', + '"--spec-draft-device"', + '"-devd"', + '"--device-draft"', + ): + assert flag in block + assert '"LLAMA_ARG_DEVICE"' in block + + +def test_resolve_cpu_moe_flag(): + # Clamp the requested MoE-layer count to the model's MoE layers, then offset + # past leading dense layers (--n-cpu-moe counts from layer 0). + R = LlamaCppBackend._resolve_cpu_moe_flag + assert R(0, 40, 0) is None # nothing requested + assert R(8, 0, 0) is None # dense model (no MoE layers) + assert R(8, 40, 0) == 8 # all-MoE: direct + assert R(100, 40, 0) == 40 # clamp to the MoE layer count + # GLM-4.7-Flash (deepseek2): block_count 47, leading_dense 1, n_moe 46. + assert R(5, 46, 1) == 6 # offset past the 1 dense layer + assert R(46, 46, 1) == 47 # all MoE on CPU == block_count + + +def test_manual_allows_tensor_parallel_via_split_mode(): + # Manual offload keeps the user's TP choice but skips the memory-based planner + # (plan_tp excludes manual, so its empty gpu set can't downgrade TP). The + # --split-mode tensor emission gates on tensor_parallel alone, so manual + # reaches it -- with tp_tensor_split None it's an even split (no + # --tensor-split). --fit off means no fit/tensor abort. + src = _load_model_source() + assert 'plan_tp = tensor_parallel and gpu_memory_mode != "manual"' in src + assert "if plan_tp:" in src + assert "if plan_tp and len(tp_gpus) < 2:" in src + sm = src.find('cmd.extend(["--split-mode", "tensor"])') + assert sm != -1, "TP must emit --split-mode tensor" + guard = src.rfind("if tensor_parallel:", 0, sm) + assert guard != -1 and sm - guard < 200, "split-mode gates on tensor_parallel" + # The tensor-split is only emitted for a planned (non-even) split, which + # manual never produces, so manual stays an even split. + assert "if tp_tensor_split and len(tp_tensor_split) > 1:" in src + + +def test_fit_sets_target_margin(): + # Manual + Auto (auto_fit) tightens the per-device VRAM margin to 512 MiB. + caps = {"supports_fit_target": True} + flags = LlamaCppBackend._ctx_integrity_flags(1, True, True, 0, 0, caps) + assert flags[flags.index("--fit-target") + 1] == "512" + # Not emitted on the legacy auto path (fit on but not auto_fit): -c 0 pins + # native there, so the tighter margin must not ride along. + assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags(1, True, False, 0, 0, caps) + # Not emitted when fit is off. + assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags(1, False, False, 0, 0, caps) + # Not emitted when the binary lacks support. + assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags( + 1, True, True, 0, 0, {"supports_fit_target": False} + ) + + +# ── GPU picker (gpu_ids -> CUDA_VISIBLE_DEVICES) ───────────────────── + + +def test_load_request_accepts_gpu_ids(): + req = LoadRequest(model_path = "owner/repo", gpu_ids = [1, 0]) + assert req.gpu_ids == [1, 0] + assert LoadRequest(model_path = "owner/repo").gpu_ids is None + + +@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse]) +def test_response_models_emit_gpu_ids(model_cls): + if model_cls is LoadResponse: + obj = model_cls(status = "loaded", model = "m", display_name = "m", inference = {}, gpu_ids = [1]) + else: + obj = model_cls(gpu_ids = [1]) + assert obj.model_dump()["gpu_ids"] == [1] + + +def test_gpu_ids_property_default_and_reset(): + backend = LlamaCppBackend() + assert backend.gpu_ids is None + backend._gpu_ids = [0, 1] + assert backend.gpu_ids == [0, 1] + backend._process = _FakeProcess() + backend.unload_model() + assert backend.gpu_ids is None + + +def _target_state_gpu_ids(backend, gpu_ids): + return backend._already_in_target_state( + gguf_path = None, + model_identifier = "owner/repo", + hf_variant = "Q4_K_M", + n_ctx = 8192, + cache_type_kv = None, + speculative_type = "auto", + chat_template_override = None, + extra_args = None, + is_vision = False, + gpu_ids = gpu_ids, + ) + + +def test_gpu_ids_reload_detection_is_order_insensitive(): + backend = _loaded_backend("auto") + backend._gpu_ids = [0, 1] + # Same set, different order -> no reload. + assert _target_state_gpu_ids(backend, [1, 0]) is True + # Different set -> reload. + assert _target_state_gpu_ids(backend, [0]) is False + # Dropping the pick (auto) -> reload. + assert _target_state_gpu_ids(backend, None) is False + + +def test_gpu_ids_reload_detection_collapses_diffusion_to_single_device(): + # The diffusion runner drives only its single lowest device, so the backend + # records [lowest]. A later multi-GPU request that still resolves to that + # same lowest device must dedupe (no needless reload); a request whose lowest + # device moves, or that drops the pick, must reload. + backend = _loaded_backend("auto") + backend._is_diffusion = True + backend._gpu_ids = [1] # loaded on the lowest of an earlier [3, 1] pick + assert _target_state_gpu_ids(backend, [3, 1]) is True + assert _target_state_gpu_ids(backend, [1]) is True + # Lowest device changes (2, not 1) -> reload. + assert _target_state_gpu_ids(backend, [3, 2]) is False + # Dropping the pick (auto) -> reload. + assert _target_state_gpu_ids(backend, None) is False + + +def test_start_diffusion_server_resets_tensor_parallel(): + # A prior tensor-parallel chat load leaves self._tensor_parallel True (load_model + # phase 1 only kills the process, it skips the unload reset). Diffusion is never + # TP, so startup must clear it -- else /status misreports TP and an identical + # diffusion re-Apply reloads against stale tensor-parallel state. + src = inspect.getsource(llama_cpp_module.LlamaCppBackend._start_diffusion_server) + assert "self._tensor_parallel = False" in src + + +def test_route_matches_loaded_settings_collapses_diffusion_gpu_ids(): + # The route-level reload dedupe mirrors the backend: for a loaded diffusion + # model it compares the request against the single recorded device, not the + # full requested list, or a same-device multi-GPU pick reloads needlessly. + route_src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8") + match_impl = route_src[route_src.index("def _request_matches_loaded_settings") :] + guard = match_impl.index("if llama_backend.is_diffusion:") + collapse = match_impl.index("[sorted(request.gpu_ids)[0]] if request.gpu_ids else None") + compare = match_impl.index("if _req_gpu_ids != llama_backend.gpu_ids:") + assert guard < collapse < compare + + +# ── Manual tensor split: child enumeration pinned to the picker's order ────── + + +def _patch_split_pin_env(monkeypatch, *, inherited, reported): + """Point the pin helper at a fake inherited mask and picker report. + ``reported`` None = enumeration unavailable (falls back to ascending).""" + import utils.hardware as hw + + monkeypatch.setattr( + LlamaCppBackend, "_resolve_visible_physical_ids", staticmethod(lambda: inherited) + ) + info = ( + {"available": False} + if reported is None + else { + "available": True, + "index_kind": "physical", + "devices": [{"index": i} for i in reported], + } + ) + monkeypatch.setattr(hw, "get_backend_visible_gpu_info", lambda: info) + + +def test_split_pin_reorders_inherited_numeric_mask(monkeypatch): + # Parent CUDA_VISIBLE_DEVICES=3,1 makes the child enumerate dev0=phys3, but + # nvidia-smi reported the picker's list ascending -- the mask must be + # re-emitted in that order or the per-GPU shares land on the wrong cards. + _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3]) + env = {"CUDA_VISIBLE_DEVICES": "3,1"} + LlamaCppBackend._pin_visible_gpu_order_for_split(env) + assert env["CUDA_DEVICE_ORDER"] == "PCI_BUS_ID" + assert env["CUDA_VISIBLE_DEVICES"] == "1,3" + + +def test_split_pin_keeps_mask_order_when_picker_reported_it(monkeypatch): + # Torch-fallback enumeration (no nvidia-smi) reports devices in inherited + # mask order, so the picker's split list follows the mask -- the pin must + # keep that order, not re-sort it into a mismatch. + _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [3, 1]) + env = {"CUDA_VISIBLE_DEVICES": "3,1"} + LlamaCppBackend._pin_visible_gpu_order_for_split(env) + assert env["CUDA_VISIBLE_DEVICES"] == "3,1" + + +def test_split_pin_falls_back_to_ascending_without_report(monkeypatch): + # Enumeration unavailable: ascending physical is the best guess (it matches + # the dominant nvidia-smi report order). + _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = None) + env = {"CUDA_VISIBLE_DEVICES": "3,1"} + LlamaCppBackend._pin_visible_gpu_order_for_split(env) + assert env["CUDA_VISIBLE_DEVICES"] == "1,3" + + +def test_split_pin_without_mask_only_sets_pci_order(monkeypatch): + # No inherited mask (or a UUID/MIG one resolving to None): enumeration order + # is fully fixed by CUDA_DEVICE_ORDER, so no mask is written. + _patch_split_pin_env(monkeypatch, inherited = None, reported = None) + env = {} + LlamaCppBackend._pin_visible_gpu_order_for_split(env) + assert env == {"CUDA_DEVICE_ORDER": "PCI_BUS_ID"} + + +def test_split_pin_mirrors_hip_mask_on_rocm(monkeypatch): + # ROCm: the pin must land in HIP_VISIBLE_DEVICES too, and an inherited ROCR + # mask is cleared so the mask can't apply twice (ROCR re-indexes, then HIP + # would index into the already-reduced set). + _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3]) + torch_stub = _types.ModuleType("torch") + torch_stub.version = _types.SimpleNamespace(hip = "6.0") + monkeypatch.setitem(sys.modules, "torch", torch_stub) + env = {"CUDA_VISIBLE_DEVICES": "3,1", "ROCR_VISIBLE_DEVICES": "3,1"} + LlamaCppBackend._pin_visible_gpu_order_for_split(env) + assert env["CUDA_VISIBLE_DEVICES"] == "1,3" + assert env["HIP_VISIBLE_DEVICES"] == "1,3" + assert "ROCR_VISIBLE_DEVICES" not in env + + +# ── Diffusion single-device selection ─────────────────────────────────────── + + +def test_diffusion_gpu_arg_uses_lowest_explicit_physical_id(monkeypatch): + monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "3,1") + monkeypatch.setenv("DG_GPU", "7") + assert LlamaCppBackend._diffusion_gpu_arg([3, 1]) == "1" + + +def test_diffusion_gpu_arg_preserves_parent_mask_order(monkeypatch): + monkeypatch.delenv("DG_GPU", raising = False) + monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "3,1") + assert LlamaCppBackend._diffusion_gpu_arg(None) == "3" + + +def test_diffusion_gpu_arg_honors_override_and_cpu_mask(monkeypatch): + monkeypatch.setenv("DG_GPU", "GPU-abc") + assert LlamaCppBackend._diffusion_gpu_arg(None) == "GPU-abc" + assert LlamaCppBackend._diffusion_gpu_arg(None, cpu_only = True) == "" + + +# ── Deliberate zero-offload (manual gpu_layers=0): training-skip flag ───────── + + +def test_zero_offload_flag_false_without_companions(): + # CPU-only by construction: False lets training skip unloading a server that + # holds no VRAM. + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", "--fit", "off"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is False + + +@pytest.mark.parametrize( + "companion", + ["--mmproj", "--model-draft", "-md", "--spec-draft-model", "-hfd"], +) +def test_zero_offload_flag_true_with_companion(companion): + # mmproj / a drafter offload to GPU regardless of --gpu-layers, so the + # server still holds VRAM and training must unload it. Drafter detection + # reuses the extras parser, so pass-through aliases count too. + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", companion, "x.gguf"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + + +def test_zero_offload_flag_true_with_inline_companion_forms(): + cmd = ["llama-server", "-m", "model.gguf", "--spec-draft-model=x.gguf"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + cmd = ["llama-server", "-m", "model.gguf", "--mmproj=proj.gguf"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + + +def test_zero_offload_flag_true_with_env_drafter(): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"] + env = {"LLAMA_ARG_SPEC_DRAFT_MODEL": "x.gguf"} + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is True + + +@pytest.mark.parametrize( + "device_args", + [ + ["--device", "CUDA0"], + ["--device=CUDA0"], + ["-dev", "CUDA0"], + ["--spec-draft-device", "CUDA0"], + ["--device-draft=CUDA0"], + ], +) +def test_zero_offload_flag_true_with_device_pin(device_args): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", *device_args] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + + +def test_zero_offload_flag_true_with_env_device_pin(): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"] + env = {"LLAMA_ARG_DEVICE": "CUDA0"} + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is True + + +@pytest.mark.parametrize( + ("device_args", "env"), + [ + (["--device", "cpu"], {}), + (["--device=none"], {}), + (["--spec-draft-device", "cpu"], {}), + ([], {"LLAMA_ARG_DEVICE": "none"}), + (["--device", "CUDA0", "--device", "cpu"], {}), + ], +) +def test_zero_offload_flag_false_with_cpu_device_pin(device_args, env): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", *device_args] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is False + + +def test_zero_offload_flag_true_with_surviving_tensor_mode(): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", "--split-mode", "tensor"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + + +def test_zero_offload_flag_true_for_unmasked_vulkan(monkeypatch): + monkeypatch.setattr(LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda: True)) + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True + + +def test_zero_offload_flag_none_without_gpus(): + cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"] + assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [], {}) is None + + +def test_cmd_has_gpu_companion_detection(): + # The env mask for CPU-only zero-offload loads keys off this scan: any + # --mmproj form or a drafter (flag aliases / env) keeps the GPUs visible. + has = LlamaCppBackend._cmd_has_gpu_companion + assert has(["llama-server", "-m", "m.gguf"], {}) is False + assert has(["llama-server", "--mmproj", "p.gguf"], {}) is True + assert has(["llama-server", "--mmproj=p.gguf"], {}) is True + assert has(["llama-server", "-md", "d.gguf"], {}) is True + assert has(["llama-server"], {"LLAMA_ARG_SPEC_DRAFT_MODEL": "d.gguf"}) is True + + +def test_cmd_companion_ignores_cpu_forced_drafter(): + # A CPU-pinned drafter holds no VRAM: the zero-offload mask may hide the GPUs + # and training may leave the server alone. + has = LlamaCppBackend._cmd_has_gpu_companion + cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-ngl", "0"] + assert has(cmd, {}) is False + cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-device", "cpu"] + assert has(cmd, {}) is False + # mmproj still counts even alongside a CPU drafter. + cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-ngl", "0", "--mmproj", "p.gguf"] + assert has(cmd, {}) is True diff --git a/studio/backend/tests/test_gpu_selection.py b/studio/backend/tests/test_gpu_selection.py index 69ad560788..d4f2fbe993 100644 --- a/studio/backend/tests/test_gpu_selection.py +++ b/studio/backend/tests/test_gpu_selection.py @@ -853,7 +853,13 @@ class TestRouteErrors(unittest.TestCase): self.assertIn("only supported on CUDA devices", str(exc_info.exception)) - def test_inference_route_rejects_gpu_ids_for_gguf(self): + def test_inference_route_validates_gpu_ids_for_gguf(self): + # gpu_ids is now SUPPORTED for GGUF (the GPU picker), but still + # validated: a rejected pick surfaces as a clean 400, not the old + # "not supported for GGUF" rejection. Patch the validator so the test + # is deterministic regardless of the host's (or a prior test's) GPU env. + import utils.hardware.hardware as hardware_mod + inference_route = _load_route_module( "inference_route_module_for_gguf_gpu_ids_test", "routes/inference.py", @@ -887,6 +893,11 @@ class TestRouteErrors(unittest.TestCase): ), patch.object(inference_route.asyncio, "to_thread", new = _inline_to_thread), patch.object(inference_route, "_hf_offline_if_dns_dead", nullcontext), + patch.object( + hardware_mod, + "resolve_requested_gpu_ids", + side_effect = ValueError("Invalid gpu_ids [0, 1]: rejected by test"), + ), ): with self.assertRaises(HTTPException) as exc_info: asyncio.run( @@ -901,8 +912,11 @@ class TestRouteErrors(unittest.TestCase): ) ) + # The validator's ValueError becomes a clean 400 (not the removed + # "not supported for GGUF" rejection). self.assertEqual(exc_info.exception.status_code, 400) - self.assertIn("GGUF", exc_info.exception.detail) + self.assertIn("gpu_ids", exc_info.exception.detail.lower()) + self.assertNotIn("not supported", exc_info.exception.detail.lower()) def test_training_route_returns_400_for_invalid_gpu_ids(self): training_route = _load_route_module( diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py index 2fff744b64..48aff29659 100644 --- a/studio/backend/tests/test_hf_xet_fallback.py +++ b/studio/backend/tests/test_hf_xet_fallback.py @@ -1,10 +1,10 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -"""Tests for the Studio shim over the shared unsloth_zoo Xet -> HTTP fallback. +"""Tests for the Unsloth shim over the shared unsloth_zoo Xet -> HTTP fallback. The transport-policy matrix is tested once in unsloth_zoo; here we assert only the -Studio seam: re-exporting the shared API and injecting the marker-aware +Unsloth seam: re-exporting the shared API and injecting the marker-aware prepare_cache_for_transport on the HTTP retry. CPU-only, no network, no real subprocess. """ @@ -69,7 +69,7 @@ def test_child_should_disable_xet_truth_table(): def test_shim_injects_studio_prepare_on_http_retry(monkeypatch): - """A Xet stall retries over HTTP and the shim runs Studio's marker-aware + """A Xet stall retries over HTTP and the shim runs Unsloth's marker-aware ``prepare_cache_for_transport(..., 'http')`` before the retry.""" _requires_shared() for var in ("UNSLOTH_DISABLE_XET", "UNSLOTH_STABLE_DOWNLOADS", "HF_HUB_DISABLE_XET"): @@ -107,11 +107,11 @@ def test_shim_injects_studio_prepare_on_http_retry(monkeypatch): out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None) assert out == "/cache/model.gguf" assert seen_disable_xet == [False, True] # Xet first, then HTTP - assert prepared == [("model", DL_REPO, "http")], "shim must run Studio's marker-aware prep" + assert prepared == [("model", DL_REPO, "http")], "shim must run Unsloth's marker-aware prep" def test_shim_snapshot_injects_studio_prepare(monkeypatch): - """The snapshot wrapper forwards Studio's marker-aware prep, like the file wrapper.""" + """The snapshot wrapper forwards Unsloth's marker-aware prep, like the file wrapper.""" captured = {} def fake_snapshot(repo_id, **kwargs): @@ -127,7 +127,7 @@ def test_shim_snapshot_injects_studio_prepare(monkeypatch): def test_degrades_gracefully_without_shared_helper(monkeypatch): - """On an older unsloth_zoo lacking the shared helper, the shim still imports (Studio + """On an older unsloth_zoo lacking the shared helper, the shim still imports (Unsloth boots) and exposes stub API doing plain HF downloads with the watchdog disabled.""" import importlib @@ -206,7 +206,7 @@ def test_degrades_gracefully_without_shared_helper(monkeypatch): def test_degrades_when_unsloth_zoo_entirely_absent(): """When unsloth_zoo is absent entirely, the import raises ModuleNotFoundError(name='unsloth_zoo') (top-level package). Guard that the shim still - degrades and does not re-raise, breaking every Studio import that pulls it in.""" + degrades and does not re-raise, breaking every Unsloth import that pulls it in.""" import importlib class _BlockZoo: @@ -248,7 +248,7 @@ def test_degrades_when_unsloth_zoo_entirely_absent(): def test_degrades_when_shared_helper_import_raises_importerror(): """unsloth_zoo can be installed yet fail to import when torch is missing (llama.cpp/GGUF-only - Studio), raising ImportError not ModuleNotFoundError. The shim must degrade for that too.""" + Unsloth), raising ImportError not ModuleNotFoundError. The shim must degrade for that too.""" import importlib class _BlockWithImportError: @@ -329,7 +329,7 @@ def test_retries_under_light_gpu_init_when_import_fails(monkeypatch): # with it set); accessing DownloadStallError drives it via __getattr__. stall_error = degraded.DownloadStallError assert seen_env == [None, "1"], seen_env - # Both attempts raised -> Studio still boots in degraded mode. + # Both attempts raised -> Unsloth still boots in degraded mode. assert issubclass(stall_error, RuntimeError) # The env override must not leak past the load. assert os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT") is None diff --git a/studio/backend/tests/test_identity.py b/studio/backend/tests/test_identity.py index 1e84ddef35..712348f7ca 100644 --- a/studio/backend/tests/test_identity.py +++ b/studio/backend/tests/test_identity.py @@ -3,7 +3,7 @@ """Tests for the server identity handshake (`GET /api/auth/identity`). -The endpoint lets a client confirm an endpoint is really this Studio install +The endpoint lets a client confirm an endpoint is really this Unsloth install before sending it a credential: the client sends a random nonce and checks the returned HMAC against one computed from the install identity secret. A process that cannot read this same-user secret cannot forge a matching proof. diff --git a/studio/backend/tests/test_index_bootstrap_origin_extra.py b/studio/backend/tests/test_index_bootstrap_origin_extra.py index feda88c14c..e1c52a653e 100644 --- a/studio/backend/tests/test_index_bootstrap_origin_extra.py +++ b/studio/backend/tests/test_index_bootstrap_origin_extra.py @@ -26,7 +26,7 @@ def _build_request( def test_is_same_origin_request_ipv6_loopback_same_origin(): - """Studio supports ``-H ::1`` binds; netloc is ``[::1]:8902``. Bare + """Unsloth supports ``-H ::1`` binds; netloc is ``[::1]:8902``. Bare ``partition(":")`` mis-parses the bracketed form and would refuse the bootstrap on legitimate same-origin navigation. """ diff --git a/studio/backend/tests/test_llama_cpp_context_fit.py b/studio/backend/tests/test_llama_cpp_context_fit.py index d3a10df8ca..2a4f6d19d2 100644 --- a/studio/backend/tests/test_llama_cpp_context_fit.py +++ b/studio/backend/tests/test_llama_cpp_context_fit.py @@ -567,7 +567,7 @@ class TestClassifyGpuOffload: assert inst._classify_gpu_offload(False, []) is None def test_user_did_not_intend_gpu_returns_none(self): - # Studio called start_llama_server without expecting GPU; don't warn. + # Unsloth called start_llama_server without expecting GPU; don't warn. inst = self._backend( [ "load_tensors: CPU_Mapped model buffer size = 21000.0 MiB", diff --git a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py index 04d4aac9e1..049058e511 100644 --- a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py +++ b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py @@ -222,7 +222,7 @@ class TestFlashAttnOff: assert _flash_off(["llama-server", "-fa=on"]) == ["llama-server", "-fa=off"] def test_flips_every_occurrence_last_wins(self): - # extra_args can re-enable FA after Studio's flag; llama.cpp is last-wins, + # extra_args can re-enable FA after Unsloth's flag; llama.cpp is last-wins, # so one leftover 'on' would re-crash the retry. Every enable must flip. cmd = ["llama-server", "--flash-attn", "on", "--mmproj", "/p", "--flash-attn", "on"] out = _flash_off(cmd) @@ -234,7 +234,7 @@ class TestFlashAttnOff: assert _flash_off(["llama-server", "--flash-attn=off"]) is None def test_none_when_user_off_wins_last(self): - # User appended 'off' after Studio's 'on'; effective (last-wins) is off, + # User appended 'off' after Unsloth's 'on'; effective (last-wins) is off, # so there is nothing to retry. assert _flash_off(["llama-server", "--flash-attn", "on", "--flash-attn", "off"]) is None diff --git a/studio/backend/tests/test_llama_cpp_mtp_detection.py b/studio/backend/tests/test_llama_cpp_mtp_detection.py index 3f9d2a8f50..8fe04c0e39 100644 --- a/studio/backend/tests/test_llama_cpp_mtp_detection.py +++ b/studio/backend/tests/test_llama_cpp_mtp_detection.py @@ -1014,7 +1014,7 @@ def test_already_in_target_state_2b_falls_back_to_ngram_below_threshold(monkeypa ) -# usage backfill from timings (Studio UI t/s widget fix). +# usage backfill from timings (Unsloth UI t/s widget fix). def test_backfill_usage_from_timings_fills_when_completion_tokens_zero(): @@ -1606,7 +1606,7 @@ def test_reload_forced_mtp_bounces_auto_mla(): ) -# ── Full named-repo resolver matrix (the shipping Studio families) ───── +# ── Full named-repo resolver matrix (the shipping Unsloth families) ───── # # Locks auto / off / forced-mtp routing for every Qwen3.5 (MTP + plain) and # gemma-4 (regular + QAT) GGUF repo, including the giant MoEs that stay diff --git a/studio/backend/tests/test_llama_cpp_no_context_shift.py b/studio/backend/tests/test_llama_cpp_no_context_shift.py index 10b1dc7ff6..662c918305 100644 --- a/studio/backend/tests/test_llama_cpp_no_context_shift.py +++ b/studio/backend/tests/test_llama_cpp_no_context_shift.py @@ -5,7 +5,7 @@ With llama-server's default context-shift behavior, the UI cannot tell the user the KV cache was rotated -- earlier turns silently vanish from the conversation. -The Studio backend always passes ``--no-context-shift`` so the server returns a +The Unsloth backend always passes ``--no-context-shift`` so the server returns a clean error instead, and the chat adapter can point the user at the ``Context Length`` input in the settings panel. @@ -118,9 +118,17 @@ def test_flag_sits_inside_the_base_cmd_list(): "conditional branch -- otherwise some code paths would still " "run with silent context shift enabled." ) - # Pin that it sits next to -c / --ctx so the grouping makes sense. - assert '"-c"' in block assert '"--flash-attn"' in block + # -c is emitted in the conditional right after the base list, not inside + # it: auto-fit (--fit on with no pinned context) must omit -c entirely, + # because "-c 0" pins the full native context and disables --fit's + # VRAM-based sizing. Pin that it still sits next to the base block so the + # context grouping stays intact. + after = rest[end_rel : end_rel + 1000] + assert '"-c"' in after, ( + "-c must still be emitted in the conditional immediately after the " + "base cmd list (omitted only in auto-fit, where --fit sizes context)." + ) def _iter_lines_with_offset(text: str): diff --git a/studio/backend/tests/test_llama_cpp_props_readback.py b/studio/backend/tests/test_llama_cpp_props_readback.py index 316956325f..fe1e67edad 100644 --- a/studio/backend/tests/test_llama_cpp_props_readback.py +++ b/studio/backend/tests/test_llama_cpp_props_readback.py @@ -4,7 +4,7 @@ """Tests for the post-launch /props context readback. llama-server's memory-fit step or --parallel slot split can allocate less -context than the requested -c while Studio keeps advertising the requested +context than the requested -c while Unsloth keeps advertising the requested value; clients sized to it then die on exceed_context_size_error 400s. ``_reconcile_effective_ctx_with_server`` must adopt the server's real ``default_generation_settings.n_ctx`` whenever it is smaller. @@ -223,33 +223,48 @@ _CAPS_NONE = {"supports_kv_unified": False, "supports_fit_ctx": False} def test_kv_unified_added_for_multi_slot(): """Explicit --parallel N disables llama-server's auto-slots kv-unified - default, splitting -c into per-slot windows of -c/N; Studio must restore + default, splitting -c into per-slot windows of -c/N; Unsloth must restore the shared pool so one request can use the full advertised context.""" - flags = LlamaCppBackend._ctx_integrity_flags(4, False, 98304, 98304, _CAPS_ALL) + flags = LlamaCppBackend._ctx_integrity_flags(4, False, False, 98304, 98304, _CAPS_ALL) assert "--kv-unified" in flags def test_kv_unified_skipped_for_single_slot_or_old_build(): assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags( - 1, False, 98304, 98304, _CAPS_ALL + 1, False, False, 98304, 98304, _CAPS_ALL ) assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags( - 4, False, 98304, 98304, _CAPS_NONE + 4, False, False, 98304, 98304, _CAPS_NONE ) def test_fit_ctx_floors_explicit_request_under_fit(): - flags = LlamaCppBackend._ctx_integrity_flags(1, True, 98304, 98304, _CAPS_ALL) + # An explicit requested ctx floors --fit-ctx at that value on any --fit + # path, including legacy auto (auto_fit False). + flags = LlamaCppBackend._ctx_integrity_flags(1, True, False, 98304, 98304, _CAPS_ALL) assert flags[flags.index("--fit-ctx") + 1] == "98304" -def test_fit_ctx_skipped_without_fit_or_explicit_ctx_or_support(): +def test_fit_ctx_skipped_without_fit_or_support(): + # No --fit on -> no --fit-ctx. assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags( - 1, False, 98304, 98304, _CAPS_ALL + 1, False, False, 98304, 98304, _CAPS_ALL ) - assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(1, True, 0, 262144, _CAPS_ALL) + # --fit on but the binary doesn't support --fit-ctx. assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags( - 1, True, 98304, 98304, _CAPS_NONE + 1, True, True, 98304, 98304, _CAPS_NONE + ) + + +def test_fit_ctx_floors_auto_request_at_8192_only_under_auto_fit(): + # Manual + Auto (auto_fit) floors the auto window at 8192 so --fit can't + # shrink it to a tiny size. + flags = LlamaCppBackend._ctx_integrity_flags(1, True, True, 0, 262144, _CAPS_ALL) + assert flags[flags.index("--fit-ctx") + 1] == "8192" + # Legacy auto (fit on but not auto_fit) emits -c 0 to pin native, so the + # 8192 floor must NOT ride along and override that pin. + assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags( + 1, True, False, 0, 262144, _CAPS_ALL ) diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py index bd2c008589..e99e227d40 100644 --- a/studio/backend/tests/test_llama_cpp_tool_loop.py +++ b/studio/backend/tests/test_llama_cpp_tool_loop.py @@ -122,7 +122,7 @@ def _structured_tool_call(tool_name: str, arguments: dict, call_id: str) -> list def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch): """llama-server may emit content first and then native delta.tool_calls. - Studio must not drop that tool call after it has streamed the preface. + Unsloth must not drop that tool call after it has streamed the preface. """ tool_call_id = "call_render_late" diff --git a/studio/backend/tests/test_llama_cpp_wait_for_health.py b/studio/backend/tests/test_llama_cpp_wait_for_health.py index 82c5b4931a..423c3dd009 100644 --- a/studio/backend/tests/test_llama_cpp_wait_for_health.py +++ b/studio/backend/tests/test_llama_cpp_wait_for_health.py @@ -224,7 +224,7 @@ class TestRetryLogFilenameUnique: class TestFitOffRetryEligible: """Gate for the one-shot --fit off startup-crash retry. - Retry only when Studio's own VRAM math placed the model and nothing + Retry only when Unsloth's own VRAM math placed the model and nothing on the command line chose the fit mode explicitly.""" def test_eligible_for_plain_ngl_launch(self): diff --git a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py index d0213f6079..b28df7ec3f 100644 --- a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py +++ b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py @@ -346,7 +346,7 @@ def test_helper_is_static_method_callable_off_class(): def test_kill_orphaned_servers_returns_count(): """The reaper reports how many owned orphans it killed, so __init__ can - arm the settle wait. Only Studio-owned llama-server procs count.""" + arm the settle wait. Only Unsloth-owned llama-server procs count.""" import os mypid = os.getpid() @@ -376,7 +376,7 @@ def test_kill_orphaned_servers_returns_count(): patch.object(LlamaCppBackend, "_pid_parent_is_alive", staticmethod(lambda pid: False)), ): n = LlamaCppBackend._kill_orphaned_servers() - assert n == 1, "only the Studio-owned orphan should be counted" + assert n == 1, "only the Unsloth-owned orphan should be counted" assert killed == [mypid + 1] # No owned orphans -> zero, so __init__ leaves the cold-start sentinel. @@ -392,8 +392,8 @@ def test_kill_orphaned_servers_returns_count(): def test_kill_orphaned_servers_spares_live_parent(): - """A Studio-owned llama-server whose parent is still running is not an - orphan (a live Studio or the user's shell owns it) and must never be + """An Unsloth-owned llama-server whose parent is still running is not an + orphan (a live Unsloth or the user's shell owns it) and must never be killed; only the true orphan (parent gone) is reaped.""" import os @@ -548,7 +548,7 @@ def test_record_then_reap_round_trip_identity_matches(tmp_path): def test_reap_recorded_pid_spares_live_server(tmp_path): - """A recorded server whose parent is still alive (the running Studio) is NEVER + """A recorded server whose parent is still alive (the running Unsloth) is NEVER reaped, and its pidfile is kept. This is the finding-3 guard: a helper backend constructed in-process must not kill the active chat server. Uses the REAL _pid_parent_is_alive (the child's parent is this live test process).""" diff --git a/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py b/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py index 957de4bad6..489d9eb8d1 100644 --- a/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py +++ b/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py @@ -3,7 +3,7 @@ """Tests for the Windows pip-nvidia DLL dir resolver. -Studio installs torch with bundled CUDA wheels (nvidia-cuda-runtime-cu13, +Unsloth installs torch with bundled CUDA wheels (nvidia-cuda-runtime-cu13, nvidia-cublas-cu13, etc.) and the prebuilt llama-server.exe must find those DLLs at runtime to load CUDA. Mirrors the Linux LD_LIBRARY_PATH block. See unslothai/unsloth#5106. diff --git a/studio/backend/tests/test_llama_server_args.py b/studio/backend/tests/test_llama_server_args.py index deeb228026..c6d16363f8 100644 --- a/studio/backend/tests/test_llama_server_args.py +++ b/studio/backend/tests/test_llama_server_args.py @@ -75,7 +75,7 @@ validate_extra_args = _lsa.validate_extra_args # Reasoning controls ["--reasoning-format", "deepseek"], ["-rea", "auto"], - # Soft-managed: user flags last-wins over Studio's auto-set version. + # Soft-managed: user flags last-wins over Unsloth's auto-set version. # --parallel / -np / --n-parallel are hard-denied (KV-cache + slot # count would desync); use `unsloth studio run --parallel N` instead. ["-c", "131072"], @@ -150,7 +150,7 @@ def test_non_flag_token_passes_through(): "--mmproj", "-mmu", "--mmproj-url", - # Networking (Studio binds + proxies) + # Networking (Unsloth binds + proxies) "--host", "--port", "--path", @@ -176,12 +176,12 @@ def test_non_flag_token_passes_through(): "--models-autoload", "--no-models-autoload", # Server-mode flips: --embedding / --rerank restrict llama-server to - # those endpoints and break Studio's chat hop. + # those endpoints and break Unsloth's chat hop. "--embedding", "--embeddings", "--rerank", "--reranking", - # llama-server's own --tools clashes with Studio's tool policy. + # llama-server's own --tools clashes with Unsloth's tool policy. "--tools", ], ) @@ -194,7 +194,7 @@ def test_denylist_rejects_all_aliases(denied): "args,offending", [ # Pass-through --parallel would last-wins-override the real slot - # count while Studio's KV-cache fit + llama_parallel_slots stay at + # count while Unsloth's KV-cache fit + llama_parallel_slots stay at # the typer value -- plan vs. process disagree. (["--parallel", "8"], "--parallel"), (["--parallel=8"], "--parallel"), @@ -656,7 +656,7 @@ def test_extra_args_disable_mmproj_last_wins(): def test_strip_shadowing_flags_drops_model_draft_with_spec(): - # --model-draft (and aliases) are Studio-managed since the separate + # --model-draft (and aliases) are Unsloth-managed since the separate # MTP drafter support: an inherited copy must not last-wins-override # the auto-detected drafter. out = strip_shadowing_flags( @@ -681,7 +681,7 @@ def test_strip_shadowing_flags_drops_model_draft_with_spec(): ) def test_strip_shadowing_flags_drops_hf_drafter_selectors_with_spec(selector): # HF drafter selectors must reset on inherit like local --model-draft, or a - # stale inherited HF drafter last-wins over Studio's re-derived spec choice. + # stale inherited HF drafter last-wins over Unsloth's re-derived spec choice. out = strip_shadowing_flags( selector + ["--top-k", "20"], strip_context = False, @@ -747,6 +747,34 @@ def test_strip_shadowing_flags_defaults_strip_split_mode_too(): assert strip_shadowing_flags(["--split-mode", "tensor"]) == [] +def test_strip_offload_is_opt_in_and_covers_moe(): + base = dict( + strip_context = False, + strip_cache = False, + strip_spec = False, + strip_template = False, + strip_split_mode = False, + ) + # Default: offload (incl. MoE) flags are NOT stripped. + assert strip_shadowing_flags(["--n-cpu-moe", "8", "--top-k", "20"], **base) == [ + "--n-cpu-moe", + "8", + "--top-k", + "20", + ] + # Opt-in strips layer AND MoE offload flags (value-aware), keeps the rest. + assert strip_shadowing_flags( + ["--n-cpu-moe", "8", "--gpu-layers", "33", "--fit", "off", "--top-k", "20"], + **base, + strip_offload = True, + ) == ["--top-k", "20"] + # Boolean --cpu-moe drops the flag only, not the following value. + assert strip_shadowing_flags(["--cpu-moe", "--seed", "-1"], **base, strip_offload = True) == [ + "--seed", + "-1", + ] + + @pytest.mark.parametrize( "args", [ @@ -769,7 +797,7 @@ def test_strip_split_mode_only_preserves_none_and_empty(): def test_strip_shadowing_flags_drops_tensor_split_with_split_mode(): # --tensor-split is coupled to the split mode: stripped together so a stale - # ratio can't override Studio's computed tensor split. Other flags survive. + # ratio can't override Unsloth's computed tensor split. Other flags survive. out = strip_shadowing_flags( ["--split-mode", "row", "--tensor-split", "1,1", "--top-k", "20"], strip_context = False, @@ -796,6 +824,23 @@ def test_strip_split_mode_only_drops_tensor_split_too(): assert strip_split_mode_only(["-sm=tensor", "-ts=3,1"]) == [] +def test_strip_tensor_split_alone_preserves_split_mode(): + # Manual mode emits its own --tensor-split, so an inherited ratio is dropped + # -- but the user's --split-mode row/none/layer choice (which the manual + # ratio toggle can't express) must survive. strip_tensor_split removes only + # the ratio, unlike strip_split_mode which removes the whole group. + out = strip_shadowing_flags( + ["--split-mode", "row", "--tensor-split", "1,1", "--top-k", "20"], + strip_context = False, + strip_cache = False, + strip_spec = False, + strip_template = False, + strip_split_mode = False, + strip_tensor_split = True, + ) + assert out == ["--split-mode", "row", "--top-k", "20"] + + def test_strip_shadowing_flags_keeps_model_draft_without_spec(): out = strip_shadowing_flags( ["--model-draft", "/custom/mtp.gguf"], diff --git a/studio/backend/tests/test_local_llama_cpp_link.py b/studio/backend/tests/test_local_llama_cpp_link.py index c78c029d91..6b44f61972 100644 --- a/studio/backend/tests/test_local_llama_cpp_link.py +++ b/studio/backend/tests/test_local_llama_cpp_link.py @@ -4,7 +4,7 @@ """Behavioral tests for the --with-llama-cpp-dir 'unmanaged local link' contract. When the canonical llama.cpp dir is a symlink (POSIX) / junction (Windows) to a -user's own checkout, Studio must treat it as externally managed: +user's own checkout, Unsloth must treat it as externally managed: - the in-app updater must not offer or apply a prebuilt over the link - orphan cleanup must not kill a llama-server the user launched from that tree @@ -67,7 +67,7 @@ def test_active_install_is_local_link(tmp_path: Path) -> None: binary = str(link / _server_subpath()) assert u._active_install_is_local_link(binary) is True - # A plain (non-link) llama.cpp dir is Studio-managed, not a local link. + # A plain (non-link) llama.cpp dir is Unsloth-managed, not a local link. plain = tmp_path / "plain" / "llama.cpp" plain.mkdir(parents = True) assert u._active_install_is_local_link(str(plain / _server_subpath())) is False diff --git a/studio/backend/tests/test_mcp_servers.py b/studio/backend/tests/test_mcp_servers.py index 6432ffb8e1..c5c37f098f 100644 --- a/studio/backend/tests/test_mcp_servers.py +++ b/studio/backend/tests/test_mcp_servers.py @@ -577,7 +577,7 @@ def test_clear_oauth_tokens_swallows_constructor_errors(tmp_path, monkeypatch): def test_tool_xml_parser_handles_hyphenated_function_names(): """Hyphenated tool names like `mcp__srv__list-issues` must parse, else the - model can call the tool but Studio can't dispatch.""" + model can call the tool but Unsloth can't dispatch.""" from core.inference.tool_call_parser import parse_tool_calls_from_text calls = parse_tool_calls_from_text( diff --git a/studio/backend/tests/test_mcp_stdio_improvements.py b/studio/backend/tests/test_mcp_stdio_improvements.py index b0bfd45135..745c2cc447 100644 --- a/studio/backend/tests/test_mcp_stdio_improvements.py +++ b/studio/backend/tests/test_mcp_stdio_improvements.py @@ -188,7 +188,7 @@ def test_validate_url_allows_url_in_argument(monkeypatch): # ── P6: Data Recipe stdio path obeys the same host gate ───────────── -# build_mcp_providers needs the Studio-only data_designer plugin; skip if absent. +# build_mcp_providers needs the Unsloth-only data_designer plugin; skip if absent. _STDIO_RECIPE = { "mcp_providers": [ diff --git a/studio/backend/tests/test_mlx_inference_backend.py b/studio/backend/tests/test_mlx_inference_backend.py index fa50cd84d6..fafaea0043 100644 --- a/studio/backend/tests/test_mlx_inference_backend.py +++ b/studio/backend/tests/test_mlx_inference_backend.py @@ -2,6 +2,7 @@ import sys import types +from contextlib import contextmanager from types import SimpleNamespace import pytest @@ -40,12 +41,16 @@ class _DummyModel: def _install_fake_mlx(monkeypatch): mlx_pkg = types.ModuleType("mlx") mlx_core = types.ModuleType("mlx.core") + mlx_utils = types.ModuleType("mlx.utils") mlx_core.metal = _DummyMetal() mlx_core.set_wired_limit = _DummyMX.set_wired_limit mlx_core.device_info = _DummyMX.device_info + mlx_utils.tree_unflatten = dict mlx_pkg.core = mlx_core + mlx_pkg.utils = mlx_utils monkeypatch.setitem(sys.modules, "mlx", mlx_pkg) monkeypatch.setitem(sys.modules, "mlx.core", mlx_core) + monkeypatch.setitem(sys.modules, "mlx.utils", mlx_utils) def _install_fake_fast_mlx(monkeypatch, calls): @@ -68,6 +73,99 @@ def _install_fake_fast_mlx(monkeypatch, calls): monkeypatch.setitem(sys.modules, "unsloth_zoo.mlx.loader", mlx_loader) +class _AdapterTree: + def __init__(self, modules): + self.modules = dict(modules) + + def named_modules(self): + return list(self.modules.items()) + + def update_modules(self, modules): + self.modules.update(modules) + + +def test_temporary_mlx_adapter_state_bypasses_and_restores_wrappers(monkeypatch): + _install_fake_mlx(monkeypatch) + from core.inference.mlx_inference import _temporary_mlx_adapter_state + + base = object() + wrapper = SimpleNamespace(lora_a = object(), lora_b = object(), linear = base, m = object()) + model = _AdapterTree({"model.layers.0.proj": wrapper}) + + with pytest.raises(RuntimeError, match = "generation failed"): + with _temporary_mlx_adapter_state(model, False): + assert model.modules["model.layers.0.proj"] is base + raise RuntimeError("generation failed") + assert model.modules["model.layers.0.proj"] is wrapper + + +def test_temporary_mlx_adapter_state_validates_requests(): + from core.inference.mlx_inference import _temporary_mlx_adapter_state + + wrapper = SimpleNamespace(lora_a = object(), lora_b = object(), embedding = object()) + model = _AdapterTree({"embed_tokens": wrapper}) + with _temporary_mlx_adapter_state(model, True): + assert model.modules["embed_tokens"] is wrapper + with pytest.raises(NotImplementedError, match = "named adapter"): + with _temporary_mlx_adapter_state(model, "other"): + pass + + base_model = _AdapterTree({"proj": object()}) + with _temporary_mlx_adapter_state(base_model, None): + pass + with _temporary_mlx_adapter_state(base_model, True): + pass + + unsupported = _AdapterTree({"proj": SimpleNamespace(lora_a = object(), lora_b = object())}) + with _temporary_mlx_adapter_state(unsupported, True): + pass + with pytest.raises(RuntimeError, match = "without their base modules"): + with _temporary_mlx_adapter_state(unsupported, False): + pass + + +def test_temporary_mlx_adapter_state_uses_real_mlx_module_tree(): + nn = pytest.importorskip("mlx.nn") + pytest.importorskip("mlx_lm") + from mlx_lm.models.switch_layers import SwitchLinear + from mlx_lm.tuner.dora import DoRALinear + from mlx_lm.tuner.lora import LoRAEmbedding, LoRALinear, LoRASwitchLinear + + from core.inference.mlx_inference import _temporary_mlx_adapter_state + + class _Layer(nn.Module): + def __init__(self): + super().__init__() + quantized = nn.QuantizedLinear.from_linear(nn.Linear(32, 32), group_size = 32, bits = 4) + self.quantized_proj = LoRALinear.from_base(quantized) + self.dora_proj = DoRALinear.from_base(nn.Linear(4, 4)) + + class _Model(nn.Module): + def __init__(self): + super().__init__() + self.layers = [_Layer()] + self.embed_tokens = LoRAEmbedding.from_base(nn.Embedding(16, 4)) + self.experts = LoRASwitchLinear.from_base(SwitchLinear(4, 4, 2)) + + model = _Model() + wrappers = { + path: module + for path, module in model.named_modules() + if hasattr(module, "lora_a") and hasattr(module, "lora_b") + } + bases = { + path: getattr(module, "linear", getattr(module, "embedding", None)) + for path, module in wrappers.items() + } + + with _temporary_mlx_adapter_state(model, False): + live = dict(model.named_modules()) + assert all(live[path] is base for path, base in bases.items()) + + restored = dict(model.named_modules()) + assert all(restored[path] is wrapper for path, wrapper in wrappers.items()) + + def test_mlx_inference_text_load_forwards_studio_settings(monkeypatch): _install_fake_mlx(monkeypatch) calls = [] @@ -138,7 +236,7 @@ def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewri _install_fake_fast_mlx(monkeypatch, calls) def _native_vlm_load(*_args, **_kwargs): - raise AssertionError("Studio MLX VLM inference must use FastMLXModel") + raise AssertionError("Unsloth MLX VLM inference must use FastMLXModel") mlx_vlm = types.ModuleType("mlx_vlm") mlx_vlm.load = _native_vlm_load @@ -333,10 +431,87 @@ def test_mlx_generate_chat_response_accepts_template_kwargs(): ), f"{name!r} must default to None so existing callers stay valid" +def test_mlx_vlm_reemits_think_prefill_inside_adapter_context(monkeypatch): + """A prefilled block must be re-emitted as the first VLM snapshot, + inside the adapter context (so unsupported requests still raise first), so + the UI renders the thinking block during prefill and a pre-first-token + cancel does not drop it. Mirrors _generate_text.""" + from core.inference import mlx_inference + + MLXInferenceBackend = mlx_inference.MLXInferenceBackend + + order = [] + + @contextmanager + def _adapter_state(_model, state): + assert backend._generation_lock.locked() + order.append("adapter_enter") + try: + yield + finally: + order.append("adapter_exit") + + monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state) + monkeypatch.setattr( + "core.inference.chat_template_helpers.detect_think_prefill", + lambda *_a, **_k: "\n", + ) + + prompt_utils = SimpleNamespace( + MODEL_CONFIG = {"deepseek_vl_v2": object()}, + apply_chat_template = lambda *_a, **_k: " model-aware", + ) + mlx_vlm = types.ModuleType("mlx_vlm") + mlx_vlm.prompt_utils = prompt_utils + + def _vlm_stream(*_a, **_k): + # The prefill must have been emitted before any generated token. + assert order[-1] == "adapter_enter" + yield SimpleNamespace(text = "ok", prompt_tokens = 3, generation_tokens = 1) + + mlx_vlm.stream_generate = _vlm_stream + monkeypatch.setitem(sys.modules, "mlx_vlm", mlx_vlm) + monkeypatch.setattr( + "core.inference.chat_template_helpers.apply_chat_template_for_generation", + lambda _t, _m, **_k: " model-aware", + ) + + backend = MLXInferenceBackend() + backend._model = SimpleNamespace(config = {"model_type": "deepseek_vl_v2"}) + backend._processor = SimpleNamespace(tokenizer = SimpleNamespace()) + args = ([{"role": "user", "content": [{"type": "image"}]}], object(), 0, 1, 0, 0, 1, 1, None) + + gen = backend._generate_vlm(*args, _adapter_state = False) + # First snapshot is the prefill alone, emitted after entering the adapter context. + assert next(gen) == "\n" + assert order == ["adapter_enter"] + # Subsequent snapshots are cumulative (prefill + generated text). + assert next(gen) == "\nok" + gen.close() + assert order == ["adapter_enter", "adapter_exit"] + + def test_mlx_vlm_generation_selects_renderer_by_capability(monkeypatch): - from core.inference.mlx_inference import MLXInferenceBackend + from core.inference import mlx_inference + + MLXInferenceBackend = mlx_inference.MLXInferenceBackend calls = {"generic": [], "model": [], "stream": []} + adapter_events = [] + adapter_active = {"value": False} + + @contextmanager + def _adapter_state(_model, state): + assert backend._generation_lock.locked() + adapter_events.append(("enter", state)) + adapter_active["value"] = True + try: + yield + finally: + adapter_active["value"] = False + adapter_events.append(("exit", state)) + + monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state) state = {"generic": "serialized", "model": " model-aware"} prompt_utils = SimpleNamespace( MODEL_CONFIG = {"deepseek_vl_v2": object()}, @@ -346,10 +521,13 @@ def test_mlx_vlm_generation_selects_renderer_by_capability(monkeypatch): ) mlx_vlm = types.ModuleType("mlx_vlm") mlx_vlm.prompt_utils = prompt_utils - mlx_vlm.stream_generate = lambda *_args, **kwargs: ( - calls["stream"].append((_args, kwargs)) - or iter([SimpleNamespace(text = "ok", prompt_tokens = 3, generation_tokens = 1)]) - ) + + def _vlm_stream(*args, **kwargs): + assert adapter_active["value"] + calls["stream"].append((args, kwargs)) + yield SimpleNamespace(text = "ok", prompt_tokens = 3, generation_tokens = 1) + + mlx_vlm.stream_generate = _vlm_stream monkeypatch.setitem(sys.modules, "mlx_vlm", mlx_vlm) def generic(_target, _messages, **kwargs): @@ -369,7 +547,11 @@ def test_mlx_vlm_generation_selects_renderer_by_capability(monkeypatch): backend._processor = SimpleNamespace(tokenizer = SimpleNamespace()) args = ([{"role": "user", "content": [{"type": "image"}]}], object(), 0, 1, 0, 0, 1, 1, None) tools = [{"function": {"name": "search"}}] - assert list(backend._generate_vlm(*args)) == ["ok"] + generator = backend._generate_vlm(*args, _adapter_state = False) + assert next(generator) == "ok" + assert adapter_active["value"] and backend._generation_lock.locked() + generator.close() + assert adapter_events == [("enter", False), ("exit", False)] assert calls["model"][0]["num_images"] == 1 assert calls["stream"][0][0][2] == " model-aware" with pytest.raises(RuntimeError, match = "dropping requested tools"): @@ -449,7 +631,10 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch): """Mac text path must route through apply_chat_template_for_generation so reasoning / tool kwargs reach the tokenizer.""" _install_fake_mlx(monkeypatch) - from core.inference.mlx_inference import MLXInferenceBackend + from core.inference import mlx_inference + + MLXInferenceBackend = mlx_inference.MLXInferenceBackend + real_adapter_state = mlx_inference._temporary_mlx_adapter_state # The text path renders once with tools, then the native-template fallback makes a second no- # tools probe call (tools=None) to detect whether the template dropped the schema. @@ -474,11 +659,31 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch): mlx_lm_sample.make_sampler = lambda **_kw: object() mlx_lm_sample.make_logits_processors = lambda **_kw: None + adapter_events = [] + adapter_active = {"value": False} + stream_state = {"fail": False} + + @contextmanager + def _adapter_state(_model, state): + assert backend._generation_lock.locked() + adapter_events.append(("enter", state)) + adapter_active["value"] = True + try: + yield + finally: + adapter_active["value"] = False + adapter_events.append(("exit", state)) + + monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state) + class _Resp: def __init__(self, tok): self.token = tok def _stream_generate(_model, _tokenizer, **_kw): + assert adapter_active["value"] + if stream_state["fail"]: + raise RuntimeError("generation failed") yield _Resp(1) mlx_lm_pkg.stream_generate = _stream_generate @@ -500,17 +705,45 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch): backend._tokenizer = _Tok() backend._is_vlm = False - out = list( - backend.generate_chat_response( - messages = [{"role": "user", "content": "ping"}], - tools = [{"function": {"name": "web_search"}}], - enable_thinking = True, - reasoning_effort = "medium", - preserve_thinking = True, - max_new_tokens = 1, - ) + generator = backend.generate_with_adapter_control( + use_adapter = False, + messages = [{"role": "user", "content": "ping"}], + tools = [{"function": {"name": "web_search"}}], + enable_thinking = True, + reasoning_effort = "medium", + preserve_thinking = True, + max_new_tokens = 1, ) - assert out == ["hi"] + assert next(generator) == "hi" + assert adapter_active["value"] and backend._generation_lock.locked() + generator.close() + assert adapter_events == [("enter", False), ("exit", False)] + stream_state["fail"] = True + with pytest.raises(RuntimeError, match = "generation failed"): + list( + backend.generate_with_adapter_control( + use_adapter = False, + messages = [{"role": "user", "content": "ping"}], + max_new_tokens = 1, + ) + ) + assert adapter_events[-2:] == [("enter", False), ("exit", False)] + assert not backend._generation_lock.locked() + + monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", real_adapter_state) + monkeypatch.setattr( + "core.inference.chat_template_helpers.detect_think_prefill", + lambda *_args, **_kwargs: "", + ) + stream_state["fail"] = False + named = backend.generate_with_adapter_control( + use_adapter = "named", + messages = [{"role": "user", "content": "ping"}], + max_new_tokens = 1, + ) + with pytest.raises(NotImplementedError, match = "named adapter"): + next(named) + assert not adapter_active["value"] and not backend._generation_lock.locked() # The toggled kwargs must reach the chat-template helper on the real render # (one of the calls carries the tools; the fallback probe passes tools=None). tool_renders = [ diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py index 365cc46410..47a695ccbd 100644 --- a/studio/backend/tests/test_mlx_repair.py +++ b/studio/backend/tests/test_mlx_repair.py @@ -103,7 +103,7 @@ def test_repair_install_pins_transformers_and_cleans_up(monkeypatch): assert mr.attempt_mlx_repair() is True cmd = captured["cmd"] # transformers is pinned via a constraint file so the mlx install cannot - # upgrade it underneath Studio, and the temp constraint file is cleaned up. + # upgrade it underneath Unsloth, and the temp constraint file is cleaned up. assert "--constraint" in cmd assert "--upgrade" in cmd reinstall_pairs = set(zip(cmd, cmd[1:])) @@ -123,7 +123,7 @@ def test_install_requires_prebuilt_wheels(monkeypatch): # A source distribution's PEP 517 build backend runs arbitrary code at install # time, before the post-install stack check. The unattended self-heal must # require pre-built wheels so a malicious resolver-selected sdist cannot execute - # during ordinary Studio startup. mlx/mlx-metal ship wheels only and + # during ordinary Unsloth startup. mlx/mlx-metal ship wheels only and # mlx-lm/mlx-vlm publish py3-none-any wheels, so a healthy self-heal still works. pytest.importorskip("transformers") captured = {} @@ -143,7 +143,7 @@ def test_install_requires_prebuilt_wheels(monkeypatch): def test_install_env_drops_secrets_and_source_redirects(monkeypatch): - # The unattended self-heal must not hand resolver/build code the full Studio + # The unattended self-heal must not hand resolver/build code the full Unsloth # environment: secrets and package-source redirects are dropped, while the # variables uv genuinely needs are forwarded. monkeypatch.setenv("HF_TOKEN", "secret-hf") diff --git a/studio/backend/tests/test_mtp_vram_budget.py b/studio/backend/tests/test_mtp_vram_budget.py index 0efbbf596d..694d60cfc6 100644 --- a/studio/backend/tests/test_mtp_vram_budget.py +++ b/studio/backend/tests/test_mtp_vram_budget.py @@ -502,7 +502,7 @@ class TestExtraArgsMtpDetection: assert _extra_args_mtp_draft_path([], env = dict(os.environ)) == "/large.gguf" def test_load_model_gates_env_spec_type_on_off_mode(self): - # LLAMA_ARG_SPEC_TYPE only reaches the child when Studio emits no spec + # LLAMA_ARG_SPEC_TYPE only reaches the child when Unsloth emits no spec # flag (UI mode "off", no user --spec-type); otherwise the emitted # --spec-type/--spec-default overrides the env, so the reserve must not # consult it or a stale MTP env over-reserves (Finding F3). Whitespace- @@ -530,8 +530,8 @@ class TestExtraArgsMtpDetection: def test_load_model_drafter_budget_precedence(self): # The budget sizes the drafter the launch actually loads: CLI extras win, - # then Studio's emitted mtp_draft_path (overrides LLAMA_ARG_SPEC_DRAFT_MODEL), - # then the env drafter -- not the env before Studio's (reviewer.py R3). + # then Unsloth's emitted mtp_draft_path (overrides LLAMA_ARG_SPEC_DRAFT_MODEL), + # then the env drafter -- not the env before Unsloth's (reviewer.py R3). compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split()) assert "_cli_draft_for_budget=_extra_args_mtp_draft_path(extra_args,env={})" in compact assert "_env_draft_for_budget=_extra_args_mtp_draft_path([],env=os.environ)" in compact @@ -732,7 +732,7 @@ class TestExtraArgsMtpDetection: assert _extra_args_n_ubatch([], env = {"LLAMA_ARG_UBATCH": "notint"}) is None def test_env_main_cache_type_for_budget(self): - # The child inherits LLAMA_ARG_CACHE_TYPE_K/_V, but Studio emits no + # The child inherits LLAMA_ARG_CACHE_TYPE_K/_V, but Unsloth emits no # --cache-type when neither param nor extras set it -> a heavier env # main KV (f32) must be adopted so the reserve matches the child. assert _env_main_cache_type_for_budget(env = {}) is None @@ -765,7 +765,7 @@ class TestExtraArgsMtpDetection: assert "cache_type_kv=_env_main_cache_type_for_budget()" in compact def test_env_split_mode_is_tensor(self): - # The child inherits LLAMA_ARG_SPLIT_MODE, but Studio emits --split-mode + # The child inherits LLAMA_ARG_SPLIT_MODE, but Unsloth emits --split-mode # only on its tensor branch -> a tensor env must flip the budget so the # heavier per-device compute buffer is reserved (not layer overhead). assert _env_split_mode_is_tensor(env = {}) is False @@ -918,7 +918,7 @@ class TestExtraArgsMtpDetection: # Cluster A: when the final decision is layer split, an inherited # non-layer LLAMA_ARG_SPLIT_MODE (and paired LLAMA_ARG_TENSOR_SPLIT) must # be popped from the child env so the child cannot run tensor/row/none - # against Studio's layer budget. Whitespace-stripped for formatter. + # against Unsloth's layer budget. Whitespace-stripped for formatter. compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split()) assert 'env.get("LLAMA_ARG_SPLIT_MODE")' in compact assert '_inherited_sm!="layer"' in compact @@ -936,10 +936,10 @@ class TestExtraArgsMtpDetection: assert "env.pop(_ct_var,None)" in compact def test_load_model_clears_tensor_split_env_in_tensor_mode(self): - # review run3 #2: Studio owns the tensor split. When it emits no + # review run3 #2: Unsloth owns the tensor split. When it emits no # --tensor-split (even split), a stale inherited LLAMA_ARG_TENSOR_SPLIT must # be cleared in the TENSOR branch too (not just the layer downgrade), or the - # child runs a split Studio didn't budget. The else (tensor) branch pops it. + # child runs a split Unsloth didn't budget. The else (tensor) branch pops it. src = inspect.getsource(LlamaCppBackend.load_model) compact = "".join(src.split()) # appears in both the layer branch and the tensor branch. @@ -1005,14 +1005,14 @@ def test_qwen36_class_regression_picks_lower_ctx_with_mtp(): def test_mtp_draft_budget_prefers_user_extras_drafter(): # A user --model-draft in extras is appended last and wins at launch, so the - # VRAM budget must size it first; then Studio's emitted mtp_draft_path (which + # VRAM budget must size it first; then Unsloth's emitted mtp_draft_path (which # overrides LLAMA_ARG_SPEC_DRAFT_MODEL), then the env drafter (load_model is too # entangled to drive end-to-end; assert the precedence at the source level). # Whitespace-stripped so the check survives any formatter line-wrapping. compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split()) - # CLI extras sized first (env={} so the env doesn't pre-empt Studio's drafter). + # CLI extras sized first (env={} so the env doesn't pre-empt Unsloth's drafter). assert "_cli_draft_for_budget=_extra_args_mtp_draft_path(extra_args,env={})" in compact - # Order: CLI extras, then Studio's mtp_draft_path, then the env drafter. + # Order: CLI extras, then Unsloth's mtp_draft_path, then the env drafter. assert "_cli_draft_for_budgetor_studio_draft_for_budgetor_env_draft_for_budget" in compact - # The env must not be consulted before Studio's resolved drafter. + # The env must not be consulted before Unsloth's resolved drafter. assert "_extra_args_mtp_draft_path(extra_args)ormtp_draft_path" not in compact diff --git a/studio/backend/tests/test_multimodal_document.py b/studio/backend/tests/test_multimodal_document.py index 5cd7c876cc..b347c4aef8 100644 --- a/studio/backend/tests/test_multimodal_document.py +++ b/studio/backend/tests/test_multimodal_document.py @@ -3,7 +3,7 @@ """Tests for PDF / document attachment translation on external providers. -Studio adds a normalised `input_document` content part on +Unsloth adds a normalised `input_document` content part on ChatCompletionRequest so the frontend needn't know the per-provider attachment shape: diff --git a/studio/backend/tests/test_nudge_tool_calls_wiring.py b/studio/backend/tests/test_nudge_tool_calls_wiring.py index e03fd0c7d7..82a6543aeb 100644 --- a/studio/backend/tests/test_nudge_tool_calls_wiring.py +++ b/studio/backend/tests/test_nudge_tool_calls_wiring.py @@ -3,7 +3,7 @@ """Wiring guard for the plan-without-action ``nudge_tool_calls`` policy. -Decided policy: the re-prompt is ALWAYS ON for the Studio inference paths +Decided policy: the re-prompt is ALWAYS ON for the Unsloth inference paths (safetensors, GGUF/llama_cpp, MLX) and OPT-IN for the API (/v1 OpenAI-compat + Anthropic-compat, controlled by the request's ``nudge_tool_calls``, default off). @@ -16,7 +16,7 @@ Mechanism (verified here without loading a model): opt-in), while the GGUF loop keeps its pre-existing default-on behaviour (``None`` keeps nudging) so an omitted flag never disables GGUF; * the API request models default the flag to ``None`` (opt-in / off); - * the Studio-facing routes forward the request's flag, and the Studio frontend + * the Unsloth-facing routes forward the request's flag, and the Unsloth frontend sends ``nudge_tool_calls: true`` -- exercised behaviourally in ``test_safetensors_tool_loop.py`` and ``test_llama_cpp_tool_loop.py``. """ @@ -87,7 +87,7 @@ def test_api_request_models_default_the_flag_off(): def test_studio_routes_forward_the_request_flag(): - # The Studio chat frontend posts to /v1/chat/completions and /v1/messages + # The Unsloth chat frontend posts to /v1/chat/completions and /v1/messages # with nudge_tool_calls=true; the route handlers forward the request value # (external API clients that omit it fall back to the opt-in default). from routes import inference as routes_inference diff --git a/studio/backend/tests/test_offline_gguf_cache_fallback.py b/studio/backend/tests/test_offline_gguf_cache_fallback.py index 0b9a1e704f..295549c443 100644 --- a/studio/backend/tests/test_offline_gguf_cache_fallback.py +++ b/studio/backend/tests/test_offline_gguf_cache_fallback.py @@ -897,7 +897,7 @@ class TestHfOfflineIfDnsDead: assert "HF_HUB_OFFLINE" not in os.environ def test_user_set_hf_hub_offline_is_preserved(self, dns, clean_offline_env, monkeypatch): - # User explicitly set offline before launching Studio. + # User explicitly set offline before launching Unsloth. monkeypatch.setenv("HF_HUB_OFFLINE", "1") dns.fail() with _hf_offline_if_dns_dead() as did_set: diff --git a/studio/backend/tests/test_offline_inference_parent.py b/studio/backend/tests/test_offline_inference_parent.py index 71331220d6..bd0014ea64 100644 --- a/studio/backend/tests/test_offline_inference_parent.py +++ b/studio/backend/tests/test_offline_inference_parent.py @@ -139,7 +139,7 @@ class TestLoraDetectOffline: monkeypatch.setenv("HF_HUB_OFFLINE", "1") - # Studio catches Exception broadly; pin that the call still happens + # Unsloth catches Exception broadly; pin that the call still happens # (so cached LoRAs aren't missed) and returns fast via the mock. class _OfflineModeIsEnabled(Exception): pass diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py index d02a2a4f7e..c4c0ce15c9 100644 --- a/studio/backend/tests/test_openai_auto_switch.py +++ b/studio/backend/tests/test_openai_auto_switch.py @@ -689,7 +689,7 @@ def test_v1_models_retrieve_is_case_insensitive(monkeypatch): def test_index_excludes_hidden_models(tmp_path, monkeypatch): # The llama.cpp validation probe and RAG embedding weights are hidden from - # Studio's pickers; they must never become auto-switch targets. + # Unsloth's pickers; they must never become auto-switch targets. from types import SimpleNamespace import routes.models as models_route @@ -697,6 +697,10 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch): normal.write_bytes(b"x" * 32) probe = tmp_path / "stories260K.gguf" # llama.cpp install-validation probe probe.write_bytes(b"x" * 32) + embedder = tmp_path / "embedding-Q8_0.gguf" + embedder.write_bytes(b"x" * 32) + local_default_embedder = tmp_path / "bge-small-en-v1.5-F16.gguf" + local_default_embedder.write_bytes(b"x" * 32) def _info(mid, path): return SimpleNamespace(id = mid, path = str(path), model_id = mid, display_name = mid) @@ -704,7 +708,22 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch): monkeypatch.setattr( models_route, "_scan_models_dir", - lambda *a, **k: [_info("org/Normal-GGUF", normal), _info("ggml-org/models", probe)], + lambda *a, **k: [ + _info("org/Normal-GGUF", normal), + _info("ggml-org/models", probe), + SimpleNamespace( + id = str(embedder), + path = str(embedder), + model_id = "unsloth/bge-small-en-v1.5-GGUF", + display_name = "embedding-Q8_0", + ), + SimpleNamespace( + id = str(local_default_embedder), + path = str(local_default_embedder), + model_id = None, + display_name = local_default_embedder.name, + ), + ], ) monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: []) monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path) @@ -713,6 +732,8 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch): index = resolver._index() assert "org/normal-gguf" in index # keys are normalized to lowercase assert "ggml-org/models" not in index + assert "unsloth/bge-small-en-v1.5-gguf" not in index + assert str(local_default_embedder).lower() not in index # And the hidden probe cannot be auto-switched to by name. resolver._scan = (0.0, {}) assert resolver.resolve_local_gguf("ggml-org/models") is None @@ -1609,11 +1630,11 @@ def test_env_idle_ttl_standalone_when_no_stored_value(monkeypatch): def test_stored_idle_value_overrides_env_and_stays_gated(monkeypatch): # An explicit stored value wins over the env default and remains gated on the # auto-switch toggle. - store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 30} + store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 90} monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d)) monkeypatch.setenv("UNSLOTH_MODEL_IDLE_TTL", "600") monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True) - assert settings.get_auto_unload_idle_seconds() == 30 # stored wins, not env + assert settings.get_auto_unload_idle_seconds() == 90 # stored wins, not env monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False) assert settings.get_auto_unload_idle_seconds() == 0 # explicit value still gated off @@ -1729,6 +1750,8 @@ def test_index_advertises_alias_not_filesystem_path(tmp_path, monkeypatch): # host path in /v1/models, yet the model stays resolvable by that path too. from types import SimpleNamespace import routes.models as models_route + from storage import studio_db + import utils.paths as paths gguf = tmp_path / "model-Q4_K_M.gguf" gguf.write_bytes(b"x" * 32) @@ -1742,6 +1765,8 @@ def test_index_advertises_alias_not_filesystem_path(tmp_path, monkeypatch): monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: []) monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path) monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False) + monkeypatch.setattr(paths, "lmstudio_model_dirs", lambda: []) + monkeypatch.setattr(studio_db, "list_scan_folders", lambda: []) resolver._scan = (0.0, {}) # The advertised id is the alias, never the absolute path. @@ -3094,3 +3119,53 @@ def test_responses_stream_hint_matches_toggle_regardless_of_active_model(monkeyp monkeypatch, enabled = False, active_model_name = "unsloth/Llama-3.2-1B-Instruct" ) assert "Model auto-switch" in non_gguf_loaded + + +def test_setter_rejects_idle_below_floor(monkeypatch): + import storage.studio_db as db + + writes = [] + monkeypatch.setattr(db, "upsert_app_settings", lambda m: writes.append(dict(m))) + settings._cache.clear() + + with pytest.raises(ValueError, match = "at least 60"): + settings.set_openai_auto_switch(True, 30) + assert writes == [] # rejected before any persist + # 0 (off) and >= 60 pass through unchanged. + assert settings.set_openai_auto_switch(True, 0)[1] == 0 + assert settings.set_openai_auto_switch(True, 60)[1] == 60 + assert settings.set_openai_auto_switch(True, 3600)[1] == 3600 + + +def test_put_route_rejects_idle_below_floor(): + import routes.settings as settings_route + from fastapi import HTTPException + + payload = settings_route.OpenAIAutoSwitchPayload(enabled = True, auto_unload_idle_seconds = 30) + with pytest.raises(HTTPException) as excinfo: + settings_route.update_openai_auto_switch(payload, "tester") + assert excinfo.value.status_code == 400 + + +def test_stored_legacy_idle_below_floor_is_clamped(monkeypatch): + # Values persisted before the floor existed are raised to it on read, for + # both the effective TTL and the value the settings UI displays. + store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 5} + monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d)) + monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True) + assert settings.get_auto_unload_idle_seconds() == 60 + assert settings.get_stored_auto_unload_idle_seconds() == 60 + store[settings.AUTO_UNLOAD_IDLE_SETTING_KEY] = 90 + assert settings.get_auto_unload_idle_seconds() == 90 + + +def test_env_idle_below_floor_is_clamped(monkeypatch): + monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: d) + monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "5") + assert settings.get_auto_unload_idle_seconds() == 60 + monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "0") + assert settings.get_auto_unload_idle_seconds() == 0 + monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "600") + assert settings.get_auto_unload_idle_seconds() == 600 + monkeypatch.delenv(settings.MODEL_IDLE_TTL_ENV_VAR) + assert settings.get_auto_unload_idle_seconds() == 0 diff --git a/studio/backend/tests/test_openai_compaction.py b/studio/backend/tests/test_openai_compaction.py index c7de0a9aed..6fad2c5eaf 100644 --- a/studio/backend/tests/test_openai_compaction.py +++ b/studio/backend/tests/test_openai_compaction.py @@ -86,7 +86,7 @@ def test_cloud_openai_sets_compaction_block(monkeypatch): def test_cloud_openai_below_default_threshold_passes_through(monkeypatch): - # Studio doesn't clamp the OpenAI side -- the API accepts whatever the + # Unsloth doesn't clamp the OpenAI side -- the API accepts whatever the # caller sends, so a small probe like 60k still goes through. captured = _capture( monkeypatch, diff --git a/studio/backend/tests/test_openai_image_generation.py b/studio/backend/tests/test_openai_image_generation.py index ace57588d3..c2eef0381f 100644 --- a/studio/backend/tests/test_openai_image_generation.py +++ b/studio/backend/tests/test_openai_image_generation.py @@ -4,7 +4,7 @@ """Unit tests for OpenAI Responses API image_generation tool wiring. The tool is a server-side Responses-API tool (``{type: "image_generation"}``); -the result comes back as an ``image_generation_call`` output item, which Studio +the result comes back as an ``image_generation_call`` output item, which Unsloth translates into ``_toolEvent`` chunks so the chat adapter renders it inline. Tests pin: the tool is added to the body only on a cloud OpenAI base when asked for, the done event produces the expected chunks, and non-cloud bases drop it. diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py index 8725b28ac8..161c8743c4 100644 --- a/studio/backend/tests/test_openai_tool_passthrough.py +++ b/studio/backend/tests/test_openai_tool_passthrough.py @@ -119,7 +119,7 @@ class TestFriendlyUpstreamError: raw = '{"error":{"code":400,"message":"Failed to initialize samplers: failed to parse grammar","type":"invalid_request_error"}}' msg = _friendly_upstream_error(raw) assert "failed to parse grammar" not in msg # raw body is not surfaced verbatim - assert "tool-calling grammar" in msg and "Update Studio" in msg + assert "tool-calling grammar" in msg and "Update Unsloth" in msg def test_failed_to_initialize_samplers_alone_matches(self): assert "tool-calling grammar" in _friendly_upstream_error("Failed to initialize samplers") @@ -262,7 +262,7 @@ class TestChatMessageToolRoles: def test_tool_empty_content_accepted(self): # Empty tool output (mkdir, git add, ...) is routine in agentic loops; - # OpenAI and llama-server both accept it, so Studio must not 400. + # OpenAI and llama-server both accept it, so Unsloth must not 400. msg = ChatMessage(role = "tool", tool_call_id = "call_1", content = "") assert msg.content == "" @@ -400,7 +400,7 @@ class TestChatCompletionRequestToolFields: assert req.session_id == "abc" def test_stream_defaults_false_matching_openai_spec(self): - # OpenAI defaults `stream` to false. Studio used to default true, + # OpenAI defaults `stream` to false. Unsloth used to default true, # breaking naive curl/.NET clients (#5047) that omit it. Pin the fix. req = self._make() assert req.stream is False @@ -664,7 +664,7 @@ class TestChatCompletionRequestToolFields: raise AssertionError("client tools must use passthrough") def generate_chat_completion_with_tools(self, **_kwargs): - raise AssertionError("Studio tool loop must stay disabled") + raise AssertionError("Unsloth tool loop must stay disabled") async def fake_passthrough(llama_backend, payload, model_name, **kwargs): captured["body"] = inference_route._build_openai_passthrough_body( @@ -707,11 +707,11 @@ class TestChatCompletionRequestToolFields: assert monitor.active_count() == 0 def test_permission_mode_does_not_reject_client_tool_passthrough(self, monkeypatch): - # A non-streaming client-tool passthrough (client tools, no Studio tool + # A non-streaming client-tool passthrough (client tools, no Unsloth tool # loop) that also carries permission_mode "ask"/"auto" must reach the # provider passthrough, not the confirm-without-stream guard: the # validator leaves confirm_tool_calls unset for passthrough, and a bare - # permission_mode only gates Studio's own local tool loop. An explicit + # permission_mode only gates Unsloth's own local tool loop. An explicit # confirm_tool_calls=True still forces the local-confirm rejection. # The pre-switch guard only runs when an automatic load may run, so force # that predicate on to exercise it against a resident passthrough backend. @@ -732,7 +732,7 @@ class TestChatCompletionRequestToolFields: raise AssertionError("client tools must use passthrough") def generate_chat_completion_with_tools(self, **_kwargs): - raise AssertionError("Studio tool loop must stay disabled") + raise AssertionError("Unsloth tool loop must stay disabled") async def fake_passthrough(llama_backend, payload, model_name, **kwargs): inference_route.api_monitor.finish(kwargs.get("monitor_id")) @@ -757,7 +757,7 @@ class TestChatCompletionRequestToolFields: return self._v1_client(monkeypatch, _GGUFBackend()) # A process --enable-tools policy must not turn a client-tool passthrough - # into a Studio local loop, so a policy of None or True both keep the + # into an Unsloth local loop, so a policy of None or True both keep the # passthrough (the guard mirrors _explicit_studio_tool_loop_requested). for policy in (None, True): for mode in ("ask", "auto"): @@ -810,7 +810,7 @@ class TestChatCompletionRequestToolFields: assert "requires stream=true" in resp.json()["error"]["message"] def test_permission_mode_policy_forced_local_loop_rejected_before_switch(self, monkeypatch): - # A process --enable-tools policy forces Studio's own tool loop on even + # A process --enable-tools policy forces Unsloth's own tool loop on even # when the request omits enable_tools and carries no client tools. A # non-streaming ask/auto request is then confirm-gated with no stream to # prompt on, so it must 400 at the pre-switch guard -- before @@ -863,7 +863,7 @@ class TestChatCompletionRequestToolFields: def test_enable_tools_on_non_tool_backend_keeps_client_tools_on_passthrough(self, monkeypatch): # DiffusionGemma forces supports_tools off while passthrough stays # available (#6851): enable_tools=True must not steal client tools - # from the passthrough into a Studio tool loop that cannot run. + # from the passthrough into an Unsloth tool loop that cannot run. import routes.inference as inference_route captured = {} @@ -883,7 +883,7 @@ class TestChatCompletionRequestToolFields: raise AssertionError("client tools must use passthrough") def generate_chat_completion_with_tools(self, **_kwargs): - raise AssertionError("Studio tool loop cannot run on a non-tool backend") + raise AssertionError("Unsloth tool loop cannot run on a non-tool backend") async def fake_passthrough(llama_backend, payload, model_name, **kwargs): captured["body"] = inference_route._build_openai_passthrough_body( @@ -2581,7 +2581,7 @@ class TestGgufVisionToolRouting: raise AssertionError("plain GGUF path should not be used") def _tools(**_kwargs): - raise AssertionError("Studio tool loop should not steal response_format") + raise AssertionError("Unsloth tool loop should not steal response_format") backend = SimpleNamespace( is_loaded = True, @@ -2654,7 +2654,7 @@ class TestGgufVisionToolRouting: raise AssertionError("plain GGUF path should not be used") def _tools(**_kwargs): - raise AssertionError("Studio tool loop should not replace client tools") + raise AssertionError("Unsloth tool loop should not replace client tools") backend = SimpleNamespace( is_loaded = True, @@ -2726,7 +2726,7 @@ class TestGgufVisionToolRouting: yield "plain response" def _tools(**_kwargs): - raise AssertionError("tool_choice='none' must not start Studio's tool loop") + raise AssertionError("tool_choice='none' must not start Unsloth's tool loop") backend = SimpleNamespace( is_loaded = True, @@ -2780,7 +2780,7 @@ class TestGgufVisionToolRouting: raise AssertionError("plain GGUF path should not be used") def _tools(**_kwargs): - raise AssertionError("enabled_tools alone must not start Studio's tool loop") + raise AssertionError("enabled_tools alone must not start Unsloth's tool loop") backend = SimpleNamespace( is_loaded = True, @@ -2844,7 +2844,7 @@ class TestGgufVisionToolRouting: raise AssertionError("plain GGUF path should not be used") def _tools(**_kwargs): - raise AssertionError("enabled_tools alone must not start Studio's tool loop") + raise AssertionError("enabled_tools alone must not start Unsloth's tool loop") backend = SimpleNamespace( is_loaded = True, diff --git a/studio/backend/tests/test_orchestrator_unload_cancel.py b/studio/backend/tests/test_orchestrator_unload_cancel.py index fb80b6d061..3a36500aee 100644 --- a/studio/backend/tests/test_orchestrator_unload_cancel.py +++ b/studio/backend/tests/test_orchestrator_unload_cancel.py @@ -34,6 +34,70 @@ def _bare_orchestrator(): return o +def test_adapter_control_raises_stream_errors(monkeypatch): + o = _bare_orchestrator() + monkeypatch.setattr( + o, + "_generate_dispatched", + lambda **_kwargs: iter([orch_mod.GenStreamError("Error: adapter failed")]), + ) + + with pytest.raises(RuntimeError, match = "adapter failed"): + list(o.generate_with_adapter_control(use_adapter = False)) + + closed = [] + + def _stream(**_kwargs): + try: + yield "token" + yield "late token" + finally: + closed.append(True) + + monkeypatch.setattr(o, "_generate_dispatched", _stream) + generator = o.generate_with_adapter_control(use_adapter = False) + assert next(generator) == "token" + generator.close() + assert closed == [True] + + +def test_worker_closes_cancelled_generator_before_gen_done(): + from core.inference.worker import _handle_generate + + events = [] + + class _Backend: + last_generation_stats = None + + def generate_with_adapter_control(self, **_kwargs): + try: + yield "token" + yield "late token" + finally: + events.append("closed") + + class _Responses: + def __init__(self): + self.items = [] + + def put(self, item): + if item["type"] == "gen_done": + assert events == ["closed"] + self.items.append(item) + + responses = _Responses() + cancel = threading.Event() + cancel.set() + _handle_generate( + _Backend(), + {"request_id": "r1", "messages": [], "use_adapter": False}, + responses, + cancel, + ) + + assert [item["type"] for item in responses.items] == ["gen_done"] + + def test_unload_cancels_inflight_generation_then_unloads(monkeypatch): o = _bare_orchestrator() monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True) diff --git a/studio/backend/tests/test_password_prompt_backstop.py b/studio/backend/tests/test_password_prompt_backstop.py index 597eac1625..3c2c1956f9 100644 --- a/studio/backend/tests/test_password_prompt_backstop.py +++ b/studio/backend/tests/test_password_prompt_backstop.py @@ -3,7 +3,7 @@ """Pre-tunnel terminal password gate: never publish a public Cloudflare URL while the seeded default admin password is active. Imports run.py directly, -so run under the Studio venv.""" +so run under the Unsloth venv.""" from __future__ import annotations diff --git a/studio/backend/tests/test_permission_mode.py b/studio/backend/tests/test_permission_mode.py index 3b7197fc49..4fc64a6291 100644 --- a/studio/backend/tests/test_permission_mode.py +++ b/studio/backend/tests/test_permission_mode.py @@ -1438,7 +1438,7 @@ def test_unknown_permission_mode_normalizes_to_ask_on_request_models(): def test_ask_auto_self_enable_confirm_on_chat_request(): # "Ask" gates every call, so a direct /chat/completions caller that requests - # ask but omits the legacy confirm flag self-enables it when Studio's own tool + # ask but omits the legacy confirm flag self-enables it when Unsloth's own tool # loop is requested. Only the router's loop-entry signals count (enable_tools / # mcp_enabled); enabled_tools alone never starts the loop. for loop in ({"enable_tools": True}, {"mcp_enabled": True}): @@ -1481,7 +1481,7 @@ def test_ask_auto_self_enable_confirm_on_chat_request(): confirm_tool_calls = False, ) assert req.confirm_tool_calls is False - # A plain client-tool passthrough (client-supplied tools that Studio does not + # A plain client-tool passthrough (client-supplied tools that Unsloth does not # execute) must NOT self-enable confirm, or the route rejects the passthrough. req = ChatCompletionRequest( messages = [{"role": "user", "content": "hi"}], diff --git a/studio/backend/tests/test_providers_api.py b/studio/backend/tests/test_providers_api.py index 5e24ed752d..7cac3a9e99 100644 --- a/studio/backend/tests/test_providers_api.py +++ b/studio/backend/tests/test_providers_api.py @@ -38,11 +38,11 @@ BASE_URL = os.getenv("STUDIO_TEST_URL", "http://localhost:8000") USERNAME = os.getenv("STUDIO_TEST_USER", "unsloth") PASSWORD = os.getenv("STUDIO_TEST_PASSWORD", "") -# Skip the whole module when no live Studio server / bootstrap password is +# Skip the whole module when no live Unsloth server / bootstrap password is # available (e.g. on CI) so pytest discovery does not error out. pytestmark = pytest.mark.skipif( not PASSWORD, - reason = "Integration test requires a running Studio server; set STUDIO_TEST_PASSWORD to enable.", + reason = "Integration test requires a running Unsloth server; set STUDIO_TEST_PASSWORD to enable.", ) # provider_type → (env var name, model for inference test) diff --git a/studio/backend/tests/test_rag_embed_llama_server.py b/studio/backend/tests/test_rag_embed_llama_server.py index 0e1f74cefe..3a332ee19b 100644 --- a/studio/backend/tests/test_rag_embed_llama_server.py +++ b/studio/backend/tests/test_rag_embed_llama_server.py @@ -149,7 +149,7 @@ def test_build_env_gpu_inherits_devices(monkeypatch): monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "0,1") b = LlamaServerBackend() env = b._build_env("/bin/llama-server", use_gpu = True) - assert env.get("CUDA_VISIBLE_DEVICES") == "0,1" # inherit Studio's selection + assert env.get("CUDA_VISIBLE_DEVICES") == "0,1" # inherit Unsloth's selection def test_use_gpu_explicit_modes(monkeypatch): diff --git a/studio/backend/tests/test_recommended_folders_permission.py b/studio/backend/tests/test_recommended_folders_permission.py index 33a457755e..b65695ad93 100644 --- a/studio/backend/tests/test_recommended_folders_permission.py +++ b/studio/backend/tests/test_recommended_folders_permission.py @@ -112,7 +112,7 @@ def test_path_under_unreadable_parent_returns_false_not_raises(tmp_path): ) def test_demonstrates_the_underlying_stdlib_regression(tmp_path): """Documents *why* _safe_is_dir exists: the old bare pattern raises on - the interpreters Studio ships on (3.12+).""" + the interpreters Unsloth ships on (3.12+).""" parent = tmp_path / "ollama" parent.mkdir() os.chmod(parent, 0o000) diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py index 46dd0d42e4..69715649b7 100644 --- a/studio/backend/tests/test_responses_tool_passthrough.py +++ b/studio/backend/tests/test_responses_tool_passthrough.py @@ -120,7 +120,7 @@ class TestResponsesRequestTools: def test_builtin_tool_type_passes_validation(self): """Non-function built-in tools (web_search, file_search, mcp, ...) must not raise at validation so SDKs that default to them don't - fail on Studio; they're filtered out during translation.""" + fail on Unsloth; they're filtered out during translation.""" req = ResponsesRequest( input = "hi", tools = [{"type": "web_search_preview"}], diff --git a/studio/backend/tests/test_rocm_oom_guard.py b/studio/backend/tests/test_rocm_oom_guard.py index 6e70c7cde4..699d0b74f5 100644 --- a/studio/backend/tests/test_rocm_oom_guard.py +++ b/studio/backend/tests/test_rocm_oom_guard.py @@ -36,7 +36,7 @@ class TestIsIntegratedSignal: """hipDeviceProp_t.integrated wins when truthy; 0/absent never downgrades. Same universal gate PR #5988's UMA safetensors fast-load uses -- keeps - Studio's two unified-memory consumers on one signal.""" + Unsloth's two unified-memory consumers on one signal.""" def test_integrated_upgrades_unknown_apu(self) -> None: # gfx1103 Phoenix iGPU: outside the hardcoded arch set, but the diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py index 915f82ac8e..31c728afca 100644 --- a/studio/backend/tests/test_safetensors_tool_loop.py +++ b/studio/backend/tests/test_safetensors_tool_loop.py @@ -2230,8 +2230,8 @@ def _reprompt_loop(*, auto_heal_tool_calls): tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}], execute_tool = exec_fn, auto_heal_tool_calls = auto_heal_tool_calls, - # Studio always nudges (always-on for the Studio inference paths); the - # API opts in per request. Model the Studio caller here. + # Unsloth always nudges (always-on for the Unsloth inference paths); the + # API opts in per request. Model the Unsloth caller here. nudge_tool_calls = True, max_tool_iterations = 3, ) @@ -3203,7 +3203,7 @@ class TestLoopBehaviour: class TestLoopRePrompt: - """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``MAX_ACT_REPROMPTS`` extra slots. Studio always nudges, so these drive the loop with ``nudge_tool_calls=True``.""" + """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``MAX_ACT_REPROMPTS`` extra slots. Unsloth always nudges, so these drive the loop with ``nudge_tool_calls=True``.""" def test_reasoning_intent_does_not_reprompt_a_visible_answer(self): generations = 0 @@ -4258,7 +4258,7 @@ class TestPlanWithoutActionReprompt: def test_omitted_nudge_flag_is_not_reprompted(self): # The retry is new on this loop: API callers who do not send the flag - # must keep today's behavior. Studio opts in explicitly. + # must keep today's behavior. Unsloth opts in explicitly. loop, exec_fn = _make_loop( turns = [ ["I'll search the web for that."], diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py index 2c13e13bbb..a8c0c2305f 100644 --- a/studio/backend/tests/test_secure_tunnel_gate.py +++ b/studio/backend/tests/test_secure_tunnel_gate.py @@ -2,7 +2,7 @@ # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 """Cloudflare tunnel start gate, incl. --secure on loopback. Imports run.py -directly, so run under the Studio venv.""" +directly, so run under the Unsloth venv.""" from __future__ import annotations diff --git a/studio/backend/tests/test_server_disk_logging.py b/studio/backend/tests/test_server_disk_logging.py index 05d03d869c..ce733c2aaa 100644 --- a/studio/backend/tests/test_server_disk_logging.py +++ b/studio/backend/tests/test_server_disk_logging.py @@ -3,7 +3,7 @@ """Tests for the server session log + native-crash capture in run.py. -Field regression: Studio "terminates without a warning" -- a native crash in +Field regression: Unsloth "terminates without a warning" -- a native crash in the GPU runtime kills the process with no Python traceback, and a desktop- shortcut console closes before anything can be read. The server must tee its console output to disk and aim faulthandler at the same file so even hard diff --git a/studio/backend/tests/test_slot_offload_fit.py b/studio/backend/tests/test_slot_offload_fit.py index ac606e4627..d354c7e113 100644 --- a/studio/backend/tests/test_slot_offload_fit.py +++ b/studio/backend/tests/test_slot_offload_fit.py @@ -3,7 +3,7 @@ """Tests for the offload-avoidance serving-slot reduction (`_slots_that_fit_on_gpu`). -When a pinned context does not fit at the requested `--parallel` slot count, Studio would +When a pinned context does not fit at the requested `--parallel` slot count, Unsloth would flip to `--fit on` and llama-server offloads layers to host RAM, collapsing decode ~3x (oobabooga #6718). Instead the loader retries the on-GPU fit at fewer slots and keeps the largest count that stays fully on GPU (`-ngl -1`). These tests drive the real helper with diff --git a/studio/backend/tests/test_studio_api.py b/studio/backend/tests/test_studio_api.py index 928b636e3e..087c00b648 100644 --- a/studio/backend/tests/test_studio_api.py +++ b/studio/backend/tests/test_studio_api.py @@ -11,7 +11,7 @@ the CLI's ``--help`` output: 1. curl -- basic chat completions (non-streaming) 2. curl -- streaming chat completions 3. Python OpenAI SDK -- streaming completions - 4. curl -- Studio server-side tools (enable_tools=true) + 4. curl -- Unsloth server-side tools (enable_tools=true) 5. curl -- Standard OpenAI function calling (non-streaming) 6. curl -- Standard OpenAI function calling (streaming) 7. curl -- Standard OpenAI function calling (multi-turn tool loop) @@ -31,7 +31,7 @@ Usage: python tests/test_studio_api.py python tests/test_studio_api.py --model unsloth/... --gguf-variant ... - # Pytest mode, external server — start a Studio server yourself, + # Pytest mode, external server — start an Unsloth server yourself, # then point pytest at it. Fastest iteration loop. unsloth studio run --model unsloth/Qwen3-1.7B-GGUF --gguf-variant UD-Q4_K_XL & export UNSLOTH_E2E_BASE_URL=http://127.0.0.1:8080 @@ -341,7 +341,7 @@ def _final_finish_reason(chunks: list[dict]) -> str | None: def test_openai_tools_nonstream(base_url: str, api_key: str): """Standard OpenAI function calling, non-streaming, tool_choice='required'. - Regression: before the fix, Studio stripped `tools` and the model + Regression: before the fix, Unsloth stripped `tools` and the model returned plain text with finish_reason='stop'. After the fix, llama-server's response is forwarded verbatim so the client sees finish_reason='tool_calls' with a structured tool_calls array and diff --git a/studio/backend/tests/test_tensor_parallel.py b/studio/backend/tests/test_tensor_parallel.py index 0d71b89d87..00c7aeac69 100644 --- a/studio/backend/tests/test_tensor_parallel.py +++ b/studio/backend/tests/test_tensor_parallel.py @@ -262,9 +262,12 @@ def test_proportional_tensor_split_is_emitted_in_tensor_mode(): src = _load_model_source() assert '"--tensor-split"' in src gate = src.find("if tensor_parallel:") - ts = src.find('"--tensor-split"') + # Find the TP block's emission (after the gate); manual mode emits its own + # --tensor-split earlier in the source from the user's per-GPU shares. + ts = src.find('"--tensor-split"', gate) nxt_else = src.find("self._tensor_parallel = False") assert 0 <= gate < ts < nxt_else, "--tensor-split must be emitted under `if tensor_parallel:`" + assert "tp_tensor_split" in src[gate:nxt_else] def test_mtp_decode_probe_wired_under_tensor_parallel(): @@ -420,7 +423,7 @@ def test_runtime_recovery_fires_for_user_env_mtp(monkeypatch): # MTP driven by user extra_args / LLAMA_ARG_SPEC_TYPE leaves _speculative_type # unset, but the launch flag still gates recovery on (pass-through MTP). b = _recovery_backend() - b._speculative_type = None # Studio stepped back; user/env owns the spec + b._speculative_type = None # Unsloth stepped back; user/env owns the spec done = threading.Event() captured = {} diff --git a/studio/backend/tests/test_tool_confirm_stream.py b/studio/backend/tests/test_tool_confirm_stream.py index b8e0472e12..0813f6b68d 100644 --- a/studio/backend/tests/test_tool_confirm_stream.py +++ b/studio/backend/tests/test_tool_confirm_stream.py @@ -3,12 +3,12 @@ """End-to-end handshake test for the tool-confirmation gate, no model. -The real Studio stream wrappers in ``routes/inference.py`` drive the +The real Unsloth stream wrappers in ``routes/inference.py`` drive the synchronous agentic generator with ``await asyncio.to_thread(next, gen, ...)`` so the blocking ``threading.Event`` wait runs off the event loop. This test rebuilds that exact pattern around the real ``state.tool_approvals`` functions, served by a real uvicorn process on -loopback (the same server Studio uses), and proves the load-bearing +loopback (the same server Unsloth uses), and proves the load-bearing property: * ``tool_start`` reaches the client before the gate blocks, and diff --git a/studio/backend/tests/test_tool_message_empty_content.py b/studio/backend/tests/test_tool_message_empty_content.py index d63b16ce80..636a35f5a9 100644 --- a/studio/backend/tests/test_tool_message_empty_content.py +++ b/studio/backend/tests/test_tool_message_empty_content.py @@ -4,7 +4,7 @@ """Empty ``role="tool"`` content must be accepted on the OpenAI-compat surface. Agentic clients send ``content: ""`` when a command produced no output; -OpenAI and llama-server both accept it. Studio used to 400, which standard +OpenAI and llama-server both accept it. Unsloth used to 400, which standard clients treat as non-retryable and kill the session. The validator must normalize empty/missing tool content to ``""`` instead of raising. """ diff --git a/studio/backend/tests/test_tp_vision_regression.py b/studio/backend/tests/test_tp_vision_regression.py index 09af876da6..d1372ca415 100644 --- a/studio/backend/tests/test_tp_vision_regression.py +++ b/studio/backend/tests/test_tp_vision_regression.py @@ -126,10 +126,21 @@ _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: tensor needs >= 2 GPUs clearing the compute-buffer reserve. Gated + # on plan_tp (not raw tensor_parallel) so manual mode skips this planner (#6414). + "plan_tp and len(tp_gpus) < 2", # Capacity: pooled usable VRAM can't hold weights + MTP reserve -> layer split. "_tp_weight_budget_mib <= _tp_required_mib", + # Manual mode, Auto layers: --fit owns memory and is incompatible with a + # tensor split, so TP is dropped (surfaced via logger.info) before the + # cache-drop, so a quantized KV survives into the --fit load (#6414). + "tensor_parallel and gpu_memory_mode == 'manual' and (gpu_layers < 0)", + # Manual mode, explicit layers: a tensor split still needs >= 2 GPUs in use. + "tensor_parallel and gpu_memory_mode == 'manual' and (gpu_layers >= 0) and (self._effective_gpu_count(sorted(gpu_ids) if gpu_ids else None) < 2)", + # Manual mode, zero layers: nothing to split on the GPU, and a tensor-mode + # launch under the CPU-only GPU mask (no visible devices) aborts the server + # instead of the intended CPU-only load (#6414). + "gpu_memory_mode == 'manual' and gpu_layers == 0", } @@ -364,7 +375,7 @@ def test_compute_buffer_downgrade_preserves_multi_gpu_intent(): 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") + gate = src.find("plan_tp 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) @@ -625,7 +636,7 @@ def _fallback_loaded_backend(layer_preserves_tensor_intent: bool) -> LlamaCppBac def test_tensor_off_echo_preserves_multi_gpu_fallback(): - """The Studio UI always sends tensor_parallel and echoes the /load response's + """The Unsloth 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).""" diff --git a/studio/backend/tests/test_trained_model_scan.py b/studio/backend/tests/test_trained_model_scan.py index 7bf572e214..5d74bb7d28 100644 --- a/studio/backend/tests/test_trained_model_scan.py +++ b/studio/backend/tests/test_trained_model_scan.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 -"""Tests for Studio trained-model discovery used by Chat.""" +"""Tests for Unsloth trained-model discovery used by Chat.""" import json from pathlib import Path diff --git a/studio/backend/tests/test_training_config_popover_source.py b/studio/backend/tests/test_training_config_popover_source.py new file mode 100644 index 0000000000..4263b012eb --- /dev/null +++ b/studio/backend/tests/test_training_config_popover_source.py @@ -0,0 +1,109 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +"""Source-level regression guards for the Training Config popover data source +(#6853). + +The live Training Progress popover used to read the editable form store +(useTrainingConfigStore) while a run was active, so it showed stale/static +values whenever the user touched the form after starting the run; only the +History view read the run's saved config snapshot. These guards pin the fixed +wiring: both views feed ProgressSection a config override mapped from +GET /api/train/runs/{id}, and ProgressSection prefers that override whenever +one is present -- not only for historical views. +""" + +from __future__ import annotations + +from pathlib import Path + +_STUDIO_FRONTEND = Path(__file__).resolve().parents[2] / "frontend" / "src" / "features" / "studio" + + +def _read(rel: str) -> str: + return (_STUDIO_FRONTEND / rel).read_text(encoding = "utf-8") + + +def test_progress_section_prefers_override_over_form_store(): + src = _read("sections/progress-section.tsx") + # Fields key on the override's presence, not isHistorical: a live view passing + # an override wins over the store; without one, live keeps the store while + # History shows blanks rather than unrelated live form values. + assert "const cfg = configOverride ?? (isHistorical ? undefined : config)" in src + assert "const cfgEpochs = cfg?.epochs" in src + assert "isHistorical ? configOverride?.epochs" not in src + + +def test_live_view_fetches_the_active_run_config(): + src = _read("live-training-view.tsx") + # Live view resolves the run's saved config snapshot by job id... + assert "getTrainingRun(" in src + assert "mapRunConfigToOverride(" in src + # ...and hands it to the popover. + assert "configOverride={runConfigOverride}" in src + + +def test_live_view_fetches_as_soon_as_the_job_id_exists(): + # start_training() inserts the run row BEFORE the pump consumes any event, so + # the saved config is available during configuring/loading/downloading. The + # job id is therefore the whole readiness condition: gating on a first step + # or a terminal phase would show the wrong config for the entire pre-step + # window of a long load, or for a run adopted from another client. + src = _read("live-training-view.tsx") + assert "if (!runtime.jobId) {" in src + assert "[runtime.jobId, fetchedRunConfig, fetchAttempt]" in src + # No step/phase readiness gate may creep back in. + assert "runRowReady" not in src + + +def test_live_view_retries_the_transient_row_miss(): + # start_training() creates the row before the pump, but a lookup racing that + # commit can still 404. Nothing else in the effect deps changes on failure, so + # the retry must be explicit and bounded, else a genuinely absent row would + # poll forever instead of falling back to the form store. + src = _read("live-training-view.tsx") + assert "RUN_CONFIG_FETCH_RETRIES" in src + assert "RUN_CONFIG_FETCH_RETRY_MS" in src + assert "setFetchAttempt(" in src + assert "attempts >= RUN_CONFIG_FETCH_RETRIES" in src + # The budget is keyed by job so a new run always starts fresh. + assert "fetchAttempt?.jobId === jobId ? fetchAttempt.count : 0" in src + # The pending retry must be cancelled with the effect. + assert "clearTimeout(retryTimer)" in src + + +def test_live_view_prefers_saved_training_method(): + # The method label / LoRA-row visibility must come from the run snapshot, + # not the editable form (which may have changed since the run started). + src = _read("live-training-view.tsx") + assert "runConfigOverride?.trainingMethod ?? config.trainingMethod" in src + + +def test_history_view_uses_the_shared_mapper(): + src = _read("historical-training-view.tsx") + # Shared mapper, not a re-inlined field-by-field copy that could drift. + assert "mapRunConfigToOverride(detail.config)" in src + assert "num_epochs" not in src + + +def test_shared_mapper_matches_backend_config_keys(): + src = _read("sections/run-config-override.ts") + # The mapper reads the run config JSON the backend snapshots at job start; + # keep the key set pinned so a silent rename breaks loudly here. + for key in ( + "training_type", + "load_in_4bit", + "num_epochs", + "batch_size", + "learning_rate", + "max_steps", + "max_seq_length", + "warmup_steps", + "optim", + "lora_r", + "lora_alpha", + "lora_dropout", + "use_rslora", + "use_loftq", + ): + assert key in src, f"run-config mapper lost backend key {key}" diff --git a/studio/backend/tests/test_training_nan_loss_handling.py b/studio/backend/tests/test_training_nan_loss_handling.py index a2dc78bee2..5a477a084d 100644 --- a/studio/backend/tests/test_training_nan_loss_handling.py +++ b/studio/backend/tests/test_training_nan_loss_handling.py @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -"""Pin Studio's behavior when a training event reports non-finite (NaN/Inf) loss. +"""Pin Unsloth's behavior when a training event reports non-finite (NaN/Inf) loss. The training event handler used to filter NaN/Inf to None silently while leaving the previous finite loss in progress.loss — so the API kept reporting diff --git a/studio/backend/tests/test_training_stop_watchdog.py b/studio/backend/tests/test_training_stop_watchdog.py index 457dfc8ea2..0cd702bce2 100644 --- a/studio/backend/tests/test_training_stop_watchdog.py +++ b/studio/backend/tests/test_training_stop_watchdog.py @@ -258,15 +258,27 @@ def test_watchdog_no_op_when_worker_superseded(monkeypatch): def test_new_run_gets_its_own_watchdog(monkeypatch): # A stale watchdog sleeping on an old proc must not stop a new run's stop from # creating its own watcher. - monkeypatch.setitem(_G, "_STOP_GRACE_S", 100.0) - monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0) b = TrainingBackend() - _record_force_terminate(monkeypatch, b) + started = [] + release = threading.Event() + + def _blocked_watchdog( + target_proc, + cancel, + watched_job_id = None, + ): + started.append(target_proc) + # No timeout: the finally always releases this, so a superseded watchdog stays + # alive through the assertions regardless of load; as a daemon it can't hang exit. + release.wait() + + monkeypatch.setattr(b, "_stop_watchdog_loop", _blocked_watchdog) old_proc = _FakeProc(alive = True) b._proc = old_proc b._start_stop_watchdog(cancel = False) first_wd = b._stop_watchdog + assert _wait_until(lambda: started == [old_proc]) # New run: fresh worker replaces the handle; its stop must get a new watcher # even though the old (superseded) watchdog is still alive. @@ -276,12 +288,12 @@ def test_new_run_gets_its_own_watchdog(monkeypatch): second_wd = b._stop_watchdog try: + assert _wait_until(lambda: started == [old_proc, new_proc]) assert first_wd.is_alive() assert second_wd is not first_wd, "a new run must get its own watchdog" assert b._stop_watchdog_proc is new_proc finally: - old_proc._alive = False - new_proc._alive = False + release.set() first_wd.join(timeout = 5) second_wd.join(timeout = 5) diff --git a/studio/backend/tests/test_transformers_latest.py b/studio/backend/tests/test_transformers_latest.py index 20616dccba..af48d674cc 100644 --- a/studio/backend/tests/test_transformers_latest.py +++ b/studio/backend/tests/test_transformers_latest.py @@ -1036,7 +1036,7 @@ def test_upgrade_check_mixed_pypi_main_reports_dev_only(monkeypatch): def test_install_endpoint_not_mounted_on_v1(): - """The consented pip-install endpoint is a Studio admin action; it must live + """The consented pip-install endpoint is an Unsloth admin action; it must live on studio_router (kept off the OpenAI-compatible /v1 mount), not router.""" from routes import inference as ri diff --git a/studio/backend/utils/_studio_release_build.py b/studio/backend/utils/_studio_release_build.py index 267197a202..07ede36912 100644 --- a/studio/backend/utils/_studio_release_build.py +++ b/studio/backend/utils/_studio_release_build.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 -"""Build-stamped Studio release metadata. +"""Build-stamped Unsloth release metadata. Release builds may rewrite this module in the build workspace before creating Python artifacts. Keep the committed value neutral so source checkouts do not diff --git a/studio/backend/utils/api_errors.py b/studio/backend/utils/api_errors.py index cae8daf287..a3686c3a26 100644 --- a/studio/backend/utils/api_errors.py +++ b/studio/backend/utils/api_errors.py @@ -20,7 +20,7 @@ client-error responses on the ``/v1/*`` surface: CRITICAL: the exception handlers installed by :func:`install_api_error_handlers` are global, but they ONLY transform responses for paths that start with ``/v1/``. For every other path (``/api/...``, frontend routes) they reproduce FastAPI's -default behavior byte-for-byte, because the Studio frontend depends on the +default behavior byte-for-byte, because the Unsloth frontend depends on the ``{"detail": ...}`` shape for ``/api/*``. Public contract (other modules depend on these): @@ -107,7 +107,7 @@ def anthropic_error_body( Returns ``{"type": "error", "request_id": None, "error": {"type", "message"}}``. ``request_id`` is a required (nullable) field on the spec's ErrorResponse; - Studio has no request-id system, so it is null. ``err_type`` defaults to + Unsloth has no request-id system, so it is null. ``err_type`` defaults to :data:`ANTHROPIC_TYPE_BY_STATUS` for ``status`` (``"api_error"`` fallback). """ return { @@ -192,7 +192,7 @@ def install_api_error_handlers(app) -> None: Both handlers are global but only transform responses for OpenAI/Anthropic- compatible surfaces (see :func:`wants_api_error_envelope`: the ``/v1/*`` mount and the preview ``/p/.../v1/*`` mount). Every other path reproduces FastAPI's - default ``{"detail": ...}`` behavior exactly so the Studio frontend keeps working. + default ``{"detail": ...}`` behavior exactly so the Unsloth frontend keeps working. """ @app.exception_handler(RequestValidationError) diff --git a/studio/backend/utils/client_ip.py b/studio/backend/utils/client_ip.py index 94acbf1809..cc48a096d2 100644 --- a/studio/backend/utils/client_ip.py +++ b/studio/backend/utils/client_ip.py @@ -4,12 +4,12 @@ """Resolve the caller's IP for rate limiting. Trust model, in order: - 1. If the operator opts in via ``UNSLOTH_STUDIO_TRUST_FORWARDED`` (Studio behind + 1. If the operator opts in via ``UNSLOTH_STUDIO_TRUST_FORWARDED`` (Unsloth behind their own reverse proxy), honor the *rightmost* ``X-Forwarded-For`` hop -- the one the trusted proxy appended. The leftmost entry is client-controlled and spoofable, so this assumes a proxy that appends (or overwrites) the header; only enable the env var behind such a proxy. - 2. If the socket peer is loopback, honor ``CF-Connecting-IP``. Studio's managed + 2. If the socket peer is loopback, honor ``CF-Connecting-IP``. Unsloth's managed Cloudflare tunnel terminates at 127.0.0.1, so every tunneled visitor would otherwise collapse onto the same socket peer (the local cloudflared process) and share one rate-limit bucket. ``CF-Connecting-IP`` is set by Cloudflare's diff --git a/studio/backend/utils/cpu_threads.py b/studio/backend/utils/cpu_threads.py index 4ed0021054..91d577408d 100644 --- a/studio/backend/utils/cpu_threads.py +++ b/studio/backend/utils/cpu_threads.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 -"""Early CPU thread-pool configuration for Studio processes.""" +"""Early CPU thread-pool configuration for Unsloth processes.""" import os from typing import MutableMapping, Optional diff --git a/studio/backend/utils/datasets/cache_safe.py b/studio/backend/utils/datasets/cache_safe.py index e629210f33..d2dc7b737a 100644 --- a/studio/backend/utils/datasets/cache_safe.py +++ b/studio/backend/utils/datasets/cache_safe.py @@ -7,7 +7,7 @@ A shared HF datasets cache can contain subtrees owned by another user (for example populated by an earlier root-run job). datasets then raises "[Errno 13] Permission denied: ..._builder.lock" while locking the cached builder, killing the training run even though the dataset itself is fine. -Retry such loads in a Studio-owned cache so the run proceeds; the worst case +Retry such loads in an Unsloth-owned cache so the run proceeds; the worst case is one rebuild of the dataset in the fallback location. """ @@ -26,7 +26,7 @@ def studio_datasets_cache() -> str: def load_dataset_cache_safe(*args, **kwargs): - """datasets.load_dataset, retried in a Studio-owned cache on EACCES.""" + """datasets.load_dataset, retried in an Unsloth-owned cache on EACCES.""" from datasets import load_dataset try: return load_dataset(*args, **kwargs) diff --git a/studio/backend/utils/hardware/VRAM_ESTIMATION.md b/studio/backend/utils/hardware/VRAM_ESTIMATION.md index a6b4de29d2..68ca1d5ffd 100644 --- a/studio/backend/utils/hardware/VRAM_ESTIMATION.md +++ b/studio/backend/utils/hardware/VRAM_ESTIMATION.md @@ -106,7 +106,7 @@ Non_flash_attention = B * num_attention_heads * S^2 * 2 * 12.0 * effective_layer Activations = max(Per_layer_with_gc, Non_flash_attention) ``` -Studio resolves the attention implementation with Unsloth's +Unsloth resolves the attention implementation with Unsloth's `resolve_attention_implementation` helper and uses that result directly. The estimator does not duplicate model-family attention policy. diff --git a/studio/backend/utils/hardware/amd.py b/studio/backend/utils/hardware/amd.py index f5b64c45d0..91a06c9a2a 100644 --- a/studio/backend/utils/hardware/amd.py +++ b/studio/backend/utils/hardware/amd.py @@ -125,7 +125,7 @@ def _run_amd_smi(*args: str, timeout: int = _AMD_SMI_DEFAULT_TIMEOUT) -> Optiona # amd-smi does not exist on Windows (neither Adrenalin nor the HIP SDK # ship a CLI) and can be absent on minimal Linux installs. Disable the # poller in one step instead of burning the 3-strike circuit breaker - # on guaranteed FileNotFoundError spawns. Studio's VRAM display falls + # on guaranteed FileNotFoundError spawns. Unsloth's VRAM display falls # back to torch mem_get_info. if not _amd_smi_disabled: logger.info( diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py index 117ad7b780..adc9a54aab 100644 --- a/studio/backend/utils/hardware/hardware.py +++ b/studio/backend/utils/hardware/hardware.py @@ -37,7 +37,7 @@ logger = get_logger(__name__) # ── GPU index ordering ────────────────────────────────────────────────────── # CUDA defaults to CUDA_DEVICE_ORDER=FASTEST_FIRST, numbering GPUs by compute -# performance. nvidia-smi -- and every free-VRAM probe in Studio -- numbers GPUs +# performance. nvidia-smi -- and every free-VRAM probe in Unsloth -- numbers GPUs # by PCI bus id instead. On a mixed-GPU host (e.g. an RTX 5090 alongside an RTX # PRO 6000) the two orderings disagree, so an index picked from nvidia-smi data # ("the emptiest card is GPU 1") gets written into CUDA_VISIBLE_DEVICES and then @@ -49,7 +49,7 @@ logger = get_logger(__name__) # and spawn workers copy os.environ. setdefault so an explicit user override wins. os.environ.setdefault("CUDA_DEVICE_ORDER", "PCI_BUS_ID") -# Studio workers can import MLX without importing unsloth first, so mirror the +# Unsloth workers can import MLX without importing unsloth first, so mirror the # package bootstrap here. Keep an explicit user value authoritative. if platform.system() == "Darwin" and platform.machine() == "arm64": os.environ.setdefault("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1") @@ -117,7 +117,7 @@ def _has_mlx() -> bool: def _has_usable_mlx_stack() -> bool: - """True only when the FULL Studio MLX training/export stack is usable + """True only when the FULL Unsloth MLX training/export stack is usable (mlx + mlx-lm + mlx-vlm at the minimum versions unsloth-zoo requires), not just a bare ``import mlx.core``. A backtracked/old mlx-vlm still imports but breaks VLM Train/Export, so the training gate must match the self-heal's own diff --git a/studio/backend/utils/helper_precache_settings.py b/studio/backend/utils/helper_precache_settings.py index db19a2d028..e7d3c0e6dd 100644 --- a/studio/backend/utils/helper_precache_settings.py +++ b/studio/backend/utils/helper_precache_settings.py @@ -32,7 +32,7 @@ def helper_model_disabled_by_env() -> bool: def get_helper_precache_enabled() -> bool: """Read the persisted startup pre-cache preference. - Missing or unreadable settings default to False so Studio startup never + Missing or unreadable settings default to False so Unsloth startup never performs optional network work unless the user explicitly opted in. """ try: @@ -45,7 +45,7 @@ def get_helper_precache_enabled() -> bool: def set_helper_precache_enabled(value: Any) -> bool: - """Persist whether Studio should pre-cache the Helper LLM at startup.""" + """Persist whether Unsloth should pre-cache the Helper LLM at startup.""" parsed = _coerce_bool(value) if parsed is None: raise ValueError("Helper LLM startup pre-cache must be true or false.") diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py index 9bc4a60fad..2628b99a2d 100644 --- a/studio/backend/utils/hf_xet_fallback.py +++ b/studio/backend/utils/hf_xet_fallback.py @@ -1,9 +1,9 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -"""Studio shim over the shared ``unsloth_zoo.hf_xet_fallback`` Xet -> HTTP stall fallback. +"""Unsloth shim over the shared ``unsloth_zoo.hf_xet_fallback`` Xet -> HTTP stall fallback. -Re-exports the shared API and injects Studio's marker-aware cache purge +Re-exports the shared API and injects Unsloth's marker-aware cache purge (``prepare_cache_for_transport``) so the download manager keeps its ``.transport`` marker semantics on the HTTP retry. @@ -68,7 +68,7 @@ def _load_shared() -> bool: _shared_available = True _shared_import_error = None return True - except Exception as exc2: # noqa: BLE001 - degrade so Studio still boots with plain HF + except Exception as exc2: # noqa: BLE001 - degrade so Unsloth still boots with plain HF _shared_import_error = exc2 _shared_available = False import logging as _logging @@ -263,7 +263,7 @@ __all__ = [ def _studio_prepare_for_http(repo_type: str, repo_id: str) -> None: - """Studio's marker-aware purge before an HTTP resume, keeping the download manager's ``.transport`` + """Unsloth's marker-aware purge before an HTTP resume, keeping the download manager's ``.transport`` accounting consistent (vs unsloth_zoo's generic default). Guarded: a purge failure is logged, not fatal to the retry.""" try: @@ -273,7 +273,7 @@ def _studio_prepare_for_http(repo_type: str, repo_id: str) -> None: try: from loggers import get_logger get_logger(__name__).debug( - "Studio prepare_cache_for_transport failed for %s: %s", repo_id, exc + "Unsloth prepare_cache_for_transport failed for %s: %s", repo_id, exc ) except ModuleNotFoundError as logger_exc: if logger_exc.name != "loggers": @@ -294,8 +294,8 @@ def hf_hub_download_with_xet_fallback( on_status: Optional[Callable[[str], None]] = None, force_download: bool = False, ) -> str: - """Single-file download via the shared fallback with Studio's marker-aware HTTP-retry prep. - ``force_download`` re-fetches a newer blob over a cached one (Studio's model-update path).""" + """Single-file download via the shared fallback with Unsloth's marker-aware HTTP-retry prep. + ``force_download`` re-fetches a newer blob over a cached one (Unsloth's model-update path).""" return _shared_hf_hub_download_with_xet_fallback( repo_id, filename, @@ -313,6 +313,6 @@ def hf_hub_download_with_xet_fallback( def snapshot_download_with_xet_fallback(repo_id: str, **kwargs: Any) -> str: - """Whole-repo download via the shared fallback with Studio's marker-aware HTTP-retry prep.""" + """Whole-repo download via the shared fallback with Unsloth's marker-aware HTTP-retry prep.""" kwargs.setdefault("prepare_for_http_fn", _studio_prepare_for_http) return _shared_snapshot_download_with_xet_fallback(repo_id, **kwargs) diff --git a/studio/backend/utils/hidden_models.py b/studio/backend/utils/hidden_models.py new file mode 100644 index 0000000000..20d0bb966e --- /dev/null +++ b/studio/backend/utils/hidden_models.py @@ -0,0 +1,142 @@ +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +"""Infra-only model detection shared by the model routes and the hub +inventory. Lives directly under ``utils`` (not ``utils.models``) so the hub +cache scanner can import it without pulling in ``utils/models/__init__.py``, +which eagerly loads the model-config/checkpoint stack, and without importing +``routes.models`` (import-time side effects, would cycle).""" + +from __future__ import annotations + +import re +from pathlib import Path +from typing import Optional + +# Hub repo id shape ("owner/name", no leading separator); anything else is +# treated as a local filesystem path. +_HF_REPO_ID_RE = re.compile(r"^[A-Za-z0-9][\w.\-]*/[\w.\-]+$") + +# The llama.cpp install-validation probe repo. Always hidden. +_PROBE_REPO_ID = "ggml-org/models" +# The probe's on-disk filename. Carries the ".gguf" so it stays specific and +# does not hide unrelated repos like ``user/stories260K-finetune-GGUF``. +_PROBE_FILENAME = "stories260k.gguf" +# Keep previously cached defaults hidden after settings changes. +_DEFAULT_EMBEDDING_REPO_IDS = { + "unsloth/bge-small-en-v1.5", + "unsloth/bge-small-en-v1.5-GGUF", +} +# Local copies do not always retain the repo id. Keep a narrow basename +# fallback for Studio's static default embedder only; configured custom repos +# remain exact-match-only. +_DEFAULT_EMBEDDING_PATH_BASENAMES = {"bge-small-en-v1.5"} + + +def _safe_resolve(path: Path) -> Optional[str]: + """resolve() to a string, or None when the path is inaccessible.""" + try: + return str(path.resolve()) + except OSError: + return None + + +def _existing_resolved_path(value: str) -> Optional[str]: + """Resolve an existing local path.""" + path = Path(value).expanduser() + try: + if not path.exists(): + return None + except OSError: + return None + return _safe_resolve(path) + + +def _path_contains_repo_id(value: str, repo_ids: set[str]) -> bool: + """Match exact repo-derived path segments.""" + parts = [part for part in value.lower().replace("\\", "/").split("/") if part] + for repo_id in repo_ids: + owner, name = repo_id.split("/", 1) + if f"models--{owner}--{name}" in parts: + return True + if any( + parts[index] == owner and parts[index + 1] == name for index in range(len(parts) - 1) + ): + return True + return False + + +def _path_basename_is_default_embedder(value: str) -> bool: + """Match a default embedder folder or a suffixed local weight filename.""" + normalized = value.lower().replace("\\", "/").rstrip("/") + basename = normalized.rsplit("/", 1)[-1] + return any( + basename == needle + or any(basename.startswith(f"{needle}{separator}") for separator in ("-", "_", ".")) + for needle in _DEFAULT_EMBEDDING_PATH_BASENAMES + ) + + +def is_hidden_model(*values: str | None) -> bool: + """True if any id/path is the RAG embedding model (the effective embedder + or its GGUF companion repo) or the llama.cpp install validation probe + (ggml-org/models / stories260K), so pickers hide them (GGUF and non-GGUF). + None are usable chat models; the probe can be cached as a side effect of + installing the prebuilt llama-server and otherwise sorts smallest, so it + would be auto-selected. + + Hub repo ids are matched EXACTLY (case-insensitive full "owner/name"), so a + custom embedder with a generic basename like "org/model" cannot substring + hide unrelated cached repos such as "user/model-chat" or "org/model-GGUF". + Existing paths take precedence over the identical ``owner/name`` repo + shape. Cache and LM Studio paths use exact repo-derived segments. Local + copies of the static default embedder also use a boundary-aware basename + fallback; configured custom repos never do.""" + from core.rag import config as rag_config + + hidden_repo_ids = { + _PROBE_REPO_ID.lower(), + *(repo_id.lower() for repo_id in _DEFAULT_EMBEDDING_REPO_IDS), + } + exact_paths: list[str] = [] + for model in { + rag_config.EMBEDDING_MODEL, + rag_config.default_gguf_repo(), + rag_config.effective_embedding_model(), + rag_config.effective_gguf_repo(), + }: + existing_path = _existing_resolved_path(model) + if existing_path: + exact_paths.append(existing_path.lower()) + elif _HF_REPO_ID_RE.match(model): + hidden_repo_ids.add(model.lower()) + else: + resolved = _safe_resolve(Path(model).expanduser()) + if resolved: + exact_paths.append(resolved.lower()) + for v in values: + if not v: + continue + low = v.lower() + if _HF_REPO_ID_RE.match(v): + # A repo id ("owner/name"): match the hidden set exactly. It is + # never a filesystem path, so skip the path/filename checks. + if low in hidden_repo_ids: + return True + continue + # Anything else is treated as a filesystem path (the cached snapshot + # path, or a local model id). Match the probe by its exact filename and + # any configured local-path embedder by exact resolved path. Split on + # both separators so a Windows-style path ("...\\stories260K.gguf") is + # matched even when this runs on a POSIX interpreter (and vice versa). + if low.replace("\\", "/").rsplit("/", 1)[-1] == _PROBE_FILENAME: + return True + if _path_basename_is_default_embedder(v): + return True + if _path_contains_repo_id(v, hidden_repo_ids): + return True + if exact_paths: + resolved = _safe_resolve(Path(v).expanduser()) + if resolved and resolved.lower() in exact_paths: + return True + return False diff --git a/studio/backend/utils/host_policy.py b/studio/backend/utils/host_policy.py index f506eadc03..55565bb338 100644 --- a/studio/backend/utils/host_policy.py +++ b/studio/backend/utils/host_policy.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 -"""Bind-host trust policy for the Studio backend. +"""Bind-host trust policy for the Unsloth backend. Stdlib only -- safe to import without the rest of the backend. diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py index f6d3635301..31dbda63ea 100644 --- a/studio/backend/utils/llama_cpp_update.py +++ b/studio/backend/utils/llama_cpp_update.py @@ -82,7 +82,7 @@ def _utcnow() -> str: def _find_binary() -> Optional[str]: """Locate the active llama-server binary via the inference backend's own - resolver, so update targets exactly what Studio runs. Lazy import keeps the + resolver, so update targets exactly what Unsloth runs. Lazy import keeps the heavy inference module off this module's import path.""" try: from core.inference.llama_cpp import LlamaCppBackend @@ -109,7 +109,7 @@ def _installer_script() -> Optional[Path]: """Locate install_llama_prebuilt.py. Honours UNSLOTH_LLAMA_INSTALLER, then searches up from this file for both ``/install_llama_prebuilt.py`` and ``/studio/install_llama_prebuilt.py`` so it works in the dev tree and - in an installed Studio layout.""" + in an installed Unsloth layout.""" env = os.environ.get("UNSLOTH_LLAMA_INSTALLER") if env and Path(env).is_file(): return Path(env) @@ -227,7 +227,7 @@ def _is_under(path: Path, root: Path) -> bool: def _llama_install_root(binary: Optional[str]) -> Optional[Path]: - """The Studio-managed llama.cpp root the active binary lives under, or None + """The Unsloth-managed llama.cpp root the active binary lives under, or None when the binary is unmanaged. Installing anywhere the active binary is not would not replace what _find_llama_server_binary runs (which prefers a pinned LLAMA_SERVER_PATH, then UNSLOTH_LLAMA_CPP_PATH, then a llama.cpp tree), so we @@ -327,7 +327,7 @@ def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]: def _is_external_link(path: Optional[Path]) -> bool: """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink or a Windows directory junction / reparse point. Such a link resolves into - the user's own llama.cpp checkout, so Studio must never auto-update it.""" + the user's own llama.cpp checkout, so Unsloth must never auto-update it.""" if path is None: return False try: @@ -635,7 +635,7 @@ def start_update() -> dict: "reason": "local_link", "message": ( "llama.cpp is a local directory linked with --with-llama-cpp-dir; " - "Studio won't replace it. Update your own llama.cpp checkout instead." + "Unsloth won't replace it. Update your own llama.cpp checkout instead." ), "job": get_update_status()["job"], } diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py index 7e1c9864c9..4ea1ec62f5 100644 --- a/studio/backend/utils/mlx_repair.py +++ b/studio/backend/utils/mlx_repair.py @@ -3,7 +3,7 @@ """Best-effort MLX self-heal for Apple Silicon. -On macOS, Studio enables Train/Export only when the MLX training/export stack is +On macOS, Unsloth enables Train/Export only when the MLX training/export stack is usable (see utils.hardware.hardware.detect_hardware -> CHAT_ONLY). MLX is pulled only transitively via unsloth-zoo, and a resolver backtrack (mlx-vlm -> transformers>=5 vs the single-env transformers pin) can silently drop it, leaving @@ -13,7 +13,7 @@ a background thread, then re-detects so the gate re-opens without a manual The install mirrors the main Apple Silicon installer (install_python_stack.py): it points UV_OVERRIDE at overrides-darwin-arm64.txt so the resolver keeps the -Studio transformers pin AND installs a current mlx-vlm, and it requires the same +Unsloth transformers pin AND installs a current mlx-vlm, and it requires the same minimum versions unsloth-zoo declares so a backtracked old mlx-vlm (which still imports but breaks VLM Train/Export) is never accepted as healthy. @@ -69,11 +69,11 @@ _MLX_REINSTALL_ARGS = tuple( # reject anything. mlx/mlx-metal ship wheels only (no sdist on PyPI) and # mlx-lm/mlx-vlm publish py3-none-any wheels, so requiring wheels does not break a # healthy self-heal; if a wheel is genuinely unavailable the install fails and -# Studio stays chat-only (the existing safe fallback) until `unsloth studio update`. +# Unsloth stays chat-only (the existing safe fallback) until `unsloth studio update`. _ONLY_BINARY_ARG = "--only-binary=:all:" # Allowlist of environment variables forwarded to the install subprocess. The # self-heal runs without confirmation on the default startup path, so it must not -# hand resolver/build code the full Studio environment. Everything outside this +# hand resolver/build code the full Unsloth environment. Everything outside this # set is dropped, which excludes three dangerous classes by construction: # * secrets (HF_TOKEN, AWS_*, WANDB_API_KEY, ...) that a malicious wheel/sdist # build hook would otherwise read straight out of os.environ; @@ -207,13 +207,13 @@ def _mlx_install_env() -> dict[str, str]: The self-heal runs without confirmation on the default startup path, so it forwards only the variables uv genuinely needs (see _MLX_ENV_ALLOWLIST) instead - of the full Studio environment: secrets and package-source redirects in + of the full Unsloth environment: secrets and package-source redirects in os.environ are dropped so a malicious resolver-selected artifact cannot read - Studio secrets or be steered to a hostile index. + Unsloth secrets or be steered to a hostile index. Mirror the main installer (install_python_stack.py) by pointing UV_OVERRIDE at overrides-darwin-arm64.txt, which relaxes mlx-vlm/mlx-lm's transformers>=5 - requirement to >=4.57.6. Without it, uv keeps the Studio transformers pin only + requirement to >=4.57.6. Without it, uv keeps the Unsloth transformers pin only by silently backtracking mlx-vlm to an old, unsupported version (uv honours UV_OVERRIDE; plain pip ignores it, so the transformers constraint below is the pip-path safety net). We set UV_OVERRIDE ourselves, so a poisoned one in the @@ -234,17 +234,17 @@ def _mlx_install_env() -> dict[str, str]: def _transformers_constraint_args() -> tuple[list[str], str | None]: """Pin transformers to the running version for the mlx install. - The install must never upgrade transformers underneath a running Studio + The install must never upgrade transformers underneath a running Unsloth (the single-env install pins transformers==4.57.6). With UV_OVERRIDE set this is belt-and-suspenders; on the plain-pip path (no UV_OVERRIDE support) it is the actual guard -- the resolver either finds an mlx build compatible with the - pin or fails, leaving us chat-only rather than breaking Studio. Returns + pin or fails, leaving us chat-only rather than breaking Unsloth. Returns (pip args, temp file path to clean up). Read the version from installed metadata rather than `import transformers`: transformers can have valid metadata yet fail to import (e.g. an incompatible huggingface_hub), and in that case we still want to pin it so the mlx install - cannot quietly upgrade it out from under Studio.""" + cannot quietly upgrade it out from under Unsloth.""" from importlib.metadata import PackageNotFoundError, version as _dist_version try: @@ -263,10 +263,10 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool: """Install a usable mlx/mlx-lm/mlx-vlm stack by name into the running venv. Best-effort; returns True iff the resulting stack meets unsloth-zoo's minimums (so a backtracked old mlx-vlm is rejected, not accepted). transformers is held - at its pinned version so the install can never upgrade it underneath Studio.""" + at its pinned version so the install can never upgrade it underneath Unsloth.""" # Prepare the constraint inside the try: this runs on a daemon thread, so an # exception here (e.g. tempfile.mkstemp failing on a full disk or bad TMPDIR) - # must leave Studio chat-only, not crash the background self-heal thread. + # must leave Unsloth chat-only, not crash the background self-heal thread. constraint_path = None try: constraint_args, constraint_path = _transformers_constraint_args() @@ -279,7 +279,7 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool: ) if cmd is None: logger.warning( - "MLX self-heal requires uv so Studio can apply dependency overrides; " + "MLX self-heal requires uv so Unsloth can apply dependency overrides; " "staying chat-only. Run `unsloth studio update` to restore uv." ) return False diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py index b6b080b1c4..f2125ad034 100644 --- a/studio/backend/utils/models/checkpoints.py +++ b/studio/backend/utils/models/checkpoints.py @@ -37,7 +37,7 @@ def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]: def _infer_base_model_from_history(checkpoint_dir: Path) -> Optional[str]: - """Best-effort base-model lookup using persisted Studio run metadata.""" + """Best-effort base-model lookup using persisted Unsloth run metadata.""" checkpoint_name = checkpoint_dir.name resolved_checkpoint_dir = str(checkpoint_dir.resolve()) diff --git a/studio/backend/utils/models/gguf_metadata.py b/studio/backend/utils/models/gguf_metadata.py index c24ec28e1d..50b3cd3513 100644 --- a/studio/backend/utils/models/gguf_metadata.py +++ b/studio/backend/utils/models/gguf_metadata.py @@ -50,9 +50,11 @@ _CACHE_MAX_ENTRIES = 4096 # keyed by (file cache key, wanted key). None = key absent / file unreadable. _BOOL_CACHE: Dict[Tuple[_CacheKey, str], Optional[bool]] = {} -# Native training context length (``{arch}.context_length``). None = absent / -# unreadable. Lets the UI show the real context ceiling before a model loads. -_CONTEXT_CACHE: Dict[_CacheKey, Optional[int]] = {} +# GGUF header dims for the staged/deferred-load UI: context_length, layer_count +# (block_count), and moe_layer_count (block_count minus leading dense layers; 0 +# if not MoE). One cached pass fills all three so the staged sheet can size every +# slider before the model loads. None = unreadable / not a GGUF. +_DIMS_CACHE: Dict[_CacheKey, Optional[Dict[str, Optional[int]]]] = {} def _cache_key(path: str) -> Optional[_CacheKey]: @@ -142,32 +144,45 @@ def _parse_gguf_header(path: str) -> Optional[Dict[str, str]]: return out -def read_gguf_context_length(path: str) -> Optional[int]: - """Return the GGUF's native training context length (``{arch}.context_length``), - or ``None`` if missing/unreadable/not a GGUF. Cached by (path, mtime, size). - Lets the UI populate the context slider before the model is loaded.""" +def read_gguf_staged_dims(path: str) -> Optional[Dict[str, Optional[int]]]: + """GGUF header dims for the staged-load UI in one cached pass: + ``{"context_length", "layer_count", "moe_layer_count"}``. Each may be None + when absent (moe_layer_count is 0 for a dense model). Returns ``None`` if not + a GGUF / unreadable. Cached by (path, mtime, size). Lets the staged sheet size + the context, GPU-layers and MoE sliders before the model loads.""" key = _cache_key(path) if key is None: return None with _CACHE_LOCK: - if key in _CONTEXT_CACHE: - return _CONTEXT_CACHE[key] - result = _parse_gguf_context_length(path) + if key in _DIMS_CACHE: + return _DIMS_CACHE[key] + result = _parse_gguf_staged_dims(path) with _CACHE_LOCK: - while len(_CONTEXT_CACHE) >= _CACHE_MAX_ENTRIES: + while len(_DIMS_CACHE) >= _CACHE_MAX_ENTRIES: try: - _CONTEXT_CACHE.pop(next(iter(_CONTEXT_CACHE))) + _DIMS_CACHE.pop(next(iter(_DIMS_CACHE))) except StopIteration: break - _CONTEXT_CACHE[key] = result + _DIMS_CACHE[key] = result return result -def _parse_gguf_context_length(path: str) -> Optional[int]: - # The context key is architecture-namespaced (``llama.context_length`` etc.), - # so we learn the key only after reading ``general.architecture``. GGUF writes - # general.* before arch.* keys, matching the loader's own parser. - ctx_key: Optional[str] = None +def read_gguf_context_length(path: str) -> Optional[int]: + """Native training context length (``{arch}.context_length``), or ``None``. + Thin accessor over read_gguf_staged_dims.""" + dims = read_gguf_staged_dims(path) + return dims["context_length"] if dims else None + + +def _parse_gguf_arch_uints(path: str, wanted_suffixes: frozenset[str]) -> Optional[Dict[str, int]]: + """Walk a GGUF header once and return the requested architecture-namespaced + uint (vtype 4/10) keys, e.g. ``{"block_count": 32}``. Keys are + ``{arch}.``; the arch is learned from ``general.architecture`` (GGUF + writes general.* before arch.* keys, matching the loader's own parser). + Returns ``None`` if not a GGUF / unreadable, else a dict (possibly empty or + partial when some keys are absent).""" + arch: Optional[str] = None + found: Dict[str, int] = {} try: with open(path, "rb") as f: head = f.read(24) @@ -204,28 +219,68 @@ def _parse_gguf_context_length(path: str) -> Optional[int]: sbytes = f.read(slen) if len(sbytes) < slen: break - ctx_key = f"{sbytes.decode('utf-8', 'replace')}.context_length" - elif ctx_key is not None and key == ctx_key and vtype in (4, 10): + arch = sbytes.decode("utf-8", "replace") + elif ( + arch is not None + and vtype in (4, 10) + and key.startswith(f"{arch}.") + and key[len(arch) + 1 :] in wanted_suffixes + ): width = 4 if vtype == 4 else 8 n_bytes = f.read(width) if len(n_bytes) < width: break - value = struct.unpack(" 0 else None + found[key[len(arch) + 1 :]] = struct.unpack( + " Optional[Dict[str, Optional[int]]]: + vals = _parse_gguf_arch_uints( + path, + frozenset( + { + "context_length", + "block_count", + "expert_count", + "leading_dense_block_count", + } + ), + ) + if vals is None: + return None + ctx = vals.get("context_length") + block = vals.get("block_count") + # A real context/layer count is positive; treat 0/garbage as absent so the + # UI never builds a slider with max < min. + context_length = ctx if ctx and ctx > 0 else None + layer_count = block if block and block > 0 else None + # MoE layer count = block_count - leading dense layers, only when experts + # exist; else 0 (dense -> slider hidden). Mirrors n_moe_layers in + # core/inference/llama_cpp.py. + if not vals.get("expert_count") or not block: + moe_layer_count: Optional[int] = 0 + else: + moe_layer_count = max(0, block - (vals.get("leading_dense_block_count") or 0)) + return { + "context_length": context_length, + "layer_count": layer_count, + "moe_layer_count": moe_layer_count, + } # Strings (8) and arrays (9) are handled inline. diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py index 284bbb5745..dadf103cea 100644 --- a/studio/backend/utils/models/model_config.py +++ b/studio/backend/utils/models/model_config.py @@ -2083,7 +2083,7 @@ def _has_model_weight_files(model_dir: Path) -> bool: def _detect_training_output_type(model_dir: Path) -> Optional[str]: - """Classify a Studio training output as LoRA or full finetune.""" + """Classify an Unsloth training output as LoRA or full finetune.""" adapter_config = model_dir / "adapter_config.json" adapter_model = model_dir / "adapter_model.safetensors" if adapter_config.exists() or adapter_model.exists(): @@ -2105,7 +2105,7 @@ def _looks_like_lora_adapter(model_dir: Path) -> bool: def scan_trained_models(outputs_dir: str = str(outputs_root())) -> List[Tuple[str, str, str]]: - """Scan outputs folder for trained Studio models. + """Scan outputs folder for trained Unsloth models. Returns: List of (display_name, model_path, model_type), where model_type is diff --git a/studio/backend/utils/openai_auto_switch_settings.py b/studio/backend/utils/openai_auto_switch_settings.py index 1689395f40..462435e5d5 100644 --- a/studio/backend/utils/openai_auto_switch_settings.py +++ b/studio/backend/utils/openai_auto_switch_settings.py @@ -8,7 +8,9 @@ Two settings, both off by default so existing API behavior is unchanged: names a downloaded local GGUF different from the loaded one transparently loads it before serving (llama-swap-style). Unknown names pass through. - ``openai_api_auto_unload_idle_seconds``: when > 0, the loaded GGUF is - unloaded after this many idle seconds to free VRAM. + unloaded after this many idle seconds to free VRAM. Enabled values have a + 60s floor (0 stays "off"): a tiny TTL tears the model down between turns of + an active chat, forcing a full weight reload + prompt re-prefill per turn. The idle TTL can also be set at startup via the ``UNSLOTH_MODEL_IDLE_TTL`` env var. Unlike the stored setting (which stays gated on auto-switch), the env value @@ -33,6 +35,7 @@ MODEL_IDLE_TTL_ENV_VAR = "UNSLOTH_MODEL_IDLE_TTL" DEFAULT_OPENAI_AUTO_SWITCH_ENABLED = False DEFAULT_AUTO_UNLOAD_IDLE_SECONDS = 0 +MIN_AUTO_UNLOAD_IDLE_SECONDS = 60 _CACHE_TTL_S = 2.0 _cache_lock = threading.Lock() @@ -58,6 +61,10 @@ def _coerce_int(value: Any) -> int | None: return None +def _apply_idle_floor(seconds: int) -> int: + return 0 if seconds <= 0 else max(MIN_AUTO_UNLOAD_IDLE_SECONDS, seconds) + + def _cached_setting(key: str, default: Any) -> Any: """Read an app setting, memoized for _CACHE_TTL_S to spare the hot path.""" now = time.monotonic() @@ -91,12 +98,34 @@ def _stored_idle_seconds() -> Optional[int]: return _coerce_int(_cached_setting(AUTO_UNLOAD_IDLE_SETTING_KEY, None)) +_env_floor_warned = False + + def _env_idle_seconds() -> Optional[int]: - """UNSLOTH_MODEL_IDLE_TTL as a non-negative seconds value, or None if unset/invalid.""" + """UNSLOTH_MODEL_IDLE_TTL as a non-negative seconds value, or None if unset/invalid. + + Floored to MIN_AUTO_UNLOAD_IDLE_SECONDS here (with a one-time warning) since + headless/container deploys have no UI to surface a validation error.""" raw = os.environ.get(MODEL_IDLE_TTL_ENV_VAR) if raw is None or not raw.strip(): return None - return _coerce_int(raw) + parsed = _coerce_int(raw) + if parsed is None: + return None + floored = _apply_idle_floor(parsed) + if floored != parsed: + global _env_floor_warned + if not _env_floor_warned: + _env_floor_warned = True + from loggers import get_logger + get_logger(__name__).warning( + "%s=%s is below the %ss minimum; using %ss", + MODEL_IDLE_TTL_ENV_VAR, + parsed, + MIN_AUTO_UNLOAD_IDLE_SECONDS, + floored, + ) + return floored def get_stored_auto_unload_idle_seconds() -> int: @@ -108,7 +137,9 @@ def get_stored_auto_unload_idle_seconds() -> int: """ stored = _stored_idle_seconds() if stored is not None: - return stored + # Floor legacy values persisted before the minimum existed, so the UI + # displays the effective TTL and round-trips it cleanly. + return _apply_idle_floor(stored) env = _env_idle_seconds() return env if env is not None else DEFAULT_AUTO_UNLOAD_IDLE_SECONDS @@ -118,8 +149,9 @@ def get_auto_unload_idle_seconds() -> int: stored = _stored_idle_seconds() if stored is not None: # An explicit UI/API value stays gated on auto-switch: off reports 0 so the - # off state is identical to pre-feature. - return stored if get_openai_auto_switch_enabled() else 0 + # off state is identical to pre-feature. Floored to cover values persisted + # before the minimum existed. + return _apply_idle_floor(stored) if get_openai_auto_switch_enabled() else 0 # No stored value: UNSLOTH_MODEL_IDLE_TTL is a standalone startup default that # enables idle-unload even with auto-switch off (headless/container deploys). env = _env_idle_seconds() @@ -136,6 +168,11 @@ def set_openai_auto_switch(enabled: Any, idle_seconds: Any) -> tuple[bool, int]: parsed_idle = _coerce_int(idle_seconds) if parsed_idle is None: raise ValueError("Auto-unload idle seconds must be a non-negative integer.") + if 0 < parsed_idle < MIN_AUTO_UNLOAD_IDLE_SECONDS: + raise ValueError( + f"Auto-unload idle seconds must be 0 (off) or at least " + f"{MIN_AUTO_UNLOAD_IDLE_SECONDS}." + ) from storage.studio_db import upsert_app_settings upsert_app_settings( diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py index 759681da3f..1faa2b1281 100644 --- a/studio/backend/utils/paths/storage_roots.py +++ b/studio/backend/utils/paths/storage_roots.py @@ -36,7 +36,7 @@ def _infer_studio_home_from_venv() -> Path | None: def studio_root() -> Path: - """Studio install root. + """Unsloth install root. Priority: UNSLOTH_STUDIO_HOME, then STUDIO_HOME alias, then sys.prefix inference, then legacy ~/.unsloth/studio. UNSLOTH_STUDIO_HOME wins if @@ -62,7 +62,7 @@ def cache_root() -> Path: def studio_bin_root() -> Path: - """Dir for Studio-managed executables (the `unsloth` shim, downloaded tools like cloudflared).""" + """Dir for Unsloth-managed executables (the `unsloth` shim, downloaded tools like cloudflared).""" return studio_root() / "bin" @@ -443,7 +443,7 @@ def resolve_export_write_dir(path_value: str | None = None) -> Path: Unlike :func:`resolve_export_dir`, this function passes absolute paths through as-is so users can target a different drive when - their Studio install lives on a constrained system volume + their Unsloth install lives on a constrained system volume (see :gh-issue:`6082`). Used only by the export write path. """ if not path_value or not str(path_value).strip(): diff --git a/studio/backend/utils/preview_rate_limit.py b/studio/backend/utils/preview_rate_limit.py index dd38cfd5e7..c59a1bf5b3 100644 --- a/studio/backend/utils/preview_rate_limit.py +++ b/studio/backend/utils/preview_rate_limit.py @@ -5,7 +5,7 @@ A signed link stops ref guessing, but anyone with a link can still drive GPU generation. This bounds sustained abuse from a single source. In-process and -single-worker only (like the login limiter in ``routes/auth.py``); Studio runs as +single-worker only (like the login limiter in ``routes/auth.py``); Unsloth runs as one uvicorn process, so a shared store isn't needed. """ diff --git a/studio/backend/utils/process_lifetime.py b/studio/backend/utils/process_lifetime.py index 3ffd54cc26..c63227ae86 100644 --- a/studio/backend/utils/process_lifetime.py +++ b/studio/backend/utils/process_lifetime.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 -"""Bind Studio child processes to the parent's lifetime so none survive an +"""Bind Unsloth child processes to the parent's lifetime so none survive an abnormal parent exit (terminal-window close, Task Manager "End Task", SIGKILL, crash) -- the cooperative shutdown path only runs on graceful exits. @@ -139,7 +139,7 @@ def _install_windows_job() -> None: kernel32.CloseHandle(job) return # AssignProcessToJobObject(parent) makes children inherit the job. May - # fail if Studio already runs inside an incompatible host job (pre-Win8); + # fail if Unsloth already runs inside an incompatible host job (pre-Win8); # degrade to the cooperative path rather than blocking startup. if not kernel32.AssignProcessToJobObject(job, kernel32.GetCurrentProcess()): kernel32.CloseHandle(job) diff --git a/studio/backend/utils/studio_version.py b/studio/backend/utils/studio_version.py index 9c18070fbb..82ade74bba 100644 --- a/studio/backend/utils/studio_version.py +++ b/studio/backend/utils/studio_version.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 -"""Network-free Studio release version resolution for display-only UI.""" +"""Network-free Unsloth release version resolution for display-only UI.""" from __future__ import annotations @@ -20,7 +20,7 @@ _MAX_VERSION_LENGTH = 64 def is_valid_studio_release_version(value: object) -> bool: - """Return True for Studio release tags such as ``v0.1.39-beta``.""" + """Return True for Unsloth release tags such as ``v0.1.39-beta``.""" if not isinstance(value, str): return False version = value.strip() @@ -102,7 +102,7 @@ def _git_branch(repo_root: Path) -> str | None: def get_studio_version(repo_root: Path | None = None) -> str: - """Return the installed Studio release tag for display, or ``dev``. + """Return the installed Unsloth release tag for display, or ``dev``. Intentionally separate from the PyPI ``unsloth`` package version used by update checks. Never performs network requests. diff --git a/studio/backend/utils/training_runs.py b/studio/backend/utils/training_runs.py index dc2535e570..dcdfa1395d 100644 --- a/studio/backend/utils/training_runs.py +++ b/studio/backend/utils/training_runs.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 -"""Helpers for naming and describing Studio training runs.""" +"""Helpers for naming and describing Unsloth training runs.""" from __future__ import annotations diff --git a/studio/backend/utils/transformers_latest.py b/studio/backend/utils/transformers_latest.py index 40c8f729a5..9f1d11be5b 100644 --- a/studio/backend/utils/transformers_latest.py +++ b/studio/backend/utils/transformers_latest.py @@ -5,7 +5,7 @@ When a model's ``model_type`` is absent from every installed transformers overlay (base 4.57.x plus the .venv_t5_530/550/510 sidecars and, if provisioned, .venv_t5_latest), -Studio cannot load it today. This module answers, without authentication, code execution, +Unsloth cannot load it today. This module answers, without authentication, code execution, or trust_remote_code: 1. Does the LATEST transformers release on PyPI ship this ``model_type``? @@ -387,7 +387,7 @@ def check_upgrade_for_model(model_name: str, hf_token: str | None = None) -> dic _SHADOWABLE_DEPS = frozenset({"tokenizers", "safetensors"}) # Provided by the sidecar recipe; checked against its pin, not the base env. _SIDECAR_PROVIDED = {"huggingface-hub": "1.8.0", "hf-xet": "1.4.2"} -# CLI-only; never imported at runtime in Studio's workers. +# CLI-only; never imported at runtime in Unsloth's workers. _IGNORED_DEPS = frozenset({"typer"}) @@ -538,7 +538,7 @@ def _install_latest_transformers_locked(version: str, before_swap = None) -> dic return { "success": False, "version": version, - "message": "Cannot install: Studio is in offline mode.", + "message": "Cannot install: Unsloth is in offline mode.", } # Re-verify against a LIVE snapshot (a release may land inside the cache TTL); # fall back to the cached one on fetch failure. @@ -573,13 +573,13 @@ def _install_latest_transformers_locked(version: str, before_swap = None) -> dic "version": version, "message": "Cannot install transformers " f"{version}: this environment does not satisfy {', '.join(blockers)}. " - "A Studio update is required first.", + "An Unsloth update is required first.", } if not ensure_latest_transformers_venv(version, extra_packages, before_swap = before_swap): return { "success": False, "version": version, - "message": f"Installing transformers {version} failed; see the Studio logs.", + "message": f"Installing transformers {version} failed; see the Unsloth logs.", } _invalidate_capability_caches() return { diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py index 9f9f8aa3de..1fbcc9f46f 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -2151,7 +2151,7 @@ def end_sidecar_swap() -> None: def sidecar_swap_in_progress() -> bool: """True while a .venv_t5_latest install or repair holds the reservation, - in this process or any other Studio process (lock file).""" + in this process or any other Unsloth process (lock file).""" return sidecar_swap_kind() is not None diff --git a/studio/backend/utils/upload_limits.py b/studio/backend/utils/upload_limits.py index c21ea69af7..fff0ac423f 100644 --- a/studio/backend/utils/upload_limits.py +++ b/studio/backend/utils/upload_limits.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 -"""Shared Studio upload/request size limits.""" +"""Shared Unsloth upload/request size limits.""" from __future__ import annotations diff --git a/studio/frontend/.npmrc b/studio/frontend/.npmrc index 19783b5ff4..414379da6e 100644 --- a/studio/frontend/.npmrc +++ b/studio/frontend/.npmrc @@ -1,4 +1,4 @@ -# Studio frontend npm configuration. +# Unsloth frontend npm configuration. # # Mini Shai-Hulud / Axios-style supply chain defense. # Requires npm >=11.10.0. Refuses tarballs published less than 7 days ago, diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx index c35706e50a..a7e9469cfc 100644 --- a/studio/frontend/src/app/provider.tsx +++ b/studio/frontend/src/app/provider.tsx @@ -412,7 +412,7 @@ function TauriWrapper({ children }: { children: ReactNode }) { {desktopBooting ? (

-
Preparing Studio
+
Preparing Unsloth
The local backend is ready. Signing in to your desktop session before loading chats. 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 8e06181585..ff4bd4bebd 100644 --- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx +++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx @@ -1679,7 +1679,7 @@ export function HubModelPicker({ const deviceType = usePlatformStore((s) => s.deviceType); const isMac = deviceType === "mac"; - // Drop models Studio can't run for chat (diffusion / image / video / etc.) + // Drop models Unsloth can't run for chat (diffusion / image / video / etc.) // using the Hub's classifier on the tags the listing already carries. const isChatSupported = useCallback( (r: HfModelResult) => @@ -1730,7 +1730,7 @@ export function HubModelPicker({ let rows = recommendedSearch.results .filter((r) => !isHiddenModelId(r.id)) .filter((r) => !isMobileVariant(r.id)); - // Drop models Studio can't run for chat (diffusion / image / video / etc.). + // Drop models Unsloth can't run for chat (diffusion / image / video / etc.). rows = rows.filter(isChatSupported); // With no explicit format, show the device-recommended formats (GGUF, plus // MLX on Mac). When the user picks a format, honor it instead so Safetensors @@ -1868,7 +1868,7 @@ export function HubModelPicker({ // eslint-disable-next-line react-hooks/exhaustive-deps [lmStudioModels, downloadedSort, formatFilter, loadTimes, localQuery], ); - // Local ./models entries. Chat-only Studio runs GGUF (any host) and MLX (Mac + // Local ./models entries. Chat-only Unsloth runs GGUF (any host) and MLX (Mac // only), so raw checkpoints there are hidden (mirrors the cached non-GGUF // rule). An MLX build a Mac user dropped in ./models stays selectable. const sortedLocalDir = useMemo( diff --git a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts index ec75b17f20..08492ab480 100644 --- a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts +++ b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts @@ -2,7 +2,9 @@ // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 // Per-model pre-load inference settings, persisted in localStorage so the load -// dialog can offer "Remember settings for ". +// dialog can offer "Remember settings for ". GGUF picks only: every +// field is a llama.cpp load knob, so all save/restore call sites gate on +// GGUF-ness (a non-GGUF blob would only snapshot leftover standing values). const KEY = "unsloth_load_settings"; @@ -12,14 +14,22 @@ export interface RememberedLoadSettings { speculativeType: string | null; specDraftNMax: number | null; tensorParallel: boolean; + // GPU Memory controls. Optional so an older blob (which lacked them) still + // parses, leaving the live knobs untouched on apply. The mode is kept with the + // manual knobs (gpuLayers/nCpuMoe are ignored outside Manual mode). A null + // selectedGpuIds is meaningful (all GPUs), so it's distinguished from absent. + // The per-GPU split ratio is deliberately NOT remembered: it's positionally + // bound to the exact GPU set/order and unvalidated, so it would mismatch. + gpuMemoryMode?: "auto" | "manual"; + gpuLayers?: number; + nCpuMoe?: number; + selectedGpuIds?: number[] | null; } -// Storage key for a pick's remembered settings. The remembered knobs are -// VRAM-budget driven (context override, KV-cache dtype, tensor-parallel), so the -// right values differ per quant. An HF repo collapses all its GGUF variants into -// one `id`, so fold the variant in to scope settings per quant. Local .gguf -// paths key by their file path (already file-specific); native drag-drop files -// key by display label, so same-named files in different folders share an entry. +// Storage key for a pick's remembered settings, scoped per quant (the VRAM-budget +// knobs differ per quant). An HF repo collapses its GGUF variants into one `id`, +// so fold the variant in. Local .gguf paths are already file-specific; native +// drag-drop files key by display label, so same-named files share an entry. export function rememberedLoadSettingsKey(selection: { id: string; ggufVariant?: string | null; diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx index 62b8af6e3a..235dcb3c3d 100644 --- a/studio/frontend/src/components/assistant-ui/thread.tsx +++ b/studio/frontend/src/components/assistant-ui/thread.tsx @@ -1970,7 +1970,7 @@ function isNativeComposing(event: Event) { } // Fallback timeout for stuck IME composition. With Chrome on Windows against -// a WSL-hosted Studio (issue #5546), `compositionend` never fires after the +// a WSL-hosted Unsloth (issue #5546), `compositionend` never fires after the // candidate commits, so `composingRef` stays true and Send stays disabled. // Every compositionupdate / non-composing input resets the timer; only a true // gap-after-commit lets it fire. 2500ms is above a normal candidate-window diff --git a/studio/frontend/src/components/ui/confetti.tsx b/studio/frontend/src/components/ui/confetti.tsx index 892bffdb18..35f5913240 100644 --- a/studio/frontend/src/components/ui/confetti.tsx +++ b/studio/frontend/src/components/ui/confetti.tsx @@ -34,7 +34,7 @@ export type ConfettiRef = Api | null; const ConfettiContext = createContext({} as Api); -// Studio CSP blocks canvas-confetti's default blob: worker, so force +// Unsloth CSP blocks canvas-confetti's default blob: worker, so force // useWorker: false. Module-scoped so the prop default keeps stable // identity across renders (`canvasRef` depends on `globalOptions`). const DEFAULT_GLOBAL_OPTIONS: ConfettiGlobalOptions = { diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx index 119471da10..73db10d41b 100644 --- a/studio/frontend/src/features/auth/components/auth-form.tsx +++ b/studio/frontend/src/features/auth/components/auth-form.tsx @@ -298,7 +298,7 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null { // reset-password"), which the installer puts on PATH on every platform. // Do NOT rewrite it to a relative Windows path like // ".\unsloth_studio\Scripts\unsloth.exe ..." -- that only resolves inside - // the Studio home dir and fails with CommandNotFoundException elsewhere. + // the Unsloth home dir and fails with CommandNotFoundException elsewhere. // Show the backend message as-is. const msg = err instanceof Error ? err.message : "Auth failed."; setError(msg); diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts index 0bf46e7343..7083f02288 100644 --- a/studio/frontend/src/features/chat/api/chat-adapter.ts +++ b/studio/frontend/src/features/chat/api/chat-adapter.ts @@ -45,12 +45,18 @@ import { import { type PendingImageEditReference, type RagAutoInject, + GPU_LAYERS_AUTO, + loadedGpuMemoryFieldsUnlessStaged, + reconcilePersistedGpuIds, resolveLoadedSpeculativeSettings, resolveSpeculativeSettingsForLoad, + persistGpuMemoryModeOnLoad, resolveToolsEnabledOnLoad, saveSpeculativeType, useChatRuntimeStore, } from "../stores/chat-runtime-store"; +import { resolveFitMaxSeqLength, resolveManualAutoCtxPin } from "../presets/preset-policy"; +import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info"; import { useExternalProvidersStore } from "../stores/external-providers-store"; import { shouldPreserveFullOutput, @@ -1489,6 +1495,13 @@ async function autoLoadSmallestModel(): Promise<{ max_seq_length: number; is_lora: boolean; gguf_variant?: string | null; + // GGUF-only: scopes the training guard to the same placement policy /load + // will use. Manual mode must match because it makes placement user-owned. + // The layer/MoE/split/KV/spec knobs are deliberately not sent: Auto mode's + // guard sizes conservatively, while Manual mode bypasses that estimate. + // The safetensors fallback omits both fields and uses HF auto-placement. + gpu_ids?: number[]; + gpu_memory_mode?: "auto" | "manual"; }): Promise { const validation = await validateModel({ ...payload, @@ -1520,12 +1533,18 @@ async function autoLoadSmallestModel(): Promise<{ return false; } const currentStore = useChatRuntimeStore.getState(); - const remembered = loadRememberedLoadSettings( - rememberedLoadSettingsKey({ - id: candidate.id, - ggufVariant: candidate.ggufVariant, - }), - ); + // Blobs are saved for GGUF picks only (the sheet gates on it), so don't + // let a legacy non-GGUF blob feed a stale context/spec choice into a + // safetensors auto-load. + const remembered = + candidate.kind === "gguf" + ? loadRememberedLoadSettings( + rememberedLoadSettingsKey({ + id: candidate.id, + ggufVariant: candidate.ggufVariant, + }), + ) + : null; const effectiveMaxSeqLength = resolveLoadMaxSeqLength({ modelId: candidate.id, ggufVariant: candidate.ggufVariant, @@ -1537,6 +1556,38 @@ async function autoLoadSmallestModel(): Promise<{ maxSeqLength: candidate.maxSeqLength, presetSource: currentStore.activePresetSource, }); + // The GPU knobs are per-model, so read them from the same remembered + // settings that fed effectiveMaxSeqLength -- on a background auto-load the + // live store holds session defaults, not the saved Manual mode / layer pin / + // GPU pick. Absent fields fall back like applyRememberedLoadSettings: the + // mode to the store (a persisted standing preference), the per-model knobs to + // their defaults. The saved GPU pick is reconciled against the GPUs present + // now, like the interactive restore. + const effectiveGpuMemoryMode = + remembered?.gpuMemoryMode ?? currentStore.gpuMemoryMode; + const effectiveGpuLayers = remembered?.gpuLayers ?? GPU_LAYERS_AUTO; + const effectiveNCpuMoe = remembered?.nCpuMoe ?? 0; + if (remembered?.selectedGpuIds != null) { + // Warm the device cache first: on a cold cache the reconcile passes the + // saved pick through unvalidated, and a stale cross-host pick then fails + // the load with the picker hidden. + await ensureGpuDeviceCache(); + } + const effectiveGpuIds = + remembered?.selectedGpuIds !== undefined + ? reconcilePersistedGpuIds(remembered.selectedGpuIds) + : null; + // Under Manual GPU memory + Auto layers, llama.cpp's --fit owns context + // sizing, so send 0 (or the pinned length). GGUF-only; a no-op otherwise. + // The context pin is per-model too, so it comes from remembered settings, + // not the live store. + const fitMaxSeqLength = resolveFitMaxSeqLength( + candidate.kind === "gguf", + effectiveGpuMemoryMode, + effectiveGpuLayers, + remembered?.contextLength ?? null, + effectiveMaxSeqLength, + ); const effectiveSpeculativeType = remembered?.speculativeType ?? specSettings.speculativeType; const effectiveSpecDraftNMax = @@ -1544,9 +1595,16 @@ async function autoLoadSmallestModel(): Promise<{ if ( !(await canAutoLoad({ model_path: candidate.id, - max_seq_length: effectiveMaxSeqLength, + max_seq_length: fitMaxSeqLength, is_lora: false, gguf_variant: candidate.ggufVariant, + // The same remembered-derived GPU pick the load below sends. + ...(candidate.kind === "gguf" + ? { + gpu_ids: effectiveGpuIds ?? undefined, + gpu_memory_mode: effectiveGpuMemoryMode, + } + : {}), })) ) { skippedAutoLoadCandidates.add( @@ -1558,7 +1616,7 @@ async function autoLoadSmallestModel(): Promise<{ const loadResp = await loadModel({ model_path: candidate.id, hf_token: hfToken, - max_seq_length: effectiveMaxSeqLength, + max_seq_length: fitMaxSeqLength, load_in_4bit: true, is_lora: false, gguf_variant: candidate.ggufVariant, @@ -1567,8 +1625,22 @@ async function autoLoadSmallestModel(): Promise<{ speculative_type: effectiveSpeculativeType, spec_draft_n_max: effectiveSpecDraftNMax, tensor_parallel: remembered?.tensorParallel ?? false, + // GGUF-only: the safetensors fallback loads via HF auto-placement (no + // explicit pins). The split ratio is deliberately never remembered + // (positionally bound to an exact GPU set), so auto-load leaves llama.cpp's + // free-VRAM default in charge rather than sending a stale store value. + ...(candidate.kind === "gguf" + ? { + gpu_memory_mode: effectiveGpuMemoryMode, + gpu_layers: effectiveGpuLayers, + n_cpu_moe: effectiveNCpuMoe, + gpu_ids: effectiveGpuIds ?? undefined, + } + : {}), }); saveSpeculativeType(effectiveSpeculativeType); + // Self-gates on is_gguf (skips diffusion), so persists only for a real GGUF load. + persistGpuMemoryModeOnLoad(loadResp, effectiveGpuMemoryMode); useChatRuntimeStore .getState() .setCheckpoint(candidate.id, candidate.ggufVariant ?? undefined); @@ -1597,6 +1669,15 @@ async function autoLoadSmallestModel(): Promise<{ store.setModels([...store.models, autoModel]); } if (candidate.kind === "gguf") { + // Keep an explicit Manual+Auto context pin the load just applied (so a + // later Apply doesn't silently revert it to auto-fit sizing), mirroring + // the interactive path's keepCustomCtx; other cases baseline on + // ggufContextLength. + const keepCustomCtx = resolveManualAutoCtxPin( + effectiveGpuMemoryMode, + effectiveGpuLayers, + remembered?.contextLength ?? null, + ); useChatRuntimeStore.setState({ ggufContextLength: loadResp.context_length ?? 131072, ggufMaxContextLength: @@ -1613,6 +1694,10 @@ async function autoLoadSmallestModel(): Promise<{ loadedKvCacheDtype: loadResp.cache_type_kv ?? null, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, + ...loadedGpuMemoryFieldsUnlessStaged(loadResp, { + customContextLength: keepCustomCtx, + }), + loadedCustomContextLength: keepCustomCtx, defaultChatTemplate: loadResp.chat_template ?? null, chatTemplateOverride: null, loadedChatTemplateOverride: null, @@ -1633,6 +1718,9 @@ async function autoLoadSmallestModel(): Promise<{ loadedKvCacheDtype: loadResp.cache_type_kv ?? null, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, + // Non-GGUF response: clears any stale GPU baseline a prior manual-GPU + // GGUF load left, matching the interactive/status sibling load paths. + ...loadedGpuMemoryFieldsUnlessStaged(loadResp), defaultChatTemplate: loadResp.chat_template ?? null, chatTemplateOverride: null, loadedChatTemplateOverride: null, @@ -1820,12 +1908,17 @@ async function autoLoadSmallestModel(): Promise<{ duration: 30000, }); try { + const rt = useChatRuntimeStore.getState(); if ( !(await canAutoLoad({ model_path: "unsloth/Qwen3.5-4B-MTP-GGUF", max_seq_length: 0, is_lora: false, gguf_variant: "UD-Q4_K_XL", + // The same live-store GPU pick the load below sends (a fresh default + // model has no remembered settings to prefer). + gpu_ids: rt.selectedGpuIds ?? undefined, + gpu_memory_mode: rt.gpuMemoryMode, })) ) { toast.dismiss(toastId); @@ -1835,6 +1928,9 @@ async function autoLoadSmallestModel(): Promise<{ const loadResp = await loadModel({ model_path: "unsloth/Qwen3.5-4B-MTP-GGUF", hf_token: hfToken, + // Model default under both modes: Auto layers + no pin means + // resolveFitMaxSeqLength returns 0 for every mode (the canAutoLoad + // preflight above sends the same). max_seq_length: 0, load_in_4bit: true, is_lora: false, @@ -1842,8 +1938,20 @@ async function autoLoadSmallestModel(): Promise<{ trust_remote_code: trustRemoteCode, speculative_type: specSettings.speculativeType, spec_draft_n_max: specSettings.specDraftNMax, + // GPU Memory mode is a standing preference, so honor it on auto-load. + // The layer/MoE/split knobs and the context pin are per-model: the live + // store may hold edits drafted for a staged pick, and a fresh default + // model has no remembered settings, so those stay at their defaults like + // the cached-candidate path. The GPU pick deliberately differs (it's the + // picker's current on-screen selection, which the canAutoLoad preflight + // above already committed to). + gpu_memory_mode: rt.gpuMemoryMode, + gpu_layers: GPU_LAYERS_AUTO, + n_cpu_moe: 0, + gpu_ids: rt.selectedGpuIds ?? undefined, }); saveSpeculativeType(specSettings.speculativeType); + persistGpuMemoryModeOnLoad(loadResp, rt.gpuMemoryMode); useChatRuntimeStore .getState() .setCheckpoint("unsloth/Qwen3.5-4B-MTP-GGUF", "UD-Q4_K_XL"); @@ -1880,6 +1988,10 @@ async function autoLoadSmallestModel(): Promise<{ loadedKvCacheDtype: loadResp.cache_type_kv ?? null, tensorParallel: loadResp.tensor_parallel ?? false, loadedTensorParallel: loadResp.tensor_parallel ?? false, + ...loadedGpuMemoryFieldsUnlessStaged(loadResp), + // Drives the GPU Memory controls' diffusion gate; set alongside the + // GPU fields on every load path so the gate can't read stale. + loadedIsDiffusion: loadResp.is_diffusion ?? false, defaultChatTemplate: loadResp.chat_template ?? null, chatTemplateOverride: null, loadedIsMultimodal: isMultimodalResponse(loadResp), diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts index ebf9461172..0f6af38033 100644 --- a/studio/frontend/src/features/chat/api/chat-api.ts +++ b/studio/frontend/src/features/chat/api/chat-api.ts @@ -127,28 +127,38 @@ export async function validateModel( native_path_lease: payload.nativePathLease ?? null, hf_token: payload.hf_token, gguf_variant: payload.gguf_variant ?? null, - // Send the intended load settings so validate's VRAM check matches the - // follow-up /load and doesn't unload for a load /load would then reject. + // Intended load settings so validate's preflight matches the follow-up + // /load. Default placement is sized against the selected GPUs. max_seq_length: payload.max_seq_length, load_in_4bit: payload.load_in_4bit, + gpu_ids: payload.gpu_ids, + // Manual placement is an explicit override: Auto layers use llama.cpp + // --fit, while a pinned layer count is owned by the user. Tell validate + // so it applies the same training-guard policy as /load. + gpu_memory_mode: payload.gpu_memory_mode, }), }); return parseJsonOrThrow(response); } /** - * Read a GGUF's native context length from its local header (no GPU load, no - * download). Returns null when the file isn't downloaded yet, the model isn't a - * GGUF, or it's gated. For a native (drag-drop / picked) file, pass - * `nativePathToken` so the backend reads the granted local path. Used by the - * deferred-load staging flow to fill the context slider before the single load. + * Read a GGUF's header dims (native context length, total layer count, MoE + * expert-layer count) from its local file (no GPU load, no download). All are + * null when the file isn't downloaded yet, the model isn't a GGUF, or it's + * gated. For a native (drag-drop / picked) file, pass `nativePathToken` so the + * backend reads the granted local path. Used by the deferred-load staging flow + * to size the context, GPU-layers and MoE sliders before the single load. */ -export async function fetchGgufContextLength(payload: { +export async function fetchGgufStagedMetadata(payload: { model_path: string; gguf_variant?: string | null; hf_token?: string | null; nativePathToken?: string | null; -}): Promise { +}): Promise<{ + contextLength: number | null; + layerCount: number | null; + moeLayerCount: number | null; +}> { let nativePathLease: string | null = null; if (payload.nativePathToken) { try { @@ -156,8 +166,8 @@ export async function fetchGgufContextLength(payload: { await consumeNativePathToken(payload.nativePathToken, "validate-model") ).nativePathLease; } catch { - // Lease expired / revoked: degrade to no context (the load can re-mint). - return null; + // Lease expired / revoked: degrade to no metadata (the load can re-mint). + return { contextLength: null, layerCount: null, moeLayerCount: null }; } } const response = await authFetch("/api/inference/validate", { @@ -172,7 +182,11 @@ export async function fetchGgufContextLength(payload: { }), }); const res = await parseJsonOrThrow(response); - return res.context_length ?? null; + return { + contextLength: res.context_length ?? null, + layerCount: res.layer_count ?? null, + moeLayerCount: res.moe_layer_count ?? null, + }; } export async function unloadModel(payload: UnloadModelRequest): Promise { diff --git a/studio/frontend/src/features/chat/artifacts/artifact-card.tsx b/studio/frontend/src/features/chat/artifacts/artifact-card.tsx index ee8c26abf1..0345dc6e2a 100644 --- a/studio/frontend/src/features/chat/artifacts/artifact-card.tsx +++ b/studio/frontend/src/features/chat/artifacts/artifact-card.tsx @@ -8,7 +8,7 @@ import { cn } from "@/lib/utils"; import { useAuiState } from "@assistant-ui/react"; import { LayoutTwoColumnIcon as Layout2ColumnIcon } from "@hugeicons/core-free-icons"; import { HugeiconsIcon } from "@hugeicons/react"; -import { useLayoutEffect, useMemo } from "react"; +import { useLayoutEffect, useMemo, useRef } from "react"; import { useChatRuntimeStore } from "../stores/chat-runtime-store"; import type { ArtifactViewMode } from "./html-frame"; import { @@ -83,15 +83,18 @@ export function ArtifactCard({ ], ); const surface = artifactThreadId ? "panel" : "overlay"; + // Once per mount, so a view-change cleanup can't re-trigger a stale open. + const autoOpenAttemptedRef = useRef(false); useLayoutEffect(() => { if (selectedArtifactId === artifact.id) { updateArtifact(artifact); } - if (!autoOpen) { + if (!autoOpen || autoOpenAttemptedRef.current) { return; } + autoOpenAttemptedRef.current = true; if (hasAutoOpenedArtifact(artifact.id)) { return; } diff --git a/studio/frontend/src/features/chat/artifacts/html-frame.tsx b/studio/frontend/src/features/chat/artifacts/html-frame.tsx index b26f2f6685..36e3ed8a5b 100644 --- a/studio/frontend/src/features/chat/artifacts/html-frame.tsx +++ b/studio/frontend/src/features/chat/artifacts/html-frame.tsx @@ -28,7 +28,7 @@ export function buildArtifactSrcDoc(code: string): string { } // Preview iframes intentionally omit allow-downloads: generated canvases can -// offer their own UI, but downloads must go through Studio's explicit +// offer their own UI, but downloads must go through Unsloth's explicit // copy/download controls outside the no-same-origin sandbox. export function ArtifactHtmlFrame({ code, diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx index 380ce0e0ab..217eaf8b6d 100644 --- a/studio/frontend/src/features/chat/chat-page.tsx +++ b/studio/frontend/src/features/chat/chat-page.tsx @@ -247,13 +247,13 @@ const SingleContent = memo(function SingleContent({ useState(false); const [isArtifactSurfaceVisible, setIsArtifactSurfaceVisible] = useState(false); + // Without a URL threadId the artifact must belong to the active thread. const showArtifactPanel = Boolean( artifact && artifactSurface === "panel" && (threadId ? !artifact.threadId || artifact.threadId === threadId - : Boolean(newThreadNonce) || - Boolean(artifact.threadId && artifact.threadId === activeThreadId)), + : Boolean(artifact.threadId && artifact.threadId === activeThreadId)), ); const artifactLayoutActive = showArtifactPanel || isArtifactPanelLayoutActive; @@ -1445,9 +1445,11 @@ export function ChatPage({ // were already seeded on stage, so keepSpeculative only when a config was // saved -- otherwise the standing speculative preference should win. autoLoadStagedRef.current = (pending) => { - const remembered = loadRememberedLoadSettings( - rememberedLoadSettingsKey(pending), - ); + // Blobs are saved for GGUF picks only (the sheet gates on it), so don't + // let a legacy non-GGUF blob claim a seeded config here. + const remembered = hasGgufSource(pending) + ? loadRememberedLoadSettings(rememberedLoadSettingsKey(pending)) + : null; void selectModel({ ...pending, isDownloaded: true, @@ -1771,10 +1773,8 @@ export function ChatPage({ useEffect(() => { if (view.mode !== "single") return; - if (view.threadId || view.newThreadNonce || !selectedArtifact) return; - // view excludes __LOCALID_ threads (they fall through to mode:"single" - // with no threadId/nonce). Don't close a canvas whose thread is the - // active local thread. + if (view.threadId || !selectedArtifact) return; + // Close any canvas that doesn't belong to the active thread. if ( selectedArtifact.threadId && selectedArtifact.threadId === activeThreadId @@ -2815,6 +2815,11 @@ export function ChatPage({ selectModel({ id: state.params.checkpoint, ggufVariant: state.activeGgufVariant ?? undefined, + // A native (drag-drop / picked) GGUF's checkpoint is only a display + // label, so the reload needs its path token to re-mint a lease -- + // else applying the now-exposed GPU/context controls can't resolve + // the file. Null for non-native loads, which reload by id as before. + nativePathToken: state.activeNativePathToken ?? undefined, forceReload: true, isDownloaded: true, loadingDescription: "Reloading with updated chat template.", diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx index 95e5cfbd79..e39955e576 100644 --- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx +++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx @@ -244,8 +244,8 @@ export function ChatProvidersSettings({ (s) => s.setConnectionsEnabled, ); const isCustomProvider = isCustomProviderType(providerType); - // Local presets (Ollama, llama.cpp) never use API keys — hide the field. - // vLLM may optionally use a bearer token on secured deployments. + // llama.cpp hides the key field. Ollama and vLLM show an optional key: + // Ollama cloud and secured vLLM need one; local servers leave it empty. const showApiKeyField = !customPresetSkipsApiKeyField(providerType); const showReasoningToggle = supportsProviderReasoningToggle(providerType); diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx index cedd298ecf..b368a811fa 100644 --- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx +++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx @@ -55,6 +55,7 @@ import { Switch } from "@/components/ui/switch"; import { Textarea } from "@/components/ui/textarea"; import { InfoHint } from "@/components/ui/info-hint"; import { Tooltip, TooltipContent } from "@/components/ui/tooltip"; +import { useGpuDevices } from "@/hooks/use-gpu-info"; import { useIsMobile } from "@/hooks/use-mobile"; import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check"; import { cn } from "@/lib/utils"; @@ -99,8 +100,11 @@ import { providerSupportsFastMode, } from "./provider-capabilities"; import { + GPU_LAYERS_AUTO, + distributeByWeight, isPendingGguf, pendingSelectionMatches, + rebalanceSplit, useChatRuntimeStore, } from "./stores/chat-runtime-store"; import { RetrievalSettingsSection } from "@/features/rag/components/retrieval-settings-section"; @@ -250,6 +254,7 @@ function ParamSlider({ displayValue, info, valueSize, + disabled, }: { label: string; value: number; @@ -260,6 +265,7 @@ function ParamSlider({ displayValue?: string; info?: ReactNode; valueSize?: number; + disabled?: boolean; }) { return (
@@ -279,6 +285,7 @@ function ParamSlider({ displayValue={displayValue} ariaLabel={label} size={valueSize ?? 4} + disabled={disabled} />
onChange(snapToStep(v, step, min, max))} className="panel-slider" + disabled={disabled} />
); @@ -540,8 +548,17 @@ export function ChatSettingsPanel({ const base = slash >= 0 ? id.slice(slash + 1) : id; return base || id; })(); + const activeNativePathToken = useChatRuntimeStore( + (s) => s.activeNativePathToken, + ); + const loadedGgufContextLength = useChatRuntimeStore((s) => s.ggufContextLength); + // A GGUF loaded from a native path / direct .gguf has no HF variant, so key + // off the same signal the status hydration uses -- variant OR native token OR + // a GGUF context -- else the GPU Memory controls hide for a loaded local GGUF. const isLoadedGguf = - useChatRuntimeStore((s) => s.activeGgufVariant) != null; + useChatRuntimeStore((s) => s.activeGgufVariant) != null || + activeNativePathToken != null || + loadedGgufContextLength != null; // While a pick is staged the sheet configures *that* model, so its GGUF-ness // (not the currently loaded model's) decides whether the GGUF-only controls // show. Otherwise a staged non-GGUF Hub repo would inherit the loaded GGUF's @@ -607,6 +624,25 @@ export function ChatSettingsPanel({ const loadedTensorParallel = useChatRuntimeStore( (s) => s.loadedTensorParallel, ); + const gpuMemoryMode = useChatRuntimeStore((s) => s.gpuMemoryMode); + const setGpuMemoryMode = useChatRuntimeStore((s) => s.setGpuMemoryMode); + const loadedGpuMemoryMode = useChatRuntimeStore((s) => s.loadedGpuMemoryMode); + const loadedIsDiffusion = useChatRuntimeStore((s) => s.loadedIsDiffusion); + const gpuLayers = useChatRuntimeStore((s) => s.gpuLayers); + const setGpuLayers = useChatRuntimeStore((s) => s.setGpuLayers); + const loadedGpuLayers = useChatRuntimeStore((s) => s.loadedGpuLayers); + const nCpuMoe = useChatRuntimeStore((s) => s.nCpuMoe); + const setNCpuMoe = useChatRuntimeStore((s) => s.setNCpuMoe); + const loadedNCpuMoe = useChatRuntimeStore((s) => s.loadedNCpuMoe); + const splitRatio = useChatRuntimeStore((s) => s.splitRatio); + const setSplitRatio = useChatRuntimeStore((s) => s.setSplitRatio); + const loadedSplitRatio = useChatRuntimeStore((s) => s.loadedSplitRatio); + const ggufLayerCount = useChatRuntimeStore((s) => s.ggufLayerCount); + const moeLayerCount = useChatRuntimeStore((s) => s.moeLayerCount); + const selectedGpuIds = useChatRuntimeStore((s) => s.selectedGpuIds); + const setSelectedGpuIds = useChatRuntimeStore((s) => s.setSelectedGpuIds); + const loadedGpuIds = useChatRuntimeStore((s) => s.loadedGpuIds); + const gpuDevices = useGpuDevices(); const chatTemplateOverride = useChatRuntimeStore( (s) => s.chatTemplateOverride, ); @@ -614,6 +650,9 @@ export function ChatSettingsPanel({ (s) => s.loadedChatTemplateOverride, ); const customContextLength = useChatRuntimeStore((s) => s.customContextLength); + const loadedCustomContextLength = useChatRuntimeStore( + (s) => s.loadedCustomContextLength, + ); const setCustomContextLength = useChatRuntimeStore( (s) => s.setCustomContextLength, ); @@ -641,10 +680,14 @@ export function ChatSettingsPanel({ : null; useEffect(() => { if (!pendingKey) return; - const saved = loadRememberedLoadSettings(pendingKey); + // GGUF-only, like the stageOrLoad / Hub restore paths: every remembered + // field is a llama.cpp knob, so a non-GGUF pick has nothing to restore -- + // and applying its blob would clobber the standing gpuMemoryMode with a + // stale snapshot (the save on Load below is gated the same way). + const saved = pendingIsGguf ? loadRememberedLoadSettings(pendingKey) : null; setRemember(saved != null); if (saved) applyRememberedLoadSettings(saved); - }, [pendingKey, applyRememberedLoadSettings]); + }, [pendingKey, pendingIsGguf, applyRememberedLoadSettings]); // While staging, the sheet reflects the STAGED model, so its header context // takes precedence over the loaded model's (which may differ or be larger). const baseContext = pendingIsGguf ? stagedContextLength : ggufContextLength; @@ -661,15 +704,132 @@ export function ChatSettingsPanel({ const ctxDisplayValue = customContextLength ?? baseContext ?? ""; const ctxMaxValue = baseNativeContext ?? baseContext ?? null; const kvDirty = kvCacheDtype !== loadedKvCacheDtype; - const ctxDirty = customContextLength !== null; + const ctxDirty = customContextLength !== loadedCustomContextLength; const specDirty = speculativeType !== loadedSpeculativeType; const specDraftDirty = specDraftNMax !== loadedSpecDraftNMax; const tpDirty = tensorParallel !== (loadedTensorParallel ?? false); + // A loaded diffusion GGUF runs mode-agnostic (pins all layers on one GPU, + // ignores --fit/--gpu-layers), so the GPU Memory mode + manual controls don't + // apply -- hide them and don't let the preserved standing mode read as dirty. + // The GPU picker still applies (diffusion pins the chosen device). A staged pick + // keeps the controls (a pending pick's diffusion-ness isn't known until load). + const gpuModeApplies = + isGguf && (pendingSelection != null || !loadedIsDiffusion); + const gpuDirty = + gpuModeApplies && gpuMemoryMode !== (loadedGpuMemoryMode ?? "auto"); + const isManual = gpuModeApplies && gpuMemoryMode === "manual"; + // Manual with the GPU Layers slider at "Auto" (leftmost): --fit owns the whole + // layout, so the offload knobs (MoE, split, TP) don't apply. + const autoLayers = isManual && gpuLayers < 0; + // GPUs actually in use: the picked subset, or all visible when none picked. + const gpusInUse = selectedGpuIds ?? gpuDevices.map((d) => d.index); + // TP is off with fewer than 2 GPUs in use (single GPU, or the picker narrowed + // to one): tensor split is a no-op there and aborts on some archs. Mirrors the + // multi-GPU gate on the GPU picker / Split ratio. (Under Auto layers the whole + // TP control is hidden -- llama.cpp's --fit aborts under --split-mode tensor.) + const tpDisabled = gpusInUse.length <= 1; + // Manual gpu-layers ceiling = model layer count + 1 (else a safe fallback): + // llama.cpp counts the output layer as one more offloadable layer past the + // repeating blocks ("offloaded 33/33" needs -ngl 33 on a 32-block model), so + // the slider max must reach it or full offload is unreachable. While staging, + // use the staged model's layer count (read from its header). + const stagedLayerCount = pendingSelection?.layerCount ?? null; + const modelLayerCount = pendingIsGguf ? stagedLayerCount : ggufLayerCount; + const gpuLayersMax = modelLayerCount != null ? modelLayerCount + 1 : 256; + // MoE-offload slider: shown only for MoE models, capped at their MoE-layer + // count. While staging, use the staged model's count (read from its header); + // otherwise the loaded model's. + const stagedMoeLayerCount = pendingSelection?.moeLayerCount ?? null; + const moeLayersMax = pendingIsGguf + ? (stagedMoeLayerCount ?? 0) + : (moeLayerCount ?? 0); + const showMoeSlider = isManual && !autoLayers && moeLayersMax > 0; + // gpuLayers always counts; MoE only with an explicit layer count (see above). + const manualDirty = + isManual && + (gpuLayers !== loadedGpuLayers || + (!autoLayers && nCpuMoe !== (loadedNCpuMoe ?? 0))); + // GPU picker: only meaningful on multi-GPU, and only when the reported + // indices are physical (relative ordinals from a parent CUDA_VISIBLE_DEVICES + // mask can't be mapped back to pin a device). null = use all (auto). + const showGpuPicker = + isGguf && + gpuDevices.length > 1 && + gpuDevices.every((d) => d.physicalIndex); + const isGpuChecked = (index: number) => + selectedGpuIds === null || selectedGpuIds.includes(index); + const toggleGpu = (index: number) => { + const all = gpuDevices.map((d) => d.index); + const current = selectedGpuIds ?? all; + const next = current.includes(index) + ? current.filter((i) => i !== index) + : [...current, index].sort((a, b) => a - b); + if (next.length === 0) return; // keep at least one GPU selected + setSelectedGpuIds(next.length === all.length ? null : next); + // The per-GPU split is positional, so any change to the set of GPUs in use + // invalidates it: drop it (the sliders fall back to the VRAM-weighted + // default). TP needs 2+ GPUs, so disable it when only one remains. + setSplitRatio(null); + if (next.length <= 1) { + setTensorParallel(false); + } + }; + const gpuIdsKey = (ids: number[] | null) => (ids === null ? "auto" : ids.join(",")); + const gpuIdsDirty = gpuIdsKey(selectedGpuIds) !== gpuIdsKey(loadedGpuIds); + // Per-GPU layer split (--tensor-split): manual + 2+ GPUs in use. One slider + // per GPU, each a layer count; together they sum to the GPU Layers total. + const showSplitRatio = + isManual && !autoLayers && showGpuPicker && gpusInUse.length > 1; + // The total the per-GPU counts sum to (the GPU Layers slider value); 0 under + // Auto, where the split is hidden. The devices behind the GPUs in use, for + // labels + the VRAM-weighted default. + const splitTotal = Math.max(0, Math.min(gpuLayers, gpuLayersMax)); + const gpusInUseDevices = gpusInUse.map( + (i) => gpuDevices.find((d) => d.index === i) ?? null, + ); + // Displayed per-GPU counts. splitRatio is a stable reference balance (only a + // slider edit changes it), rescaled to the current total; deriving rather than + // mutating it on GPU Layers changes keeps the balance intact when the total + // passes through low values or Auto. No saved split: free-VRAM-weighted default + // (llama.cpp's unset default splits by free VRAM, so the first edit starts from + // the default's placement, not a total-VRAM ratio that can land layers on a + // busy GPU). A genuine 0 (a full GPU) is a real weight, not missing data: the + // probe's no-data case degrades to the total server-side, and an all-zero list + // falls back to an even split in distributeByWeight. Not yet sent. + const splitCounts = + splitRatio && splitRatio.length === gpusInUse.length + ? distributeByWeight(splitTotal, splitRatio) + : distributeByWeight( + splitTotal, + gpusInUseDevices.map((d) => d?.memoryFreeGb ?? d?.memoryTotalGb ?? 1), + ); + const setSplitCount = (k: number, v: number) => + setSplitRatio(rebalanceSplit(splitTotal, splitCounts, k, v)); + const splitRatioDirty = + isManual && + !autoLayers && + JSON.stringify(splitRatio ?? null) !== JSON.stringify(loadedSplitRatio ?? null); + // Auto-fit context (Manual + Auto layers): <= 0 means "Auto" (--fit sizes it); + // a positive value pins it. Surface the length --fit chose once it's loaded. + const fitCtxAuto = autoLayers && (customContextLength ?? 0) <= 0; + const loadedAutoLayers = + loadedGpuMemoryMode === "manual" && (loadedGpuLayers ?? GPU_LAYERS_AUTO) < 0; + const fitResolvedCtx = + fitCtxAuto && loadedAutoLayers ? ggufContextLength : null; // A saved chat-template override is a reload-time setting too, so surface // Apply for a template-only edit (otherwise it could never be applied). const templateDirty = chatTemplateOverride !== loadedChatTemplateOverride; const modelSettingsDirty = - kvDirty || ctxDirty || specDirty || specDraftDirty || tpDirty || templateDirty; + kvDirty || + ctxDirty || + specDirty || + specDraftDirty || + tpDirty || + gpuDirty || + manualDirty || + gpuIdsDirty || + splitRatioDirty || + templateDirty; const [presetNameInput, setPresetNameInput] = useState(activePreset); const [systemPromptEditorOpen, setSystemPromptEditorOpen] = useState(false); const [systemPromptDraft, setSystemPromptDraft] = useState(""); @@ -980,7 +1140,64 @@ export function ChatSettingsPanel({ )} {isGguf && ( <> - {showContextControl && ( + {showContextControl && (autoLayers ? ( +
+
+
+ + Context Length + + + Auto: llama.cpp's --fit sizes the context to fit VRAM. + Set a length to pin it instead -- --fit then optimizes + GPU layer offload around it. The length --fit chose + shows here after loading. + +
+ { + setCustomContextLength(v > 0 ? v : null); + }} + ariaLabel="Context Length" + size={8} + disabled={modelControlsDisabled} + /> +
+ { + // Far-left snaps to Auto; otherwise to the nearest 1024. + if (v < 512) { + setCustomContextLength(null); + } else { + setCustomContextLength(Math.round(v / 1024) * 1024); + } + }} + className="panel-slider" + disabled={modelControlsDisabled} + /> + {fitResolvedCtx != null && ( +

+ llama.cpp loaded {fitResolvedCtx.toLocaleString()} tokens. +

+ )} +
+ ) : (
@@ -1036,7 +1253,7 @@ export function ChatSettingsPanel({

)}
- )} + ))}
@@ -1191,6 +1408,163 @@ export function ChatSettingsPanel({ )} )} + {gpuModeApplies && ( +
+
+ + GPU Memory + + +
+
+ Default: Unsloth + fits the model and context to your GPUs. +
+
+ Manual: set GPU + Layers yourself. Leave it on Auto to let llama.cpp size + the context and offload overflow (including MoE experts) + to RAM. +
+
+ +
+
+ +
+
+ )} + {isManual && ( + <> + + Layers to keep on the GPU (--gpu-layers); the rest run + on CPU. Auto lets llama.cpp size the split (and the + context) to fit VRAM. At the maximum, the whole model + is on the GPU. + + } + /> + {showMoeSlider && ( + + Keep the experts of this many MoE layers on the CPU + (--n-cpu-moe) to save VRAM. 0 = all experts on the + GPU; at the maximum, all are on the CPU. + + } + /> + )} + {showSplitRatio && ( +
+
+ + Layers per GPU + + + Splits GPU Layers across GPUs (--tensor-split). + Without Tensor Parallelism each value is the layer + count on that GPU; with it, every GPU holds a slice + of each layer, so the values are only a ratio. + +
+ {gpusInUseDevices.map((d, k) => ( + setSplitCount(k, v)} + valueSize={6} + disabled={modelControlsDisabled} + /> + ))} +
+ )} + + )} + {showGpuPicker && ( +
+
+ + GPUs + + + Which GPUs this model may use. Unchecked GPUs are hidden + from llama.cpp (CUDA_VISIBLE_DEVICES, or + HIP_VISIBLE_DEVICES on ROCm). Leave all checked to use + every GPU. + +
+
+ {gpuDevices.map((d) => ( +
+ + GPU {d.index}: {d.name} + {d.memoryTotalGb + ? ` · ${Math.round(d.memoryTotalGb)} GB` + : ""} + + toggleGpu(d.index)} + data-test-id={`gpu-pick-${d.index}`} + disabled={modelControlsDisabled} + /> +
+ ))} +
+
+ )} + {gpuModeApplies && !autoLayers && (
@@ -1206,10 +1580,11 @@ export function ChatSettingsPanel({ className="panel-switch shrink-0" checked={tensorParallel} onCheckedChange={setTensorParallel} - disabled={modelControlsDisabled} + disabled={tpDisabled || modelControlsDisabled} data-test-id="tensor-parallel-switch" />
+ )} )} {/* No persistent "enable custom code" toggle: it is consented per model @@ -1228,14 +1603,21 @@ export function ChatSettingsPanel({ {Math.round((stagedDownloadFraction ?? 0) * 100)}%

)} - + {/* GGUF picks only: a non-GGUF pick shows none of the load + knobs the blob captures, so there is nothing to remember. */} + {pendingIsGguf && ( + + )} {stagedLoading ? ( // Mid-load: nothing to load or abandon until it settles, so disable.
) : null} - + {/* The template override is a load-time knob too (applied on the next + reload) and the in-flight load already snapshotted it, so lock its + editors like the sibling controls -- a mid-load save would be + silently clobbered by the load response despite its toast. */} +
)} @@ -2086,7 +2477,7 @@ function BypassPermissionsToggle() { ); } -function ChatTemplateFields() { +function ChatTemplateFields({ disabled = false }: { disabled?: boolean }) { const defaultTemplate = useChatRuntimeStore((s) => s.defaultChatTemplate); const override = useChatRuntimeStore((s) => s.chatTemplateOverride); const setOverride = useChatRuntimeStore((s) => s.setChatTemplateOverride); @@ -2120,7 +2511,8 @@ function ChatTemplateFields() { @@ -2131,7 +2523,8 @@ function ChatTemplateFields() { -
diff --git a/studio/frontend/src/features/chat/external-providers.ts b/studio/frontend/src/features/chat/external-providers.ts index bc718abbba..eb9e4656b0 100644 --- a/studio/frontend/src/features/chat/external-providers.ts +++ b/studio/frontend/src/features/chat/external-providers.ts @@ -184,11 +184,12 @@ export function supportsRemoteModelCatalog( ); } -/** Presets that skip the API-key field (local servers with no auth by default). */ +/** Presets that hide the API-key field. Ollama is not skipped: Ollama cloud + * requires a key; local servers leave the optional field empty. */ export function customPresetSkipsApiKeyField( providerType: string | null | undefined, ): boolean { - return providerType === "ollama" || providerType === "llama_cpp"; + return providerType === "llama_cpp"; } /** Catalog load plus optional manual model IDs. */ diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts index 10e0904e4f..3003b52230 100644 --- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts +++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts @@ -29,9 +29,14 @@ import { } from "../api/chat-api"; import { formatEta, formatRate } from "../utils/format-transfer"; import { + GPU_LAYERS_AUTO, isLocalModelPath, + loadedGpuMemoryFields, + loadedGpuMemoryFieldsUnlessStaged, pendingSelectionMatches, + persistGpuMemoryModeOnLoad, readPersistedSpeculativeType, + reconcilePersistedGpuIds, resolveToolsEnabledOnLoad, saveSpeculativeType, useChatRuntimeStore, @@ -46,9 +51,12 @@ import { } from "../lib/apply-inference-status-to-store"; import { mergeBackendRecommendedInference, + resolveFitMaxSeqLength, resolveLoadMaxSeqLength, + resolveManualAutoCtxPin, } from "../presets/preset-policy"; import { recordLastLocalModelLoad } from "../utils/last-local-model-load"; +import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info"; import { isMultimodalResponse, } from "../types/api"; @@ -291,9 +299,12 @@ async function syncInferenceStatusToStore(options?: { if (statusRes.active_model && !isExternalSelectionActive) { const checkpointId = resolveInferenceCheckpointId(statusRes); if (checkpointId) { + const previousGgufVariant = + useChatRuntimeStore.getState().activeGgufVariant; setCheckpoint(checkpointId, statusRes.gguf_variant); applyActiveModelStatusToStore(statusRes, { previousCheckpoint: selectedCheckpoint, + previousGgufVariant, }); // setModels(listRes...) above used catalog data, which omits audio // capability. Re-apply live status so attach gates survive a refresh. @@ -511,7 +522,11 @@ export function useChatModelRuntime() { typeof selection === "string" ? false : selection.isDownloaded ?? false; const model = models.find((entry) => entry.id === modelId); const lora = loras.find((entry) => entry.id === modelId); - const isGguf = explicitIsGguf ?? model?.isGguf ?? false; + // A native path-token selection is a local GGUF by construction (the + // native model intents only grant .gguf files), but its id is a display + // label that need not end in ".gguf" -- without this, Manual + Auto + // layers would pin the UI context instead of letting --fit size it. + const isGguf = explicitIsGguf ?? model?.isGguf ?? nativePathToken != null; const loraIsAdapter = lora?.exportType === "lora"; const isLora = explicitIsLora ?? model?.isLora ?? loraIsAdapter ?? false; @@ -578,18 +593,27 @@ export function useChatModelRuntime() { let trustRemoteCode = stateBeforeUnload.params.trustRemoteCode ?? false; let approvedRemoteCodeFingerprint: string | null = null; const maxSeqLength = stateBeforeUnload.params.maxSeqLength; + const previousActiveNativePathToken = + stateBeforeUnload.activeNativePathToken; const previousIsGguf = previousModel?.isGguf === true || previousVariant != null + || previousActiveNativePathToken != null || (previousCheckpoint?.toLowerCase().endsWith(".gguf") ?? false); - const rollbackMaxSeqLength = previousIsGguf - ? (stateBeforeUnload.ggufContextLength ?? 0) - : maxSeqLength; + // Respect the rolled-back model's auto-layers mode: a Manual+Auto model + // with an unpinned (auto) context must reload with 0 (so --fit + // re-auto-sizes), not the positive context it happened to pick (which + // the backend would treat as a pin). + const rollbackMaxSeqLength = resolveFitMaxSeqLength( + previousIsGguf, + stateBeforeUnload.loadedGpuMemoryMode ?? "auto", + stateBeforeUnload.loadedGpuLayers ?? GPU_LAYERS_AUTO, + stateBeforeUnload.loadedCustomContextLength, + previousIsGguf ? (stateBeforeUnload.ggufContextLength ?? 0) : maxSeqLength, + ); const hfToken = stateBeforeUnload.hfToken || null; const previousModelRequiresTrustRemoteCode = stateBeforeUnload.modelRequiresTrustRemoteCode; - const previousActiveNativePathToken = - stateBeforeUnload.activeNativePathToken; // Snapshot the load settings at click time, before the awaits below // (validation, the trust dialog, unload). For a staged Load these knobs // stay editable and a sheet-close revert (abandonStagedModel) can fire @@ -598,11 +622,29 @@ export function useChatModelRuntime() { // updates this snapshot in lock-step so non-staged loads are unchanged. const loadChatTemplateOverride = stateBeforeUnload.chatTemplateOverride; const loadKvCacheDtype = stateBeforeUnload.kvCacheDtype; - const loadCustomContextLength = stateBeforeUnload.customContextLength; + // gpuMemoryMode is a standing preference (kept across a model switch); + // the rest are per-model knobs the reset below clears, so they are + // re-baselined there in lock-step with the store. + let loadCustomContextLength = stateBeforeUnload.customContextLength; const loadGgufContextLength = stateBeforeUnload.ggufContextLength; const loadTensorParallel = stateBeforeUnload.tensorParallel; const loadActivePresetSource = stateBeforeUnload.activePresetSource; const loadActiveGgufVariant = stateBeforeUnload.activeGgufVariant; + const loadGpuMemoryMode = stateBeforeUnload.gpuMemoryMode; + let loadGpuLayers = stateBeforeUnload.gpuLayers; + let loadNCpuMoe = stateBeforeUnload.nCpuMoe; + let loadSplitRatio = stateBeforeUnload.splitRatio; + // Reconcile the persisted pick against the GPUs present now, so a stale + // cross-host / now-hidden pick is dropped before /load rather than + // rejected there. Warm the device cache first: load-on-selection can + // run before any GPU hook mounted, and a cold cache would pass the + // pick through unvalidated. validateGpuIds derives from this too. + if (stateBeforeUnload.selectedGpuIds != null) { + await ensureGpuDeviceCache(); + } + let loadSelectedGpuIds = reconcilePersistedGpuIds( + stateBeforeUnload.selectedGpuIds, + ); let loadSpeculativeType = stateBeforeUnload.speculativeType; let loadSpecDraftNMax = stateBeforeUnload.specDraftNMax; try { @@ -615,16 +657,47 @@ export function useChatModelRuntime() { // context can exceed maxSeqLength, so sizing on raw maxSeqLength could // pass, unload, then have /load refuse it. Uses the click-time // snapshot (same values loadModel uses below), so the two agree. - const validateMaxSeqLength = resolveLoadMaxSeqLength({ - modelId, - ggufVariant, - customContextLength: loadCustomContextLength, - ggufContextLength: loadGgufContextLength, - currentCheckpoint, - activeGgufVariant: loadActiveGgufVariant, - maxSeqLength, - presetSource: loadActivePresetSource, - }); + // Mirror what /load does on a cross-model switch: the reset below + // clears the per-model Auto-layers context pin + GPU pick, and + // Manual+Auto sizes context through resolveFitMaxSeqLength. + // gpuMemoryMode is a standing preference, kept across the switch. + // A same-repo quant switch (same checkpoint, different gguf_variant) + // is a different model for per-model knobs: the pinned context, + // gpuLayers, GPU pick, and MoE offload are scoped per variant, so + // treat a variant change like a model switch and re-baseline them. + const switchingModelOrVariant = + currentCheckpoint !== modelId || + (loadActiveGgufVariant ?? null) !== (ggufVariant ?? null); + const resetsPerModelSettings = Boolean( + currentCheckpoint && switchingModelOrVariant && !keepSpeculative, + ); + const validateCustomContextLength = resetsPerModelSettings + ? null + : loadCustomContextLength; + const validateGpuIds = resetsPerModelSettings + ? null + : loadSelectedGpuIds; + // The reset below re-baselines gpuLayers to Auto; mirror it here. + const validateGpuLayers = resetsPerModelSettings + ? GPU_LAYERS_AUTO + : loadGpuLayers; + const validateMaxSeqLength = resolveFitMaxSeqLength( + isGguf, + loadGpuMemoryMode, + validateGpuLayers, + validateCustomContextLength, + resolveLoadMaxSeqLength({ + modelId, + ggufVariant, + isGguf, + customContextLength: validateCustomContextLength, + ggufContextLength: loadGgufContextLength, + currentCheckpoint, + activeGgufVariant: loadActiveGgufVariant, + maxSeqLength, + presetSource: loadActivePresetSource, + }), + ); const validation = await validateModel({ model_path: modelId, nativePathLease: validateNativePathLease, @@ -633,6 +706,8 @@ export function useChatModelRuntime() { load_in_4bit: true, is_lora: isLora, gguf_variant: ggufVariant ?? null, + gpu_ids: validateGpuIds ?? undefined, + ...(isGguf ? { gpu_memory_mode: loadGpuMemoryMode } : {}), }); // Upgrade consent runs before the security dialogs; Accept installs and the load continues. if (validation.requires_transformers_upgrade) { @@ -697,18 +772,52 @@ export function useChatModelRuntime() { // keepSpeculative skips this for a staged Load: the user picked the // mode for this model on the sidebar, so honor it (the backend still // falls back at runtime if the model has no MTP head). - if (currentCheckpoint && currentCheckpoint !== modelId && !keepSpeculative) { + if (resetsPerModelSettings) { const persistedSpeculativeType = readPersistedSpeculativeType(); useChatRuntimeStore.setState({ speculativeType: persistedSpeculativeType, loadedSpeculativeType: persistedSpeculativeType, specDraftNMax: null, loadedSpecDraftNMax: null, + // Per-model GPU knobs must not follow onto a different model + // (gpuMemoryMode is a standing preference and is kept). + selectedGpuIds: null, + gpuLayers: GPU_LAYERS_AUTO, + nCpuMoe: 0, + splitRatio: null, + // A Manual+Auto context pin is per-model; clear it so a different + // model loads at Auto/native, not the previous model's pin. + customContextLength: null, }); loadSpeculativeType = persistedSpeculativeType; loadSpecDraftNMax = null; + // Keep the click-time snapshot in lock-step with the store reset so + // the load below sizes against the cleared per-model knobs, not the + // previous model's (gpuMemoryMode is standing, so left as captured). + loadCustomContextLength = null; + loadSelectedGpuIds = null; + loadGpuLayers = GPU_LAYERS_AUTO; + loadNCpuMoe = 0; + loadSplitRatio = null; } + // Pinning layers on the SAME model keeps the currently resolved + // context: with no explicit pin, a manual+pinned reload would send 0, + // which the backend's --fit off branch treats as the NATIVE context -- + // far larger than the sheet shows when the load was fit-sized (Default + // or Manual + Auto layers may auto-reduce context to fit VRAM), a + // likely OOM. ggufContextLength is that resolved value; a model already + // at native reloads unchanged, so this is safe for any prior mode. + if ( + isGguf && + !switchingModelOrVariant && + loadGpuMemoryMode === "manual" && + loadGpuLayers >= 0 && + loadCustomContextLength == null && + (loadGgufContextLength ?? 0) > 0 + ) { + loadCustomContextLength = loadGgufContextLength; + } const effectiveMaxSeqLength = resolveLoadMaxSeqLength({ modelId, ggufVariant, @@ -720,13 +829,20 @@ export function useChatModelRuntime() { maxSeqLength, presetSource: loadActivePresetSource, }); + const loadMaxSeqLength = resolveFitMaxSeqLength( + isGguf, + loadGpuMemoryMode, + loadGpuLayers, + loadCustomContextLength, + effectiveMaxSeqLength, + ); const effectiveChatTemplateOverride = loadChatTemplateOverride?.trim() ? loadChatTemplateOverride : null; const loadResponse = await loadModel({ model_path: modelId, nativePathLease: loadNativePathLease, hf_token: hfToken, - max_seq_length: effectiveMaxSeqLength, + max_seq_length: loadMaxSeqLength, load_in_4bit: true, is_lora: isLora, gguf_variant: ggufVariant ?? null, @@ -737,6 +853,11 @@ export function useChatModelRuntime() { speculative_type: loadSpeculativeType, spec_draft_n_max: loadSpecDraftNMax, tensor_parallel: loadTensorParallel, + gpu_memory_mode: loadGpuMemoryMode, + gpu_layers: loadGpuLayers, + n_cpu_moe: loadNCpuMoe, + tensor_split: loadSplitRatio ?? undefined, + gpu_ids: loadSelectedGpuIds ?? undefined, }); // If cancelled while loading, don't update UI to show @@ -747,6 +868,9 @@ export function useChatModelRuntime() { // preference now (the requested intent, not the resolved echo; // saveSpeculativeType keeps only the universal auto/ngram/off). saveSpeculativeType(loadSpeculativeType); + // Persist the GPU Memory mode only on a successful load (not on + // dropdown change), so an abandoned selection doesn't stick. + persistGpuMemoryModeOnLoad(loadResponse, loadGpuMemoryMode); const currentParams = useChatRuntimeStore.getState().params; setParams( @@ -782,9 +906,13 @@ export function useChatModelRuntime() { const reportedNativeCtx = loadResponse.is_gguf ? (loadResponse.native_context_length ?? null) : null; - // A successful reload has applied settings, so clear pending custom - // context state and display the backend-reported effective context. - const keepCustomCtx = null; + // Keep an explicit Manual+Auto context pin (so a later Apply doesn't + // revert it to Auto); other cases baseline on ggufContextLength. + const keepCustomCtx = resolveManualAutoCtxPin( + loadGpuMemoryMode, + loadGpuLayers, + loadCustomContextLength, + ); const reasoningAlwaysOn = loadResponse.reasoning_always_on ?? false; const reasoningStyle = loadResponse.reasoning_style ?? "enable_thinking"; const supportsReasoning = loadResponse.supports_reasoning ?? false; @@ -837,11 +965,13 @@ export function useChatModelRuntime() { loadedKvCacheDtype: loadedKv, tensorParallel: loadedTp, loadedTensorParallel: loadedTp, + ...loadedGpuMemoryFields(loadResponse), speculativeType: loadedSpec, loadedSpeculativeType: loadedSpec, specDraftNMax: loadResponse.spec_draft_n_max ?? null, loadedSpecDraftNMax: loadResponse.spec_draft_n_max ?? null, customContextLength: keepCustomCtx, + loadedCustomContextLength: keepCustomCtx, defaultChatTemplate: loadResponse.chat_template ?? null, chatTemplateOverride: effectiveChatTemplateOverride, loadedChatTemplateOverride: effectiveChatTemplateOverride, @@ -938,7 +1068,7 @@ export function useChatModelRuntime() { } } try { - await loadModel({ + const rollbackResponse = await loadModel({ model_path: previousCheckpoint, nativePathLease: rollbackNativePathLease, hf_token: hfToken, @@ -951,14 +1081,51 @@ export function useChatModelRuntime() { // Resend the previous model's pinned approval so restoring it is not re-blocked. approved_remote_code_fingerprint: approvedRemoteCodeFingerprints.get(previousCheckpoint) ?? null, + chat_template_override: + stateBeforeUnload.loadedChatTemplateOverride, + cache_type_kv: stateBeforeUnload.loadedKvCacheDtype, + speculative_type: + stateBeforeUnload.loadedSpeculativeType, + spec_draft_n_max: + stateBeforeUnload.loadedSpecDraftNMax, // Restore the previous model in the split mode it was running, // not the default layer split. tensor_parallel: stateBeforeUnload.loadedTensorParallel ?? false, + gpu_memory_mode: stateBeforeUnload.loadedGpuMemoryMode ?? "auto", + gpu_layers: stateBeforeUnload.loadedGpuLayers ?? -1, + n_cpu_moe: stateBeforeUnload.loadedNCpuMoe ?? 0, + tensor_split: stateBeforeUnload.loadedSplitRatio ?? undefined, + gpu_ids: stateBeforeUnload.loadedGpuIds ?? undefined, }); + const rollbackSpeculativeType = normalizeSpeculativeType( + rollbackResponse.speculative_type, + ); useChatRuntimeStore.setState({ activeNativePathToken: previousActiveNativePathToken ?? null, - loadedSpeculativeType: null, - loadedSpecDraftNMax: null, + loadedSpeculativeType: rollbackSpeculativeType, + loadedSpecDraftNMax: + rollbackResponse.spec_draft_n_max ?? null, + loadedKvCacheDtype: rollbackResponse.cache_type_kv ?? null, + loadedChatTemplateOverride: + stateBeforeUnload.loadedChatTemplateOverride, + // Re-baseline the GPU knobs from the rolled-back load's own + // response (the shared seeding every load path uses): the + // refresh() below can't do it, since the status reseed is + // gated off while modelLoading is still true. A failed staged + // Load stays staged for retry, so the staged hold applies. + ...loadedGpuMemoryFieldsUnlessStaged(rollbackResponse, { + tensorParallel: rollbackResponse.tensor_parallel ?? false, + loadedTensorParallel: + rollbackResponse.tensor_parallel ?? false, + // refresh() is held while modelLoading remains true, so + // restore the rolled-back model's context pin directly. + customContextLength: + stateBeforeUnload.loadedCustomContextLength, + }), + loadedTensorParallel: + rollbackResponse.tensor_parallel ?? false, + loadedCustomContextLength: + stateBeforeUnload.loadedCustomContextLength, }); await refresh(); } catch { diff --git a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts index 0a0df1139b..bfb3eeb14c 100644 --- a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts +++ b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts @@ -45,7 +45,7 @@ export function groupThreads( for (const t of threads) { // Coerce archived to a boolean before comparing. Legacy threads (from the - // older browser-only Studio, or any record predating the archived field) + // older browser-only Unsloth, or any record predating the archived field) // can have archived === undefined or null; a raw `!== archived` comparison // would drop those from BOTH the Recents (archived=false) and Archived // (archived=true) lists, hiding existing chats. Treat missing as false. diff --git a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts index d8076c720b..a3e7a2d264 100644 --- a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts +++ b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts @@ -7,7 +7,7 @@ import { useRepoDownload } from "@/features/hub/download-manager/use-repo-downlo import type { DownloadJob } from "@/features/hub/download-manager/use-repo-download"; import { useLatestRef } from "@/features/hub/hooks/use-latest-ref"; -import { fetchGgufContextLength } from "../api/chat-api"; +import { fetchGgufStagedMetadata } from "../api/chat-api"; import { isPendingGguf, pendingSelectionMatches, @@ -46,8 +46,16 @@ export function useStagedModelPreparation(opts?: { const pendingDownloaded = useChatRuntimeStore( (s) => s.pendingSelection?.isDownloaded ?? false, ); - const pendingHasContext = useChatRuntimeStore( - (s) => s.pendingSelection?.contextLength != null, + // "Already probed" must key off layerCount / moeLayerCount, which only the + // full header probe fills (it sets all three together, so either is a + // reliable marker). contextLength alone can be list-seeded from + // /gguf-variants, which returns no layer/MoE counts -- treating it as + // complete would skip the probe and leave the GPU Layers slider at its 256 + // fallback and the MoE slider hidden until the model loads. + const pendingHasMetadata = useChatRuntimeStore( + (s) => + s.pendingSelection?.layerCount != null || + s.pendingSelection?.moeLayerCount != null, ); const setPendingSelection = useChatRuntimeStore((s) => s.setPendingSelection); const onAutoLoadRef = useLatestRef(opts?.onAutoLoad); @@ -69,25 +77,31 @@ export function useStagedModelPreparation(opts?: { if (!current?.id || !isPendingGguf(current)) return; const { id, ggufVariant, nativePathToken } = current; try { - const contextLength = await fetchGgufContextLength({ - model_path: id, - gguf_variant: ggufVariant, - hf_token: useChatRuntimeStore.getState().hfToken || null, - nativePathToken, - }); + const { contextLength, layerCount, moeLayerCount } = + await fetchGgufStagedMetadata({ + model_path: id, + gguf_variant: ggufVariant, + hf_token: useChatRuntimeStore.getState().hfToken || null, + nativePathToken, + }); // Apply only if the same model is still staged (the user may have switched // picks or loaded/cancelled while the request was in flight). const latest = useChatRuntimeStore.getState().pendingSelection; if ( latest && - contextLength != null && - pendingSelectionMatches(latest, { id, ggufVariant, nativePathToken }) + pendingSelectionMatches(latest, { id, ggufVariant, nativePathToken }) && + (contextLength != null || layerCount != null || moeLayerCount != null) ) { - setPendingSelection({ ...latest, contextLength }); + setPendingSelection({ + ...latest, + contextLength, + layerCount, + moeLayerCount, + }); } } catch { - // Leave contextLength null: the context slider stays hidden and the user - // can still load (context fills in from the load response afterwards). + // Leave metadata null: the context/MoE sliders stay hidden and the user + // can still load (they fill in from the load response afterwards). } }, [setPendingSelection]); @@ -125,7 +139,7 @@ export function useStagedModelPreparation(opts?: { if ( !pendingId || (!pendingIsGguf && !pendingIsHubRepo) || - pendingHasContext + pendingHasMetadata ) { return; } @@ -146,7 +160,7 @@ export function useStagedModelPreparation(opts?: { pendingIsGguf, pendingIsHubRepo, pendingDownloaded, - pendingHasContext, + pendingHasMetadata, startDownloadRef, fetchMetadataRef, ]); diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts index a4b5f848e2..69bb38bbbe 100644 --- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts +++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts @@ -2,13 +2,17 @@ // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 import { getInferenceStatus } from "../api/chat-api"; -import { mergeBackendRecommendedInference } from "../presets/preset-policy"; +import { + mergeBackendRecommendedInference, + resolveManualAutoCtxPin, +} from "../presets/preset-policy"; import { clampReasoningEffortToLevels } from "../provider-capabilities"; import { CHAT_REASONING_ENABLED_KEY, type ReasoningEffort, type ReasoningStyle, loadOptionalBool, + loadedGpuMemoryFields, resolveToolsEnabledOnLoad, useChatRuntimeStore, } from "../stores/chat-runtime-store"; @@ -20,6 +24,10 @@ import type { ChatModelSummary } from "../types/runtime"; type LocalReasoningEffort = Extract; +function sameArray(a: T[] | null, b: T[] | null): boolean { + return JSON.stringify(a) === JSON.stringify(b); +} + // Canonicalises backend / persisted speculative mode values onto the UI modes. export function normalizeSpeculativeType( v: string | null | undefined, @@ -119,6 +127,10 @@ function ensureActiveModelInStoreList( export type ApplyInferenceStatusOptions = { previousCheckpoint?: string; + /** activeGgufVariant BEFORE the caller's setCheckpoint synced it to the + * status -- without it a variant-only switch underneath the tab reads as + * steady state and the hydration reseed keeps the old quant's baselines. */ + previousGgufVariant?: string | null; }; /** Mirror refresh() hydration so adopted CLI models get reasoning/tools flags. */ @@ -144,9 +156,13 @@ export function applyActiveModelStatusToStore( ); } + const previousGgufVariant = + options.previousGgufVariant !== undefined + ? options.previousGgufVariant + : store.activeGgufVariant; const hydratingExistingModel = previousCheckpoint !== checkpointId || - store.activeGgufVariant !== (status.gguf_variant ?? null); + previousGgufVariant !== (status.gguf_variant ?? null); const supportsReasoning = status.supports_reasoning ?? false; const reasoningAlwaysOn = status.reasoning_always_on ?? false; const reasoningStyle = status.reasoning_style ?? "enable_thinking"; @@ -185,6 +201,66 @@ export function applyActiveModelStatusToStore( // While a load is in flight, performLoad owns the load params. Seeding them // from a stale poll here would clobber the values the load dialog just set. const seedLoadParams = !prevState.modelLoading; + // A Manual + Auto-layers load sent its positive context pin as max_seq_length, + // and status only exposes the RESOLVED context; re-seed the pin from the + // requested value (parity with the load paths' keepCustomCtx). Baselines + // unconditionally: anything but an applicable pin is null, so a previous + // model's pin can't survive a model change underneath and reload at the old length. + const gpuPin = status.is_gguf + ? resolveManualAutoCtxPin( + status.gpu_memory_mode ?? "auto", + status.gpu_layers ?? -1, + status.requested_context_length ?? null, + ) + : null; + const incomingGpuMode = status.is_gguf + ? (status.gpu_memory_mode ?? "auto") + : null; + const incomingGpuLayers = + incomingGpuMode === "manual" ? (status.gpu_layers ?? null) : null; + const incomingNCpuMoe = + incomingGpuMode === "manual" ? (status.n_cpu_moe ?? null) : null; + const incomingSplit = + incomingGpuMode === "manual" ? (status.tensor_split ?? null) : null; + const incomingGpuIds = status.is_gguf ? (status.gpu_ids ?? null) : null; + const gpuStatusChanged = + prevState.loadedGpuMemoryMode !== incomingGpuMode || + prevState.loadedGpuLayers !== incomingGpuLayers || + prevState.loadedNCpuMoe !== incomingNCpuMoe || + !sameArray(prevState.loadedSplitRatio, incomingSplit) || + !sameArray(prevState.loadedGpuIds, incomingGpuIds) || + prevState.loadedCustomContextLength !== gpuPin; + const gpuMemoryEditsPending = + (prevState.loadedGpuMemoryMode !== null && + prevState.gpuMemoryMode !== prevState.loadedGpuMemoryMode) || + (prevState.loadedGpuMemoryMode === "manual" && + (prevState.gpuLayers !== prevState.loadedGpuLayers || + prevState.nCpuMoe !== prevState.loadedNCpuMoe || + !sameArray(prevState.splitRatio, prevState.loadedSplitRatio))) || + prevState.customContextLength !== prevState.loadedCustomContextLength; + const gpuIdsEditPending = !sameArray( + prevState.selectedGpuIds, + prevState.loadedGpuIds, + ); + const incomingGpuFields = loadedGpuMemoryFields(status); + // A same-model reload from another client advances every loaded baseline. + // Preserve each editable group only when this tab has an unapplied change. + const preserveSameModelEdits = gpuStatusChanged && !hydratingExistingModel; + const gpuStatusFields = { + ...incomingGpuFields, + customContextLength: gpuPin, + loadedCustomContextLength: gpuPin, + ...(preserveSameModelEdits && + gpuMemoryEditsPending && { + gpuMemoryMode: prevState.gpuMemoryMode, + gpuLayers: prevState.gpuLayers, + nCpuMoe: prevState.nCpuMoe, + splitRatio: prevState.splitRatio, + customContextLength: prevState.customContextLength, + }), + ...(preserveSameModelEdits && + gpuIdsEditPending && { selectedGpuIds: prevState.selectedGpuIds }), + }; useChatRuntimeStore.setState({ supportsReasoning, @@ -215,30 +291,51 @@ export function applyActiveModelStatusToStore( loadedIsMultimodal: isMultimodalResponse(status), loadedIsDiffusion: status.is_diffusion ?? false, specFallbackReason: status.spec_fallback_reason ?? null, + // The spec / KV seeds share the GPU-fields reseed mechanism below: a + // non-GGUF status leaves their loaded baselines null, so the "unseeded" + // guard re-fires every refresh -- hold them too while a staged pick's + // settings are being edited, or the refresh resets the staged edit. + // hydratingExistingModel reopens every load-param seed: when the active + // model changed underneath this tab (auto-switch, another client), the + // old model's baselines are stale and must adopt the new status. ...(seedLoadParams && - prevState.loadedSpeculativeType === null && { + prevState.pendingSelection == null && + (prevState.loadedSpeculativeType === null || hydratingExistingModel) && { speculativeType: currentSpecType, loadedSpeculativeType: currentSpecType, }), ...(seedLoadParams && + prevState.pendingSelection == null && status.spec_draft_n_max !== undefined && - prevState.loadedSpecDraftNMax === null && - prevState.specDraftNMax === null && { + (hydratingExistingModel || + (prevState.loadedSpecDraftNMax === null && + prevState.specDraftNMax === null)) && { specDraftNMax: status.spec_draft_n_max ?? null, loadedSpecDraftNMax: status.spec_draft_n_max ?? null, }), ...(seedLoadParams && + prevState.pendingSelection == null && status.cache_type_kv !== undefined && - prevState.loadedKvCacheDtype === null && { + (prevState.loadedKvCacheDtype === null || hydratingExistingModel) && { kvCacheDtype: status.cache_type_kv, loadedKvCacheDtype: status.cache_type_kv, }), ...(seedLoadParams && + prevState.pendingSelection == null && status.tensor_parallel !== undefined && - prevState.loadedTensorParallel === null && { + (prevState.loadedTensorParallel === null || hydratingExistingModel) && { tensorParallel: status.tensor_parallel, loadedTensorParallel: status.tensor_parallel, }), + // Re-seed on first hydration, model/variant changes, or a same-model backend + // placement change. gpuStatusFields preserves dirty local edits in the last + // case while advancing their loaded baselines. + ...(seedLoadParams && + prevState.pendingSelection == null && + (prevState.loadedGpuMemoryMode === null || + hydratingExistingModel || + gpuStatusChanged) && + gpuStatusFields), ...(status.chat_template_override !== undefined && prevState.loadedChatTemplateOverride === null && prevState.chatTemplateOverride === null && { @@ -298,7 +395,11 @@ export async function tryAdoptServerActiveModel(): Promise { if (previousCheckpoint) { return true; } + const previousGgufVariant = useChatRuntimeStore.getState().activeGgufVariant; store.setCheckpoint(checkpointId, status.gguf_variant); - applyActiveModelStatusToStore(status, { previousCheckpoint }); + applyActiveModelStatusToStore(status, { + previousCheckpoint, + previousGgufVariant, + }); return true; } diff --git a/studio/frontend/src/features/chat/lib/friendly-names.ts b/studio/frontend/src/features/chat/lib/friendly-names.ts index 79744b3181..bc9c77b12d 100644 --- a/studio/frontend/src/features/chat/lib/friendly-names.ts +++ b/studio/frontend/src/features/chat/lib/friendly-names.ts @@ -5,7 +5,7 @@ * Friendly default names for auto-created OpenAI shell containers, used by the * chat-adapter's lazy-create path (Code pill on, no thread container, non-default * TTL). Goal: a memorable label like "otter" instead of "chat-abc12345"; users - * can still rename via the Studio alias map. + * can still rename via the Unsloth alias map. * * The list is curated to be unambiguous, non-offensive nouns from natural * categories (animals, plants, geography, materials, weather), avoid diff --git a/studio/frontend/src/features/chat/permission-mode-select.tsx b/studio/frontend/src/features/chat/permission-mode-select.tsx index 4277c1bfcf..a9cb8ce5d1 100644 --- a/studio/frontend/src/features/chat/permission-mode-select.tsx +++ b/studio/frontend/src/features/chat/permission-mode-select.tsx @@ -278,28 +278,23 @@ export function PermissionModeComposerPill({ data-pill-label={active.label} data-active={fullAccess ? "true" : "false"} data-variant={fullAccess ? "danger" : undefined} + data-keep-label="true" aria-label="Permission level for tool calls" title={`${active.label}: ${active.description}`} > {/* The icon doubles as an off switch (mirrors the MCP pill): hover swaps it to an X; clicking it turns bypass permissions Off (no - prompts, sandbox on) without opening the menu. In compact - icon-only mode the glyph is the whole button, so clicks fall - through and open the menu instead. */} + prompts, sandbox on) without opening the menu. data-keep-label + exempts this pill from compact icon-only mode, so the off switch + stays clickable even while the other pills are collapsed. */} { - if (e.currentTarget.closest('[data-pill-compact="true"]')) { - return; - } e.stopPropagation(); }} onClick={(e) => { - if (e.currentTarget.closest('[data-pill-compact="true"]')) { - return; - } e.stopPropagation(); setPermissionMode("off"); }} diff --git a/studio/frontend/src/features/chat/presets/preset-policy.ts b/studio/frontend/src/features/chat/presets/preset-policy.ts index f96ee91f1b..23d79a35e1 100644 --- a/studio/frontend/src/features/chat/presets/preset-policy.ts +++ b/studio/frontend/src/features/chat/presets/preset-policy.ts @@ -339,3 +339,34 @@ export function resolveLoadMaxSeqLength({ } return maxSeqLength; } + +/** + * Adjust a resolved max-seq-length for the GPU Memory mode. Under Manual + Auto + * layers (GGUF, gpuLayers < 0) llama.cpp's --fit owns context sizing, so send 0 + * (the backend omits -c) unless the user pinned a length; every other case keeps + * the resolved fallback. Shared by every GGUF load path so they can't drift. + */ +export function resolveFitMaxSeqLength( + isGguf: boolean | null | undefined, + gpuMemoryMode: "auto" | "manual", + gpuLayers: number, + customContextLength: number | null, + fallback: number, +): number { + if (!isGguf || gpuMemoryMode !== "manual" || gpuLayers >= 0) return fallback; + return customContextLength && customContextLength > 0 ? customContextLength : 0; +} + +// A Manual + Auto-layers load sends its positive context pin as max_seq_length; +// keep it across a status reseed/Apply so the model isn't reverted to auto-fit +// sizing. Anything else (Auto mode, pinned layers, no pin) baselines to null. +// The caller keeps its own isGguf/targetIsGguf guard inline. +export function resolveManualAutoCtxPin( + gpuMemoryMode: "auto" | "manual", + gpuLayers: number, + customContextLength: number | null, +): number | null { + return gpuMemoryMode === "manual" && gpuLayers < 0 && (customContextLength ?? 0) > 0 + ? customContextLength + : null; +} diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts index 79c9a3205c..ec251cdada 100644 --- a/studio/frontend/src/features/chat/provider-capabilities.ts +++ b/studio/frontend/src/features/chat/provider-capabilities.ts @@ -409,7 +409,7 @@ function isGeminiImageModel(modelId: string): boolean { * Whether the saved Gemini connection points at a custom OpenAI-compat gateway * (any non-Google host). The backend `_is_openai_compatible` routes these * through `/chat/completions` instead of the native translator, so native Gemini - * tool envelopes never reach them. Hide the matching Studio pills here so the + * tool envelopes never reach them. Hide the matching Unsloth pills here so the * request, builder, and UI agree. */ export function isGeminiCustomOpenAICompatBase( diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx index a0813fe27b..31e50ee60c 100644 --- a/studio/frontend/src/features/chat/shared-composer.tsx +++ b/studio/frontend/src/features/chat/shared-composer.tsx @@ -84,6 +84,8 @@ import { useTransformersUpgradeDialogStore, } from "@/features/transformers-upgrade"; import { loadModel, validateModel } from "./api/chat-api"; +import { resolveFitMaxSeqLength, resolveManualAutoCtxPin } from "./presets/preset-policy"; +import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info"; import { parseExternalModelId, providerTypeSupportsVision, @@ -95,8 +97,11 @@ import { usePlusMenuPrefsStore, } from "./stores/plus-menu-prefs-store"; import { + loadedGpuMemoryFieldsUnlessStaged, type ReasoningEffort, + reconcilePersistedGpuIds, resolveLoadedSpeculativeSettings, + persistGpuMemoryModeOnLoad, resolveSpeculativeSettingsForLoad, saveSpeculativeType, useChatRuntimeStore, @@ -1037,10 +1042,32 @@ export function SharedComposer({ return parts[parts.length - 1] || id; } + // Warm the device cache before the snapshot below reconciles the GPU + // pick: on a cold cache the reconcile passes a stale pick through. + if (store.selectedGpuIds != null) { + await ensureGpuDeviceCache(); + } + // The GPU/offload knobs both compare loads must use, snapshotted at Send. + // ensureModelLoaded runs sequentially and the first load's response echo + // (loadedGpuMemoryFields) rewrites the live store -- a non-GGUF or Auto + // first model resets gpuLayers/nCpuMoe/split/pick to defaults -- so + // reading the store per load would hand model 2 the first model's echoed + // defaults instead of the settings the user pressed Send with. + const compareLoadKnobs = { + gpuMemoryMode: store.gpuMemoryMode, + gpuLayers: store.gpuLayers, + nCpuMoe: store.nCpuMoe, + splitRatio: store.splitRatio, + // Reconcile the pick against the GPUs present now, like the model-switch + // path: an early remember-restore can hold a stale cross-host pick that + // /load would reject (the device cache is populated by send time). + selectedGpuIds: reconcilePersistedGpuIds(store.selectedGpuIds), + tensorParallel: store.tensorParallel, + customContextLength: store.customContextLength, + }; // Set when an accepted transformers install unloaded the active model // server-side; a later failure must then clear the stale checkpoint. let upgradeUnloadedActive = false; - // Helper: load a model and update store checkpoint async function ensureModelLoaded( sel: CompareModelSelection, @@ -1057,15 +1084,35 @@ export function SharedComposer({ if (isAlreadyActive) { return "ready"; } + const targetIsGguf = + sel.id.toLowerCase().endsWith(".gguf") || sel.ggufVariant != null; + // Size validation exactly as the load below, so the training-guard + // preflight checks the footprint that actually loads (under Manual + Auto + // layers the load sends 0 / the pinned context, not raw maxSeqLength). + const compareMaxSeqLength = resolveFitMaxSeqLength( + targetIsGguf, + compareLoadKnobs.gpuMemoryMode, + compareLoadKnobs.gpuLayers, + compareLoadKnobs.customContextLength, + maxSeqLength, + ); const validation = await validateModel({ model_path: sel.id, hf_token: currentStore.hfToken || null, - max_seq_length: maxSeqLength, + max_seq_length: compareMaxSeqLength, load_in_4bit: true, is_lora: sel.isLora, gguf_variant: sel.ggufVariant ?? null, trust_remote_code: loadTrustRemoteCode, chat_template_override: effectiveChatTemplateOverride, + // Scope the validate to the picked GPUs. GGUF-only, like the load + // below: a non-GGUF target must not inherit a hidden GGUF GPU pick. + ...(targetIsGguf + ? { + gpu_ids: compareLoadKnobs.selectedGpuIds ?? undefined, + gpu_memory_mode: compareLoadKnobs.gpuMemoryMode, + } + : {}), }); // Upgrade dialog first (mirrors the primary load path). if (validation.requires_transformers_upgrade) { @@ -1114,7 +1161,7 @@ export function SharedComposer({ const resp = await loadModel({ model_path: sel.id, hf_token: useChatRuntimeStore.getState().hfToken || null, - max_seq_length: maxSeqLength, + max_seq_length: compareMaxSeqLength, load_in_4bit: true, is_lora: sel.isLora, gguf_variant: sel.ggufVariant ?? null, @@ -1123,10 +1170,25 @@ export function SharedComposer({ chat_template_override: effectiveChatTemplateOverride, speculative_type: specSettings.speculativeType, spec_draft_n_max: specSettings.specDraftNMax, - // Honor the Tensor Parallelism toggle on compare loads too. - tensor_parallel: currentStore.tensorParallel, + // Honor the Tensor Parallelism + GPU Memory choices on compare loads. + // GGUF-only, like the auto-load path: the picker is a GGUF control, + // so a non-GGUF target loads via HF auto-placement instead of being + // pinned to a leftover GGUF pick it can't even show. + tensor_parallel: compareLoadKnobs.tensorParallel, + ...(targetIsGguf + ? { + gpu_memory_mode: compareLoadKnobs.gpuMemoryMode, + gpu_layers: compareLoadKnobs.gpuLayers, + n_cpu_moe: compareLoadKnobs.nCpuMoe, + tensor_split: compareLoadKnobs.splitRatio ?? undefined, + gpu_ids: compareLoadKnobs.selectedGpuIds ?? undefined, + } + : {}), }); saveSpeculativeType(specSettings.speculativeType); + // Persist the GPU Memory mode on a non-diffusion GGUF compare-load too, + // so an applied manual choice survives a restart. + persistGpuMemoryModeOnLoad(resp, compareLoadKnobs.gpuMemoryMode); upgradeUnloadedActive = false; const store = useChatRuntimeStore.getState(); store.setCheckpoint( @@ -1136,6 +1198,17 @@ export function SharedComposer({ store.setModelRequiresTrustRemoteCode( resp.requires_trust_remote_code ?? false, ); + // Keep an explicit Manual+Auto context pin the load just applied (so a + // later Apply/Reset doesn't silently revert the model to auto-fit + // sizing), mirroring the interactive path's keepCustomCtx. Non-GGUF + // compare loads don't send the pin, so their baseline clears. + const keepCustomCtx = targetIsGguf + ? resolveManualAutoCtxPin( + compareLoadKnobs.gpuMemoryMode, + compareLoadKnobs.gpuLayers, + compareLoadKnobs.customContextLength, + ) + : null; useChatRuntimeStore.setState({ supportsReasoning: resp.supports_reasoning ?? false, reasoningAlwaysOn: resp.reasoning_always_on ?? false, @@ -1144,6 +1217,32 @@ export function SharedComposer({ supportsTools: resp.supports_tools ?? false, tensorParallel: resp.tensor_parallel ?? false, loadedTensorParallel: resp.tensor_parallel ?? false, + customContextLength: keepCustomCtx, + loadedCustomContextLength: keepCustomCtx, + // Seed the loaded GGUF context (interactive/auto-load parity): the + // settings sheet keys the GGUF GPU controls off it for a direct .gguf + // with no variant, and a later Apply reads it as the resolved context. + ...(targetIsGguf + ? { + ggufContextLength: resp.context_length ?? 131072, + ggufMaxContextLength: + resp.max_context_length ?? resp.context_length ?? 131072, + ggufNativeContextLength: resp.native_context_length ?? null, + } + : { ggufContextLength: null }), + // Compare loads resolve by id (HF repo / local path), never through a + // native-path lease, so a token left by a previously loaded native + // GGUF is stale here -- isLoadedGguf keys off it, and a stale token + // would dress a non-GGUF compare load in GGUF controls. Mirror the + // interactive path, which writes it on every load success. + activeNativePathToken: null, + // Held under an open staged pick: setCheckpoint preserves a stage on + // the empty->active transition, so a compare load can complete with + // staged GPU edits still on screen. + ...loadedGpuMemoryFieldsUnlessStaged(resp), + // Drives the GPU Memory controls' diffusion gate; set alongside the + // GPU fields on every load path so the gate can't read stale. + loadedIsDiffusion: resp.is_diffusion ?? false, loadedIsMultimodal: isMultimodalResponse(resp), ...resolveLoadedSpeculativeSettings(resp), }); diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts index af99458349..5786947118 100644 --- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts +++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts @@ -7,6 +7,10 @@ import { mirrorHfTokenInto, useHfTokenStore, } from "@/features/hub"; +import { + cachedPinnableGpuIndices, + ensureGpuDeviceCache, +} from "@/hooks/use-gpu-info"; import { toast } from "@/lib/toast"; import { create } from "zustand"; import { isExternalModelId, parseExternalModelId } from "../external-providers"; @@ -74,6 +78,7 @@ export const CHAT_RAG_AUTOINJECT_MIN_SCORE_KEY = export const CHAT_RAG_OCR_KEY = "unsloth_chat_rag_ocr_scanned"; export const CHAT_RAG_CAPTION_KEY = "unsloth_chat_rag_caption_figures"; export const CHAT_SPECULATIVE_TYPE_KEY = "unsloth_chat_speculative_type"; +export const CHAT_GPU_MEMORY_MODE_KEY = "unsloth_chat_gpu_memory_mode"; // Persist only the model-agnostic intents (auto/ngram/off). MTP modes // (mtp/mtp+ngram) and spec_draft_n_max stay session-only: a persisted MTP @@ -497,6 +502,213 @@ export function saveSpeculativeType(value: string | null): void { } } +// GPU Memory strategy is a standing preference (like speculative type), not a +// per-model setting: a "manual" choice persists across model switches and reloads. +export function readPersistedGpuMemoryMode(): "auto" | "manual" { + return loadString(CHAT_GPU_MEMORY_MODE_KEY, "auto") === "manual" ? "manual" : "auto"; +} + +export function saveGpuMemoryMode(value: "auto" | "manual"): void { + saveString(CHAT_GPU_MEMORY_MODE_KEY, value); +} + +/** Persist the GPU Memory mode after a load, but only for a non-diffusion GGUF: + * non-GGUF has no such mode, and diffusion runs mode-agnostic (reports "auto"), + * so neither must clobber the standing manual preference. */ +export function persistGpuMemoryModeOnLoad( + resp: { is_gguf?: boolean; is_diffusion?: boolean }, + mode: "auto" | "manual", +): void { + if (resp.is_gguf && !resp.is_diffusion) saveGpuMemoryMode(mode); +} + +// Manual-mode gpu_layers sentinel: -1 = Auto (hand layer + context sizing to +// llama.cpp's --fit). The Manual default; "all on GPU" is the slider's max. +export const GPU_LAYERS_AUTO = -1; + +// Round real-valued shares to integers summing exactly to `total`, giving the +// leftover units to the largest fractional parts (largest-remainder method). +function largestRemainder(shares: number[], total: number): number[] { + const out = shares.map((x) => Math.floor(x)); + let rem = total - out.reduce((a, b) => a + b, 0); + const byFrac = shares + .map((x, i) => ({ i, frac: x - Math.floor(x) })) + .sort((a, b) => b.frac - a.frac); + for (let k = 0; rem > 0 && k < byFrac.length; k++, rem--) out[byFrac[k].i] += 1; + return out; +} + +// Spread `total` layers across GPUs in proportion to `weights` (e.g. per-GPU +// VRAM), as integers summing exactly to `total`; even split for all-zero/empty +// weights. Default per-GPU layer split before the user edits it (mirrors +// llama.cpp's free-VRAM default). +export function distributeByWeight(total: number, weights: number[]): number[] { + if (weights.length === 0) return []; + const t = Math.max(0, Math.floor(total)); + const sum = weights.reduce((a, b) => a + b, 0); + const w = sum > 0 ? weights : weights.map(() => 1); + const wSum = w.reduce((a, b) => a + b, 0); + return largestRemainder( + w.map((x) => (t * x) / wSum), + t, + ); +} + +// Set GPU `index` to `value` and rebalance the rest so per-GPU counts still sum +// to `total`; others absorb the remainder in proportion to their counts (evenly +// if all zero). The --tensor-split editor: counts are sent verbatim, and +// llama.cpp gives each GPU exactly its count when gpu_layers == sum(counts). +export function rebalanceSplit( + total: number, + counts: number[], + index: number, + value: number, +): number[] { + const v = Math.max(0, Math.min(value, total)); + const out = counts.slice(); + const otherIdx = counts.map((_, i) => i).filter((i) => i !== index); + // No other GPU to absorb the remainder: this one holds everything. + if (otherIdx.length === 0) { + out[index] = total; + return out; + } + out[index] = v; + const dist = distributeByWeight( + total - v, + otherIdx.map((i) => counts[i]), + ); + otherIdx.forEach((i, k) => (out[i] = dist[k])); + return out; +} + +// Validate a persisted gpu_ids pick against the GPUs present right now, before +// restoring it from remembered settings. Returns null (= automatic) when the +// pick is stale (none of the saved ids exist, or the host can't pin a multi-GPU +// set), so a saved [1] on a now-1-GPU host doesn't get sent and rejected with no +// way to clear it. A null pick (= automatic) passes through unchanged, and an +// unpopulated device cache leaves the pick alone (the backend still guards). +export function reconcilePersistedGpuIds( + ids: number[] | null, +): number[] | null { + if (ids == null) return ids; + const pinnable = cachedPinnableGpuIndices(); + if (pinnable === null) return ids; // cache not ready: can't validate, keep it + const kept = ids.filter((i) => pinnable.includes(i)); + return kept.length > 0 ? kept : null; +} + +// Store fields derived from a load/status response's GPU-memory settings. +// Shared by every load path so the manual-knob round-trip can't drift. +export function loadedGpuMemoryFields(resp: { + is_gguf?: boolean; + is_diffusion?: boolean; + gpu_memory_mode?: "auto" | "manual"; + gpu_layers?: number; + n_cpu_moe?: number; + tensor_split?: number[] | null; + n_layers?: number | null; + n_moe_layers?: number; + gpu_ids?: number[] | null; +}) { + // GPU-memory state is meaningful only for a GGUF chat load. A non-GGUF response + // still carries gpu_memory_mode (its default "auto" is serialized), so gate on + // the authoritative is_gguf flag, not the field's presence -- otherwise loading + // a transformers model would reset the standing manual preference. + if (!resp.is_gguf) { + // Clear the GPU pick / offload baseline a prior GGUF load may have left, so it + // reflects the non-GGUF model (no pin) -- else a stale loadedGpuIds reads as + // dirty (gpuIdsDirty is ungated) and Reset restores it while the picker is + // hidden. gpuMemoryMode (the standing preference) is kept, but its loaded + // baseline clears to null so Reset preserves the preference, not a stale mode. + return { + selectedGpuIds: null, + loadedGpuIds: null, + loadedGpuMemoryMode: null, + gpuLayers: GPU_LAYERS_AUTO, + loadedGpuLayers: null, + nCpuMoe: 0, + loadedNCpuMoe: null, + splitRatio: null, + loadedSplitRatio: null, + ggufLayerCount: null, + moeLayerCount: null, + }; + } + const mode = resp.gpu_memory_mode ?? "auto"; + const gpuIds = resp.gpu_ids ?? null; + // Layer/MoE/split knobs apply (and are reported) only in manual mode; in auto + // the server ignores them, so don't seed the loaded baseline or the editable + // knobs with values it never applied. In manual, the server reports gpu_layers + // = -1 under Auto, which round-trips the slider back to its Auto position. + const manualKnobs = + mode === "manual" + ? { + loadedGpuLayers: resp.gpu_layers ?? null, + loadedNCpuMoe: resp.n_cpu_moe ?? null, + loadedSplitRatio: resp.tensor_split ?? null, + gpuLayers: resp.gpu_layers ?? GPU_LAYERS_AUTO, + nCpuMoe: resp.n_cpu_moe ?? 0, + splitRatio: resp.tensor_split ?? null, + } + : { + loadedGpuLayers: null, + loadedNCpuMoe: null, + loadedSplitRatio: null, + // Auto ignores these, so reset the editable knobs too (not just the + // loaded baseline) -- else a later switch back to Manual would snapshot + // and send a previous model's stale gpuLayers/nCpuMoe/split that this + // load never applied. Mirrors the non-GGUF branch above. + gpuLayers: GPU_LAYERS_AUTO, + nCpuMoe: 0, + splitRatio: null, + }; + return { + // A diffusion GGUF runs mode-agnostic (pins all layers on one GPU, reports + // "auto"), so adopt everything a chat GGUF does EXCEPT the live standing + // preference -- the next chat load must still honor the user's manual choice. + // The loaded baseline is still "auto", but the UI hides mode controls for a + // loaded diffusion model so it can't read as dirty against the preference. + ...(resp.is_diffusion ? {} : { gpuMemoryMode: mode }), + loadedGpuMemoryMode: mode, + ggufLayerCount: resp.n_layers ?? null, + // MoE expert-layer count: the n_cpu_moe slider max, and 0 hides the slider. + moeLayerCount: resp.n_moe_layers ?? null, + // The picker reflects what loaded (the request sent the user's pick). + selectedGpuIds: gpuIds, + loadedGpuIds: gpuIds, + ...manualKnobs, + }; +} + +/** loadedGpuMemoryFields (plus any seedExtras), unless a staged pick is open. + * + * With a staged pick open (the load fired mid-staging), preserve its editable + * GPU knobs and seedExtras, but still advance every loaded baseline. Otherwise + * cancelling the stage restores its edits onto the newly loaded model. The + * status reseed cannot repair that while pendingSelection holds it off. + */ +export function loadedGpuMemoryFieldsUnlessStaged( + resp: Parameters[0], + seedExtras?: T, +) { + const fields = loadedGpuMemoryFields(resp); + if (useChatRuntimeStore.getState().pendingSelection != null) { + return { + loadedGpuMemoryMode: fields.loadedGpuMemoryMode, + loadedGpuLayers: fields.loadedGpuLayers, + loadedNCpuMoe: fields.loadedNCpuMoe, + loadedSplitRatio: fields.loadedSplitRatio, + loadedGpuIds: fields.loadedGpuIds, + // These are metadata ceilings for the model that actually loaded, not + // editable values from the open stage. Advance them with the baselines + // so abandoning the stage cannot expose the previous model's limits. + ggufLayerCount: fields.ggufLayerCount, + moeLayerCount: fields.moeLayerCount, + }; + } + return { ...fields, ...seedExtras }; +} + /** A local model staged for a deferred load (see `pendingSelection`). Shape is * a subset of the load hook's `SelectedModelInput`, structurally assignable. */ export type PendingModelSelection = { @@ -515,6 +727,13 @@ export type PendingModelSelection = { * Scoped here (not the shared `ggufContextLength`) so a staged model's * metadata never pollutes the currently-loaded model's context display. */ contextLength?: number | null; + /** Total layer count (GGUF block_count); the manual gpu-layers ceiling is + * this + 1 (llama.cpp counts the output layer as offloadable too); + * scoped here like contextLength. */ + layerCount?: number | null; + /** MoE expert-layer count from the GGUF header (manual --n-cpu-moe ceiling); + * 0 for dense models, scoped here like contextLength. */ + moeLayerCount?: number | null; /** "Load on selection" on + un-cached GGUF: download via the manager (global * indicator) without opening the sheet, then load once the download finishes. */ autoLoad?: boolean; @@ -671,7 +890,7 @@ type ChatRuntimeStore = { // Describe figures/charts at ingest time (vision model required). ragCaptionFigures: boolean; /** - * When on, local Studio tool calls pause for an explicit allow/deny in the + * When on, local Unsloth tool calls pause for an explicit allow/deny in the * chat before they run. */ confirmToolCalls: boolean; @@ -743,6 +962,32 @@ type ChatRuntimeStore = { tensorParallel: boolean; /** Backend-reported tensor-parallel state; null until first hydrated. */ loadedTensorParallel: boolean | null; + /** GPU memory strategy for GGUF loads. "auto" = Unsloth picks GPUs and context + * to fit; "manual" = you own the offload (gpuLayers < 0 = Auto/--fit, >= 0 + * pins layers + nCpuMoe). */ + gpuMemoryMode: "auto" | "manual"; + /** Backend-reported gpu memory mode; null until first hydrated. */ + loadedGpuMemoryMode: "auto" | "manual" | null; + /** Manual mode: layers to offload to GPU. -1 = Auto (--fit); >= model layer + * count = all. */ + gpuLayers: number; + loadedGpuLayers: number | null; + /** Manual mode: MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none. */ + nCpuMoe: number; + loadedNCpuMoe: number | null; + /** Manual mode: per-GPU layer counts (--tensor-split), in GPU-in-use order; + * null = unset (llama.cpp splits by free VRAM). */ + splitRatio: number[] | null; + /** Backend-reported per-GPU split ratio (--tensor-split); null = unset. */ + loadedSplitRatio: number[] | null; + /** Model layer count (GGUF block_count); the manual gpu-layers ceiling is + * this + 1 (the output layer is offloadable too). */ + ggufLayerCount: number | null; + /** MoE expert-layer count: the nCpuMoe slider max; 0/null hides the slider. */ + moeLayerCount: number | null; + /** Picked physical GPU indices (null = use all / automatic). */ + selectedGpuIds: number[] | null; + loadedGpuIds: number[] | null; /** Persisted: when false, picking a local model stages it as * `pendingSelection` (and opens settings) instead of loading immediately, * so load settings can be set before the single load. */ @@ -766,6 +1011,9 @@ type ChatRuntimeStore = { * per step, cleared when the run ends, never persisted into the transcript. */ activeDiffusionCanvas: DiffusionCanvasFrame | null; customContextLength: number | null; + /** The pinned context the loaded model used (null = Auto), so dirty-tracking + * and a later fit Apply can tell an explicit pin apart from Auto. */ + loadedCustomContextLength: number | null; defaultChatTemplate: string | null; chatTemplateOverride: string | null; loadedChatTemplateOverride: string | null; @@ -884,6 +1132,11 @@ type ChatRuntimeStore = { * which skip the sheet but must still honor a saved config. */ applyRememberedLoadSettings: (settings: RememberedLoadSettings) => void; setTensorParallel: (value: boolean) => void; + setGpuMemoryMode: (mode: "auto" | "manual") => void; + setGpuLayers: (value: number) => void; + setNCpuMoe: (value: number) => void; + setSplitRatio: (value: number[] | null) => void; + setSelectedGpuIds: (ids: number[] | null) => void; setLoadOnSelection: (value: boolean) => void; setExpandQuantizations: (value: boolean) => void; setShowAllQuantizations: (value: boolean) => void; @@ -1101,11 +1354,12 @@ function setScalarSettingVersion( /** The "revert to the loaded model" baseline for the editable load knobs. * Shared by resetModelSettingsToLoaded (full revert) and stageModel (which - * overrides speculative to start a fresh pick from the standing default). */ + * overrides speculative and the per-model GPU knobs to start a fresh pick). */ function loadedBaselineSettings(s: ChatRuntimeStore) { const hasLoadedModel = Boolean(s.params.checkpoint); return { - customContextLength: null, + // Revert to the loaded model's pin (null = Auto), not a blanket Auto. + customContextLength: s.loadedCustomContextLength, kvCacheDtype: s.loadedKvCacheDtype, tensorParallel: s.loadedTensorParallel ?? false, speculativeType: hasLoadedModel @@ -1113,6 +1367,20 @@ function loadedBaselineSettings(s: ChatRuntimeStore) { : readPersistedSpeculativeType(), specDraftNMax: hasLoadedModel ? s.loadedSpecDraftNMax : null, chatTemplateOverride: s.loadedChatTemplateOverride, + // GPU memory mode is a standing preference; revert to the loaded model's + // mode (or the persisted default when nothing is loaded). Manual knobs and + // the GPU pick are per-model and revert to their loaded baseline. A loaded + // model with no applicable mode -- diffusion ("auto" baseline) or non-GGUF + // (null baseline) -- keeps the live preference so Reset can't drop it. + gpuMemoryMode: !hasLoadedModel + ? readPersistedGpuMemoryMode() + : s.loadedIsDiffusion + ? s.gpuMemoryMode + : (s.loadedGpuMemoryMode ?? s.gpuMemoryMode), + gpuLayers: s.loadedGpuLayers ?? GPU_LAYERS_AUTO, + nCpuMoe: s.loadedNCpuMoe ?? 0, + splitRatio: s.loadedSplitRatio ?? null, + selectedGpuIds: s.loadedGpuIds, }; } @@ -1213,6 +1481,18 @@ export const useChatRuntimeStore = create((set, get) => ({ loadedSpecDraftNMax: null, tensorParallel: false, loadedTensorParallel: null, + gpuMemoryMode: readPersistedGpuMemoryMode(), + loadedGpuMemoryMode: null, + gpuLayers: GPU_LAYERS_AUTO, + loadedGpuLayers: null, + nCpuMoe: 0, + loadedNCpuMoe: null, + splitRatio: null, + loadedSplitRatio: null, + ggufLayerCount: null, + moeLayerCount: null, + selectedGpuIds: null, + loadedGpuIds: null, loadOnSelection: loadBool(CHAT_LOAD_ON_SELECTION_KEY, true), expandQuantizations: loadBool(CHAT_EXPAND_QUANTIZATIONS_KEY, false), showAllQuantizations: loadBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, true), @@ -1221,6 +1501,7 @@ export const useChatRuntimeStore = create((set, get) => ({ loadedIsMultimodal: false, loadedIsDiffusion: false, customContextLength: null, + loadedCustomContextLength: null, defaultChatTemplate: null, chatTemplateOverride: null, loadedChatTemplateOverride: null, @@ -1455,9 +1736,23 @@ export const useChatRuntimeStore = create((set, get) => ({ loadedSpecDraftNMax: null, tensorParallel: false, loadedTensorParallel: null, + // Standing preference: survives unload, unlike the per-model knobs above. + gpuMemoryMode: readPersistedGpuMemoryMode(), + loadedGpuMemoryMode: null, + gpuLayers: GPU_LAYERS_AUTO, + loadedGpuLayers: null, + nCpuMoe: 0, + loadedNCpuMoe: null, + splitRatio: null, + loadedSplitRatio: null, + ggufLayerCount: null, + moeLayerCount: null, + selectedGpuIds: null, + loadedGpuIds: null, loadedIsMultimodal: false, loadedIsDiffusion: false, customContextLength: null, + loadedCustomContextLength: null, defaultChatTemplate: null, chatTemplateOverride: null, loadedChatTemplateOverride: null, @@ -1753,17 +2048,67 @@ export const useChatRuntimeStore = create((set, get) => ({ setSpeculativeType: (speculativeType) => set({ speculativeType }), setSpecDraftNMax: (specDraftNMax) => set({ specDraftNMax }), setTensorParallel: (tensorParallel) => set({ tensorParallel }), + // Standing preference, but persisted only on a successful load (see + // use-chat-model-runtime), not on selection -- so an unapplied pick the user + // resets/abandons doesn't stick to the next session. + setGpuMemoryMode: (gpuMemoryMode) => set({ gpuMemoryMode }), + setGpuLayers: (gpuLayers) => set({ gpuLayers }), + setNCpuMoe: (nCpuMoe) => set({ nCpuMoe }), + setSplitRatio: (splitRatio) => set({ splitRatio }), + setSelectedGpuIds: (selectedGpuIds) => set({ selectedGpuIds }), resetModelSettingsToLoaded: () => set((s) => loadedBaselineSettings(s)), - applyRememberedLoadSettings: (settings) => + applyRememberedLoadSettings: (settings) => { + const gpuCacheWasCold = cachedPinnableGpuIndices() === null; + const restoredGpuIds = + settings.selectedGpuIds !== undefined + ? reconcilePersistedGpuIds(settings.selectedGpuIds) + : undefined; // Coalesce every field: a blob persisted by an older/newer build can omit // keys, and a raw spread would push `undefined` into fields typed non-null. + // The GPU knobs are spread only when present, but first reset the per-model + // ones to defaults: this path (load-on-selection) starts from the loaded + // model's baseline and skips the model-switch reset, so a blob omitting + // gpuLayers/nCpuMoe/selectedGpuIds (older build) or splitRatio (never + // remembered) must not inherit the previous model's placement. gpuMemoryMode + // (standing preference) is NOT reset, only applied when the blob carries it; + // selectedGpuIds keeps a meaningful null (all GPUs), so it keys off undefined. set({ + gpuLayers: GPU_LAYERS_AUTO, + nCpuMoe: 0, + splitRatio: null, + selectedGpuIds: null, customContextLength: settings.contextLength ?? null, kvCacheDtype: settings.kvCacheDtype ?? null, speculativeType: settings.speculativeType ?? "auto", specDraftNMax: settings.specDraftNMax ?? null, tensorParallel: settings.tensorParallel ?? false, - }), + ...(settings.gpuMemoryMode != null && { + gpuMemoryMode: settings.gpuMemoryMode, + }), + ...(settings.gpuLayers != null && { gpuLayers: settings.gpuLayers }), + ...(settings.nCpuMoe != null && { nCpuMoe: settings.nCpuMoe }), + ...(restoredGpuIds !== undefined && { + // Reconcile against the GPUs present now (see reconcilePersistedGpuIds): + // a saved [1] on a 1-GPU host (or under relative/UUID visibility) would + // hide the picker yet still send gpu_ids, which the backend rejects. + selectedGpuIds: restoredGpuIds, + }), + }); + // A cold cache makes the synchronous restore provisional. Reconcile again + // when the shared fetch completes, but only if this exact restored array is + // still current so a user edit, stage change, or load cannot be overwritten. + if (gpuCacheWasCold && restoredGpuIds != null) { + void ensureGpuDeviceCache().then(() => { + set((state) => { + if (state.selectedGpuIds !== restoredGpuIds) return state; + const reconciled = reconcilePersistedGpuIds(restoredGpuIds); + return reconciled === restoredGpuIds + ? state + : { selectedGpuIds: reconciled }; + }); + }); + } + }, setLoadOnSelection: (loadOnSelection) => { saveBool(CHAT_LOAD_ON_SELECTION_KEY, loadOnSelection); set({ loadOnSelection }); @@ -1798,6 +2143,22 @@ export const useChatRuntimeStore = create((set, get) => ({ // Load's keepSpeculative) a forced MTP mode onto a model that may lack it. speculativeType: readPersistedSpeculativeType(), specDraftNMax: null, + // Keep the on-screen GPU Memory selection (loadedBaselineSettings would + // otherwise revert it to the loaded model's mode, dropping a Manual choice + // just made). Use the live store value, not the persisted one, which can + // lag a mode hydrated from an out-of-band load. + gpuMemoryMode: s.gpuMemoryMode, + // Per-model GPU knobs start from defaults too so a fresh pick doesn't + // inherit the loaded model's layer/MoE/split/GPU choices, matching the + // immediate-switch reset. + gpuLayers: GPU_LAYERS_AUTO, + nCpuMoe: 0, + splitRatio: null, + selectedGpuIds: null, + // Fresh pick starts at Auto context (loadedBaselineSettings would + // otherwise restore the current model's pin). Leaves the baseline + // intact, like the GPU knobs, so abandoning restores the loaded pin. + customContextLength: null, }; }); }, @@ -1824,7 +2185,7 @@ export const useChatRuntimeStore = create((set, get) => ({ setContextUsage: (contextUsage) => set({ contextUsage }), })); -// Mirror token edits made through the shared store (e.g. Studio's field). +// Mirror token edits made through the shared store (e.g. Unsloth's field). const unsubscribeHfTokenMirror = mirrorHfTokenInto(useChatRuntimeStore); if (import.meta.hot) { import.meta.hot.dispose(unsubscribeHfTokenMirror); diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts index d72c406fdd..c24ddde5f5 100644 --- a/studio/frontend/src/features/chat/types/api.ts +++ b/studio/frontend/src/features/chat/types/api.ts @@ -65,6 +65,18 @@ export interface LoadModelRequest { * of by layer for GGUF models. Multi-GPU only; no effect on a single GPU. */ tensor_parallel?: boolean | null; + /** GPU memory strategy for GGUF models. "auto" (default): Unsloth selects GPUs + * and caps context to fit VRAM. "manual": you own the offload -- gpu_layers + * -1 (Auto) hands sizing to llama.cpp's --fit, >= 0 pins layers/n_cpu_moe. */ + gpu_memory_mode?: "auto" | "manual"; + /** Manual mode: layers to offload to GPU (--gpu-layers, --fit off); -1 = Auto (--fit). */ + gpu_layers?: number; + /** Manual mode: MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none. */ + n_cpu_moe?: number; + /** Manual mode: relative model share per GPU (--tensor-split), in GPU order. */ + tensor_split?: number[] | null; + /** Picked physical GPU indices (omit/empty = automatic). */ + gpu_ids?: number[]; } export interface ValidateModelResponse { @@ -80,6 +92,13 @@ export interface ValidateModelResponse { requires_security_review?: boolean; /** Native context length from the local GGUF header; null until downloaded. */ context_length?: number | null; + /** Total layer count (GGUF block_count); the manual gpu-layers ceiling is + * this + 1 (llama.cpp counts the output layer as offloadable too); null + * until downloaded. */ + layer_count?: number | null; + /** MoE expert-layer count from the GGUF header (manual --n-cpu-moe ceiling); + * 0 for dense models, null until downloaded. */ + moe_layer_count?: number | null; /** Architecture only shipped by a newer transformers; UI pauses on the upgrade dialog. */ requires_transformers_upgrade?: boolean; /** Set only when requires_transformers_upgrade. */ @@ -159,6 +178,14 @@ export interface LoadModelResponse { spec_draft_n_max?: number | null; /** Whether tensor-parallel split (--split-mode tensor) is active. */ tensor_parallel?: boolean; + gpu_memory_mode?: "auto" | "manual"; + gpu_layers?: number; + n_cpu_moe?: number; + tensor_split?: number[] | null; + n_layers?: number | null; + /** Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not MoE. */ + n_moe_layers?: number; + gpu_ids?: number[] | null; } export interface UnloadModelRequest { @@ -203,6 +230,17 @@ export interface InferenceStatusResponse { spec_draft_n_max?: number | null; /** Whether tensor-parallel split (--split-mode tensor) is active. */ tensor_parallel?: boolean; + gpu_memory_mode?: "auto" | "manual"; + gpu_layers?: number; + n_cpu_moe?: number; + tensor_split?: number[] | null; + /** n_ctx the active GGUF load was invoked with (0 = Auto); re-seeds a + * Manual + Auto-layers context pin on hydration. Null for non-GGUF. */ + requested_context_length?: number | null; + gpu_ids?: number[] | null; + n_layers?: number | null; + /** Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not MoE. */ + n_moe_layers?: number; /** * Why MTP was disabled on the loaded model despite being requested. * "binary_no_mtp" / "binary_outdated" -> updating llama.cpp would re-enable diff --git a/studio/frontend/src/features/chat/utils/chat-history-storage.ts b/studio/frontend/src/features/chat/utils/chat-history-storage.ts index 00df3657b1..2ed17c26a0 100644 --- a/studio/frontend/src/features/chat/utils/chat-history-storage.ts +++ b/studio/frontend/src/features/chat/utils/chat-history-storage.ts @@ -324,7 +324,7 @@ async function importLegacyChatsIfNeeded(): Promise { if (legacyChatImportPromise) return legacyChatImportPromise; legacyChatImportPromise = (async () => { - // Fast-path: no Dexie DB -- new user, never had browser-only Studio. + // Fast-path: no Dexie DB -- new user, never had browser-only Unsloth. if (await dexieDbAbsent()) { markLegacyChatImportDone(); return; diff --git a/studio/frontend/src/features/hub/download-manager/api.ts b/studio/frontend/src/features/hub/download-manager/api.ts index 3373a47b04..2c55b90edd 100644 --- a/studio/frontend/src/features/hub/download-manager/api.ts +++ b/studio/frontend/src/features/hub/download-manager/api.ts @@ -12,7 +12,7 @@ function parseErrorText(status: number, body: unknown): string { const detail = (body as { detail?: unknown }).detail; const formatted = formatFastApiDetail(detail); if (status === 405) { - return `${formatted || "Method Not Allowed"} - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.`; + return `${formatted || "Method Not Allowed"} - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.`; } if (formatted) return formatted; const message = (body as { message?: unknown }).message; diff --git a/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts b/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts new file mode 100644 index 0000000000..f78679310f --- /dev/null +++ b/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts @@ -0,0 +1,46 @@ +// 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 { loadEmbeddingModelSettings } from "@/features/settings"; +import { useEffect, useState } from "react"; +import { useInventoryVersion } from "../stores/inventory-events"; + +/** Backend-resolved embedding repos that optimistic inventory rows must hide. */ +export function useHiddenEmbeddingModelIds( + enabled: boolean, +): ReadonlySet { + const inventoryVersion = useInventoryVersion(); + const [hiddenIds, setHiddenIds] = useState>( + () => new Set(), + ); + + // biome-ignore lint/correctness/useExhaustiveDependencies: inventory invalidation must reload backend-resolved embedder ids + useEffect(() => { + if (!enabled) { + return; + } + let cancelled = false; + loadEmbeddingModelSettings() + .then((settings) => { + if (cancelled) { + return; + } + setHiddenIds( + new Set( + [ + settings.embeddingModel, + settings.embeddingGgufRepo, + settings.defaultEmbeddingModel, + settings.defaultEmbeddingGgufRepo, + ].map((value) => value.trim().toLowerCase()), + ), + ); + }) + .catch(() => undefined); + return () => { + cancelled = true; + }; + }, [enabled, inventoryVersion]); + + return hiddenIds; +} diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx index 630daa48ad..d57f9636fe 100644 --- a/studio/frontend/src/features/hub/hub-page.tsx +++ b/studio/frontend/src/features/hub/hub-page.tsx @@ -63,6 +63,7 @@ import { useDiscoverSearch } from "./hooks/use-discover-search"; import { useFeedWriteBack } from "./hooks/use-feed-write-back"; import { useHubFeed } from "./hooks/use-hub-feed"; import { useHubModelVram } from "./hooks/use-hub-model-vram"; +import { useHiddenEmbeddingModelIds } from "./hooks/use-hidden-embedding-models"; import { useModelsSelection } from "./hooks/use-models-selection"; import { CHANNEL_TO_SECTION, @@ -73,7 +74,10 @@ import { SECTION_TO_CHANNEL, findChannel, } from "./lib/channels"; -import { isHiddenModelId } from "./lib/hidden-models"; +import { + isConfiguredHiddenModelId, + isHiddenModelId, +} from "./lib/hidden-models"; import { inventoryRowMatches, tokenizeQuery } from "./lib/inventory-search"; import { resolveOwnerProviderLogo } from "./lib/provider-logos"; import { @@ -386,6 +390,7 @@ export function ModelsPage() { useState("all"); const isDiscoverTab = tab === "discover"; const isDatasetMode = resourceType === "datasets"; + const hiddenEmbeddingModelIds = useHiddenEmbeddingModelIds(!isDatasetMode); const urlSection = hubSearch.section ?? null; const isModelDiscover = isDiscoverTab && !isDatasetMode; const sectionChannelId: ChannelId | null = urlSection @@ -700,6 +705,7 @@ export function ModelsPage() { return discoverRows.filter( (row) => !isHiddenModelId(row.id) && + !isConfiguredHiddenModelId(hiddenEmbeddingModelIds, row.id) && // The default feed only shows models with a provider logo. (!isFeedMode || resolveOwnerProviderLogo(row.owner, row.repo) !== null) && @@ -714,6 +720,7 @@ export function ModelsPage() { ); }, [ discoverRows, + hiddenEmbeddingModelIds, isDatasetMode, isFeedMode, effectiveDiscoverFormat, @@ -739,7 +746,11 @@ export function ModelsPage() { effectiveCachedRows, effectiveLocalRows, ) - .filter((row) => !isHiddenModelId(row.id)) + .filter( + (row) => + !isHiddenModelId(row.id) && + !isConfiguredHiddenModelId(hiddenEmbeddingModelIds, row.id), + ) .filter((row) => matchesFormat(row.result.isGguf, "gguf")) // Same fit filter as the main Discover list, so the feed carousel // honors the toggle too. @@ -751,6 +762,7 @@ export function ModelsPage() { ), [ hubFeed.trending.results, + hiddenEmbeddingModelIds, modelDiscoveryInventorySignature, fitOnDeviceOnly, gpu, @@ -778,22 +790,29 @@ export function ModelsPage() { () => (isDiscoverTab ? [] : tokenizeQuery(deferredDebouncedQuery)), [isDiscoverTab, deferredDebouncedQuery], ); - // Hide infra models (e.g. the RAG embedder bge-small-en-v1.5) from the On - // Device list like Discover, but reveal a row when a query matches it so the - // user can confirm it is already downloaded. + // Server cache rows already apply variant-aware infra hiding. Optimistic + // rows are not server-confirmed, so apply the client filter first. const isVisibleInventoryRow = useCallback( - (row: CachedInventoryRow | LocalInventoryRow) => - // Local rows can have a null repoId and an id that is a hash rather than - // the file path/name, so also check path/title (the backend's - // _is_hidden_model checks the on-disk path for the same reason). - !isHiddenModelId( - row.id, - row.repoId, - row.kind !== "cache" ? row.path : undefined, - row.kind !== "cache" ? row.title : undefined, - ) || - (inventoryTokens.length > 0 && inventoryRowMatches(row, inventoryTokens)), - [inventoryTokens], + (row: CachedInventoryRow | LocalInventoryRow) => { + if (row.kind === "cache") { + return ( + !row.optimistic || + (!isHiddenModelId(row.id, row.repoId, row.cachePath) && + !isConfiguredHiddenModelId( + hiddenEmbeddingModelIds, + row.id, + row.repoId, + row.cachePath, + )) + ); + } + // Local rows may lack a repo id, so also check path and title. + return ( + !isHiddenModelId(row.id, row.repoId, row.path, row.title) || + (inventoryTokens.length > 0 && inventoryRowMatches(row, inventoryTokens)) + ); + }, + [hiddenEmbeddingModelIds, inventoryTokens], ); // Format filter is a deliberate scope narrowing, so hard-filter it out. The // text query instead drives dim-not-filter on On Device (see ModelsCatalog) so diff --git a/studio/frontend/src/features/hub/index.ts b/studio/frontend/src/features/hub/index.ts index 3515f6ca76..5d4151e87d 100644 --- a/studio/frontend/src/features/hub/index.ts +++ b/studio/frontend/src/features/hub/index.ts @@ -2,6 +2,7 @@ // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 export { cancelStagedModelDownload } from "./download-manager"; +export { bumpInventoryVersion } from "./stores/inventory-events"; export { getHfToken, mirrorHfTokenInto, diff --git a/studio/frontend/src/features/hub/inventory/inventory-hints.ts b/studio/frontend/src/features/hub/inventory/inventory-hints.ts index af9f254ab3..5e202e3150 100644 --- a/studio/frontend/src/features/hub/inventory/inventory-hints.ts +++ b/studio/frontend/src/features/hub/inventory/inventory-hints.ts @@ -12,6 +12,7 @@ export type InventoryHintRow = { repo_id: string; size_bytes: number; partial?: boolean; + optimistic?: boolean; }; export type InventoryHintReconciliation = { @@ -41,6 +42,7 @@ function optimisticRow(hint: InventoryHint): InventoryHintRow { repo_id: hint.repoId, size_bytes: hint.bytes ?? 0, partial: false, + optimistic: true, }; } @@ -101,9 +103,14 @@ function mergeInventoryHint( if (idx === -1) { return [...rows, seed]; } + const serverRow = rows[idx]; const merged = { - ...rows[idx], - ...seed, + ...serverRow, + // A completed hint may arrive before a partial server scan catches up. In + // that case keep the synthetic row non-runnable. A complete server row is + // already authoritative even when its runnable-weight size is smaller than + // the hint's full-snapshot byte count, so do not mark that merge optimistic. + ...(serverRow.partial ? seed : { optimistic: false }), size_bytes: Math.max(rowSizeBytes(rows[idx]), rowSizeBytes(seed)), }; return [...rows.slice(0, idx), merged, ...rows.slice(idx + 1)]; diff --git a/studio/frontend/src/features/hub/inventory/types.ts b/studio/frontend/src/features/hub/inventory/types.ts index c86ffb1d86..6f65a56037 100644 --- a/studio/frontend/src/features/hub/inventory/types.ts +++ b/studio/frontend/src/features/hub/inventory/types.ts @@ -54,6 +54,7 @@ export interface CachedInventoryRow { libraryName?: string | null; quantMethod?: string | null; liveDownload?: boolean; + optimistic?: boolean; } export interface LocalInventoryRow { diff --git a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts index a7dc7ae3f3..fea7b3d331 100644 --- a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts +++ b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts @@ -204,6 +204,7 @@ function liveDownloadInventoryRows( size_bytes: job.displayBytes, partial: true, partial_transport: null, + optimistic: true, }, modelFormat, ), diff --git a/studio/frontend/src/features/hub/inventory/view-models.ts b/studio/frontend/src/features/hub/inventory/view-models.ts index 63d70418be..334050fab4 100644 --- a/studio/frontend/src/features/hub/inventory/view-models.ts +++ b/studio/frontend/src/features/hub/inventory/view-models.ts @@ -176,6 +176,7 @@ export function buildCachedInventoryRow( runtime?: string | null; format_variant?: string | null; capabilities?: BackendModelCapabilities | null; + optimistic?: boolean; }, fallbackFormat: ModelInventoryFormat, ): CachedInventoryRow { @@ -185,6 +186,15 @@ export function buildCachedInventoryRow( const inferredFromEndpoint = rawModelFormat === "unknown" && modelFormat !== "unknown"; const requiresVariant = modelFormat === "gguf"; + const capabilities = normalizeCapabilities( + inferredFromEndpoint ? null : row.capabilities, + modelFormat, + row.partial ?? false, + requiresVariant, + ); + if (row.optimistic) { + capabilities.canChat = false; + } return { kind: "cache", id: @@ -202,12 +212,7 @@ export function buildCachedInventoryRow( modelFormat, ), formatVariant: row.format_variant ?? null, - capabilities: normalizeCapabilities( - inferredFromEndpoint ? null : row.capabilities, - modelFormat, - row.partial ?? false, - requiresVariant, - ), + capabilities, bytes: row.size_bytes, cachePath: row.cache_path ?? null, partial: row.partial ?? false, @@ -216,6 +221,7 @@ export function buildCachedInventoryRow( tags: row.tags, libraryName: row.library_name ?? null, quantMethod: row.quant_method ?? null, + optimistic: row.optimistic, }; } diff --git a/studio/frontend/src/features/hub/lib/hidden-models.ts b/studio/frontend/src/features/hub/lib/hidden-models.ts index 2dbe257947..634a061e0c 100644 --- a/studio/frontend/src/features/hub/lib/hidden-models.ts +++ b/studio/frontend/src/features/hub/lib/hidden-models.ts @@ -1,11 +1,13 @@ // SPDX-License-Identifier: AGPL-3.0-only // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -// Infra models hidden from every browse/preview list (Hub discover and the chat -// model selector). Mirrors the backend `_is_hidden_model`: the RAG embedding -// model and the llama.cpp validation probe are not usable chat models. Per-repo -// file/download views are NOT filtered, so a reinstall still shows the model as -// already downloaded. +// Infra models hidden from browse/preview lists (Hub Discover, the chat model +// selector, and local on-device rows). Mirrors the backend +// `utils.hidden_models`: the RAG embedding model and the llama.cpp validation +// probe are not usable chat models. Server-confirmed cache rows are trusted +// because the backend applies variant-aware filtering. Optimistic cache rows +// still use these needles until the server confirms them. Per-repo views are +// not filtered, so reinstall flows still show downloaded files. const HIDDEN_NEEDLES = [ "bge-small-en-v1.5", // RAG embedder: unsloth/bge-small-en-v1.5[-GGUF] "ggml-org/models", // llama.cpp validation probe repo @@ -17,8 +19,20 @@ export function isHiddenModelId( ...values: (string | null | undefined)[] ): boolean { return values.some((v) => { - if (!v) return false; + if (!v) { + return false; + } const lower = v.toLowerCase(); return HIDDEN_NEEDLES.some((needle) => lower.includes(needle)); }); } + +/** Exact-match configured infra repos without hiding similarly named models. */ +export function isConfiguredHiddenModelId( + configuredIds: ReadonlySet, + ...values: (string | null | undefined)[] +): boolean { + return values.some( + (value) => value != null && configuredIds.has(value.trim().toLowerCase()), + ); +} diff --git a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx index a5ce25dd32..e60786decb 100644 --- a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx +++ b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx @@ -113,7 +113,7 @@ export function ModelSelectionStep() { return applyPriorityOrdering(ids); }, [hfResults]); - // Match Studio: only show exception signals (OOM/TIGHT) in training flows. + // Match Unsloth: only show exception signals (OOM/TIGHT) in training flows. const vramMap = useMemo(() => { const fitMap = buildModelVramMap( hfResults, diff --git a/studio/frontend/src/features/settings/api/embedding-model.ts b/studio/frontend/src/features/settings/api/embedding-model.ts index 9a61142f73..cc21559f38 100644 --- a/studio/frontend/src/features/settings/api/embedding-model.ts +++ b/studio/frontend/src/features/settings/api/embedding-model.ts @@ -2,11 +2,14 @@ // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 import { authFetch } from "@/features/auth"; +import { bumpInventoryVersion } from "@/features/hub"; import { readFastApiError } from "@/lib/format-fastapi-error"; export type EmbeddingModelSettings = { embeddingModel: string; + embeddingGgufRepo: string; defaultEmbeddingModel: string; + defaultEmbeddingGgufRepo: string; isCustom: boolean; }; @@ -14,8 +17,12 @@ type ApiEmbeddingModelSettings = { // biome-ignore lint/style/useNamingConvention: API schema embedding_model: string; // biome-ignore lint/style/useNamingConvention: API schema + embedding_gguf_repo: string; + // biome-ignore lint/style/useNamingConvention: API schema default_embedding_model: string; // biome-ignore lint/style/useNamingConvention: API schema + default_embedding_gguf_repo: string; + // biome-ignore lint/style/useNamingConvention: API schema is_custom: boolean; }; @@ -30,7 +37,9 @@ export class EmbeddingModelBlockedError extends Error {} function fromApi(settings: ApiEmbeddingModelSettings): EmbeddingModelSettings { return { embeddingModel: settings.embedding_model, + embeddingGgufRepo: settings.embedding_gguf_repo, defaultEmbeddingModel: settings.default_embedding_model, + defaultEmbeddingGgufRepo: settings.default_embedding_gguf_repo, isCustom: settings.is_custom, }; } @@ -75,7 +84,9 @@ export async function updateEmbeddingModelSettings( await readFastApiError(res, "Failed to save embedding model"), ); } - return fromApi(await res.json()); + const settings = fromApi(await res.json()); + bumpInventoryVersion(); + return settings; } export async function resetEmbeddingModelSettings(): Promise { @@ -87,5 +98,7 @@ export async function resetEmbeddingModelSettings(): Promise( @@ -45,13 +48,16 @@ export function ModelAutoSwitchSection() { }; }, [t]); - // Parse the idle-seconds draft to a non-negative integer; empty/invalid -> null. + // Parse the idle-seconds draft: 0 (off) or >= MIN_IDLE_SECONDS; else null. const parseIdleSeconds = (): number | null => { if (!draftIdleSeconds.trim()) { return null; } const parsed = Number(draftIdleSeconds); - return Number.isInteger(parsed) && parsed >= 0 ? parsed : null; + if (!Number.isInteger(parsed)) { + return null; + } + return parsed === 0 || parsed >= MIN_IDLE_SECONDS ? parsed : null; }; const persist = async ( diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx index 86125585cc..4ccc43d16c 100644 --- a/studio/frontend/src/features/settings/components/usage-examples.tsx +++ b/studio/frontend/src/features/settings/components/usage-examples.tsx @@ -229,12 +229,12 @@ curl.exe ${base}/v1/chat/completions \` } // A second OpenAI call naming a different downloaded GGUF: with auto-switch on, -// Studio loads it before serving, so the model field selects the served model. +// Unsloth loads it before serving, so the model field selects the served model. function pythonSwitchDemo(): string { return ` # "Switch model by request" is on: replace the model below with another GGUF you -# have downloaded and Studio loads it before serving. Unknown names keep serving +# have downloaded and Unsloth loads it before serving. Unknown names keep serving # the current model. response = client.chat.completions.create( model=${j(SWITCH_MODEL)}, @@ -349,7 +349,7 @@ function javascriptSwitchDemo(): string { return ` // "Switch model by request" is on: replace the model below with another GGUF you -// have downloaded and Studio loads it before serving. Unknown names keep serving +// have downloaded and Unsloth loads it before serving. Unknown names keep serving // the current model. const switchResponse = await client.chat.completions.create({ model: ${j(SWITCH_MODEL)}, @@ -512,7 +512,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) { if (!localAgentDetection) { setDetectedAgents([]); // A previously auto-picked agent was only ever verified against the - // Studio backend's PATH, which is meaningless now that this panel no + // Unsloth backend's PATH, which is meaningless now that this panel no // longer targets a loopback base -- don't leave it selected, but // never touch a choice the user made by hand. if (!agentPickedByUserRef.current) { diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts index 1e4d9d8844..f27100a322 100644 --- a/studio/frontend/src/features/settings/index.ts +++ b/studio/frontend/src/features/settings/index.ts @@ -2,6 +2,7 @@ // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 export { SettingsDialog } from "./settings-dialog"; +export { loadEmbeddingModelSettings } from "./api/embedding-model"; export { loadPersonalization, savePersonalization, diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx index 2f80fc29ca..b6ec06b06a 100644 --- a/studio/frontend/src/features/studio/historical-training-view.tsx +++ b/studio/frontend/src/features/studio/historical-training-view.tsx @@ -8,6 +8,7 @@ import { parseBackendTrainingMethod } from "@/features/training/lib/training-met import { type ReactElement, useEffect, useState } from "react"; import { ChartsSection } from "./sections/charts-section"; import { ProgressSection } from "./sections/progress-section"; +import { mapRunConfigToOverride } from "./sections/run-config-override"; import { translate, useT } from "@/i18n"; type StudioT = ReturnType; @@ -147,25 +148,7 @@ export function HistoricalTrainingView({ } const viewData = mapToViewData(detail, t); - const configOverride = detail.config - ? { - epochs: detail.config.num_epochs as number | undefined, - batchSize: detail.config.batch_size as number | undefined, - learningRate: detail.config.learning_rate as string | undefined, - maxSteps: detail.config.max_steps as number | undefined, - contextLength: detail.config.max_seq_length as number | undefined, - warmupSteps: detail.config.warmup_steps as number | undefined, - optimizerType: detail.config.optim as string | undefined, - loraRank: detail.config.lora_r as number | undefined, - loraAlpha: detail.config.lora_alpha as number | undefined, - loraDropout: detail.config.lora_dropout as number | undefined, - loraVariant: detail.config.use_rslora - ? "rslora" - : detail.config.use_loftq - ? "loftq" - : "lora", - } - : undefined; + const configOverride = mapRunConfigToOverride(detail.config); return (
diff --git a/studio/frontend/src/features/studio/live-training-view.tsx b/studio/frontend/src/features/studio/live-training-view.tsx index cce39adbf4..0aecc7030e 100644 --- a/studio/frontend/src/features/studio/live-training-view.tsx +++ b/studio/frontend/src/features/studio/live-training-view.tsx @@ -1,18 +1,42 @@ // 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 { cn } from "@/lib/utils"; import { + getTrainingRun, useTrainingConfigStore, useTrainingRuntimeStore, } from "@/features/training"; import type { TrainingViewData } from "@/features/training"; +import { cn } from "@/lib/utils"; import type { ReactElement } from "react"; +import { useEffect, useState } from "react"; import { useShallow } from "zustand/react/shallow"; import { ChartsSection } from "./sections/charts-section"; import { ProgressSection } from "./sections/progress-section"; +import { + type RunConfigOverride, + mapRunConfigToOverride, +} from "./sections/run-config-override"; import { TrainingStartOverlay } from "./training-start-overlay"; +/** Retry budget for the run-config lookup. The row is inserted at + * start_training(), but a lookup issued in the same instant can still miss it; + * a few short retries cover that without polling a genuinely absent row. */ +const RUN_CONFIG_FETCH_RETRIES = 5; +const RUN_CONFIG_FETCH_RETRY_MS = 1000; + +/** The fetched run config only applies while it belongs to the active job; + * a stale record from a previous run falls back to the form store. */ +function activeRunOverride( + fetched: { jobId: string; override: RunConfigOverride | undefined } | null, + jobId: string | null, +): RunConfigOverride | undefined { + if (fetched === null || fetched.jobId !== jobId) { + return undefined; + } + return fetched.override; +} + export function LiveTrainingView(): ReactElement { const runtime = useTrainingRuntimeStore( useShallow((state) => ({ @@ -52,6 +76,59 @@ export function LiveTrainingView(): ReactElement { })), ); + // Show the ACTIVE run's saved config, not the editable form store the user may + // have changed since starting (#6853). start_training() commits the run row + // before the pump, so the job id alone gates the fetch; the bounded retry below + // covers the narrow uncommitted window, and until it loads ProgressSection falls + // back to the form store. The result is keyed by job id and filtered at render. + const [fetchedRunConfig, setFetchedRunConfig] = useState<{ + jobId: string; + override: RunConfigOverride | undefined; + } | null>(null); + // Retry budget for the transient 404 below, keyed by job so a new run always + // starts with a fresh budget. + const [fetchAttempt, setFetchAttempt] = useState<{ + jobId: string; + count: number; + } | null>(null); + useEffect(() => { + if (!runtime.jobId) { + return; + } + const jobId = runtime.jobId; + if (fetchedRunConfig !== null && fetchedRunConfig.jobId === jobId) { + return; // already resolved for this job + } + const attempts = fetchAttempt?.jobId === jobId ? fetchAttempt.count : 0; + const controller = new AbortController(); + let retryTimer: ReturnType | undefined; + getTrainingRun(jobId, controller.signal) + .then((detail) => { + setFetchedRunConfig({ + jobId, + override: mapRunConfigToOverride(detail.config), + }); + }) + .catch(() => { + // A lookup racing the row commit can miss transiently; nothing else in + // the deps changes on failure, so retry explicitly. Bounded so a genuinely + // absent row falls back to the form store instead of polling forever. + if (controller.signal.aborted || attempts >= RUN_CONFIG_FETCH_RETRIES) { + return; + } + retryTimer = setTimeout(() => { + setFetchAttempt({ jobId, count: attempts + 1 }); + }, RUN_CONFIG_FETCH_RETRY_MS); + }); + return () => { + controller.abort(); + if (retryTimer !== undefined) { + clearTimeout(retryTimer); + } + }; + }, [runtime.jobId, fetchedRunConfig, fetchAttempt]); + const runConfigOverride = activeRunOverride(fetchedRunConfig, runtime.jobId); + const activeProjectName = runtime.startProjectName !== null ? runtime.startProjectName.trim() || null @@ -76,7 +153,11 @@ export function LiveTrainingView(): ReactElement { isTrainingRunning: runtime.isTrainingRunning, modelName: runtime.startModelName ?? config.selectedModel ?? "", projectName: activeProjectName, - trainingMethod: config.trainingMethod ?? "", + // Prefer the saved run's method: the form may have been edited (e.g. LoRA + // -> Full) after the run started, which would relabel the run and hide its + // saved LoRA rows in the popover. + trainingMethod: + runConfigOverride?.trainingMethod ?? config.trainingMethod ?? "", lossHistory: runtime.lossHistory, lrHistory: runtime.lrHistory, gradNormHistory: runtime.gradNormHistory, @@ -105,7 +186,11 @@ export function LiveTrainingView(): ReactElement { )} >
- +
{ // Hub-style sliding-pill segmented control, matching the Hub tabs // via the shared .hub-tab-toggle / .hub-tab-toggle-pill classes. + // flex-auto buttons share leftover space equally so padding stays + // equal for all labels; the pill sits inside the active button so + // it always matches its bounds. const sourceTabs: { value: "huggingface" | "upload" | "s3"; label: string; @@ -696,24 +704,12 @@ export function DatasetSection() { ? [] : [{ value: "s3" as const, label: "Amazon S3" }]), ]; - const activeIndex = Math.max( - 0, - sourceTabs.findIndex((item) => item.value === datasetSource), - ); return (
-
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx index 1aaa3b6889..37bec84a48 100644 --- a/studio/frontend/src/features/studio/sections/progress-section.tsx +++ b/studio/frontend/src/features/studio/sections/progress-section.tsx @@ -28,6 +28,7 @@ import { } from "@/features/training"; import { getTrainingMethodLabel } from "@/features/training/lib/training-methods"; import type { TrainingViewData } from "@/features/training"; +import type { RunConfigOverride } from "./run-config-override"; import { useGpuUtilization } from "@/hooks"; import type { GpuUtilization } from "@/hooks/use-gpu-utilization"; import { cn } from "@/lib/utils"; @@ -81,19 +82,7 @@ function configRow( interface ProgressSectionProps { data: TrainingViewData; isHistorical?: boolean; - configOverride?: { - epochs?: number; - batchSize?: number; - learningRate?: string; - maxSteps?: number; - contextLength?: number; - warmupSteps?: number; - optimizerType?: string; - loraRank?: number; - loraAlpha?: number; - loraDropout?: number; - loraVariant?: string; - }; + configOverride?: RunConfigOverride; } export function ProgressSection({ @@ -183,17 +172,20 @@ export function ProgressSection({ ? data.currentGradNorm : (lastValue(data.gradNormHistory) ?? data.currentGradNorm); - const cfgEpochs = isHistorical ? configOverride?.epochs : config.epochs; - const cfgBatchSize = isHistorical ? configOverride?.batchSize : config.batchSize; - const cfgLearningRate = isHistorical ? configOverride?.learningRate : config.learningRate; - const cfgMaxSteps = isHistorical ? configOverride?.maxSteps : config.maxSteps; - const cfgContextLength = isHistorical ? configOverride?.contextLength : config.contextLength; - const cfgWarmupSteps = isHistorical ? configOverride?.warmupSteps : config.warmupSteps; - const cfgOptimizerType = isHistorical ? configOverride?.optimizerType : config.optimizerType; - const cfgLoraRank = isHistorical ? configOverride?.loraRank : config.loraRank; - const cfgLoraAlpha = isHistorical ? configOverride?.loraAlpha : config.loraAlpha; - const cfgLoraDropout = isHistorical ? configOverride?.loraDropout : config.loraDropout; - const cfgLoraVariant = isHistorical ? configOverride?.loraVariant : config.loraVariant; + // Prefer the run's saved snapshot when present (#6853). Live falls back to the + // editable form store until it loads; History shows blanks, never live form values. + const cfg = configOverride ?? (isHistorical ? undefined : config); + const cfgEpochs = cfg?.epochs; + const cfgBatchSize = cfg?.batchSize; + const cfgLearningRate = cfg?.learningRate; + const cfgMaxSteps = cfg?.maxSteps; + const cfgContextLength = cfg?.contextLength; + const cfgWarmupSteps = cfg?.warmupSteps; + const cfgOptimizerType = cfg?.optimizerType; + const cfgLoraRank = cfg?.loraRank; + const cfgLoraAlpha = cfg?.loraAlpha; + const cfgLoraDropout = cfg?.loraDropout; + const cfgLoraVariant = cfg?.loraVariant; const optimizerLabel = OPTIMIZER_OPTIONS.find((o) => o.value === cfgOptimizerType)?.label ?? diff --git a/studio/frontend/src/features/studio/sections/run-config-override.ts b/studio/frontend/src/features/studio/sections/run-config-override.ts new file mode 100644 index 0000000000..a1272bfeb0 --- /dev/null +++ b/studio/frontend/src/features/studio/sections/run-config-override.ts @@ -0,0 +1,54 @@ +// SPDX-License-Identifier: AGPL-3.0-only +// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 + +import { parseBackendTrainingMethod } from "@/features/training"; + +/** Shape of the Training Config popover's data when it is driven by a saved + * run snapshot instead of the editable form store. */ +export interface RunConfigOverride { + trainingMethod?: string; + epochs?: number; + batchSize?: number; + learningRate?: string; + maxSteps?: number; + contextLength?: number; + warmupSteps?: number; + optimizerType?: string; + loraRank?: number; + loraAlpha?: number; + loraDropout?: number; + loraVariant?: string; +} + +/** Map a saved run's config (GET /api/train/runs/{id} `detail.config`) into the + * Training Config popover's override shape. Shared by the History view and the + * live Current Run view so both read the same authoritative run snapshot + * instead of the editable form store (#6853). */ +export function mapRunConfigToOverride( + config: Record | null | undefined, +): RunConfigOverride | undefined { + if (!config) { + return undefined; + } + return { + trainingMethod: parseBackendTrainingMethod( + config.training_type, + config.load_in_4bit, + ), + epochs: config.num_epochs as number | undefined, + batchSize: config.batch_size as number | undefined, + learningRate: config.learning_rate as string | undefined, + maxSteps: config.max_steps as number | undefined, + contextLength: config.max_seq_length as number | undefined, + warmupSteps: config.warmup_steps as number | undefined, + optimizerType: config.optim as string | undefined, + loraRank: config.lora_r as number | undefined, + loraAlpha: config.lora_alpha as number | undefined, + loraDropout: config.lora_dropout as number | undefined, + loraVariant: config.use_rslora + ? "rslora" + : config.use_loftq + ? "loftq" + : "lora", + }; +} diff --git a/studio/frontend/src/features/training/index.ts b/studio/frontend/src/features/training/index.ts index 553dcc2af5..a0d249ff1b 100644 --- a/studio/frontend/src/features/training/index.ts +++ b/studio/frontend/src/features/training/index.ts @@ -12,6 +12,7 @@ export { getTrainingRunDisplayTitle, getTrainingRunModelSubtitle, } from "./lib/run-display"; +export { parseBackendTrainingMethod } from "./lib/training-methods"; 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/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts index a927f83fd8..d204e24f47 100644 --- a/studio/frontend/src/features/training/stores/training-config-store.ts +++ b/studio/frontend/src/features/training/stores/training-config-store.ts @@ -1004,7 +1004,7 @@ export const useTrainingConfigStore = create()( } if (version < 12) { // hfToken moved to the shared hf-token-store; seed it once so an - // existing Studio-only token isn't lost. + // existing Unsloth-only token isn't lost. const legacyToken = typeof s.hfToken === "string" ? s.hfToken.trim() : ""; if (legacyToken && !getHfToken()) { useHfTokenStore.getState().setToken(legacyToken); diff --git a/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx b/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx index 98c590e5e4..3e39a8ff1c 100644 --- a/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx +++ b/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx @@ -82,7 +82,7 @@ export function TransformersUpgradeDialog() { <> Even the latest transformers release on PyPI does not support it yet: the architecture is only available on the - transformers development branch (main), and Studio does + transformers development branch (main), and Unsloth does not install development builds. Support arrives with the next transformers release on PyPI. diff --git a/studio/frontend/src/hooks/use-gpu-info.ts b/studio/frontend/src/hooks/use-gpu-info.ts index 1e313acdf3..db2cc021be 100644 --- a/studio/frontend/src/hooks/use-gpu-info.ts +++ b/studio/frontend/src/hooks/use-gpu-info.ts @@ -15,6 +15,19 @@ export interface GpuInfo { systemRamTotalGb: number } +export interface SystemGpuDevice { + index: number; + name: string; + memoryTotalGb: number; + /** Free VRAM at fetch time. Degrades to the total when the utilization + * probe had no usage data; 0 only when the total is unknown too. */ + memoryFreeGb: number; + /** "physical" = `index` is a stable physical/PCI id safe to pin via gpu_ids; + * "relative" = an ordinal into a parent CUDA_VISIBLE_DEVICES mask, which the + * backend can't map back, so the picker must not offer it. */ + physicalIndex: boolean; +} + const DEFAULT_GPU: GpuInfo = { available: false, name: "Unknown", @@ -25,70 +38,135 @@ const DEFAULT_GPU: GpuInfo = { systemRamTotalGb: 0 }; -// Module-level cache so multiple components share one fetch. -let cachedGpu: GpuInfo | null = null; -let fetchPromise: Promise | null = null; +// One module-level cache so every GPU hook shares a single /api/system fetch. +let cachedSystem: SystemInfoResponse | null = null; +let systemPromise: Promise | null = null; -async function fetchGpuOnce(): Promise { - if (cachedGpu) return cachedGpu; - if (fetchPromise) return fetchPromise; - - fetchPromise = (async () => { +async function fetchSystemOnce(): Promise { + if (cachedSystem) return cachedSystem; + if (systemPromise) return systemPromise; + systemPromise = (async () => { try { const res = await authFetch("/api/system"); if (!res.ok) throw new Error(`HTTP ${res.status}`); - - const data = await res.json() as SystemInfoResponse; - const gpuData = data?.gpu; - - // CPU/RAM exist even on hosts without a GPU, so populate them on every path. - // No discrete GPU (e.g. Mac): still surface system RAM so memory math - // (unified memory) has a budget to work with. - const base = { - cpuCore: data?.cpu?.physical_count ?? 0, - cpuThread: data?.cpu?.logical_count ?? 0, - systemRamAvailableGb: data?.memory?.available_gb ?? 0, - systemRamTotalGb: data?.memory?.total_gb ?? 0, - }; - - const devices = gpuData?.devices ?? []; - const info: GpuInfo = - gpuData?.available && devices.length - ? { - ...base, - available: true, - name: devices[0]?.name ?? "Unknown", - memoryTotalGb: devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0), - } - : { ...DEFAULT_GPU, ...base }; - cachedGpu = info; - return info; + cachedSystem = (await res.json()) as SystemInfoResponse; + return cachedSystem; } catch { - // Reset promise so subsequent calls retry (e.g. backend wasn't ready) - fetchPromise = null; - return DEFAULT_GPU; + systemPromise = null; // reset so a later call retries (backend not ready) + return null; } })(); + return systemPromise; +} - return fetchPromise; +function toGpuInfo(data: SystemInfoResponse | null): GpuInfo { + // CPU/RAM exist even on GPU-less hosts (e.g. Mac), so populate them on every + // path: unified-memory math still needs a RAM budget to work with. + const base = { + cpuCore: data?.cpu?.physical_count ?? 0, + cpuThread: data?.cpu?.logical_count ?? 0, + systemRamAvailableGb: data?.memory?.available_gb ?? 0, + systemRamTotalGb: data?.memory?.total_gb ?? 0, + }; + const gpuData = data?.gpu; + const devices = gpuData?.devices ?? []; + if (!gpuData?.available || !devices.length) { + return { ...DEFAULT_GPU, ...base }; + } + return { + ...base, + available: true, + name: devices[0]?.name ?? "Unknown", + memoryTotalGb: devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0), + }; +} + +function toGpuDevices(data: SystemInfoResponse | null): SystemGpuDevice[] { + // Unpinnable configurations must hide every pick surface: XPU indices are + // torch-xpu ordinals no applicator speaks, and Vulkan-only builds pin ggml's + // own ordinals -- /load and /validate 400 picks on both, so the backend + // reports gpu.gguf_gpu_ids_supported and every gate keyed on physicalIndex + // (picker, persisted-pick reconcile) follows it. The device flavor lives on + // the TOP-LEVEL device_backend field; absent support info defaults to + // pinnable (older backend). + const pinnableBackend = + data?.device_backend !== "xpu" && + data?.gpu?.gguf_gpu_ids_supported !== false; + return (data?.gpu?.devices ?? []) + .filter((d) => typeof d.index === "number") + .map((d) => ({ + index: d.index as number, + name: d.name ?? `GPU ${d.index}`, + memoryTotalGb: d.memory_total_gb ?? 0, + memoryFreeGb: d.vram_free_gb ?? 0, + physicalIndex: pinnableBackend && d.index_kind === "physical", + })); +} + +/** Aggregate GPU info from /api/system; shares one module-level fetch across all GPU hooks. */ +export function useGpuInfo(): GpuInfo { + const [gpu, setGpu] = useState( + cachedSystem ? toGpuInfo(cachedSystem) : DEFAULT_GPU, + ); + useEffect(() => { + // No early return on cachedSystem: a consumer mounting as the cache fills + // (between render and effect) would otherwise stay stuck at the default. + let cancelled = false; + fetchSystemOnce().then((d) => { + if (!cancelled) setGpu(toGpuInfo(d)); + }); + return () => { + cancelled = true; + }; + }, []); + return gpu; +} + +/** All backend-visible GPUs (index, name, total VRAM); shares the same fetch. */ +export function useGpuDevices(): SystemGpuDevice[] { + const [devices, setDevices] = useState( + cachedSystem ? toGpuDevices(cachedSystem) : [], + ); + useEffect(() => { + // No early return on cachedSystem: a consumer mounting as the cache fills + // (between render and effect) would otherwise stay stuck at the default. + let cancelled = false; + fetchSystemOnce().then((d) => { + if (!cancelled) setDevices(toGpuDevices(d)); + }); + return () => { + cancelled = true; + }; + }, []); + return devices; } /** - * Fetch GPU info from /api/system. Cached at module level, so only one request - * is made no matter how many components call this hook. + * Await the shared /api/system fetch so cachedPinnableGpuIndices (and the + * store's reconcilePersistedGpuIds) can validate a persisted pick before a + * load path sends it -- on a cold cache the reconcile passes ids through + * unvalidated, and a stale cross-host pick then fails /load with the picker + * hidden. Resolves immediately once the module cache is warm; a failed fetch + * keeps the cache cold, preserving the "can't validate, backend guards" + * degradation. */ -export function useGpuInfo(): GpuInfo { - const [gpu, setGpu] = useState(cachedGpu ?? DEFAULT_GPU); +export async function ensureGpuDeviceCache(): Promise { + await fetchSystemOnce(); +} - useEffect(() => { - if (cachedGpu) return; - - let cancelled = false; - fetchGpuOnce().then((info) => { - if (!cancelled) setGpu(info); - }); - return () => { cancelled = true; }; - }, []); - - return gpu; -} \ No newline at end of file +/** + * Pinnable physical GPU indices from the already-fetched /api/system cache, for + * non-React code (the store) that needs to validate a persisted `gpu_ids` pick + * without triggering a fetch. Returns: + * - `null` when the cache isn't populated yet (caller can't validate, so keep + * the pick and let the backend guard reject a truly bad one); + * - `[]` when the host has no pinnable multi-GPU set (single GPU, or relative/ + * UUID-masked indices) -- the picker is hidden, so any saved pick is stale; + * - the physical indices otherwise. + */ +export function cachedPinnableGpuIndices(): number[] | null { + if (!cachedSystem) return null; + const physical = toGpuDevices(cachedSystem).filter((d) => d.physicalIndex); + // Mirrors the sheet's showGpuPicker gate: only a 2+ physical-GPU host can pin. + return physical.length > 1 ? physical.map((d) => d.index) : []; +} diff --git a/studio/frontend/src/hooks/use-system.ts b/studio/frontend/src/hooks/use-system.ts index a135cce86e..8cfe2bace4 100644 --- a/studio/frontend/src/hooks/use-system.ts +++ b/studio/frontend/src/hooks/use-system.ts @@ -40,6 +40,9 @@ export interface SystemInfoResponse { gpu: { available: boolean; backend?: string; + /** Whether GGUF loads accept an explicit gpu_ids pick (false on XPU hosts + * and Vulkan-only builds, where /load and /validate 400 picks). */ + gguf_gpu_ids_supported?: boolean; backend_cuda_visible_devices?: string | null; parent_visible_gpu_ids?: number[]; index_kind?: string; diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts index db996768a6..53122864e7 100644 --- a/studio/frontend/src/hooks/use-tauri-backend.ts +++ b/studio/frontend/src/hooks/use-tauri-backend.ts @@ -71,8 +71,8 @@ function externalConflictMessage(preflight: DesktopPreflightResult) { } return preflight.port - ? `A Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.` - : "A Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app."; + ? `An Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.` + : "An 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 waitForManagedServerPort( diff --git a/studio/frontend/src/i18n/README.md b/studio/frontend/src/i18n/README.md index ba8cce5d8d..7594e4723f 100644 --- a/studio/frontend/src/i18n/README.md +++ b/studio/frontend/src/i18n/README.md @@ -8,5 +8,5 @@ - Preserve interpolation variables exactly, for example `{count}`, `{model}`, and `{provider}`. - Keep product and technical names unchanged unless there is an established localized name, for example `Unsloth Studio`, `LoRA`, `GGUF`, and `Hugging Face`. - Keep translation changes small and reviewable. Prefer separate commits for runtime changes, UI migration, and locale text. -- When adding user-facing Studio UI text, add the English message key first and add non-English overrides only when the translation is clear. +- When adding user-facing Unsloth UI text, add the English message key first and add non-English overrides only when the translation is clear. - Run `npx tsx src/i18n/check-parity.ts` before committing to ensure there are no shape mismatches or placeholder discrepancies in the non-English overlays. diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts index 28f404a384..4bd4328cce 100644 --- a/studio/frontend/src/i18n/locales/ar.ts +++ b/studio/frontend/src/i18n/locales/ar.ts @@ -113,7 +113,7 @@ export const ar = { showToken: "إظهار التوكن", tokenSaved: "تم حفظ التوكن", password: "كلمة المرور", - passwordDescription: "تغيير كلمة المرور لحساب Studio هذا.", + passwordDescription: "تغيير كلمة المرور لحساب Unsloth هذا.", passwordDialog: { trigger: "تغيير كلمة المرور", title: "تغيير كلمة المرور", @@ -155,14 +155,14 @@ export const ar = { "عندما يسمّي طلب متوافق مع OpenAI ملف GGUF مُنزّلاً مختلفًا، يتم تحميله قبل الخدمة. مُعطّل افتراضيًا؛ الأسماء غير المعروفة تُبقي على النموذج المُحمَّل.", idleUnload: "الإلغاء التلقائي عند الخمول", idleUnloadDescription: - "إلغاء تحميل النموذج بعد هذا العدد من ثواني الخمول لتحرير الـ VRAM؛ الطلب التالي يعيد تحميله. القيمة 0 تُبقيه محمَّلاً.", + "إلغاء تحميل النموذج بعد هذا العدد من ثواني الخمول لتحرير الـ VRAM؛ الطلب التالي يعيد تحميله. القيمة 0 تُبقيه محمَّلاً. الحد الأدنى 60 ثانية.", idleNeedsEnable: "فعّل تبديل النموذج حسب الطلب حتى يعاد تحميل النموذج غير المحمَّل عند الاستخدام التالي.", idleActiveViaEnv: "الإلغاء التلقائي عند الخمول مُفعَّل عبر متغير البيئة UNSLOTH_MODEL_IDLE_TTL.", loadError: "فشل تحميل إعدادات التبديل التلقائي للنموذج.", saveError: "فشل حفظ إعدادات التبديل التلقائي للنموذج.", - idleError: "أدخل عددًا صحيحًا من الثواني (0 أو أكثر).", + idleError: "أدخل 0 لإبقاء النموذج محمَّلاً، أو 60 ثانية على الأقل.", }, previewSharing: { sectionTitle: "مشاركة المعاينة", @@ -284,7 +284,7 @@ export const ar = { }, resources: { title: "النظام", - description: "مراقبة أجهزة خادم Studio هذا وتخزينه.", + description: "مراقبة أجهزة خادم Unsloth هذا وتخزينه.", liveUpdates: "التحديثات المباشرة", floatingWindow: "نافذة عائمة", disableOverlay: "تعطيل التراكب", diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts index e38cdbfa0e..c28d07790f 100644 --- a/studio/frontend/src/i18n/locales/de.ts +++ b/studio/frontend/src/i18n/locales/de.ts @@ -115,7 +115,7 @@ export const de = { tokenSaved: "Token gespeichert", password: "Passwort", passwordDescription: - "Ändern Sie das Passwort für dieses Studio-Konto.", + "Ändern Sie das Passwort für dieses Unsloth-Konto.", passwordDialog: { trigger: "Passwort ändern", title: "Passwort ändern", @@ -158,7 +158,7 @@ export const de = { "Wenn eine OpenAI-kompatible Anfrage ein anderes heruntergeladenes GGUF nennt, wird dieses vor der Auslieferung geladen. Standardmäßig aus; unbekannte Namen liefern weiterhin das geladene Modell aus.", idleUnload: "Automatisches Entladen bei Inaktivität", idleUnloadDescription: - "Entlädt das Modell nach dieser Anzahl inaktiver Sekunden, um VRAM freizugeben; die nächste Anfrage lädt es erneut. 0 hält es geladen.", + "Entlädt das Modell nach dieser Anzahl inaktiver Sekunden, um VRAM freizugeben; die nächste Anfrage lädt es erneut. 0 hält es geladen. Minimum 60 Sekunden.", idleNeedsEnable: "Aktivieren Sie \"Modell je Anfrage wechseln\", damit ein entladenes Modell bei der nächsten Nutzung erneut geladen wird.", idleActiveViaEnv: @@ -167,7 +167,7 @@ export const de = { "Einstellungen für automatischen Modellwechsel konnten nicht geladen werden.", saveError: "Einstellungen für automatischen Modellwechsel konnten nicht gespeichert werden.", - idleError: "Geben Sie eine ganze Anzahl an Sekunden ein (0 oder mehr).", + idleError: "Geben Sie 0 ein, um das Modell geladen zu halten, oder mindestens 60 Sekunden.", }, previewSharing: { sectionTitle: "Vorschau-Freigabe", @@ -295,7 +295,7 @@ export const de = { resources: { title: "System", description: - "Überwachen Sie Hardware und Speicher dieses Studio-Servers.", + "Überwachen Sie Hardware und Speicher dieses Unsloth-Servers.", liveUpdates: "Live-Updates", floatingWindow: "Schwebendes Fenster", disableOverlay: "Overlay deaktivieren", diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts index fe3a6f8542..de8ac17c29 100644 --- a/studio/frontend/src/i18n/locales/en.ts +++ b/studio/frontend/src/i18n/locales/en.ts @@ -224,14 +224,14 @@ export const en = { "When an OpenAI-compatible request names a different downloaded GGUF, load it before serving. Off by default; unknown names keep serving the loaded model.", idleUnload: "Idle auto-unload", idleUnloadDescription: - "Unload the model after this many idle seconds to free VRAM; the next request reloads it. 0 keeps it loaded.", + "Unload the model after this many idle seconds to free VRAM; the next request reloads it. 0 keeps it loaded. Minimum 60 seconds.", idleNeedsEnable: "Turn on Switch model by request so an unloaded model reloads on next use.", idleActiveViaEnv: "Idle auto-unload is active via the UNSLOTH_MODEL_IDLE_TTL environment variable.", loadError: "Failed to load model auto-switch settings.", saveError: "Failed to save model auto-switch settings.", - idleError: "Enter a whole number of seconds (0 or more).", + idleError: "Enter 0 to keep the model loaded, or at least 60 seconds.", }, previewSharing: { sectionTitle: "Preview sharing", diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts index e5f9650bef..b7dfee10b8 100644 --- a/studio/frontend/src/i18n/locales/es.ts +++ b/studio/frontend/src/i18n/locales/es.ts @@ -115,7 +115,7 @@ export const es = { tokenSaved: "Token guardado", password: "Contraseña", passwordDescription: - "Cambia la contraseña de esta cuenta de Studio.", + "Cambia la contraseña de esta cuenta de Unsloth.", passwordDialog: { trigger: "Cambiar contraseña", title: "Cambiar contraseña", @@ -157,7 +157,7 @@ export const es = { "Cuando una solicitud compatible con OpenAI nombra un GGUF descargado distinto, se carga antes de responder. Desactivado por defecto; los nombres desconocidos siguen usando el modelo cargado.", idleUnload: "Descarga automática por inactividad", idleUnloadDescription: - "Descarga el modelo tras este número de segundos inactivo para liberar VRAM; la siguiente solicitud lo recarga. 0 lo mantiene cargado.", + "Descarga el modelo tras este número de segundos inactivo para liberar VRAM; la siguiente solicitud lo recarga. 0 lo mantiene cargado. Mínimo 60 segundos.", idleNeedsEnable: "Activa Cambiar de modelo según la solicitud para que un modelo descargado se recargue en el próximo uso.", idleActiveViaEnv: @@ -166,7 +166,7 @@ export const es = { "No se pudo cargar la configuración de cambio automático de modelo.", saveError: "No se pudo guardar la configuración de cambio automático de modelo.", - idleError: "Introduce un número entero de segundos (0 o más).", + idleError: "Introduce 0 para mantener el modelo cargado, o al menos 60 segundos.", }, previewSharing: { sectionTitle: "Compartir vista previa", @@ -294,7 +294,7 @@ export const es = { resources: { title: "Sistema", description: - "Monitorea el hardware y el almacenamiento de este servidor de Studio.", + "Monitorea el hardware y el almacenamiento de este servidor de Unsloth.", liveUpdates: "Actualizaciones en vivo", floatingWindow: "Ventana flotante", disableOverlay: "Desactivar superposición", diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts index 190284175d..6105cd8ccf 100644 --- a/studio/frontend/src/i18n/locales/fr.ts +++ b/studio/frontend/src/i18n/locales/fr.ts @@ -115,7 +115,7 @@ export const fr = { tokenSaved: "Token enregistré", password: "Mot de passe", passwordDescription: - "Changez le mot de passe de ce compte Studio.", + "Changez le mot de passe de ce compte Unsloth.", passwordDialog: { trigger: "Changer le mot de passe", title: "Changer le mot de passe", @@ -157,7 +157,7 @@ export const fr = { "Lorsqu'une requête compatible OpenAI nomme un autre GGUF téléchargé, le charger avant de répondre. Désactivé par défaut ; les noms inconnus continuent de servir le modèle chargé.", idleUnload: "Déchargement automatique en cas d'inactivité", idleUnloadDescription: - "Décharger le modèle après ce nombre de secondes d'inactivité pour libérer la VRAM ; la requête suivante le recharge. 0 le maintient chargé.", + "Décharger le modèle après ce nombre de secondes d'inactivité pour libérer la VRAM ; la requête suivante le recharge. 0 le maintient chargé. Minimum 60 secondes.", idleNeedsEnable: "Activez Changer de modèle par requête pour qu'un modèle déchargé se recharge à la prochaine utilisation.", idleActiveViaEnv: @@ -166,7 +166,7 @@ export const fr = { "Échec du chargement des paramètres de changement automatique de modèle.", saveError: "Échec de l'enregistrement des paramètres de changement automatique de modèle.", - idleError: "Saisissez un nombre entier de secondes (0 ou plus).", + idleError: "Saisissez 0 pour garder le modèle chargé, ou au moins 60 secondes.", }, previewSharing: { sectionTitle: "Partage de l'aperçu", @@ -291,7 +291,7 @@ export const fr = { resources: { title: "Système", description: - "Surveillez le matériel et le stockage de ce serveur Studio.", + "Surveillez le matériel et le stockage de ce serveur Unsloth.", liveUpdates: "Mises à jour en direct", floatingWindow: "Fenêtre flottante", disableOverlay: "Désactiver la superposition", diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts index 97f55251c5..732ae1d7fa 100644 --- a/studio/frontend/src/i18n/locales/hi.ts +++ b/studio/frontend/src/i18n/locales/hi.ts @@ -113,7 +113,7 @@ export const hi = { showToken: "token दिखाएं", tokenSaved: "Token सहेजा गया", password: "पासवर्ड", - passwordDescription: "इस Studio खाते के लिए पासवर्ड बदलें।", + passwordDescription: "इस Unsloth खाते के लिए पासवर्ड बदलें।", passwordDialog: { trigger: "पासवर्ड बदलें", title: "पासवर्ड बदलें", @@ -154,14 +154,14 @@ export const hi = { "जब कोई OpenAI-संगत अनुरोध किसी अन्य डाउनलोड किए गए GGUF का नाम लेता है, तो सर्व करने से पहले उसे लोड करें। डिफ़ॉल्ट रूप से बंद; अज्ञात नाम लोड किए गए मॉडल को सर्व करते रहते हैं।", idleUnload: "निष्क्रिय ऑटो-अनलोड", idleUnloadDescription: - "VRAM मुक्त करने के लिए इतने निष्क्रिय सेकंड के बाद मॉडल को अनलोड करें; अगला अनुरोध इसे फिर से लोड करता है। 0 इसे लोड रखता है।", + "VRAM मुक्त करने के लिए इतने निष्क्रिय सेकंड के बाद मॉडल को अनलोड करें; अगला अनुरोध इसे फिर से लोड करता है। 0 इसे लोड रखता है। न्यूनतम 60 सेकंड।", idleNeedsEnable: "अनुरोध के अनुसार मॉडल बदलें चालू करें ताकि अनलोड किया गया मॉडल अगले उपयोग पर फिर से लोड हो।", idleActiveViaEnv: "निष्क्रिय ऑटो-अनलोड UNSLOTH_MODEL_IDLE_TTL एनवायरनमेंट वेरिएबल के माध्यम से सक्रिय है।", loadError: "मॉडल ऑटो-स्विच सेटिंग्स लोड करने में विफल।", saveError: "मॉडल ऑटो-स्विच सेटिंग्स सहेजने में विफल।", - idleError: "सेकंड की पूरी संख्या दर्ज करें (0 या अधिक)।", + idleError: "मॉडल को लोड रखने के लिए 0 दर्ज करें, या कम से कम 60 सेकंड।", }, previewSharing: { sectionTitle: "पूर्वावलोकन साझाकरण", @@ -283,7 +283,7 @@ export const hi = { }, resources: { title: "सिस्टम", - description: "इस Studio सर्वर के हार्डवेयर और स्टोरेज की निगरानी करें।", + description: "इस Unsloth सर्वर के हार्डवेयर और स्टोरेज की निगरानी करें।", liveUpdates: "लाइव अपडेट", floatingWindow: "फ्लोटिंग विंडो", disableOverlay: "ओवरले अक्षम करें", diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts index 23752b6c26..de5e93c672 100644 --- a/studio/frontend/src/i18n/locales/ja.ts +++ b/studio/frontend/src/i18n/locales/ja.ts @@ -149,12 +149,12 @@ export const ja = { enable: "リクエストごとにモデルを切り替え", enableDescription: "OpenAI互換のリクエストが別のダウンロード済み GGUF を指定した場合、応答する前にそのモデルを読み込みます。デフォルトはオフです。不明な名前の場合は、読み込み済みのモデルで応答を続けます。", idleUnload: "アイドル時の自動アンロード", - idleUnloadDescription: "指定した秒数だけアイドル状態が続くとモデルをアンロードして VRAM を解放します。次のリクエストで再読み込みされます。0 にすると読み込んだままにします。", + idleUnloadDescription: "指定した秒数だけアイドル状態が続くとモデルをアンロードして VRAM を解放します。次のリクエストで再読み込みされます。0 にすると読み込んだままにします。最小 60 秒。", idleNeedsEnable: "アンロードされたモデルが次回使用時に再読み込みされるように、「リクエストごとにモデルを切り替え」をオンにしてください。", idleActiveViaEnv: "アイドル時の自動アンロードは UNSLOTH_MODEL_IDLE_TTL 環境変数によって有効になっています。", loadError: "モデル自動切り替え設定の読み込みに失敗しました。", saveError: "モデル自動切り替え設定の保存に失敗しました。", - idleError: "秒数を整数（0 以上）で入力してください。", + idleError: "モデルを読み込んだままにするには 0 を、それ以外は 60 秒以上を入力してください。", }, previewSharing: { sectionTitle: "プレビュー共有", @@ -360,7 +360,7 @@ export const ja = { }, resources: { title: "システム", - description: "この Studio サーバーのハードウェアとストレージを監視します。", + description: "この Unsloth サーバーのハードウェアとストレージを監視します。", liveUpdates: "リアルタイム更新", floatingWindow: "フローティングウィンドウ", disableOverlay: "オーバーレイを無効化", diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts index a11a94faf2..6ff9cdbb5e 100644 --- a/studio/frontend/src/i18n/locales/ko.ts +++ b/studio/frontend/src/i18n/locales/ko.ts @@ -113,7 +113,7 @@ export const ko = { showToken: "토큰 표시", tokenSaved: "토큰이 저장되었습니다", password: "비밀번호", - passwordDescription: "이 Studio 계정의 비밀번호를 변경합니다.", + passwordDescription: "이 Unsloth 계정의 비밀번호를 변경합니다.", passwordDialog: { trigger: "비밀번호 변경", title: "비밀번호 변경", @@ -153,14 +153,14 @@ export const ko = { "OpenAI 호환 요청이 다운로드된 다른 GGUF를 지정하면, 응답하기 전에 해당 모델을 불러옵니다. 기본값은 꺼짐이며, 알 수 없는 이름은 불러온 모델을 계속 제공합니다.", idleUnload: "유휴 시 자동 해제", idleUnloadDescription: - "지정한 유휴 시간(초)이 지나면 모델을 해제하여 VRAM을 확보합니다. 다음 요청 시 다시 불러옵니다. 0으로 설정하면 계속 로드된 상태로 유지됩니다.", + "지정한 유휴 시간(초)이 지나면 모델을 해제하여 VRAM을 확보합니다. 다음 요청 시 다시 불러옵니다. 0으로 설정하면 계속 로드된 상태로 유지됩니다. 최소 60초입니다.", idleNeedsEnable: "해제된 모델이 다음 사용 시 다시 로드되도록 하려면 요청에 따라 모델 전환을 켜세요.", idleActiveViaEnv: "유휴 시 자동 해제가 UNSLOTH_MODEL_IDLE_TTL 환경 변수를 통해 활성화되어 있습니다.", loadError: "모델 자동 전환 설정을 불러오지 못했습니다.", saveError: "모델 자동 전환 설정을 저장하지 못했습니다.", - idleError: "정수(초)를 입력하세요(0 이상).", + idleError: "모델을 로드 상태로 유지하려면 0을, 그렇지 않으면 60초 이상을 입력하세요.", }, previewSharing: { sectionTitle: "미리보기 공유", @@ -282,7 +282,7 @@ export const ko = { }, resources: { title: "시스템", - description: "이 Studio 서버의 하드웨어와 저장소를 모니터링합니다.", + description: "이 Unsloth 서버의 하드웨어와 저장소를 모니터링합니다.", liveUpdates: "실시간 업데이트", floatingWindow: "플로팅 창", disableOverlay: "오버레이 비활성화", diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts index 07832ef1d0..e6d2347c10 100644 --- a/studio/frontend/src/i18n/locales/pt-br.ts +++ b/studio/frontend/src/i18n/locales/pt-br.ts @@ -157,14 +157,14 @@ export const ptBR = { "Quando uma requisição compatível com OpenAI nomear um GGUF baixado diferente, carrega-o antes de responder. Desativado por padrão; nomes desconhecidos continuam usando o modelo carregado.", idleUnload: "Descarregamento automático por inatividade", idleUnloadDescription: - "Descarrega o modelo após esta quantidade de segundos de inatividade para liberar VRAM; a próxima requisição o recarrega. 0 mantém o modelo carregado.", + "Descarrega o modelo após esta quantidade de segundos de inatividade para liberar VRAM; a próxima requisição o recarrega. 0 mantém o modelo carregado. Mínimo de 60 segundos.", idleNeedsEnable: "Ative Trocar de modelo por requisição para que um modelo descarregado seja recarregado no próximo uso.", idleActiveViaEnv: "O descarregamento automático por inatividade está ativo por meio da variável de ambiente UNSLOTH_MODEL_IDLE_TTL.", loadError: "Falha ao carregar as configurações de troca automática de modelo.", saveError: "Falha ao salvar as configurações de troca automática de modelo.", - idleError: "Insira um número inteiro de segundos (0 ou mais).", + idleError: "Insira 0 para manter o modelo carregado, ou pelo menos 60 segundos.", }, previewSharing: { sectionTitle: "Compartilhamento de pré-visualização", diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts index 81d20cc2ea..60e939bb0d 100644 --- a/studio/frontend/src/i18n/locales/ru.ts +++ b/studio/frontend/src/i18n/locales/ru.ts @@ -113,7 +113,7 @@ export const ru = { showToken: "Показать токен", tokenSaved: "Токен сохранён", password: "Пароль", - passwordDescription: "Изменить пароль для этого аккаунта Studio.", + passwordDescription: "Изменить пароль для этого аккаунта Unsloth.", passwordDialog: { trigger: "Изменить пароль", title: "Изменить пароль", @@ -154,14 +154,14 @@ export const ru = { "Когда OpenAI-совместимый запрос указывает другую загруженную GGUF, загружать её перед обслуживанием. По умолчанию выключено; неизвестные имена продолжают обслуживать загруженную модель.", idleUnload: "Автовыгрузка при простое", idleUnloadDescription: - "Выгружать модель после указанного числа секунд простоя, чтобы освободить VRAM; следующий запрос загрузит её снова. 0 оставляет модель загруженной.", + "Выгружать модель после указанного числа секунд простоя, чтобы освободить VRAM; следующий запрос загрузит её снова. 0 оставляет модель загруженной. Минимум 60 секунд.", idleNeedsEnable: "Включите «Переключать модель по запросу», чтобы выгруженная модель загружалась при следующем использовании.", idleActiveViaEnv: "Автовыгрузка при простое активна через переменную окружения UNSLOTH_MODEL_IDLE_TTL.", loadError: "Не удалось загрузить настройки автопереключения модели.", saveError: "Не удалось сохранить настройки автопереключения модели.", - idleError: "Введите целое число секунд (0 или больше).", + idleError: "Введите 0, чтобы модель оставалась загруженной, или не менее 60 секунд.", }, previewSharing: { sectionTitle: "Публикация предпросмотра", @@ -283,7 +283,7 @@ export const ru = { }, resources: { title: "Система", - description: "Мониторинг оборудования и хранилища этого сервера Studio.", + description: "Мониторинг оборудования и хранилища этого сервера Unsloth.", liveUpdates: "Обновления в реальном времени", floatingWindow: "Плавающее окно", disableOverlay: "Отключить оверлей", diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts index 4c51755244..37c73086cf 100644 --- a/studio/frontend/src/i18n/locales/zh-CN.ts +++ b/studio/frontend/src/i18n/locales/zh-CN.ts @@ -152,14 +152,14 @@ export const zhCN = { "当兼容 OpenAI 的请求指定了另一个已下载的 GGUF 时，先加载它再提供服务。默认关闭；未知名称将继续使用已加载的模型。", idleUnload: "空闲自动卸载", idleUnloadDescription: - "空闲达到该秒数后卸载模型以释放 VRAM；下次请求会重新加载。设为 0 则保持加载。", + "空闲达到该秒数后卸载模型以释放 VRAM；下次请求会重新加载。设为 0 则保持加载。最小 60 秒。", idleNeedsEnable: "开启“按请求切换模型”，以便已卸载的模型在下次使用时重新加载。", idleActiveViaEnv: "空闲自动卸载已通过 UNSLOTH_MODEL_IDLE_TTL 环境变量启用。", loadError: "加载模型自动切换设置失败。", saveError: "保存模型自动切换设置失败。", - idleError: "请输入整数秒数（0 或以上）。", + idleError: "输入 0 保持模型加载，或输入至少 60 秒。", }, previewSharing: { sectionTitle: "预览分享", @@ -375,7 +375,7 @@ export const zhCN = { }, resources: { title: "系统", - description: "监控此 Studio 服务器的硬件和存储。", + description: "监控此 Unsloth 服务器的硬件和存储。", liveUpdates: "实时更新", floatingWindow: "悬浮窗口", disableOverlay: "禁用悬浮层", diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css index 333ed70f6a..a43e3d974b 100644 --- a/studio/frontend/src/index.css +++ b/studio/frontend/src/index.css @@ -1467,7 +1467,9 @@ html[data-chat-font] .aui-root { } /* With more than 4 tools on, drop pill labels to icons only to cut clutter. - Compare keeps its label via data-keep-label. */ + Compare and the bypass-permissions pill keep their labels via + data-keep-label; the permission pill sits before the collapsed icons, so + they line up to its right. */ [data-pill-compact="true"] .composer-pill-btn:not([data-keep-label]) > span:not(.composer-pill-glyph) { @@ -2597,7 +2599,7 @@ html[data-chat-font] .aui-root { * the documented WCAG outcome (motion is "minimised, not removed"). * * .animate-spin and generated image loading dots are the exceptions: loading - * indicators across Studio (tool execution loaders, sonner toasts, Tauri + * indicators across Unsloth (tool execution loaders, sonner toasts, Tauri * startup / update screens, the primitive, and image generation * cards). Freezing them removes the only visual signal that work is in flight, * so they keep animating. diff --git a/studio/frontend/src/lib/tauri-diagnostics.ts b/studio/frontend/src/lib/tauri-diagnostics.ts index 5c07931a86..2b687478fe 100644 --- a/studio/frontend/src/lib/tauri-diagnostics.ts +++ b/studio/frontend/src/lib/tauri-diagnostics.ts @@ -62,7 +62,7 @@ export function redactDiagnosticsText(text: string): string { "$1$2", ); - // Redact Studio paths before broader home-directory paths. + // Redact Unsloth paths before broader home-directory paths. redacted = redacted.replace( /(?:\/Users|\/home)\/[^\s/]+\/\.unsloth\/studio/gi, "", diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py index 6b80ec0c23..d7bdb81f14 100644 --- a/studio/install_llama_prebuilt.py +++ b/studio/install_llama_prebuilt.py @@ -2908,7 +2908,7 @@ def _pick_rocm_gfx_target(out: str) -> str | None: break if _vis_raw is not None: _vis = _vis_raw.strip() - # Empty or "-1" means "no AMD GPU visible" (matches the rest of Studio). + # Empty or "-1" means "no AMD GPU visible" (matches the rest of Unsloth). if _vis == "" or _vis == "-1": return None _first = _vis.split(",")[0].strip() @@ -4448,7 +4448,7 @@ def ensure_diffusion_visual_server( approved_checksums: ApprovedReleaseChecksums, ) -> None: """Best-effort placement of the DiffusionGemma visual-server binary next to - llama-server in the install tree, so Studio can serve DiffusionGemma GGUFs + llama-server in the install tree, so Unsloth can serve DiffusionGemma GGUFs without any DG_* env. This is an Unsloth artifact (not a ggml-org one), so it is optional: if it is already present we just make it executable, otherwise we try the published release and quietly skip on absence. A source build @@ -4730,7 +4730,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]: # repackage the SO/DLL set (e.g. ggml-org/llama.cpp#23462 split the # per-binary entry code into paired ``lib-impl.so`` shared # libraries between b9279 and b9283) without us re-enumerating - # every new file. Studio invokes llama-server, llama-quantize, and the + # every new file. Unsloth invokes llama-server, llama-quantize, and the # DiffusionGemma visual-server (when the bundle ships it, for native # DiffusionGemma serving); other CLIs upstream ships (llama-cli, # llama-bench, ...) are skipped. @@ -6370,7 +6370,7 @@ def _linux_published_attempts(host: HostInfo, bundle: PublishedReleaseBundle) -> from asset names.""" attempts: list[AssetChoice] = [] if host.has_usable_nvidia: - # Prefer the cudart major Studio loads at runtime (torch's bundled + # Prefer the cudart major Unsloth loads at runtime (torch's bundled # libcudart), not the newest detected on disk. Without this a stray # cuda13 runtime outranks the torch cuda12 the binary links against. torch_preference = detect_torch_cuda_runtime_preference(host) diff --git a/studio/install_node_prebuilt.py b/studio/install_node_prebuilt.py index 4038216eef..fb40634e95 100644 --- a/studio/install_node_prebuilt.py +++ b/studio/install_node_prebuilt.py @@ -6,7 +6,7 @@ Downloads an official Node.js archive from nodejs.org into an isolated ``/node`` and never touches the system Node/npm. Pinning Node 24+ -LTS clears the Studio frontend build floor (Vite 8: Node ^20.19 || >=22.12, +LTS clears the Unsloth frontend build floor (Vite 8: Node ^20.19 || >=22.12, npm >= 11) with the npm it bundles. Archives are verified against sha256 digests pinned in ``node_prebuilt_pins.json`` diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py index 29331ea36c..93344a309c 100644 --- a/studio/install_python_stack.py +++ b/studio/install_python_stack.py @@ -1008,7 +1008,7 @@ def _install_bnb_windows_rocm() -> bool: # `hipinfo.exe` at import time to detect the GPU arch and logs a scary # (harmless) ERROR + WARNING on every import when it is missing. The venv # Scripts dir is on PATH only when the venv is activated, which neither - # Studio nor the installer's child processes ever do. + # Unsloth nor the installer's child processes ever do. _scripts_dir = os.path.dirname(sys.executable) if os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe")) and not shutil.which( "hipinfo.exe" @@ -2329,7 +2329,7 @@ def install_python_stack() -> int: _progress("dependency overrides (skipped, no torch)") elif _rocm_windows_torch_installed or _installed_torch_is_windows_rocm(): # No working Windows ROCm torchao build: it imports an absent c10d backend - # and crashes transformers.quantizers. Studio stubs it at runtime, so + # and crashes transformers.quantizers. Unsloth stubs it at runtime, so # installing it only ships a package that crashes on import -- skip it. _progress("dependency overrides (skipped, Windows ROCm)") _safe_print(" Windows ROCm -- skipping torchao (no working build; stubbed at runtime)") @@ -2382,7 +2382,7 @@ def install_python_stack() -> int: # "https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/unsloth/save.py", # ) - # 8. Studio dependencies + # 8. Unsloth dependencies _progress("studio deps") pip_install( "Installing studio dependencies", diff --git a/studio/setup.ps1 b/studio/setup.ps1 index dab1e1e73f..98e801cd3c 100644 --- a/studio/setup.ps1 +++ b/studio/setup.ps1 @@ -322,7 +322,7 @@ function Write-CudaDriverToolkitMismatch { $driverMajor = $DriverMaxCuda.Split('.')[0] substep "CUDA Toolkit $ToolkitVersion is a major-version mismatch: toolkit major $toolkitMajor exceeds driver CUDA major $driverMajor ($DriverMaxCuda)." $Color substep "Update the NVIDIA GPU driver to run CUDA Toolkit $ToolkitVersion, or install a CUDA $driverMajor.x toolkit." $Color - substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." $Color + substep "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit." $Color } # Detect CUDA Compute Capability via nvidia-smi. @@ -937,13 +937,13 @@ function Show-NpmRegistryHint { Write-Host "" step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "Yellow" if ($mirror) { - substep "Studio pins the public npm registry; your mirror is being ignored." + substep "Unsloth pins the public npm registry; your mirror is being ignored." substep "Detected a registry in your npm config:" substep " $mirror" - substep "Re-run pointing Studio at it:" + substep "Re-run pointing Unsloth at it:" substep " `$env:UNSLOTH_NPM_REGISTRY='$mirror'; .\install.ps1 --local" } else { - substep "If you use a private mirror/proxy, point Studio at it and re-run:" + substep "If you use a private mirror/proxy, point Unsloth at it and re-run:" substep " `$env:UNSLOTH_NPM_REGISTRY='https://your-mirror.example/api/npm/'; .\install.ps1 --local" } substep "(min-release-age and save-exact stay enforced.)" @@ -1070,7 +1070,7 @@ if (-not $HasNvidiaSmi) { } # ── Helper: run amd-smi without triggering a UAC elevation prompt ── # amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing -# DiskPart UAC prompt mid-install (Studio backend amd.py hits the same). RunAsInvoker +# DiskPart UAC prompt mid-install (Unsloth backend amd.py hits the same). RunAsInvoker # forces it (and helpers it spawns) to run un-elevated; on failure the WMI name -> # gfx fallback still resolves the arch. function Invoke-AmdSmiNoElevate { @@ -1132,7 +1132,7 @@ if (-not $HasNvidiaSmi) { if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false } # VenvDir/VIRTUAL_ENV can be unset this early (the update flow probes before # VenvDir is set), so also derive the venv from the setup python + default - # Studio home, else the venv hipInfo isn't caught. + # Unsloth home, else the venv hipInfo isn't caught. $venvRoots = @() if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV } $vd = Get-Variable -Name VenvDir -ValueOnly -ErrorAction SilentlyContinue @@ -1141,7 +1141,7 @@ if (-not $HasNvidiaSmi) { try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {} } if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") } - # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the + # A custom Unsloth home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the # venv off the default path; seed it too or its hipInfo escapes the filter. $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null } if ($studioHomeEnv) { @@ -1860,7 +1860,7 @@ $SysNpmVersion = "" $NodeSource = $null if (-not $IsPipInstall) { - # Put Node beside the Studio root. OXC can still need npm when the + # Put Node beside the Unsloth root. OXC can still need npm when the # frontend build is skipped. if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $NodeOverride = $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $NodeOverride = $env:STUDIO_HOME.Trim() } @@ -2137,17 +2137,17 @@ if ($NeedNodeForSetup) { step "frontend" "skipped (no suitable Node; system left untouched)" "Yellow" } $NeedFrontendBuild = $false - substep "found Node='$SysNodeVersion' npm='$SysNpmVersion'; Studio needs Node >=20.19/22.12/23 and npm >= 11" "Yellow" + substep "found Node='$SysNodeVersion' npm='$SysNpmVersion'; Unsloth needs Node >=20.19/22.12/23 and npm >= 11" "Yellow" substep "install a suitable Node + npm, or unset UNSLOTH_SKIP_NODE_INSTALL to let Unsloth manage an isolated Node" "Yellow" } elseif ($NodeSource -eq "bundled") { New-Item -ItemType Directory -Force -Path $NodeParent -ErrorAction SilentlyContinue | Out-Null - # Minimal ownership guard for a custom-home dir (the full Studio-owned + # Minimal ownership guard for a custom-home dir (the full Unsloth-owned # helpers are defined later); never os.replace over a user-owned dir. if ($NodeOverride -and (Test-Path -LiteralPath $NodeDir -PathType Container)) { $nodeOwnedMarker = Join-Path $NodeDir ".unsloth-studio-owned" $nodeMeta = Join-Path $NodeDir "UNSLOTH_NODE_PREBUILT_INFO.json" if (-not (Test-Path -LiteralPath $nodeOwnedMarker) -and -not (Test-Path -LiteralPath $nodeMeta)) { - Write-Host "[ERROR] $NodeDir already exists and is not a Studio-owned Node install." -ForegroundColor Red + Write-Host "[ERROR] $NodeDir already exists and is not an Unsloth-owned Node install." -ForegroundColor Red Write-Host " Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." -ForegroundColor Yellow exit 1 } @@ -2160,7 +2160,7 @@ if ($NeedNodeForSetup) { $nodeExit = $LASTEXITCODE if ($nodeExit -eq 3) { Write-Host $nodeOut -ForegroundColor DarkGray - step "node" "install blocked by another active Studio install" "Red" + step "node" "install blocked by another active Unsloth install" "Red" exit 3 } elseif ($nodeExit -ne 0) { Write-Host $nodeOut -ForegroundColor DarkGray @@ -2486,7 +2486,7 @@ if (Test-Path -LiteralPath $LegacyStudioHome -PathType Container) { $LegacyStudioHome = (Resolve-Path -LiteralPath $LegacyStudioHome).Path } $StudioHomeIsCustom = ($_studioHomeCanon -ne $LegacyStudioHome) -# Directory-local evidence that Studio created $Path, used to adopt a custom-home +# Directory-local evidence that Unsloth created $Path, used to adopt a custom-home # llama.cpp predating the .unsloth-studio-owned marker (see setup.sh). Only the # prebuilt UNSLOTH_PREBUILT_INFO.json counts; source builds are indistinguishable # from a user clone on Windows and stay under the strict guard. @@ -2506,7 +2506,7 @@ function Assert-StudioOwnedOrAbsent { Mark-StudioOwned $Path return } - Write-Host "[ERROR] $Path already exists and is not marked as a Studio-owned $Label." -ForegroundColor Red + Write-Host "[ERROR] $Path already exists and is not marked as an Unsloth-owned $Label." -ForegroundColor Red Write-Host " Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." -ForegroundColor Yellow exit 1 } @@ -2576,7 +2576,7 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode $reason = if ($installedTorchTag) { "torch $installedTorchTag != required $expectedTorchTag" } else { "torch could not be imported" } if ($InstallerManagedSetup) { substep "Stale venv detected ($reason)." "Yellow" - Write-Host " [ERROR] The existing Studio environment needs repair." -ForegroundColor Red + Write-Host " [ERROR] The existing Unsloth environment needs repair." -ForegroundColor Red Write-Host " Re-run install.ps1 so it can replace the environment safely with rollback." -ForegroundColor Yellow exit 1 } @@ -2598,7 +2598,7 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode Remove-Item -LiteralPath $VenvDir -Recurse -Force -ErrorAction Stop } catch { Write-Host " [ERROR] Could not remove stale venv: $($_.Exception.Message)" -ForegroundColor Red - Write-Host " Close any running Studio/Python processes and re-run setup." -ForegroundColor Red + Write-Host " Close any running Unsloth/Python processes and re-run setup." -ForegroundColor Red exit 1 } } @@ -2764,7 +2764,7 @@ if ($script:UnslothVerbose) { # The CUDA tag is chosen based on the driver's max supported CUDA version. # Triton/inductor filenames are long and can hit Windows MAX_PATH (260). With long -# paths on, cache under Studio home; else use a short drive-root dir for headroom. +# paths on, cache under Unsloth home; else use a short drive-root dir for headroom. if ($LongPathsEnabled) { $TorchCacheDir = Join-Path $StudioHome "TORCHINDUCTOR_CACHE_DIR" } else { @@ -2793,7 +2793,7 @@ $ROCmIndexUrl = $null # Install AMD ROCm PyTorch wheels when ROCm is confirmed OR a gfx arch is known # (name-inferred on Adrenalin-only hosts). The per-arch wheels bundle the runtime # (rocm-sdk-libraries-), so torch.cuda.is_available() is True without a HIP -# SDK -- which flips Studio out of chat-only (CHAT_ONLY) and enables Train/Export. +# SDK -- which flips Unsloth out of chat-only (CHAT_ONLY) and enables Train/Export. # Gating on $HasROCm alone left Strix Halo / Radeon 8060S on CPU torch; a failed # ROCm install still falls back to CPU below, so this is safe. if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") { @@ -3231,7 +3231,7 @@ if ($LocalLlamaCppSrc) { # Reusing a local dir disables both the prebuilt download and the source # build, so a runnable llama-server.exe must already be present. Accept any # layout LlamaCppBackend._layout_candidates() resolves (root-level, build\bin, - # or build\bin\Release) so the flag never rejects a tree Studio could run. + # or build\bin\Release) so the flag never rejects a tree Unsloth could run. $LocalLlamaServerFound = $false foreach ($_cand in @( (Join-Path $ResolvedLocal "llama-server.exe"), @@ -3253,7 +3253,7 @@ if ($LocalLlamaCppSrc) { } } else { # Fail clearly rather than junction an unbuilt or wrong-platform checkout - # and leave Studio with no usable binary. + # and leave Unsloth with no usable binary. if (-not $LocalLlamaServerFound) { step "llama.cpp" "no llama-server.exe under $ResolvedLocal (looked for .\llama-server.exe, .\build\bin and .\build\bin\Release) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "Red" exit 1 @@ -3281,7 +3281,7 @@ if ($LocalLlamaCppSrc) { # prebuilt path's active-process handling and stop with a clear message. if (Test-Path -LiteralPath $LlamaCppDir) { step "llama.cpp" "install blocked by active llama.cpp process" "Yellow" - substep "Close Studio or other llama.cpp users and retry" "Yellow" + substep "Close Unsloth or other llama.cpp users and retry" "Yellow" exit 3 } } @@ -3414,7 +3414,7 @@ if ($LocalLlamaCppLinked) { if (Test-Path -LiteralPath $LlamaCppDir) { substep "Existing install was restored" "Yellow" } - substep "Close Studio or other llama.cpp users and retry" "Yellow" + substep "Close Unsloth or other llama.cpp users and retry" "Yellow" exit 3 } else { step "llama.cpp" "prebuilt install failed (continuing)" "Yellow" @@ -4003,7 +4003,7 @@ if ($LocalLlamaCppLinked) { } # -- Step E: Build the DiffusionGemma visual server (optional, best-effort) -- - # An example target present on llama.cpp PR #24423; lets Studio serve + # An example target present on llama.cpp PR #24423; lets Unsloth serve # DiffusionGemma GGUFs without DG_VISUAL_BIN. No-op when not configured. if ($BuildOk) { $null = cmake --build $BuildDir --config Release --target llama-diffusion-gemma-visual-server -j $NumCpu 2>&1 | Out-String diff --git a/studio/setup.sh b/studio/setup.sh index 3d67db3da7..8d47eecfda 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -127,13 +127,13 @@ _suggest_npm_registry() { printf '\n' >&2 step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "$C_WARN" >&2 if [ -n "$_mirror" ]; then - substep "Studio pins the public npm registry; your mirror is being ignored." >&2 + substep "Unsloth pins the public npm registry; your mirror is being ignored." >&2 substep "Detected a registry in your npm config:" >&2 substep " $_mirror" >&2 - substep "Re-run pointing Studio at it:" >&2 + substep "Re-run pointing Unsloth at it:" >&2 substep " UNSLOTH_NPM_REGISTRY=$_mirror ./install.sh --local" >&2 else - substep "If you use a private mirror/proxy, point Studio at it and re-run:" >&2 + substep "If you use a private mirror/proxy, point Unsloth at it and re-run:" >&2 substep " UNSLOTH_NPM_REGISTRY=https://your-mirror.example/api/npm/ ./install.sh --local" >&2 fi substep "(min-release-age and save-exact stay enforced.)" >&2 @@ -396,7 +396,7 @@ _print_cuda_driver_toolkit_mismatch() { local _driver_major=${_driver_version%%.*} substep "CUDA Toolkit $_toolkit_version is a major-version mismatch: toolkit major $_toolkit_major exceeds driver CUDA major $_driver_major ($_driver_version)." "$C_WARN" substep "Update the NVIDIA GPU driver to run CUDA Toolkit $_toolkit_version, or install a CUDA $_driver_major.x toolkit." "$C_WARN" - substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." "$C_WARN" + substep "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit." "$C_WARN" } print_llama_error_log() { @@ -531,7 +531,7 @@ _STUDIO_HOME_IS_CUSTOM=false if [ "$_studio_home_canon" != "$_LEGACY_STUDIO_HOME" ]; then _STUDIO_HOME_IS_CUSTOM=true fi -# Directory-local evidence Studio created "$1": only prebuilt-installer metadata +# Directory-local evidence Unsloth created "$1": only prebuilt-installer metadata # counts (UNSLOTH_PREBUILT_INFO.json for llama.cpp, UNSLOTH_NODE_PREBUILT_INFO.json # for Node), both written only by our installers. Mirrors the setup.ps1 Node guard. # A markerless source build stays strict since this runs right before an rm -rf. @@ -549,7 +549,7 @@ _assert_studio_owned_or_absent() { : > "$_aso_dir/$_STUDIO_OWNED_MARKER" 2>/dev/null || true return 0 fi - echo "ERROR: $_aso_dir already exists and is not marked as a Studio-owned $_aso_label." >&2 + echo "ERROR: $_aso_dir already exists and is not marked as an Unsloth-owned $_aso_label." >&2 echo " Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." >&2 exit 1 fi @@ -585,7 +585,7 @@ if [ "$_NEED_FRONTEND_BUILD" = false ] && [ ! -d "$_OXC_DIR" ]; then else # ── Node (isolated; never touches the system Node/npm) ── -# Studio's frontend (Vite 8) needs Node ^20.19 || >=22.12 || >=23 and npm >= 11. +# Unsloth's frontend (Vite 8) needs Node ^20.19 || >=22.12 || >=23 and npm >= 11. # Three sources: # system -- system Node + npm already satisfy both; used read-only. # bundled -- install a pinned isolated Node under $UNSLOTH_HOME/node, build-only. @@ -666,9 +666,9 @@ elif [ "$NODE_SOURCE" = bundled ]; then fi set -e if [ "$_NODE_STATUS" -eq 3 ]; then - step "node" "install blocked by another active Studio install" "$C_ERR" + step "node" "install blocked by another active Unsloth install" "$C_ERR" sed 's/^/ | /' "$_NODE_LOG" >&2; rm -f "$_NODE_LOG" - substep "close other Studio installs and retry" + substep "close other Unsloth installs and retry" exit 3 elif [ "$_NODE_STATUS" -ne 0 ]; then step "node" "isolated Node install failed" "$C_ERR" @@ -692,7 +692,7 @@ elif [ "$NODE_SOURCE" = bundled ]; then else _FRONTEND_SKIP=true step "frontend" "skipped (no suitable Node; system left untouched)" "$C_WARN" - substep "found Node='${_SYS_NODE_VER:-none}' npm='${_SYS_NPM_VER:-none}'; Studio needs Node >=20.19/22.12/23 and npm >= 11" + substep "found Node='${_SYS_NODE_VER:-none}' npm='${_SYS_NPM_VER:-none}'; Unsloth needs Node >=20.19/22.12/23 and npm >= 11" substep "install a suitable Node + npm, or unset UNSLOTH_SKIP_NODE_INSTALL to let Unsloth manage an isolated Node" fi verbose_substep "node source: $NODE_SOURCE (sys node=${_SYS_NODE_VER:-none} npm=${_SYS_NPM_VER:-none}) dir=$NODE_DIR" @@ -873,11 +873,11 @@ _remove_agent_instruction_files \ _COLAB_NO_VENV=false if [ ! -x "$VENV_DIR/bin/python" ]; then if [ "$IS_COLAB" = true ]; then - # On Colab there is no Studio venv -- install backend deps into system Python. + # On Colab there is no Unsloth venv -- install backend deps into system Python. # Strip all version constraints so pip keeps Colab's pre-installed # packages (huggingface-hub, datasets, transformers) and only pulls # in genuinely missing ones (structlog, fastapi, etc.). - substep "Colab detected, installing Studio backend dependencies..." + substep "Colab detected, installing Unsloth backend dependencies..." _COLAB_REQS_TMP="$(mktemp)" sed 's/[><=!~;].*//' "$SCRIPT_DIR/backend/requirements/studio.txt" \ | grep -v '^#' | grep -v '^$' > "$_COLAB_REQS_TMP" @@ -1254,7 +1254,7 @@ _link_local_llama_quantize_shim() { } # Accept any layout LlamaCppBackend._layout_candidates() resolves so the flag -# never rejects a tree Studio could actually run: a root-level llama-server (a +# never rejects a tree Unsloth could actually run: a root-level llama-server (a # `make` build or a flat-extracted release) or the CMake build/bin/llama-server. _has_local_llama_server() { [ -x "$1/llama-server" ] || [ -x "$1/build/bin/llama-server" ] @@ -1298,13 +1298,13 @@ if [ -n "${UNSLOTH_LOCAL_LLAMA_CPP_DIR:-}" ]; then # Reusing disables BOTH the prebuilt download and the source build, so the # linked tree must already contain a runnable llama-server in one of the # layouts the backend resolves (root-level or build/bin/). Fail clearly - # rather than link an unbuilt or wrong-platform checkout and leave Studio + # rather than link an unbuilt or wrong-platform checkout and leave Unsloth # with no usable binary. if ! _has_local_llama_server "$_RESOLVED_LOCAL"; then step "llama.cpp" "no llama-server under $_RESOLVED_LOCAL (looked for ./llama-server and ./build/bin/llama-server) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "$C_ERR" exit 1 fi - # A stale link from a previous --with-llama-cpp-dir run isn't Studio-owned + # A stale link from a previous --with-llama-cpp-dir run isn't Unsloth-owned # content; drop it before the ownership check so re-runs stay idempotent # for a custom UNSLOTH_STUDIO_HOME (the assert would otherwise follow the # link into the user's dir and reject it as unowned). @@ -1389,7 +1389,7 @@ else if [ -d "$LLAMA_CPP_DIR" ]; then substep "existing install was restored" fi - substep "close Studio or other llama.cpp users and retry" + substep "close Unsloth or other llama.cpp users and retry" exit 3 else step "llama.cpp" "prebuilt install failed (continuing)" "$C_WARN" @@ -1886,7 +1886,7 @@ else ln -sf build/bin/llama-quantize "$LLAMA_CPP_DIR/llama-quantize" fi # DiffusionGemma visual server, if it was built (PR #24423): link next to - # llama-server so Studio serves DiffusionGemma GGUFs without DG_VISUAL_BIN. + # llama-server so Unsloth serves DiffusionGemma GGUFs without DG_VISUAL_BIN. if [ -f "$LLAMA_CPP_DIR/build/bin/llama-diffusion-gemma-visual-server" ]; then ln -sf build/bin/llama-diffusion-gemma-visual-server "$LLAMA_CPP_DIR/llama-diffusion-gemma-visual-server" fi @@ -1983,7 +1983,7 @@ echo "" # When called from install.sh (SKIP_STUDIO_BASE=1), exit non-zero so the # installer can report the GGUF failure after finishing PATH/shortcut setup. # When called directly via 'unsloth studio update', keep the install -# successful -- the footer above already reports the limitation and Studio +# successful -- the footer above already reports the limitation and Unsloth # is still usable for non-GGUF workflows. if [ "$_LLAMA_CPP_DEGRADED" = true ] && [ "${SKIP_STUDIO_BASE:-0}" = "1" ]; then exit 1 diff --git a/studio/src-tauri/src/commands.rs b/studio/src-tauri/src/commands.rs index 6bc2116786..5d41cb9217 100644 --- a/studio/src-tauri/src/commands.rs +++ b/studio/src-tauri/src/commands.rs @@ -25,12 +25,12 @@ fn should_emit_repair_failed(msg: &str) -> bool { fn external_conflict_message(conflict: &crate::preflight::ExternalBackendConflict) -> String { if conflict.reason == "desktop_owned_backend_active" { return format!( - "A desktop-owned Studio server for this install is already running on port {}. Quit the other desktop app instance, then try again.", + "A desktop-owned Unsloth server for this install is already running on port {}. Quit the other desktop app instance, then try again.", conflict.port ); } format!( - "A Studio server for this install is already running from a terminal on port {}. Stop that server, or run `unsloth studio update` from that terminal before using desktop repair/update.", + "An Unsloth server for this install is already running from a terminal on port {}. Stop that server, or run `unsloth studio update` from that terminal before using desktop repair/update.", conflict.port ) } @@ -475,7 +475,7 @@ pub async fn start_backend_update( .map_err(|e| format!("Update task panicked: {e}"))? } -/// Repair a stale managed Studio install. +/// Repair a stale managed Unsloth install. #[tauri::command] pub async fn start_managed_repair( app: AppHandle, @@ -522,7 +522,7 @@ pub async fn start_managed_repair( let repair_group_id = install::take_pending_repair_group_for_resume(&install_state) .unwrap_or_else(|| diagnostics::begin_repair_group(&diagnostics_state)); - let _ = app.emit("repair-progress", "Updating existing Studio install..."); + let _ = app.emit("repair-progress", "Updating existing Unsloth install..."); let update_app = app.clone(); let update_state = update_state.inner().clone(); let update_diagnostics = diagnostics_state.clone(); @@ -549,7 +549,7 @@ pub async fn start_managed_repair( warn!("Managed repair update finished, but preflight is still not ready; falling back to installer"); let _ = app.emit( "repair-progress", - "Update finished, but Studio is still not ready. Running bundled installer...", + "Update finished, but Unsloth is still not ready. Running bundled installer...", ); } Err(msg) => { @@ -627,7 +627,7 @@ pub async fn start_managed_repair( return Ok(()); } - let msg = "Repair finished, but Studio install is still not desktop-ready.".to_string(); + let msg = "Repair finished, but Unsloth install is still not desktop-ready.".to_string(); error!("{}", msg); diagnostics::finish_repair_group( &diagnostics_state, diff --git a/studio/src-tauri/src/desktop_auth.rs b/studio/src-tauri/src/desktop_auth.rs index 33605f6c65..db65b8796f 100644 --- a/studio/src-tauri/src/desktop_auth.rs +++ b/studio/src-tauri/src/desktop_auth.rs @@ -201,7 +201,7 @@ async fn exchange_desktop_secret( if response.status() == reqwest::StatusCode::NOT_FOUND { return Err(AuthError::StaleResponder( - "Running Studio backend is too old for this desktop app. Update that backend and restart." + "Running Unsloth backend is too old for this desktop app. Update that backend and restart." .to_string(), )); } @@ -364,7 +364,7 @@ async fn desktop_auth_inner( } Err( - "Desktop auth failed. Update or repair the managed Studio install, then restart Studio." + "Desktop auth failed. Update or repair the managed Unsloth install, then restart Unsloth." .to_string(), ) } @@ -465,7 +465,7 @@ mod tests { .message(); assert_eq!( error, - "Running Studio backend is too old for this desktop app. Update that backend and restart." + "Running Unsloth backend is too old for this desktop app. Update that backend and restart." ); } } diff --git a/studio/src-tauri/src/main.rs b/studio/src-tauri/src/main.rs index 4ed12051ed..4724317601 100644 --- a/studio/src-tauri/src/main.rs +++ b/studio/src-tauri/src/main.rs @@ -107,7 +107,7 @@ fn cleanup_child_processes(app: &tauri::AppHandle) { } fn setup_tray(app: &tauri::App) -> Result<(), Box> { - let open = MenuItemBuilder::with_id("open", "Open Studio").build(app)?; + let open = MenuItemBuilder::with_id("open", "Open Unsloth").build(app)?; let toggle = MenuItemBuilder::with_id("toggle", "Start/Stop Server").build(app)?; let quit = MenuItemBuilder::with_id("quit", "Quit").build(app)?; let menu = MenuBuilder::new(app) diff --git a/studio/src-tauri/src/native_path_policy.rs b/studio/src-tauri/src/native_path_policy.rs index b2ebb34621..b82e516e7a 100644 --- a/studio/src-tauri/src/native_path_policy.rs +++ b/studio/src-tauri/src/native_path_policy.rs @@ -191,7 +191,7 @@ fn reject_sensitive_artifact(path: &Path) -> Result<(), String> { "\\pid", ] { if lowered.contains(needle) { - return Err("Sensitive Studio state cannot be registered as an artifact.".to_string()); + return Err("Sensitive Unsloth state cannot be registered as an artifact.".to_string()); } } if let Some(ext) = path.extension().and_then(|ext| ext.to_str()) { diff --git a/tests/python/test_e2e_no_torch_sandbox.py b/tests/python/test_e2e_no_torch_sandbox.py index f412de2063..bb61af462d 100644 --- a/tests/python/test_e2e_no_torch_sandbox.py +++ b/tests/python/test_e2e_no_torch_sandbox.py @@ -30,7 +30,7 @@ VLM_PROCESSING = DATASETS_DIR / "vlm_processing.py" ITERABLE = DATASETS_DIR / "iterable.py" HARDWARE_PY = HARDWARE_DIR / "hardware.py" -# Studio venv for server tests +# Unsloth venv for server tests STUDIO_VENV = Path.home() / ".unsloth" / "studio" / "unsloth_studio" sys.path.insert(0, str(STUDIO_DIR)) @@ -957,20 +957,20 @@ server = pytest.mark.server @server class TestLiveServerStartup: - """Live server startup against the existing Studio venv with torch made unimportable (pytest -m server).""" + """Live server startup against the existing Unsloth venv with torch made unimportable (pytest -m server).""" @pytest.fixture(autouse = True) def _check_studio_venv(self): py = _studio_venv_python() if py is None: - pytest.skip("Studio venv not found at ~/.unsloth/studio/unsloth_studio") + pytest.skip("Unsloth venv not found at ~/.unsloth/studio/unsloth_studio") @pytest.fixture(scope = "class") def server_process(self): """Start the studio backend server without torch, yield (proc, port), then stop.""" py = _studio_venv_python() if py is None: - pytest.skip("Studio venv not found") + pytest.skip("Unsloth venv not found") port = _server_port() backend_dir = BACKEND_DIR diff --git a/tests/python/test_studio_import_no_torch.py b/tests/python/test_studio_import_no_torch.py index c4efbc8cea..f551519de9 100644 --- a/tests/python/test_studio_import_no_torch.py +++ b/tests/python/test_studio_import_no_torch.py @@ -1,4 +1,4 @@ -"""Sandbox tests: Studio dataset modules load/run in isolated no-torch venvs.""" +"""Sandbox tests: Unsloth dataset modules load/run in isolated no-torch venvs.""" from __future__ import annotations diff --git a/tests/python/test_tokenizers_and_torch_constraint.py b/tests/python/test_tokenizers_and_torch_constraint.py index 4322f0c7d6..c58808689b 100644 --- a/tests/python/test_tokenizers_and_torch_constraint.py +++ b/tests/python/test_tokenizers_and_torch_constraint.py @@ -69,6 +69,21 @@ class TestStructuralTorchConstraint: def test_tightened_assignment_exists(self): assert 'TORCH_CONSTRAINT="torch>=2.6,<2.11.0"' in self._sh + def test_cuda_constraint_widened_to_2_12(self): + """A fresh CUDA install widens the ceiling to <2.12.0 so cu12x/cu13x + land torch 2.11.x (matches the base image and _CUDA_TORCH_PKG_SPEC); + without it cu128/cu130 resolves torch 2.10.x.""" + assert 'TORCH_CONSTRAINT="torch>=2.4,<2.12.0"' in self._sh + + def test_cuda_case_widens_via_index_leaf(self): + """The cu* branch of the _torch_index_leaf case sets the widened + constraint (parallel to rocm7.2), anchored on the leaf.""" + m = re.search( + r'cu\[0-9\]\*\)\s*TORCH_CONSTRAINT="torch>=2\.4,<2\.12\.0"', + self._sh, + ) + assert m is not None, "CUDA (cu*) TORCH_CONSTRAINT widening case not found" + def test_variable_used_in_pip_install(self): """$TORCH_CONSTRAINT must appear in a uv pip install line.""" assert '"$TORCH_CONSTRAINT"' in self._sh @@ -384,6 +399,71 @@ class TestTorchConstraintShell: logged = log_file.read_text() assert "torch>=2.4,<2.11.0" in logged, f"uv log: {logged}" + # Mirrors the _torch_index_leaf case in install.sh: rocm7.2 -> 2.11.x floor, + # CUDA -> widened <2.12.0 ceiling, else (CPU/older ROCm) -> default. Anchored + # on the final path segment, so a mirror base path containing cu*/rocm7.2 but + # ending in a cpu/older-rocm leaf keeps the default. + _INDEX_SNIPPET = textwrap.dedent(r""" + #!/bin/bash + set -e + TORCH_INDEX_URL="{index_url}" + TORCH_CONSTRAINT="torch>=2.4,<2.11.0" + _torch_index_leaf="${TORCH_INDEX_URL%/}" + _torch_index_leaf="${_torch_index_leaf##*/}" + case "$_torch_index_leaf" in + rocm7.2) TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0" ;; + cu[0-9]*) TORCH_CONSTRAINT="torch>=2.4,<2.12.0" ;; + esac + echo "$TORCH_CONSTRAINT" + """).strip() + + def _resolve_index(self, tmp_path: pathlib.Path, index_url: str) -> str: + script_file = tmp_path / "index_snippet.sh" + script_file.write_text(self._INDEX_SNIPPET.replace("{index_url}", index_url)) + script_file.chmod(0o755) + result = subprocess.run( + ["bash", str(script_file)], + capture_output = True, + text = True, + timeout = 10, + ) + assert result.returncode == 0, f"Script failed: {result.stderr}" + return result.stdout.strip() + + @pytest.mark.parametrize("leaf", ["cu118", "cu124", "cu126", "cu128", "cu130"]) + def test_cuda_index_widens_to_2_12(self, tmp_path, leaf): + url = f"https://download.pytorch.org/whl/{leaf}" + assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.12.0" + + def test_rocm72_index_uses_211_floor(self, tmp_path): + url = "https://download.pytorch.org/whl/rocm7.2" + assert self._resolve_index(tmp_path, url) == "torch>=2.11.0,<2.12.0" + + def test_cpu_index_keeps_default(self, tmp_path): + # /cpu must NOT match the */cu[0-9]* branch. + url = "https://download.pytorch.org/whl/cpu" + assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0" + + def test_older_rocm_index_keeps_default(self, tmp_path): + url = "https://download.pytorch.org/whl/rocm7.1" + assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0" + + def test_cuda_index_custom_mirror_widens(self, tmp_path): + url = "https://internal.example.com/pytorch/cu128" + assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.12.0" + + @pytest.mark.parametrize( + "url", + [ + "https://internal.example.com/pytorch/cu128/cpu", + "https://internal.example.com/cu128/whl/rocm7.1", + ], + ) + def test_cuda_in_mirror_path_but_noncuda_leaf_keeps_default(self, tmp_path, url): + # A cu128 in the mirror base path must not widen when the leaf is cpu / + # older ROCm: the case anchors on _torch_index_leaf, not the whole URL. + assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0" + # Group 3 -- E2E tokenizers fix (requires network, ~2-5 min) @pytest.mark.e2e diff --git a/tests/saving/test_fix_sentencepiece_tokenizer_guard.py b/tests/saving/test_fix_sentencepiece_tokenizer_guard.py new file mode 100644 index 0000000000..1ee523d57b --- /dev/null +++ b/tests/saving/test_fix_sentencepiece_tokenizer_guard.py @@ -0,0 +1,307 @@ +# SPDX-License-Identifier: AGPL-3.0-only +import gc +import os + +os.environ.setdefault("PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION", "python") + +import transformers +from transformers.utils import sentencepiece_model_pb2 + +from unsloth.tokenizer_utils import fix_sentencepiece_tokenizer + + +NORMAL, CONTROL = 1, 3 + + +def _spm_bytes(pieces): + m = sentencepiece_model_pb2.ModelProto() + for piece, score, typ in pieces: + p = m.pieces.add() + p.piece = piece + p.score = score + p.type = typ + return m.SerializeToString() + + +def _read_pieces(path): + m = sentencepiece_model_pb2.ModelProto() + with open(path, "rb") as f: + m.ParseFromString(f.read()) + return [p.piece for p in m.pieces] + + +class _FakeTokenizer: + """Minimal stand-in for a sentencepiece-backed slow tokenizer. + + ``save_pretrained`` writes a tokenizer.model, which is what the real slow + tokenizers do and what fix_sentencepiece_tokenizer reads back. + """ + + def __init__( + self, + name, + spm_bytes = None, + vocab = None, + ): + self.name = name + self.eos_token = "" + self.pad_token = "" + self._spm_bytes = spm_bytes + self._vocab = vocab or {} + self.saved_to = [] + + def save_pretrained(self, location): + self.saved_to.append(location) + os.makedirs(location, exist_ok = True) + if self._spm_bytes is not None: + with open(os.path.join(location, "tokenizer.model"), "wb") as f: + f.write(self._spm_bytes) + + def __call__( + self, + texts, + add_special_tokens = False, + ): + class _Encoded: + pass + + encoded = _Encoded() + encoded.input_ids = [[self._vocab[text]] for text in texts] + return encoded + + +def _tokenizers(): + pieces = [("", 0.0, CONTROL), ("a", -1.0, NORMAL), ("", 0.0, CONTROL)] + old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"": 2}) + new = _FakeTokenizer("new") + return old, new + + +class _ReloadedTokenizer: + """Weakref-able stand-in for the tokenizer AutoTokenizer.from_pretrained returns.""" + + def __init__(self, location): + self.location = location + + +def _stub_auto_tokenizer(monkeypatch): + """fix_sentencepiece_tokenizer reloads the patched directory through + AutoTokenizer at the end; that needs a full tokenizer on disk, which is + out of scope here. Record the reload location and hand back a sentinel. + """ + loaded = [] + + class _StubAutoTokenizer: + @staticmethod + def from_pretrained(location, **kwargs): + loaded.append(location) + return _ReloadedTokenizer(location) + + monkeypatch.setattr(transformers, "AutoTokenizer", _StubAutoTokenizer) + return loaded + + +def test_old_tokenizer_is_saved_so_its_model_can_be_read(tmp_path, monkeypatch): + """The guard must not skip the body on a fresh temporary directory. + + fix_sentencepiece_tokenizer creates its scratch directory itself and then + checks for a tokenizer.model inside it, but that file only appears once + old_tokenizer.save_pretrained() has run. + """ + _stub_auto_tokenizer(monkeypatch) + old, new = _tokenizers() + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + fix_sentencepiece_tokenizer(old, new, {"": "<|im_end|>"}, temporary_location = location) + + assert old.saved_to, "old tokenizer was never saved: the body did not run" + + +def test_token_mapping_is_applied_to_the_sentencepiece_model(tmp_path, monkeypatch): + loaded = _stub_auto_tokenizer(monkeypatch) + old, new = _tokenizers() + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + # Hold the returned tokenizer so its scratch dir survives until we read it. + tok = fix_sentencepiece_tokenizer(old, new, {"": "<|im_end|>"}, temporary_location = location) + + assert "<|im_end|>" in _read_pieces(f"{loaded[-1]}/tokenizer.model") + assert tok is not None + + +def test_tokenizer_without_a_sentencepiece_model_is_returned_untouched(tmp_path, monkeypatch): + """A fast-only tokenizer writes no tokenizer.model, so the guard still + short-circuits and the caller gets new_tokenizer back unchanged. Its scratch + dir is unreferenced and reclaimed immediately. + """ + _stub_auto_tokenizer(monkeypatch) + old = _FakeTokenizer("old", spm_bytes = None) + new = _FakeTokenizer("new") + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + result = fix_sentencepiece_tokenizer( + old, new, {"": "<|im_end|>"}, temporary_location = location + ) + + assert result is new + assert not any( + name.startswith("tokenizer_") for name in os.listdir(location) + ), "the fast-only scratch dir was not reclaimed" + + +def test_each_call_uses_a_fresh_isolated_subdirectory(tmp_path, monkeypatch): + """Each call must work in its own unique subdirectory, so concurrent or + repeated calls never share scratch files, stale artifacts never leak into + the reload, and nothing the caller left in the scratch location is deleted. + """ + loaded = _stub_auto_tokenizer(monkeypatch) + location = str(tmp_path / "_unsloth_sentencepiece_temp") + os.makedirs(location, exist_ok = True) + + # A pre-existing artifact in the shared scratch location. + marker = os.path.join(location, "leftover.json") + with open(marker, "w") as f: + f.write("{}") + + old1, new1 = _tokenizers() + old2, new2 = _tokenizers() + # Hold both returned tokenizers so their scratch dirs stay alive. + tok1 = fix_sentencepiece_tokenizer( + old1, new1, {"": "<|im_end|>"}, temporary_location = location + ) + tok2 = fix_sentencepiece_tokenizer( + old2, new2, {"": "<|im_end|>"}, temporary_location = location + ) + + work1, work2 = loaded[0], loaded[1] + assert work1 != work2, "two calls reused the same directory" + assert os.path.dirname(work1) == location and os.path.dirname(work2) == location + assert os.path.isdir(work1) and os.path.isdir(work2) + # Nothing the caller left behind is deleted, and it never leaks into a work dir. + assert os.path.isfile(marker), "a pre-existing scratch file was deleted" + assert not os.path.isfile(os.path.join(work1, "leftover.json")) + assert not os.path.isfile(os.path.join(work2, "leftover.json")) + assert tok1 is not None and tok2 is not None + + +def test_sentencepiece_scratch_dir_is_reclaimed_once_the_tokenizer_is_gone(tmp_path, monkeypatch): + """The scratch dir must live as long as the returned tokenizer (its vocab_file + points there), then be reclaimed when the tokenizer is garbage collected. + """ + loaded = _stub_auto_tokenizer(monkeypatch) + old, new = _tokenizers() + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + tok = fix_sentencepiece_tokenizer(old, new, {"": "<|im_end|>"}, temporary_location = location) + work = loaded[-1] + assert os.path.isdir(work), "scratch dir vanished while the tokenizer was alive" + + del tok + gc.collect() + assert not os.path.isdir(work), "scratch dir was not reclaimed after the tokenizer was freed" + + +class _CopyFromSubdirTokenizer: + """A slow tokenizer whose sentencepiece source lives elsewhere (like the + tokenizers convert_to_fast_tokenizer produces under {location}/{name}). + save_pretrained copies that source into the destination, as HF slow + tokenizers copy their vocab_file. + """ + + def __init__(self, source_model_path): + self.eos_token = "" + self.pad_token = "" + self._source_model_path = source_model_path + + def save_pretrained(self, location): + os.makedirs(location, exist_ok = True) + if os.path.isfile(self._source_model_path): + with open(self._source_model_path, "rb") as src: + data = src.read() + with open(os.path.join(location, "tokenizer.model"), "wb") as dst: + dst.write(data) + + def __call__( + self, + texts, + add_special_tokens = False, + ): + class _Encoded: + pass + + encoded = _Encoded() + encoded.input_ids = [[2] for _ in texts] + return encoded + + +def test_source_vocab_outside_the_work_directory_is_not_disturbed(tmp_path, monkeypatch): + """A tokenizer whose sentencepiece source lives elsewhere (e.g. the subtree + convert_to_fast_tokenizer created) is copied into the fresh work directory + and patched there; the original source is left untouched. + """ + loaded = _stub_auto_tokenizer(monkeypatch) + location = str(tmp_path / "_unsloth_sentencepiece_temp") + subdir = os.path.join(location, "some_model") + os.makedirs(subdir, exist_ok = True) + + pieces = [("", 0.0, CONTROL), ("a", -1.0, NORMAL), ("", 0.0, CONTROL)] + source_model = os.path.join(subdir, "tokenizer.model") + with open(source_model, "wb") as f: + f.write(_spm_bytes(pieces)) + + old = _CopyFromSubdirTokenizer(source_model) + new = _FakeTokenizer("new") + tok = fix_sentencepiece_tokenizer(old, new, {"": "<|im_end|>"}, temporary_location = location) + + assert _read_pieces(source_model) == [ + "", + "a", + "", + ], "the original source vocab was modified" + assert "<|im_end|>" in _read_pieces(f"{loaded[-1]}/tokenizer.model") + assert tok is not None + + +def test_swap_mapping_swaps_both_pieces_without_duplicating(tmp_path, monkeypatch): + """When the caller swaps eos and stop_word in the fast JSON it must pass both + directions here; a one-way mapping would leave two stop_word pieces and no eos. + """ + loaded = _stub_auto_tokenizer(monkeypatch) + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + pieces = [("", 0.0, CONTROL), ("<|im_end|>", -1.0, NORMAL), ("", 0.0, CONTROL)] + old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"": 2, "<|im_end|>": 1}) + new = _FakeTokenizer("new") + + tok = fix_sentencepiece_tokenizer( + old, new, {"": "<|im_end|>", "<|im_end|>": ""}, temporary_location = location + ) + + result = _read_pieces(f"{loaded[-1]}/tokenizer.model") + assert result.count("<|im_end|>") == 1 and result.count("") == 1, result + assert tok is not None + + +def test_only_applied_mappings_are_patched(tmp_path, monkeypatch): + """When the caller skips a mapping whose target already exists, it must not + pass that mapping here, or the skipped source token gets renamed anyway and + duplicates the existing target in the model. + """ + loaded = _stub_auto_tokenizer(monkeypatch) + location = str(tmp_path / "_unsloth_sentencepiece_temp") + + pieces = [ + ("", 0.0, CONTROL), + ("aa", -1.0, NORMAL), + ("bb", -1.0, NORMAL), + ("X", -1.0, NORMAL), + ] + old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"aa": 1, "bb": 2}) + new = _FakeTokenizer("new") + + # Caller skipped aa->X (X already exists) and applied bb->Y, so only bb->Y is passed. + tok = fix_sentencepiece_tokenizer(old, new, {"bb": "Y"}, temporary_location = location) + + result = _read_pieces(f"{loaded[-1]}/tokenizer.model") + assert result.count("X") == 1 and "Y" in result and "aa" in result, result + assert tok is not None diff --git a/tests/saving/test_prewarm_base_model_hub_cache.py b/tests/saving/test_prewarm_base_model_hub_cache.py index cbb52863ba..4269f9d61c 100644 --- a/tests/saving/test_prewarm_base_model_hub_cache.py +++ b/tests/saving/test_prewarm_base_model_hub_cache.py @@ -4,7 +4,7 @@ """Regression tests for #6890: repeated base-model downloads across checkpoint exports. merge_and_overwrite_lora downloads missing 16-bit shards with hf_hub_download(local_dir), -which never populates the persistent HF hub cache; a temporary merge directory (Studio +which never populates the persistent HF hub cache; a temporary merge directory (Unsloth GGUF exports delete it) means every checkpoint export re-downloads the full base model. _prewarm_base_model_hub_cache snapshot-downloads the base into the hub cache first so the zoo's cache-copy fast path is hit on later exports. diff --git a/tests/sh/test_previous_torch_pin.sh b/tests/sh/test_previous_torch_pin.sh new file mode 100644 index 0000000000..253ede8a27 --- /dev/null +++ b/tests/sh/test_previous_torch_pin.sh @@ -0,0 +1,101 @@ +#!/bin/bash +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 +# Unit tests for install.sh's _previous_torch_pin, which keeps the previous +# venv's torch release on a re-run (curl | sh over an existing install) instead +# of silently moving the user to a newer release. Helpers are extracted from +# install.sh and sourced. +set -e + +SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" +INSTALL_SH="$SCRIPT_DIR/../../install.sh" +PASS=0 +FAIL=0 + +# Extract _previous_torch_pin and its dependencies _torch_flavor_tag and +# _torch_release_in_window. +_FUNC_FILE=$(mktemp) +{ + sed -n '/^_torch_flavor_tag()/,/^}/p' "$INSTALL_SH" + echo "" + sed -n '/^_torch_release_in_window()/,/^}/p' "$INSTALL_SH" + echo "" + sed -n '/^_previous_torch_pin()/,/^}/p' "$INSTALL_SH" +} > "$_FUNC_FILE" +# shellcheck disable=SC1090 +. "$_FUNC_FILE" +rm -f "$_FUNC_FILE" + +assert_eq() { + _label="$1"; _expected="$2"; _actual="$3" + if [ "$_actual" = "$_expected" ]; then + echo " PASS: $_label"; PASS=$((PASS + 1)) + else + echo " FAIL: $_label (expected '$_expected', got '$_actual')"; FAIL=$((FAIL + 1)) + fi +} + +unset UNSLOTH_TORCH_UPGRADE + +echo "=== _previous_torch_pin: matching flavor keeps the release ===" +assert_eq "cu126 wheel on cu126 leaf" "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cu126' 'cu126' 'torch>=2.4,<2.12.0')" +assert_eq "cu130 wheel on cu130 leaf" "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cu130' 'cu130' 'torch>=2.4,<2.12.0')" +assert_eq "cpu wheel on cpu leaf" "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cpu' 'cpu' 'torch>=2.4,<2.12.0')" +assert_eq "untagged wheel on cpu leaf" "torch==2.10.0" "$(_previous_torch_pin '2.10.0' 'cpu' 'torch>=2.4,<2.12.0')" +assert_eq "local suffix stripped" "torch==2.9.1" "$(_previous_torch_pin '2.9.1+cu128' 'cu128' 'torch>=2.4,<2.12.0')" + +echo "=== _previous_torch_pin: flavor change installs the new build ===" +assert_eq "cu126 wheel on cu130 leaf" "" "$(_previous_torch_pin '2.10.0+cu126' 'cu130' 'torch>=2.4,<2.12.0')" +assert_eq "cpu wheel on cu126 leaf" "" "$(_previous_torch_pin '2.10.0+cpu' 'cu126' 'torch>=2.4,<2.12.0')" +assert_eq "cu126 wheel on cpu leaf" "" "$(_previous_torch_pin '2.10.0+cu126' 'cpu' 'torch>=2.4,<2.12.0')" + +echo "=== _previous_torch_pin: rocm and unknown leaves never pin ===" +assert_eq "rocm7.2 leaf keeps its floor" "" "$(_previous_torch_pin '2.11.0+rocm7.2' 'rocm7.2' 'torch>=2.4,<2.12.0')" +assert_eq "gfx leaf keeps its floor" "" "$(_previous_torch_pin '2.11.0+rocm7.2' 'gfx120X-all' 'torch>=2.4,<2.12.0')" +assert_eq "unknown mirror leaf" "" "$(_previous_torch_pin '2.10.0+cu126' 'simple' 'torch>=2.4,<2.12.0')" + +echo "=== _previous_torch_pin: probe noise never becomes a pin ===" +assert_eq "empty version" "" "$(_previous_torch_pin '' 'cu126' 'torch>=2.4,<2.12.0')" +assert_eq "garbage version" "" "$(_previous_torch_pin 'not-a-version' 'cpu' 'torch>=2.4,<2.12.0')" +assert_eq "traceback fragment" "" "$(_previous_torch_pin "ModuleNotFoundError: No module named 'torch'" 'cpu' 'torch>=2.4,<2.12.0')" + +echo "=== _previous_torch_pin: out-of-window releases never pin ===" +assert_eq "2.3.x below the cu floor" "" "$(_previous_torch_pin '2.3.1+cu118' 'cu118' 'torch>=2.4,<2.12.0')" +assert_eq "2.12.x above the cu ceiling" "" "$(_previous_torch_pin '2.12.0+cu130' 'cu130' 'torch>=2.4,<2.12.0')" +assert_eq "floor boundary 2.4.0 kept" "torch==2.4.0" "$(_previous_torch_pin '2.4.0+cu126' 'cu126' 'torch>=2.4,<2.12.0')" +assert_eq "ceiling-adjacent 2.11.x kept" "torch==2.11.1" "$(_previous_torch_pin '2.11.1+cu130' 'cu130' 'torch>=2.4,<2.12.0')" +assert_eq "cpu window excludes 2.11.x" "" "$(_previous_torch_pin '2.11.0+cpu' 'cpu' 'torch>=2.4,<2.11.0')" +assert_eq "mac floor excludes 2.5.x" "" "$(_previous_torch_pin '2.5.1' 'cpu' 'torch>=2.6,<2.11.0')" +assert_eq "malformed window never pins" "" "$(_previous_torch_pin '2.10.0+cu126' 'cu126' 'torch')" +assert_eq "empty window never pins" "" "$(_previous_torch_pin '2.10.0+cu126' 'cu126' '')" + +echo "=== _torch_release_in_window ===" +assert_eq "in window" "yes" "$(_torch_release_in_window '2.10.0' 'torch>=2.4,<2.12.0')" +assert_eq "at floor" "yes" "$(_torch_release_in_window '2.4.0' 'torch>=2.4,<2.12.0')" +assert_eq "below floor" "no" "$(_torch_release_in_window '2.3.1' 'torch>=2.4,<2.12.0')" +assert_eq "at ceiling" "no" "$(_torch_release_in_window '2.12.0' 'torch>=2.4,<2.12.0')" +assert_eq "next major" "no" "$(_torch_release_in_window '3.0.0' 'torch>=2.4,<2.12.0')" +assert_eq "patch-level floor" "yes" "$(_torch_release_in_window '2.11.5' 'torch>=2.11.0,<2.12.0')" +assert_eq "no ceiling -> no" "no" "$(_torch_release_in_window '2.10.0' 'torch>=2.4')" +assert_eq "garbage minor -> no" "no" "$(_torch_release_in_window '2.x' 'torch>=2.4,<2.12.0')" + +echo "=== _previous_torch_pin: UNSLOTH_TORCH_UPGRADE=1 opts out ===" +assert_eq "upgrade env set" "" "$(UNSLOTH_TORCH_UPGRADE=1 _previous_torch_pin '2.10.0+cu126' 'cu126' 'torch>=2.4,<2.12.0')" +assert_eq "upgrade env 0" "torch==2.10.0" "$(UNSLOTH_TORCH_UPGRADE=0 _previous_torch_pin '2.10.0+cu126' 'cu126' 'torch>=2.4,<2.12.0')" + +echo "=== install.sh wiring ===" +# The probe must run against the OLD venv, before it is moved aside for rollback. +_probe_line=$(grep -n '_PREV_TORCH_VER=\$(' "$INSTALL_SH" | head -1 | cut -d: -f1) +_move_line=$(grep -n '_start_studio_venv_replacement "\$VENV_DIR"' "$INSTALL_SH" | head -1 | cut -d: -f1) +assert_eq "probe exists" "yes" "$([ -n "$_probe_line" ] && echo yes)" +assert_eq "probe before venv replacement" "yes" "$([ -n "$_probe_line" ] && [ -n "$_move_line" ] && [ "$_probe_line" -lt "$_move_line" ] && echo yes)" +# A kept release that vanished from the index must fall back to the supported range. +assert_eq "resolve-failure fallback wired" "yes" "$(grep -q 'TORCH_CONSTRAINT="\$_PREV_FALLBACK_CONSTRAINT"' "$INSTALL_SH" && echo yes)" +assert_eq "pin gated on SKIP_TORCH" "yes" "$(grep -q 'if \[ "\$SKIP_TORCH" = false \]; then' "$INSTALL_SH" && echo yes)" + +echo "" +if [ "$FAIL" -gt 0 ]; then + echo "$FAIL check(s) FAILED" + exit 1 +fi +echo "All $PASS checks passed" diff --git a/tests/sh/test_torch_constraint.sh b/tests/sh/test_torch_constraint.sh index 293a709360..d60dfc9f90 100644 --- a/tests/sh/test_torch_constraint.sh +++ b/tests/sh/test_torch_constraint.sh @@ -108,6 +108,20 @@ assert_eq "\$TORCH_CONSTRAINT used in pip install" "yes" "$_has_var" _hardcoded=$(grep -c '"torch>=2.4,<2.11.0"' "$INSTALL_SH" || true) assert_eq "hardcoded torch>=2.4 appears exactly once" "1" "$_hardcoded" +# A fresh CUDA install widens the ceiling to <2.12.0 so cu12x/cu13x land torch +# 2.11.x (matches the base image and _CUDA_TORCH_PKG_SPEC). +_cuda_widen=$(grep -c 'TORCH_CONSTRAINT="torch>=2.4,<2.12.0"' "$INSTALL_SH" || true) +assert_eq "CUDA TORCH_CONSTRAINT widened to <2.12.0" "1" "$_cuda_widen" + +# Widening keys off the final leaf (_torch_index_leaf), not the full URL, so a +# mirror base path with cu*/rocm7.2 but a cpu/older-rocm leaf is not mis-widened. +_cuda_case=$(grep -c 'cu\[0-9\]\*)' "$INSTALL_SH" || true) +_has_cuda_case=$([ "$_cuda_case" -ge 1 ] && echo "yes" || echo "no") +assert_eq "cu* index case adjusts TORCH_CONSTRAINT" "yes" "$_has_cuda_case" +_leaf_case=$(grep -c 'case "\$_torch_index_leaf" in' "$INSTALL_SH" || true) +_has_leaf_constraint=$([ "$_leaf_case" -ge 2 ] && echo "yes" || echo "no") +assert_eq "constraint case anchors on _torch_index_leaf" "yes" "$_has_leaf_constraint" + echo "" echo "=== Structural: tokenizers in no-torch-runtime.txt ===" diff --git a/tests/sh/test_unsloth_torch_override.sh b/tests/sh/test_unsloth_torch_override.sh new file mode 100644 index 0000000000..7e8e3f5b5b --- /dev/null +++ b/tests/sh/test_unsloth_torch_override.sh @@ -0,0 +1,131 @@ +#!/bin/bash +# SPDX-License-Identifier: AGPL-3.0-only +# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 +# Tests for the torch-trio --overrides guard on the Step-2 unsloth installs in +# install.sh. A released unsloth wheel can pin an older torch (2026.7.2 declares +# torch<2.11.0); without the overrides file a with-deps PyPI resolve downgrades +# the trio Step 1 installed, and the flavor guard misses it (PyPI's torch 2.10 +# default is itself cu128-flavored). Same assertion pattern as test_torch_constraint.sh. +set -e + +SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" +INSTALL_SH="$SCRIPT_DIR/../../install.sh" +PASS=0 +FAIL=0 + +assert_true() { + _label="$1"; _ok="$2" + if [ "$_ok" = "0" ]; then + echo " PASS: $_label" + PASS=$((PASS + 1)) + else + echo " FAIL: $_label" + FAIL=$((FAIL + 1)) + fi +} + +echo "=== test_unsloth_torch_override ===" + +# 1. Every with-deps unsloth install carries the overrides expansion (local, +# generic, migrated); the --no-deps no-torch paths need no guard. +_local_block=$(grep -A2 '"install unsloth (local)"' "$INSTALL_SH") +printf '%s' "$_local_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"' +assert_true "local (with-deps) unsloth install passes --overrides" "$?" + +_generic_block=$(grep -A2 '"install unsloth" uv pip install' "$INSTALL_SH") +printf '%s' "$_generic_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"' +assert_true "generic (with-deps) unsloth install passes --overrides" "$?" + +_migrated_block=$(grep -A3 '"install unsloth (migrated)"' "$INSTALL_SH") +printf '%s' "$_migrated_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"' +assert_true "migrated (with-deps) unsloth install passes --overrides" "$?" + +_no_torch_block=$(grep -A2 '"install unsloth (no-torch)"' "$INSTALL_SH") +if printf '%s' "$_no_torch_block" | grep -q -- '--overrides'; then _rc=1; else _rc=0; fi +assert_true "no-torch (--no-deps) unsloth install has no overrides" "$_rc" + +_migrated_nt_block=$(grep -A2 '"install unsloth (migrated no-torch)"' "$INSTALL_SH") +if printf '%s' "$_migrated_nt_block" | grep -q -- '--overrides'; then _rc=1; else _rc=0; fi +assert_true "migrated no-torch (--no-deps) unsloth install has no overrides" "$_rc" + +# 2. The overrides file is only built when SKIP_TORCH=false. +grep -B2 '_torch_trio_pins=\$(' "$INSTALL_SH" | grep -q 'SKIP_TORCH" = false' +assert_true "overrides file build is gated on SKIP_TORCH=false" "$?" + +# 3. The pin-collection snippet emits exact ==pins for the installed trio (run +# the embedded python against this test's interpreter). +_snippet=$(sed -n '/_torch_trio_pins=\$("\$_VENV_PY" -c "/,/^" 2>\/dev\/null)/p' "$INSTALL_SH" \ + | sed '1s/.*-c "//' | sed '$d') +_out=$(python3 -c "$_snippet" 2>&1) || true +# torch may or may not be importable on the test host; the snippet must not +# crash and every line it does emit must be an exact pkg==version pin. +if [ -n "$_out" ]; then + printf '%s\n' "$_out" | grep -vqE '^(torch|torchvision|torchaudio)==.+$' && _rc=1 || _rc=0 +else + _rc=0 +fi +assert_true "pin snippet emits only exact trio ==pins (or nothing)" "$_rc" + +# 4. The temp overrides file is cleaned up after Step 2. +grep -q 'rm -f "\$_UNSLOTH_TORCH_OVERRIDES"' "$INSTALL_SH" +assert_true "overrides temp file is removed after the unsloth installs" "$?" + +# 5. Any UV_OVERRIDE env file is folded in (the CLI --overrides flag would +# otherwise replace it, dropping e.g. the macOS arm64 darwin overrides). +grep -q 'for _ov_file in \${UV_OVERRIDE:-}' "$INSTALL_SH" +assert_true "UV_OVERRIDE env files are merged into the overrides file" "$?" + +# 6. The EXIT trap also removes the overrides file, so a failed Step 2 (set -e +# fires before the normal-path rm) cannot leak it. +sed -n '/_on_install_exit() {/,/^}/p' "$INSTALL_SH" \ + | grep -q 'rm -f "\$_UNSLOTH_TORCH_OVERRIDES"' +assert_true "EXIT trap removes the overrides temp file on failure" "$?" + +# 7. The UV_OVERRIDE fold filters inherited files instead of cat-ing them (run +# the extracted awk program on sample files): (a) inherited torch-trio lines +# are dropped so the generated exact pins win (uv intersects duplicates); +# (b) every line is newline-terminated so an unterminated file cannot join +# two requirements into one. +_awk_prog=$(sed -n "s/.*awk '\(.*\)' \"\$_ov_file\".*/\1/p" "$INSTALL_SH") +[ -n "$_awk_prog" ] +assert_true "UV_OVERRIDE fold uses the trio-filtering awk program" "$?" + +_ov_dir=$(mktemp -d) +printf '%s' 'transformers>=4.57.6' > "$_ov_dir/ov1.txt" # no trailing newline +cat > "$_ov_dir/ov2.txt" <<'EOF' +# comment survives +torch<2.11.0 +torchvision==0.25.0 +torchaudio!=2.11.0 +torchmetrics==1.0 +anyio<4.14.0 +EOF +_merged="$_ov_dir/merged.txt" +printf '%s\n' 'torch==2.11.0+cu128' > "$_merged" +for _f in "$_ov_dir/ov1.txt" "$_ov_dir/ov2.txt"; do + awk "$_awk_prog" "$_f" >> "$_merged" +done + +grep -qx 'transformers>=4.57.6' "$_merged" +assert_true "no-trailing-newline override stays a separate requirement line" "$?" + +if grep -qx 'torchmetrics==1.0' "$_merged" && grep -qx 'anyio<4.14.0' "$_merged"; then + _rc=0 +else + _rc=1 +fi +assert_true "unrelated inherited overrides are preserved" "$_rc" + +if grep -qE '^(torch|torchvision|torchaudio)([[:space:]<>=!~;@[]|$)' "$_merged" \ + && [ "$(grep -cE '^(torch|torchvision|torchaudio)([[:space:]<>=!~;@[]|$)' "$_merged")" != "1" ]; then + _rc=1 +else + _rc=0 +fi +grep -qx 'torch==2.11.0+cu128' "$_merged" || _rc=1 +assert_true "inherited torch-trio lines are dropped; generated pin wins" "$_rc" +rm -rf "$_ov_dir" + +echo "" +echo "Results: $PASS passed, $FAIL failed" +[ "$FAIL" -eq 0 ] || exit 1 diff --git a/tests/studio/_playwright_robust.py b/tests/studio/_playwright_robust.py index a4590066d4..b2f6df751e 100644 --- a/tests/studio/_playwright_robust.py +++ b/tests/studio/_playwright_robust.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 -"""Shared CI-runner workarounds for the Studio Playwright tests (Chromium flags, +"""Shared CI-runner workarounds for the Unsloth Playwright tests (Chromium flags, view-transition killer, page recovery, post-action response wait). Imported directly by the standalone scripts; does NOT depend on pytest. """ @@ -130,7 +130,7 @@ def wait_for_health( timeout = 3.0, ) last_status, last_body = status, body - # Accept any 200 -- different Studio builds report status differently. + # Accept any 200 -- different Unsloth builds report status differently. if status == 200: if info is not None: info(f"health pre-flight OK: status=200, body keys={list((body or {}).keys())}") diff --git a/tests/studio/install/smoke_test_parallel_studio_home.py b/tests/studio/install/smoke_test_parallel_studio_home.py index 9d840355e1..55b6df5190 100644 --- a/tests/studio/install/smoke_test_parallel_studio_home.py +++ b/tests/studio/install/smoke_test_parallel_studio_home.py @@ -80,7 +80,7 @@ def _launch_backend( env = os.environ.copy() env["HOME"] = str(fake_home) # Pin UNSLOTH_STUDIO_HOME and clear the alias so the child can't inherit a - # Studio root from the caller's shell and resolve to the wrong install. + # Unsloth root from the caller's shell and resolve to the wrong install. env["UNSLOTH_STUDIO_HOME"] = str(studio_home) env.pop("STUDIO_HOME", None) # Popen dups stdout into the child, so closing the parent's handle here is safe. diff --git a/tests/studio/install/test_launch_studio_launcher.py b/tests/studio/install/test_launch_studio_launcher.py index a7396aaf5d..1c12024f6d 100644 --- a/tests/studio/install/test_launch_studio_launcher.py +++ b/tests/studio/install/test_launch_studio_launcher.py @@ -1,4 +1,4 @@ -"""Guard install.ps1's Studio launcher against the AV-heuristic shape (Kaspersky +"""Guard install.ps1's Unsloth launcher against the AV-heuristic shape (Kaspersky HEUR:Trojan.VBS.Agent.gen): a WScript .vbs spawning a hidden ExecutionPolicy-Bypass PowerShell. The shortcut must stay windowless via powershell.exe -WindowStyle Hidden over launch-studio.ps1, never a .vbs/WScript.Shell.Run wrapper, and any pre-existing .vbs must be deleted on upgrade.""" diff --git a/tests/studio/install/test_managed_node_runtime.py b/tests/studio/install/test_managed_node_runtime.py index 0dbc6788ca..17c7e3e60f 100644 --- a/tests/studio/install/test_managed_node_runtime.py +++ b/tests/studio/install/test_managed_node_runtime.py @@ -3,7 +3,7 @@ """Tests for the runtime managed-Node resolver (studio/backend/utils/node_runtime.py). -The Studio frontend installer may provision an isolated Node under +The Unsloth frontend installer may provision an isolated Node under ``/node`` that is never added to the user's PATH. The backend OXC validator must still find a usable Node at runtime: a version-adequate system Node, else the managed isolated one. These tests pin that resolution and the diff --git a/tests/studio/install/test_pr5940_followups.py b/tests/studio/install/test_pr5940_followups.py index ac6a96167a..d6dc8b2f8e 100644 --- a/tests/studio/install/test_pr5940_followups.py +++ b/tests/studio/install/test_pr5940_followups.py @@ -420,7 +420,7 @@ def test_ps_installers_gate_amd_smi_on_windows(): assert ( "UNSLOTH_SETUP_PYTHON" in text ), f"{ps.name} venv-internal check must seed the venv root from UNSLOTH_SETUP_PYTHON" - # A custom Studio home moves the venv off the default path; it must be + # A custom Unsloth home moves the venv off the default path; it must be # seeded too or its hipInfo escapes the filter and reopens the gate. assert ( "UNSLOTH_STUDIO_HOME" in text @@ -429,7 +429,7 @@ def test_ps_installers_gate_amd_smi_on_windows(): @pytest.mark.parametrize("ps", [_INSTALL_PS1, _SETUP_PS1], ids = ["install.ps1", "setup.ps1"]) def test_ps_venv_probe_expands_tilde_for_custom_studio_home(ps): - # The probe seeds the venv root from a custom Studio home; a ~\studio form + # The probe seeds the venv root from a custom Unsloth home; a ~\studio form # must expand to USERPROFILE like the canonical resolver, else GetFullPath # keeps the literal ~ (cwd-relative) and the hipInfo escapes the filter. text = ps.read_text(encoding = "utf-8") @@ -439,7 +439,7 @@ def test_ps_venv_probe_expands_tilde_for_custom_studio_home(ps): block = text[i:j] assert "USERPROFILE" in block and ".Substring(1)" in block, ( f"{ps.name}: the venv-internal probe must expand a leading ~ in the custom " - "Studio home before seeding the venv root (mirroring the canonical resolver)" + "Unsloth home before seeding the venv root (mirroring the canonical resolver)" ) # The ~ expansion must be guarded on a non-empty USERPROFILE; otherwise # Join-Path $env:USERPROFILE throws on a service/SYSTEM account with no profile, diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py index 5cabf41f57..e7ac0ec82d 100644 --- a/tests/studio/install/test_rocm_support.py +++ b/tests/studio/install/test_rocm_support.py @@ -360,7 +360,7 @@ class TestRuntimePatterns: install_kind = "windows-hip", ) patterns = runtime_patterns_for_choice(choice) - # Narrowed from "*.exe" to the two binaries Studio actually invokes. + # Narrowed from "*.exe" to the two binaries Unsloth actually invokes. assert "llama-server.exe" in patterns assert "llama-quantize.exe" in patterns assert "*.dll" in patterns @@ -378,7 +378,7 @@ class TestRuntimePatterns: assert "lib*.dylib" in patterns def test_diffusion_visual_server_kept(self): - # The DiffusionGemma visual-server must survive the prune so Studio can + # The DiffusionGemma visual-server must survive the prune so Unsloth can # serve DiffusionGemma GGUFs natively. for kind, name in ( ("linux-cuda", "llama-diffusion-gemma-visual-server"), diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py index 7575dea581..e372a3bbeb 100644 --- a/tests/studio/install/test_selection_logic.py +++ b/tests/studio/install/test_selection_logic.py @@ -270,13 +270,13 @@ def mock_windows_runtime(monkeypatch, lines): # =========================================================================== -# Studio run.py localhost warning +# Unsloth run.py localhost warning # =========================================================================== class TestStudioLocalhostIpv6Warning: def _prepare_loopback(self, run_module, monkeypatch): - # Studio confirmed answering on the IPv4 loopback. + # Unsloth confirmed answering on the IPv4 loopback. monkeypatch.setattr( run_module, "_working_local_url", @@ -327,7 +327,7 @@ class TestStudioLocalhostIpv6Warning: assert "http://localhost:8888" in captured.out def test_ipv6_listener_does_not_suppress_warning(self, monkeypatch): - # A process on ::1 is NOT Studio (binds 127.0.0.1 only), so the warning must + # A process on ::1 is NOT Unsloth (binds 127.0.0.1 only), so the warning must # still fire -- that is exactly when http://localhost opens the wrong service. run_module = load_studio_run_module(monkeypatch) self._prepare_loopback(run_module, monkeypatch) @@ -356,7 +356,7 @@ class TestStudioLocalhostIpv6Warning: assert run_module._localhost_ipv6_mismatch_url("127.0.0.1", port) is None def test_ipv4_not_answering_suppresses_warning(self, monkeypatch): - # Studio not confirmed on 127.0.0.1 -> no warning. + # Unsloth not confirmed on 127.0.0.1 -> no warning. run_module = load_studio_run_module(monkeypatch) monkeypatch.setattr(run_module, "_working_local_url", lambda port: None) self._set_getaddrinfo(monkeypatch, [self._ipv6()]) @@ -3668,7 +3668,7 @@ class TestCpuFallback: # =========================================================================== -@pytest.mark.skipif(sys.platform == "win32", reason = "bash-only Studio installer tests") +@pytest.mark.skipif(sys.platform == "win32", reason = "bash-only Unsloth installer tests") class TestCudaDriverToolkitMismatchMessage: _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh" _SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1" @@ -3873,7 +3873,7 @@ class TestCudaDriverToolkitMismatchMessage: "or install a CUDA $driverMajor.x toolkit." in source ) assert ( - "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." + "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit." ) in source assert ( "Write-CudaDriverToolkitMismatch -ToolkitVersion $IncompatibleToolkit " diff --git a/tests/studio/playwright_chat_ime_i18n.py b/tests/studio/playwright_chat_ime_i18n.py index 9c01e95fd4..5bebf0a9e8 100644 --- a/tests/studio/playwright_chat_ime_i18n.py +++ b/tests/studio/playwright_chat_ime_i18n.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 -"""Studio chat composer IME + multilingual regression smoke. +"""Unsloth chat composer IME + multilingual regression smoke. Covers: stuck IME composition (#5318 / PR #5327), multilingual paste round-trip, stuck compositionend (#5546), and Mac input-method switch recovery (keydown/blur). diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py index 065ba7a745..35b18756ff 100644 --- a/tests/studio/playwright_chat_ui.py +++ b/tests/studio/playwright_chat_ui.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 -"""Comprehensive Studio chat UI test, run locally + in CI.""" +"""Comprehensive Unsloth chat UI test, run locally + in CI.""" import json import os @@ -156,7 +156,7 @@ with sync_playwright() as p: # pointer events and break Playwright's actionability check. reduced_motion = "reduce", ) - # Hard-disable CSS view-transitions: Studio's theme toggle + sidebar + # Hard-disable CSS view-transitions: Unsloth's theme toggle + sidebar # collapse run startViewTransition() which can leave intercepting # pointer events for a beat after each route swap. See _playwright_robust.py. install_view_transition_killer(ctx) @@ -477,7 +477,7 @@ with sync_playwright() as p: fail(f"/api/inference/load returned {load_resp['status']}: {load_resp.get('body')!r}") info(f"loaded model: {(load_resp['body'] or {}).get('display_name')}") - # Studio caches model state in zustand; reload so the composer picks + # Unsloth caches model state in zustand; reload so the composer picks # up the loaded model. page.reload() composer = page.locator('textarea[aria-label="Message input"]') @@ -493,7 +493,7 @@ with sync_playwright() as p: # (app-sidebar.tsx) -- as stable as anything in the codebase. picker_btn = page.locator('[data-tour="chat-model-selector"]').first if picker_btn.count() == 0: - # Fall back to text-based locators for older Studio builds. + # Fall back to text-based locators for older Unsloth builds. picker_btn = page.locator( 'button:has-text("gemma-3-270m"), ' 'button:has-text("Gemma 3"), ' @@ -893,7 +893,7 @@ with sync_playwright() as p: if len(observed) < 3: soft_fail(f"theme toggle ran only {len(observed)} cycle(s), expected 3") # Don't strict-fail on both polarities: the runner's - # prefers-color-scheme + Studio's "system" default can collapse + # prefers-color-scheme + Unsloth's "system" default can collapse # to one polarity even when .dark toggles correctly. The 3-cycle # completion above is the real invariant. if light_seen and dark_seen: diff --git a/tests/studio/playwright_extra_ui.py b/tests/studio/playwright_extra_ui.py index 209a8a06f1..dde6c5d635 100644 --- a/tests/studio/playwright_extra_ui.py +++ b/tests/studio/playwright_extra_ui.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 -"""Studio extra-UI Playwright test: Compare tab, Recipes editor, /export, /studio, Settings tabs.""" +"""Unsloth extra-UI Playwright test: Compare tab, Recipes editor, /export, /studio, Settings tabs.""" import json import os @@ -90,11 +90,11 @@ with sync_playwright() as p: ) install_view_transition_killer(ctx) page = ctx.new_page() - # 60s default for slow macos-14 --single-process Chromium (second Studio boot of the job). + # 60s default for slow macos-14 --single-process Chromium (second Unsloth boot of the job). page.set_default_timeout(60_000) page_errors = [] - # Filter out known-benign React errors (timing artefacts on slow CI runners, not Studio bugs); + # Filter out known-benign React errors (timing artefacts on slow CI runners, not Unsloth bugs); # shared base list lives in _playwright_robust.BENIGN_PAGE_ERROR_PATTERNS. def _on_pageerror(e): msg = str(e) @@ -451,9 +451,9 @@ with sync_playwright() as p: ) # ───────────────────────────────────────────────────── - # 4. Studio training route. + # 4. Unsloth training route. # ───────────────────────────────────────────────────── - step(f"Studio route ({'chat-only redirect' if chat_only else 'tabs + sections'})") + step(f"Unsloth route ({'chat-only redirect' if chat_only else 'tabs + sections'})") page.goto(f"{BASE}/studio") page.wait_for_timeout(1500) shoot("08-studio") diff --git a/tests/studio/run_real_mlx_smoke.py b/tests/studio/run_real_mlx_smoke.py index 275fe7ac57..63bc0dbba9 100644 --- a/tests/studio/run_real_mlx_smoke.py +++ b/tests/studio/run_real_mlx_smoke.py @@ -106,7 +106,7 @@ def _compute_loss_and_grad_norm(model, tokenizer, text: str) -> tuple[float, flo import mlx.nn as nn from mlx.utils import tree_flatten - # Match Studio's text dataset path: no EOS appended behind the user's back. + # Match Unsloth's text dataset path: no EOS appended behind the user's back. ids = list(tokenizer.encode(text)) if len(ids) < 2: raise RuntimeError(f"text too short to compute loss: {len(ids)} tokens") diff --git a/tests/studio/studio_api_smoke.py b/tests/studio/studio_api_smoke.py index 845a9ed021..d30bd11dca 100644 --- a/tests/studio/studio_api_smoke.py +++ b/tests/studio/studio_api_smoke.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 -"""End-to-end Studio API & Auth HTTP integration tests against an externally-booted Studio.""" +"""End-to-end Unsloth API & Auth HTTP integration tests against an externally-booted Unsloth.""" import json import os @@ -571,7 +571,7 @@ EXPECTED_AUTH_ENDPOINTS = [ for method, path in EXPECTED_AUTH_ENDPOINTS: if (method, path) in PUBLIC: continue - # Don't actually shut Studio down: an unauthenticated call must 401/403 before the trigger fires. + # Don't actually shut Unsloth down: an unauthenticated call must 401/403 before the trigger fires. if path == "/api/shutdown": code, _ = http(method, path) if code in (401, 403): @@ -610,6 +610,6 @@ if _failed: sys.exit(1) _emit( "", - "PASS all Studio API & Auth assertions" + "PASS all Unsloth API & Auth assertions" + (f" ({len(_warned)} audit findings logged)" if _warned else ""), ) diff --git a/tests/studio/test_auth_form_input_count.py b/tests/studio/test_auth_form_input_count.py index 4d5d72d20b..75e6cfd1fb 100644 --- a/tests/studio/test_auth_form_input_count.py +++ b/tests/studio/test_auth_form_input_count.py @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -"""Fast source and runtime contracts for Studio's frontend authentication flows. +"""Fast source and runtime contracts for Unsloth's frontend authentication flows. PR #5490 added a third "Current password" input, regressing first-boot UX to three inputs; PR #5545 restores two by rendering it only when BOOTSTRAP is absent. diff --git a/tests/studio/test_chat_title_generation.py b/tests/studio/test_chat_title_generation.py index b568a51400..6a47cfbce4 100644 --- a/tests/studio/test_chat_title_generation.py +++ b/tests/studio/test_chat_title_generation.py @@ -1,7 +1,7 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. -"""Regression checks for Studio chat title generation context.""" +"""Regression checks for Unsloth chat title generation context.""" from __future__ import annotations diff --git a/tests/studio/test_cli_studio_stop_windows.py b/tests/studio/test_cli_studio_stop_windows.py index 778679c73b..2267d7feda 100644 --- a/tests/studio/test_cli_studio_stop_windows.py +++ b/tests/studio/test_cli_studio_stop_windows.py @@ -5,7 +5,7 @@ `stop` once used `os.kill(pid, 0)`, which raises WinError 87 on Windows before reaching taskkill; the fix adds cross-platform `_pid_alive` (tasklist on Windows, -signal-0 elsewhere). AST + mock-only; no real processes, no Studio deps imported. +signal-0 elsewhere). AST + mock-only; no real processes, no Unsloth deps imported. """ import ast diff --git a/tests/studio/test_hardware_dispatch_matrix.py b/tests/studio/test_hardware_dispatch_matrix.py index bccddac967..19619bc318 100644 --- a/tests/studio/test_hardware_dispatch_matrix.py +++ b/tests/studio/test_hardware_dispatch_matrix.py @@ -1,5 +1,5 @@ # SPDX-License-Identifier: AGPL-3.0-only -"""Studio hardware dispatch matrix: spoofs platform/torch/mlx per PROFILES to exercise CUDA/ROCm/XPU/MLX/CPU paths without real hardware.""" +"""Unsloth hardware dispatch matrix: spoofs platform/torch/mlx per PROFILES to exercise CUDA/ROCm/XPU/MLX/CPU paths without real hardware.""" from __future__ import annotations @@ -31,9 +31,9 @@ class HardwareProfile: mps_available: bool # torch.backends.mps.is_available() value expect_is_mlx: bool # unsloth._IS_MLX - expect_device_type: str # Studio DeviceType (uppercased name: "CUDA"/"XPU"/"MLX"/"CPU") - expect_is_rocm: bool # Studio IS_ROCM - expect_apple_silicon: bool # Studio is_apple_silicon() + expect_device_type: str # Unsloth DeviceType (uppercased name: "CUDA"/"XPU"/"MLX"/"CPU") + expect_is_rocm: bool # Unsloth IS_ROCM + expect_apple_silicon: bool # Unsloth is_apple_silicon() extra_notes: str = "" @@ -66,7 +66,7 @@ PROFILES = [ expect_is_rocm = True, expect_apple_silicon = False, extra_notes = "PyTorch ROCm reuses torch.cuda.* over HIP; " - "Studio still uses DeviceType.CUDA but flips IS_ROCM=True.", + "Unsloth still uses DeviceType.CUDA but flips IS_ROCM=True.", ), HardwareProfile( name = "intel_xpu", @@ -154,7 +154,7 @@ def spoof_hardware(monkeypatch): import platform import torch - # platform spoof (used by both the unsloth gate and Studio's helpers) + # platform spoof (used by both the unsloth gate and Unsloth's helpers) monkeypatch.setattr(platform, "system", lambda: profile.system) monkeypatch.setattr(platform, "machine", lambda: profile.machine) @@ -227,7 +227,7 @@ def spoof_hardware(monkeypatch): monkeypatch.setattr(importlib.util, "find_spec", _no_mlx) - # Studio's _has_mlx() does `import mlx.core`, not find_spec; block it + # Unsloth's _has_mlx() does `import mlx.core`, not find_spec; block it # with a meta_path finder that raises ImportError for mlx.*. class _BlockMLXFinder: def find_spec( @@ -266,7 +266,7 @@ def _evaluate_unsloth_is_mlx_gate() -> bool: def _import_studio_hardware_module(): - """Lazy-load Studio's hardware module under the bare-imports layout.""" + """Lazy-load Unsloth's hardware module under the bare-imports layout.""" if str(STUDIO_BACKEND) not in sys.path: sys.path.insert(0, str(STUDIO_BACKEND)) # Fresh import so detect_hardware re-runs under the current spoofs. @@ -290,7 +290,7 @@ def test_unsloth_is_mlx_gate_matches_profile(profile, spoof_hardware): @pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS) def test_studio_detect_hardware_matches_profile(profile, spoof_hardware): - """Studio's detect_hardware() routes to the right DeviceType per profile.""" + """Unsloth's detect_hardware() routes to the right DeviceType per profile.""" spoof_hardware(profile) hw = _import_studio_hardware_module() detected = hw.detect_hardware() @@ -306,7 +306,7 @@ def test_studio_detect_hardware_matches_profile(profile, spoof_hardware): @pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS) def test_studio_is_apple_silicon_matches_profile(profile, spoof_hardware): - """Studio's is_apple_silicon() helper agrees with platform spoof.""" + """Unsloth's is_apple_silicon() helper agrees with platform spoof.""" spoof_hardware(profile) hw = _import_studio_hardware_module() assert hw.is_apple_silicon() is profile.expect_apple_silicon, ( diff --git a/tests/studio/test_is_mlx_dispatch_gate.py b/tests/studio/test_is_mlx_dispatch_gate.py index f31f6d1655..0e5de1b789 100644 --- a/tests/studio/test_is_mlx_dispatch_gate.py +++ b/tests/studio/test_is_mlx_dispatch_gate.py @@ -1,6 +1,6 @@ # SPDX-License-Identifier: AGPL-3.0-only -"""Regression tests for the CUDA-vs-MLX dispatch gates Studio relies on. +"""Regression tests for the CUDA-vs-MLX dispatch gates Unsloth relies on. Two gates: (1) ``unsloth._IS_MLX`` (import-time, delegates to the zoo MLX runtime gate behind a local precheck barrier); (2) @@ -147,7 +147,7 @@ def test_is_mlx_gate_false_on_non_apple_silicon(): def _import_studio_hardware(): - """Lazy import of the Studio hardware module (studio/backend on sys.path).""" + """Lazy import of the Unsloth hardware module (studio/backend on sys.path).""" studio_backend = REPO_ROOT / "studio" / "backend" if str(studio_backend) not in sys.path: sys.path.insert(0, str(studio_backend)) diff --git a/tests/studio/test_llama_cpp_wall_clock_cap.py b/tests/studio/test_llama_cpp_wall_clock_cap.py index f173cfbc55..b7b6917092 100644 --- a/tests/studio/test_llama_cpp_wall_clock_cap.py +++ b/tests/studio/test_llama_cpp_wall_clock_cap.py @@ -1,4 +1,4 @@ -"""Timeout policy checks for Studio's local llama-server path.""" +"""Timeout policy checks for Unsloth's local llama-server path.""" from __future__ import annotations diff --git a/tests/studio/test_locale_root_direction_contract.py b/tests/studio/test_locale_root_direction_contract.py index 1baabdbbca..20d10aa805 100644 --- a/tests/studio/test_locale_root_direction_contract.py +++ b/tests/studio/test_locale_root_direction_contract.py @@ -1,4 +1,4 @@ -"""Regression guard for locale changes affecting the entire Studio layout.""" +"""Regression guard for locale changes affecting the entire Unsloth layout.""" from pathlib import Path diff --git a/tests/studio/test_node_decision.ps1 b/tests/studio/test_node_decision.ps1 index bd5d5c8677..44f3ef0e1e 100644 --- a/tests/studio/test_node_decision.ps1 +++ b/tests/studio/test_node_decision.ps1 @@ -60,7 +60,7 @@ $globalBunOffset = $source.IndexOf('npm install -g bun') Check "NodeSource initialized before SKIP_STUDIO_FRONTEND branch" ( $nodeSourceOffset -ge 0 -and $skipFrontendBranchOffset -ge 0 -and $nodeSourceOffset -lt $skipFrontendBranchOffset ) -Check "custom Studio home validated before Node parent creation" ( +Check "custom Unsloth home validated before Node parent creation" ( $customHomeErrorOffset -ge 0 -and $nodeParentMkdirOffset -ge 0 -and $customHomeErrorOffset -lt $nodeParentMkdirOffset ) Check "bundled Node pins npm prefix and clears NODE_PATH" ( diff --git a/tests/studio/test_studio_gguf_export_script_pin.py b/tests/studio/test_studio_gguf_export_script_pin.py index 1f7e7adaa4..defd0d49d4 100644 --- a/tests/studio/test_studio_gguf_export_script_pin.py +++ b/tests/studio/test_studio_gguf_export_script_pin.py @@ -1,4 +1,4 @@ -"""Studio GGUF export pins convert_hf_to_gguf.py via UNSLOTH_LLAMA_CPP_SCRIPTS_DIR, with a once-per-process warning fallback when unsloth_zoo lacks the local-script resolver.""" +"""Unsloth GGUF export pins convert_hf_to_gguf.py via UNSLOTH_LLAMA_CPP_SCRIPTS_DIR, with a once-per-process warning fallback when unsloth_zoo lacks the local-script resolver.""" from __future__ import annotations diff --git a/tests/studio/test_studio_text_descender_clipping.py b/tests/studio/test_studio_text_descender_clipping.py index 7cad6cacfe..359ca4873a 100644 --- a/tests/studio/test_studio_text_descender_clipping.py +++ b/tests/studio/test_studio_text_descender_clipping.py @@ -1,4 +1,4 @@ -"""Regression guard: Studio text spans must not pair `leading-none` with +"""Regression guard: Unsloth text spans must not pair `leading-none` with `truncate`, which clips glyph descenders (g, p, q, y, j) in visible labels. """ @@ -34,17 +34,22 @@ def test_model_selector_trigger_label_uses_leading_tight(): def test_sidebar_account_block_uses_leading_tight(): src = _read(APP_SIDEBAR) - # Match the account-block parent div regardless of its gap utility; this - # guard is about the leading-* class, not the spacing. - pattern = re.compile( - r'', - ) - matches = pattern.findall(src) + class_names = re.findall(r' subprocess.CompletedProcess: + """Run ``body`` in a fresh interpreter that first imports this directory's + ``conftest`` so it inherits the same GPU-free harness the pytest session + uses (device_type stubs plus torch.cuda probe patches). Without it, + ``import unsloth.registry`` raises ``NotImplementedError`` from + ``unsloth_zoo.device_type`` on no-accelerator CI runners, so the child + would exit non-zero and the test would fail even though the registry code + is correct. A fresh process also keeps each check independent of any + ``register_models()`` calls other tests make on the shared registry. + """ + tests_dir = os.path.dirname(os.path.abspath(__file__)) + prelude = ( + f"import sys; sys.path.insert(0, {tests_dir!r})\n" + "try:\n" + " import conftest # noqa: F401 GPU-free harness on no-accelerator runners\n" + "except Exception:\n" + " pass\n" + ) + return subprocess.run( + [sys.executable, "-c", prelude + body], + capture_output = True, + text = True, + check = False, + ) + + +def test_importing_registry_does_not_register_models(): + """Importing the registry must not populate MODEL_REGISTRY on its own. + + ``_deepseek`` used to call ``register_deepseek_models(...)`` at module + scope, so merely importing ``unsloth.registry`` registered models as an + import side effect, unlike every other family which only registers on + demand. + """ + result = _run_registry_child( + "import unsloth.registry\n" + "from unsloth.registry.registry import MODEL_REGISTRY\n" + "print('REGISTRY_SIZE', len(MODEL_REGISTRY))" + ) + assert result.returncode == 0, ( + f"registry import subprocess exited {result.returncode}\n" + f"stdout:\n{result.stdout}\nstderr:\n{result.stderr}" + ) + size_lines = [line for line in result.stdout.splitlines() if line.startswith("REGISTRY_SIZE")] + assert size_lines == ["REGISTRY_SIZE 0"], result.stdout + result.stderr + + +def test_register_models_registers_no_upstream_originals(): + """``register_models()`` must register each family's ``unsloth``-org models + and must NOT leak upstream vendor "original" models. + + Before the fix, ``_deepseek``'s import-time + ``register_deepseek_models(include_original_model = True)`` set the + ``_IS_DEEPSEEK_*_REGISTERED`` guards, so the later default + ``register_models()`` early-returned for deepseek and its 10 ``deepseek-ai`` + originals leaked permanently (129 -> 139). This asserts the whole registry + is ``unsloth``-org after ``register_models()`` while deepseek is still + registered via the normal path. Runs in a fresh interpreter so it is + independent of other tests' registry mutations. + """ + result = _run_registry_child( + "import unsloth.registry\n" + "from unsloth.registry import register_models\n" + "from unsloth.registry.registry import MODEL_REGISTRY\n" + "register_models()\n" + "orgs = sorted({m.org for m in MODEL_REGISTRY.values()})\n" + "deepseek = [k for k in MODEL_REGISTRY if 'deepseek' in k.lower()]\n" + "print('ORGS', orgs)\n" + "print('NUM_DEEPSEEK', len(deepseek))" + ) + assert result.returncode == 0, ( + f"register_models subprocess exited {result.returncode}\n" + f"stdout:\n{result.stdout}\nstderr:\n{result.stderr}" + ) + out = result.stdout + # Every registered model is unsloth-org: no upstream "original" leaked. + assert "ORGS ['unsloth']" in out, out + result.stderr + # Deepseek is still registered via the normal path, just without originals. + deepseek_lines = [line for line in out.splitlines() if line.startswith("NUM_DEEPSEEK")] + assert deepseek_lines and int(deepseek_lines[0].split()[1]) > 0, out + result.stderr diff --git a/tests/test_raw_text.py b/tests/test_raw_text.py index ba16e0cfc4..18549adfe8 100644 --- a/tests/test_raw_text.py +++ b/tests/test_raw_text.py @@ -295,7 +295,25 @@ def test_smart_chunk_text_single_chunk_no_eos_returns_plain_list(): return True +def test_load_from_file_skips_non_object_json_lines(): + """Non-object .jsonl lines (valid JSON, not dicts) are skipped, not fatal.""" + # "context" contains "text", ["text"] holds it, 42 isn't iterable -- each + # would reach data[field] and raise TypeError without the isinstance guard. + with tempfile.NamedTemporaryFile("w", suffix = ".jsonl", delete = False) as f: + f.write('"context"\n["text", "x"]\n42\n{"text": "keep this"}\n') + path = f.name + try: + text = RawTextDataLoader(None)._read_file_by_format(path, "json_lines") + assert text == "keep this", text + finally: + os.unlink(path) + + print("test_load_from_file_skips_non_object_json_lines passed") + return True + + if __name__ == "__main__": success = test_raw_text_loader() success = test_smart_chunk_text_single_chunk_no_eos_returns_plain_list() and success + success = test_load_from_file_skips_non_object_json_lines() and success sys.exit(0 if success else 1) diff --git a/tests/test_studio_install_workspace_guard.py b/tests/test_studio_install_workspace_guard.py index 89836cfb0d..ea5b34672c 100644 --- a/tests/test_studio_install_workspace_guard.py +++ b/tests/test_studio_install_workspace_guard.py @@ -1,4 +1,4 @@ -"""install.sh/install.ps1 must refuse to rm -rf an existing Studio venv in env-mode without a sentinel.""" +"""install.sh/install.ps1 must refuse to rm -rf an existing Unsloth venv in env-mode without a sentinel.""" from __future__ import annotations @@ -127,7 +127,7 @@ def test_install_ps1_has_matching_env_mode_guard(): ), "install.ps1 must gate Remove-Item $VenvDir on env-mode" assert "share\\studio.conf" in block, "install.ps1 guard must check share\\studio.conf sentinel" assert "bin\\unsloth.exe" in block, "install.ps1 guard must check bin\\unsloth.exe sentinel" - assert "Refusing to delete non-Studio venv" in block + assert "Refusing to delete non-Unsloth venv" in block def test_setup_ps1_has_writability_probe(): @@ -160,7 +160,7 @@ def test_env_mode_blocks_when_bin_unsloth_is_a_directory(tmp_path): capture_output = True, ) assert res.returncode != 0, ( - "directory at bin/unsloth must NOT satisfy the Studio sentinel; " + "directory at bin/unsloth must NOT satisfy the Unsloth sentinel; " f"stdout={res.stdout!r} stderr={res.stderr!r}" ) assert (venv / "important.txt").is_file(), "unrelated workspace data must survive" @@ -205,7 +205,7 @@ def test_install_ps1_sentinel_uses_pathtype_leaf(): def test_setup_ps1_stale_venv_has_env_mode_guard(): - """setup.ps1 stale-venv branch must gate Remove-Item $VenvDir on a custom-root Studio sentinel.""" + """setup.ps1 stale-venv branch must gate Remove-Item $VenvDir on a custom-root Unsloth sentinel.""" src = SETUP_PS1.read_text() idx = src.index("Stale venv detected") block = src[idx : idx + 1500] diff --git a/tests/test_studio_root_resilience.py b/tests/test_studio_root_resilience.py index 0dfb826376..779ff2f3f1 100644 --- a/tests/test_studio_root_resilience.py +++ b/tests/test_studio_root_resilience.py @@ -1,4 +1,4 @@ -"""Studio install-root inference must not crash under hostile filesystem conditions (PermissionError/OSError swallowed; custom root kept when resolve() fails).""" +"""Unsloth install-root inference must not crash under hostile filesystem conditions (PermissionError/OSError swallowed; custom root kept when resolve() fails).""" from __future__ import annotations diff --git a/tests/test_studio_shutdown_thread_wait.py b/tests/test_studio_shutdown_thread_wait.py index 4ec2afc0f0..8299116d9a 100644 --- a/tests/test_studio_shutdown_thread_wait.py +++ b/tests/test_studio_shutdown_thread_wait.py @@ -130,4 +130,4 @@ def test_cli_entrypoints_wait_before_returning_to_shell(): assert ( _calls_shutdown_wait_getattr(tree) >= 3 - ), "Studio CLI terminal paths must wait for the backend thread after requesting shutdown" + ), "Unsloth CLI terminal paths must wait for the backend thread after requesting shutdown" diff --git a/tests/version_compat/test_trl_fake_train_cpu.py b/tests/version_compat/test_trl_fake_train_cpu.py index 4dae696282..2f7b469428 100644 --- a/tests/version_compat/test_trl_fake_train_cpu.py +++ b/tests/version_compat/test_trl_fake_train_cpu.py @@ -158,6 +158,37 @@ except Exception: _MODEL = "hf-internal-testing/tiny-random-LlamaForCausalLM" +def _guard_finite_logits(model): + """Keep the LM head logits finite so GRPO sampling can't crash. + + ``test_grpo_trains_on_cpu`` samples completions from a tiny, *untrained* + random model on CPU. Driven autoregressively -- and nudged by the fake + reward's optimizer step between the two train steps -- such a model can emit + non-finite logits, so ``torch.multinomial`` inside ``generate()`` + intermittently raises "probability tensor contains either `inf`, `nan` or + element < 0". That is a well-known nondeterministic sampling failure, not an + Unsloth/TRL regression: the Trainer already fixes the seed, but CPU reduction + order is not bit-reproducible, so the blow-up still surfaces every so often. + + Sanitize the logits to a finite, bounded range (out of place, so autograd + stays valid) before they reach the sampler. This test asserts the train loop + runs end to end, not the (deliberately meaningless) numerics, so bounding the + logits changes nothing it checks while making the run reliable. + """ + + def _finite_logits_hook(_module, _inputs, output): + logits = getattr(output, "logits", None) + if logits is None: + return output + # nan_to_num maps nan -> 0 and the infinities to large finite values; + # clamp then bounds everything to [-30, 30]. + output.logits = torch.nan_to_num(logits).clamp(-30.0, 30.0) + return output + + model.register_forward_hook(_finite_logits_hook) + return model + + def _load_plain(): """Tiny plain HF model + tokenizer on CPU. Skips (not fails) if the model cannot be fetched -- that is a network/hub issue, not an unsloth regression.""" @@ -233,6 +264,11 @@ def test_grpo_trains_on_cpu(tmp_path): assert GRPOTrainer.__name__ == "UnslothGRPOTrainer", "GRPO patch did not apply" model, tok = _load_plain() + # GRPO is the only canary that autoregressively samples completions, so it is + # the only one that can hit the non-finite-logits multinomial crash. Install + # the guard here (not in _load_plain) so the SFT/DPO canaries keep asserting + # against the model's true, unclamped outputs. + _guard_finite_logits(model) ds = Dataset.from_list([{"prompt": "hi there"}] * 4) cfg = GRPOConfig( output_dir = str(tmp_path / "ci_grpo"), diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py index 2d3674fb04..f47c78ba80 100644 --- a/unsloth/chat_templates.py +++ b/unsloth/chat_templates.py @@ -1929,6 +1929,9 @@ def get_chat_template( string_vocab = tokenizer._tokenizer.to_str() skipped = 0 + # Only mirror applied mappings into the spm model; a skipped one would + # rename a piece the JSON never changed and desync the two. + applied_mapping = {} for old_token, new_token in token_mapping.items(): old_count = string_vocab.count(f'"{old_token}"') new_count = string_vocab.count(f'"{new_token}"') @@ -1939,6 +1942,7 @@ def get_chat_template( raise RuntimeError(f"{old_token} was not part of the tokenizer!") else: string_vocab = string_vocab.replace(f'"{old_token}"', f'"{new_token}"') + applied_mapping[old_token] = new_token pass pass @@ -1973,7 +1977,7 @@ def get_chat_template( # Must fix the sentence piece tokenizer since there's no tokenizer.model file! from .tokenizer_utils import fix_sentencepiece_tokenizer - tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,) + tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, applied_mapping,) else: pass @@ -1997,8 +2001,11 @@ def get_chat_template( string_vocab = string_vocab.replace(old_eos_token, temporary_stop_token) string_vocab = string_vocab.replace(stop_word, old_eos_token) string_vocab = string_vocab.replace(temporary_stop_token, stop_word) + # JSON swapped both, so swap both here too; a one-way map leaves two stop_word pieces. + sentencepiece_mapping = { old_eos_token : stop_word, stop_word : old_eos_token, } else: string_vocab = string_vocab.replace(old_eos_token, stop_word) + sentencepiece_mapping = { old_eos_token : stop_word, } pass new_tokenizer = tokenizer._tokenizer.from_str(string_vocab) @@ -2017,9 +2024,8 @@ def get_chat_template( ) # Must fix the sentence piece tokenizer since there's no tokenizer.model file! - token_mapping = { old_eos_token : stop_word, } from .tokenizer_utils import fix_sentencepiece_tokenizer - tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,) + tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, sentencepiece_mapping,) pass else: @@ -2041,7 +2047,7 @@ def get_chat_template( .replace("'assistant'", "'" + mapping["assistant"] + "'") if use_zoo_tokenizer_patch: - # Studio MLX avoids the model-utils tokenizer wrapper because that + # Unsloth MLX avoids the model-utils tokenizer wrapper because that # import path pulls in Torch/GPU-specific modules before MLX training. from unsloth_zoo.tokenizer_utils import patch_tokenizer else: diff --git a/unsloth/dataprep/raw_text.py b/unsloth/dataprep/raw_text.py index 128d966ecd..8623285a25 100644 --- a/unsloth/dataprep/raw_text.py +++ b/unsloth/dataprep/raw_text.py @@ -236,6 +236,10 @@ class RawTextDataLoader: def _extract_text_from_json(self, data): """Extract text from JSON object using common field names.""" + # Skip non-object lines (str/list/number): `field in data` would be a + # substring/membership test, not a key lookup, and `data[field]` raises. + if not isinstance(data, dict): + return "" for field in self._TEXT_FIELDS: if field in data and isinstance(data[field], str): return data[field] diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py index 09de248c7b..5d54815705 100644 --- a/unsloth/import_fixes.py +++ b/unsloth/import_fixes.py @@ -3008,7 +3008,7 @@ def maybe_set_windows_rocm_bnb_version(): No-op unless ALL of: Windows, a real HIP torch build (env hints like HIP_PATH do not count), a ROCm DLL installed, and no explicit user value. - Linux is untouched. Values seeded by Studio's venv sitecustomize.py + Linux is untouched. Values seeded by Unsloth's venv sitecustomize.py (marked ``UNSLOTH_BNB_ROCM_VERSION_SOURCE=sitecustomize``) are redetectable defaults, not overrides; ``UNSLOTH_SKIP_BNB_ROCM_VERSION=1`` opts out and drops a seeded default. Returns the value set, else None. diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py index 1e9cc641e9..7661b0d714 100644 --- a/unsloth/models/loader_utils.py +++ b/unsloth/models/loader_utils.py @@ -821,7 +821,7 @@ def _exclude_rope_inv_freq_from_ddp(model): # ============================================================================= # Offline loading - single source of truth (shared by vision.py, loader.py and -# the Studio exporter). Decide offline ONCE at the load boundary and force it +# the Unsloth exporter). Decide offline ONCE at the load boundary and force it # ONCE around the whole load, so every nested HF call inherits it. # ============================================================================= diff --git a/unsloth/registry/_deepseek.py b/unsloth/registry/_deepseek.py index e29190f0f2..618453fb82 100644 --- a/unsloth/registry/_deepseek.py +++ b/unsloth/registry/_deepseek.py @@ -171,8 +171,6 @@ def _list_deepseek_r1_distill_models(): return distill_models -register_deepseek_models(include_original_model = True) - if __name__ == "__main__": from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info diff --git a/unsloth/save.py b/unsloth/save.py index 7d5774aa97..0e2650b174 100644 --- a/unsloth/save.py +++ b/unsloth/save.py @@ -228,7 +228,7 @@ def _loaded_via_remote_code(obj): Transformers loads auto_map code into the ``transformers_modules`` package, so a ``transformers_modules`` class proves the original load actually ran that remote code - (which the caller's / Studio's consent gate scans at load time). Export paths derive their + (which the caller's / Unsloth's consent gate scans at load time). Export paths derive their reload trust_remote_code from this - the already approved load decision - instead of from a checkpoint's static ``auto_map``: a model that loads with built-in classes must not have its unvetted remote code run when it is re-read during quantization export. Walks PEFT / wrapper @@ -3858,7 +3858,7 @@ def _prewarm_base_model_hub_cache( from huggingface_hub import HfFileSystem, hf_hub_download, snapshot_download # Resolve the cache from the live env like the merge, not huggingface_hub's frozen - # constants: a runtime cache redirect (read-only default, Studio) would else miss (#6890). + # constants: a runtime cache redirect (read-only default, Unsloth) would else miss (#6890). try: from unsloth_zoo.hf_cache import _active_caches _hub_cache = _active_caches()[1] diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py index 3a91ef188d..c7f61288d5 100644 --- a/unsloth/tokenizer_utils.py +++ b/unsloth/tokenizer_utils.py @@ -17,6 +17,9 @@ from transformers.convert_slow_tokenizer import convert_slow_tokenizer from transformers import PreTrainedTokenizerFast import re import os +import shutil +import tempfile +import weakref from transformers.models.llama.modeling_llama import logger from peft import PeftModelForCausalLM import torch @@ -370,13 +373,19 @@ def fix_sentencepiece_tokenizer( if not os.path.exists(temporary_location): os.makedirs(temporary_location) - # Check if tokenizer.model exists - if not os.path.isfile(f"{temporary_location}/tokenizer.model"): - return new_tokenizer + # Fresh per-call subdir so concurrent/repeated calls can't clobber each other's + # tokenizer.model or leak stale files, without deleting anything the caller owns. + temporary_location = tempfile.mkdtemp(prefix = "tokenizer_", dir = temporary_location) # First save the old tokenizer old_tokenizer.save_pretrained(temporary_location) + # Only sentencepiece tokenizers write tokenizer.model, so check after the save. + if not os.path.isfile(f"{temporary_location}/tokenizer.model"): + # new_tokenizer was built in memory and never references this dir, so drop it. + shutil.rmtree(temporary_location, ignore_errors = True) + return new_tokenizer + tokenizer_file = sentencepiece_model_pb2.ModelProto() tokenizer_file.ParseFromString(open(f"{temporary_location}/tokenizer.model", "rb").read()) @@ -414,6 +423,9 @@ def fix_sentencepiece_tokenizer( eos_token = new_tokenizer.eos_token, pad_token = new_tokenizer.pad_token, ) + # vocab_file points here, so the dir must outlive the tokenizer (a later + # save_pretrained copies the patched tokenizer.model from it); reclaim it on GC. + weakref.finalize(tokenizer, shutil.rmtree, temporary_location, ignore_errors = True) return tokenizer @@ -1464,7 +1476,7 @@ def get_tokenizer_info(tokenizer) -> dict: """Return a concise diagnostic summary of a tokenizer instance. Collects key properties into a JSON-safe dict for logging, debugging, or the - Studio UI. Missing attributes fall back to ``None`` rather than raising. + Unsloth UI. Missing attributes fall back to ``None`` rather than raising. Example output:: diff --git a/unsloth_cli/__init__.py b/unsloth_cli/__init__.py index b3831f5314..121b26f03f 100644 --- a/unsloth_cli/__init__.py +++ b/unsloth_cli/__init__.py @@ -81,7 +81,7 @@ app.add_typer(studio_app, name = "studio", help = "Unsloth Studio commands.") app.add_typer( start_app, name = "start", - help = "Start a coding agent (Claude, Codex, OpenClaw, OpenCode, Hermes, Pi) against Studio.", + help = "Start a coding agent (Claude, Codex, OpenClaw, OpenCode, Hermes, Pi) against Unsloth.", ) # Backwards-compatible hidden alias: `unsloth connect` routes to `unsloth start`. app.add_typer( diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py index c3b188710e..551bef4787 100644 --- a/unsloth_cli/_inference.py +++ b/unsloth_cli/_inference.py @@ -18,7 +18,7 @@ _THINK_OPEN = "" _THINK_BLOCK = re.compile(rf"{re.escape(_THINK_OPEN)}.*?", re.DOTALL) _STREAMED_ERROR_PREFIX = "Error: " -# Cloudflare (in front of remote Studio proxies like RunPod) 403s the default +# Cloudflare (in front of remote Unsloth proxies like RunPod) 403s the default # "Python-urllib/X.Y" User-Agent as a bot; send a real one on every request. _USER_AGENT = "unsloth-cli" _MPI_ENV_PAIRS = ( @@ -36,7 +36,7 @@ _no_redirect_opener = None def urlopen_no_redirect(request, timeout): """urlopen that errors on any redirect: following a 3xx would send a bearer token (or accept an identity proof) to a base we never vetted, letting a port - squatter relay a real Studio's response.""" + squatter relay a real Unsloth's response.""" global _no_redirect_opener if _no_redirect_opener is None: import urllib.error @@ -540,7 +540,7 @@ def find_studio_server(timeout: float = 3.0) -> Optional[str]: def is_loopback_url(base: str) -> bool: """True only when *base* resolves to loopback. find_studio_server() trusts a base after only a health probe, so credentials are auto-sent only to loopback - (a local Studio or an SSH tunnel on 127.0.0.1), the targets the auto flows mean.""" + (a local Unsloth or an SSH tunnel on 127.0.0.1), the targets the auto flows mean.""" from urllib.parse import urlparse host = (urlparse(base).hostname or "").lower() @@ -554,7 +554,7 @@ def is_loopback_url(base: str) -> bool: def verify_studio_identity(base: str, timeout: float = 3.0) -> bool: - """Confirm `base` is really this machine's Studio before sending a secret. + """Confirm `base` is really this machine's Unsloth before sending a secret. Send a random nonce to /api/auth/identity and check the returned HMAC against the one computed from the local same-user secret; an endpoint without that @@ -578,7 +578,7 @@ def verify_studio_identity(base: str, timeout: float = 3.0) -> bool: port = parsed.port or (443 if parsed.scheme == "https" else 80) # Resolve to one concrete address and talk to *that* address, then bind the # proof to (address, port). A name like localhost can resolve to a squatter on - # ::1 while the real Studio is on 127.0.0.1; connecting to the resolved IP and + # ::1 while the real Unsloth is on 127.0.0.1; connecting to the resolved IP and # binding to it means a proof relayed from a different address/port won't match. try: ip = socket.getaddrinfo(host, port, type = socket.SOCK_STREAM)[0][4][0] @@ -592,7 +592,7 @@ def verify_studio_identity(base: str, timeout: float = 3.0) -> bool: headers = {"User-Agent": _USER_AGENT, "Host": parsed.netloc}, ) try: - # No redirects: a 302 could relay a real Studio's proof (see urlopen_no_redirect). + # No redirects: a 302 could relay a real Unsloth's proof (see urlopen_no_redirect). # Cap the read: the server is still unverified, so don't trust its length. with urlopen_no_redirect(request, timeout = timeout) as response: proof = json.loads(response.read(65536).decode() or "{}").get("proof") @@ -623,7 +623,7 @@ def _studio_token() -> Optional[str]: class HttpChatBackend: - """Chat against a running Studio server over its OpenAI-compatible API. + """Chat against a running Unsloth server over its OpenAI-compatible API. close() leaves the model loaded on purpose — the next session (or the UI) starts instantly. @@ -666,7 +666,7 @@ class HttpChatBackend: tensor_parallel: bool = False, llama_extra_args: Optional[List[str]] = None, ) -> None: - typer.echo(f"Loading {model} on the Studio server", err = True) + typer.echo(f"Loading {model} on the Unsloth server", err = True) payload = { "model_path": model, "hf_token": hf_token, @@ -769,7 +769,7 @@ def connect_studio_server( tensor_parallel: bool = False, llama_extra_args: Optional[List[str]] = None, ): - """Backend on a running Studio server, or None (caller loads locally).""" + """Backend on a running Unsloth server, or None (caller loads locally).""" base_url = find_studio_server() if not base_url: return None @@ -782,20 +782,20 @@ def connect_studio_server( if not explicit: return None typer.echo( - f"Can't attach to the Studio server at {base_url}: {reason} Run Studio " + f"Can't attach to the Unsloth server at {base_url}: {reason} Run Unsloth " "on this machine, or unset UNSLOTH_STUDIO_URL to load the model locally.", err = True, ) raise typer.Exit(code = 1) # Only hand the self-issued JWT (signed with the local secret) to loopback: a - # remote URL is unverified and a real remote Studio would reject it anyway. + # remote URL is unverified and a real remote Unsloth would reject it anyway. if not is_loopback_url(base_url): return _refuse( - "it isn't a local Studio, so a self-issued token can't " + "it isn't a local Unsloth, so a self-issued token can't " "authenticate to it and must not be sent to it." ) - # Confirm the loopback responder is really our Studio (not a port squatter). + # Confirm the loopback responder is really our Unsloth (not a port squatter). if not verify_studio_identity(base_url): return _refuse( "its identity couldn't be verified (it may be running as a " @@ -803,7 +803,7 @@ def connect_studio_server( ) token = _studio_token() if not token: - return _refuse("couldn't self-issue a Studio token (is Studio set up here?).") + return _refuse("couldn't self-issue an Unsloth token (is Unsloth set up here?).") backend = HttpChatBackend(base_url, token) backend.ensure_loaded( model, diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py index bc4a72f36c..bba5fab08e 100644 --- a/unsloth_cli/commands/chat.py +++ b/unsloth_cli/commands/chat.py @@ -206,7 +206,7 @@ def chat( no_server: bool = typer.Option( False, "--no-server", - help = "Load the model in-process even if a Studio server is running.", + help = "Load the model in-process even if an Unsloth server is running.", ), ): """Start an interactive chat with a model (loads once, stays warm).""" @@ -262,14 +262,14 @@ def chat( llama_extra_args = llama_extra_args, ) - # Prefer a running Studio server: instant starts, model shared with the UI. + # Prefer a running Unsloth server: instant starts, model shared with the UI. chat_backend = ( None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts) ) server_mode = chat_backend is not None if server_mode and should_print: console.print( - "(Studio server connected — model stays warm after /exit)", + "(Unsloth server connected — model stays warm after /exit)", style = "bright_black", ) else: diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py index 84a126163e..524d8fd015 100644 --- a/unsloth_cli/commands/inference.py +++ b/unsloth_cli/commands/inference.py @@ -67,7 +67,7 @@ def inference( no_server: bool = typer.Option( False, "--no-server", - help = "Load the model in-process even if a Studio server is running.", + help = "Load the model in-process even if an Unsloth server is running.", ), ): """Run a single inference using the specified model.""" @@ -85,7 +85,7 @@ def inference( ) raise typer.Exit(code = 1) - # A running Studio server keeps the model warm between runs. Under + # A running Unsloth server keeps the model warm between runs. Under # mlx.launch, every rank must enter the local MLX path instead of rank 0 # alone talking to a server. load_opts = dict( diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py index fa39fdf761..9257a7fbcb 100644 --- a/unsloth_cli/commands/start.py +++ b/unsloth_cli/commands/start.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 -"""`unsloth start` — launch a coding agent against a running Studio server.""" +"""`unsloth start` — launch a coding agent against a running Unsloth server.""" import atexit import contextlib @@ -35,7 +35,7 @@ from unsloth_cli._inference import ( ) start_app = typer.Typer( - help = "Start a coding agent against a running Studio server.", + help = "Start a coding agent against a running Unsloth server.", no_args_is_help = True, context_settings = {"help_option_names": ["-h", "--help"]}, ) @@ -75,14 +75,14 @@ _CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN") # Shared by every agent command; only the config/env/command differ. _MODEL_OPTION = typer.Option( - None, "--model", "-m", help = "Model for the agent; defaults to the one loaded in Studio." + None, "--model", "-m", help = "Model for the agent; defaults to the one loaded in Unsloth." ) _KEY_OPTION = typer.Option( None, "--api-key", envvar = "UNSLOTH_API_KEY", help = ( - "Studio API key. For a local Studio it is minted automatically and " + "Unsloth API key. For a local Unsloth it is minted automatically and " "remembered per server. For a remote server, pass one with --api-key " "(or UNSLOTH_API_KEY); it is remembered for next time." ), @@ -96,7 +96,7 @@ _SERVE_OPTION = typer.Option( True, "--serve/--no-serve", help = ( - "If no Studio server is running, auto-start one for --model and stop it when the " + "If no Unsloth server is running, auto-start one for --model and stop it when the " "agent exits. --no-serve keeps the old behavior of erroring out." ), ) @@ -347,7 +347,7 @@ def _shutdown_auto_served() -> None: global _auto_served_server server, _auto_served_server = _auto_served_server, None if server is not None and server.poll() is None: - typer.echo("Stopping the auto-started Studio server…") + typer.echo("Stopping the auto-started Unsloth server…") _shutdown_server(server) @@ -381,7 +381,7 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess log_path = Path(tempfile.gettempdir()) / f"unsloth-start-server-{os.getpid()}.log" typer.echo( - f"No Studio server at {base}. Starting one for {model} (loading the model can take a while)…" + f"No Unsloth server at {base}. Starting one for {model} (loading the model can take a while)…" ) typer.echo(f"Server log: {log_path}") # 0600: the `unsloth run` banner in this log carries the minted sk-unsloth- key, and @@ -408,16 +408,16 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess if server.poll() is not None: tail = _log_tail(log_path) _shutdown_auto_served() - _fail(f"The Studio server stopped before it was ready. Last log lines:\n{tail}") + _fail(f"The Unsloth server stopped before it was ready. Last log lines:\n{tail}") # `unsloth run` prints the minted key only after the server is up AND the model is # loaded, so it is the fully-ready signal (same contract serve-unsloth-run.sh uses). if _studio_healthy(base) and "sk-unsloth-" in _log_tail(log_path, lines = 400): - typer.echo(f"Studio server ready at {base}.") + typer.echo(f"Unsloth server ready at {base}.") return server time.sleep(2.0) _shutdown_auto_served() _fail( - f"The Studio server didn't become ready within {_SERVER_START_TIMEOUT_S}s. See {log_path}." + f"The Unsloth server didn't become ready within {_SERVER_START_TIMEOUT_S}s. See {log_path}." ) @@ -463,7 +463,7 @@ def _require_studio( return expected, _start_studio_server(expected, model, load or LoadOptions()) model_hint = "" if model else " Pass --model to have it start one for you, or" _fail( - f"No running Studio server found at {expected}.{model_hint} start one with " + f"No running Unsloth server found at {expected}.{model_hint} start one with " "`unsloth studio`, or point UNSLOTH_STUDIO_URL at a remote server." ) @@ -567,12 +567,12 @@ def _key_accepted(base: str, key: str) -> bool: if exc.code in (401, 403): return False _fail( - f"Studio server error while checking an API key ({exc.code}). " + f"Unsloth server error while checking an API key ({exc.code}). " "The server may be starting up or unhealthy; try again shortly." ) except (urllib.error.URLError, TimeoutError) as exc: _fail( - "Couldn't reach the Studio server while checking an API key: " + "Couldn't reach the Unsloth server while checking an API key: " f"{getattr(exc, 'reason', None) or exc}" ) @@ -592,10 +592,10 @@ def _agent_api_key( # UNSLOTH_API_KEY meant for some other server must not fail the # launch: the loopback mint path below is guaranteed to work. # (An explicit key that the fresh server accepts, e.g. one persisted - # in this Studio home's auth db, is still honored above.) + # in this Unsloth home's auth db, is still honored above.) # Replay a key the user saved for *this exact* server first (scoped per base, - # so it only goes back there -- including a remote/SSH-tunnelled Studio whose + # so it only goes back there -- including a remote/SSH-tunnelled Unsloth whose # secret the local handshake can't match). Skip ones the server rejects. for key in _cached_keys(cache, base, "saved"): if _key_accepted(base, key): @@ -608,15 +608,15 @@ def _agent_api_key( if not is_loopback_url(base): _fail( f"No saved API key for {base} and automatic minting only runs against " - "a local Studio. Create an API key in Studio → Settings → API and " + "a local Unsloth. Create an API key in Unsloth → Settings → API and " "pass it with --api-key (it is remembered per server), or set " "UNSLOTH_API_KEY." ) if not verify_studio_identity(base): _fail( - f"Couldn't verify that {base} is your Studio (it may be running as a " + f"Couldn't verify that {base} is your Unsloth (it may be running as a " "different OS user, or another process took the port). Create an API " - "key in Studio → Settings → API and pass it with --api-key, or set " + "key in Unsloth → Settings → API and pass it with --api-key, or set " "UNSLOTH_API_KEY." ) @@ -630,8 +630,8 @@ def _agent_api_key( token = _studio_token() if token is None: _fail( - "Couldn't authenticate with the Studio server automatically. Create " - "an API key in Studio → Settings → API and pass it with --api-key, " + "Couldn't authenticate with the Unsloth server automatically. Create " + "an API key in Unsloth → Settings → API and pass it with --api-key, " "or set UNSLOTH_API_KEY." ) key = _http_json( @@ -664,7 +664,7 @@ def _is_hub_model_id(value: object) -> bool: return False # A hub id is exactly "namespace/name" over a restricted charset. Anything with # extra path segments (e.g. a server-side relative path such as - # models/Llama/Foo.gguf on a remote Studio) is not a hub id and must not be + # models/Llama/Foo.gguf on a remote Unsloth) is not a hub id and must not be # casefold-matched against a differently cased path on a case-sensitive # filesystem. This is host independent, unlike the existence probe below which # cannot see a path that only exists on the server. @@ -690,8 +690,8 @@ def _model_id_matches( if actual == requested: return True # Case-insensitive matching is only safe when the local existence probe in - # _is_hub_model_id is authoritative, i.e. against a loopback Studio on this host. - # Against a remote Studio a two-segment string is indistinguishable from a + # _is_hub_model_id is authoritative, i.e. against a loopback Unsloth on this host. + # Against a remote Unsloth a two-segment string is indistinguishable from a # server-side relative path (e.g. Models/Foo vs models/foo), so casefolding it # could attach to the wrong model on a case-sensitive server; defer to an exact # match there and let the load endpoint resolve the requested path. @@ -709,7 +709,7 @@ def _resolve_model( load: LoadOptions = LoadOptions(), ) -> dict: models = _loaded_models(base, key) - # Only casefold-match ids against a loopback Studio, where _is_hub_model_id's + # Only casefold-match ids against a loopback Unsloth, where _is_hub_model_id's # local existence probe can actually reject a server-side path; see the note there. allow_casefold = is_loopback_url(base) # /v1/models reports the model id but not the active GGUF variant or runtime load @@ -741,7 +741,7 @@ def _resolve_model( typer.echo( f"Ensuring {requested} is loaded with the requested settings…" if load_has_overrides - else f"Loading {requested} on the Studio server (this can take a while)…" + else f"Loading {requested} on the Unsloth server (this can take a while)…" ) # Mirror `unsloth run`'s load knobs; keep the default payload as just # model_path so a bare `--model` load is unchanged. @@ -762,7 +762,7 @@ def _resolve_model( timeout = 3600, error = "Model load failed", ) - # Studio registers the model under a canonical id (resolved identifier, + # Unsloth registers the model under a canonical id (resolved identifier, # casing) that /v1/models echoes but which may differ from the path we # passed; match on the id the load reports so we don't silently fall # through to models[0] and connect to a different loaded model. @@ -783,22 +783,22 @@ def _resolve_model( if match is not None: return match if requested: - # We asked Studio to load it and it didn't surface in /v1/models; don't + # We asked Unsloth to load it and it didn't surface in /v1/models; don't # silently hand back an unrelated loaded model. _fail( - f"Studio didn't report '{requested}' as loaded. Double-check the model " + f"Unsloth didn't report '{requested}' as loaded. Double-check the model " "id, or load it from the model dropdown in the UI." ) if not models: _fail( - "No model is loaded in Studio. Load one from the model dropdown in " + "No model is loaded in Unsloth. Load one from the model dropdown in " "the UI, or pass --model to load it from here." ) return models[0] def _require_gguf_for_codex(base: str, key: str, model_id: str) -> None: - # Codex always streams, and Studio only streams /v1/responses from llama-server. + # Codex always streams, and Unsloth only streams /v1/responses from llama-server. try: status = _http_json("GET", f"{base}/api/inference/status", key) except urllib.error.HTTPError as exc: @@ -901,7 +901,7 @@ def _codex_supports_model_catalog() -> bool: def _codex_model_catalog(model: dict) -> dict: - """Return conservative metadata for a Studio model unknown to Codex's built-in catalog.""" + """Return conservative metadata for an Unsloth model unknown to Codex's built-in catalog.""" model_id = model["id"] window = model.get("context_length") or model.get("max_context_length") entry = { @@ -1202,7 +1202,7 @@ def _connect( # `--model org/name:QUANT` is shorthand for `--model org/name --gguf-variant QUANT`. # Split it before we match/serve so the attach path resolves against the already-loaded # `org/name` (listed without the suffix) instead of reloading a `:`-suffixed repo id -- - # which Studio rejects and which would evict a model another session is using. + # which Unsloth rejects and which would evict a model another session is using. if model: repo, variant = _split_repo_variant(model) if variant: @@ -1240,7 +1240,7 @@ def _run( # --no-launch recipes stay intact. if launch and clear_screen: click.clear() - typer.echo(f"Studio {base} · model {entry['id']}") + typer.echo(f"Unsloth {base} · model {entry['id']}") wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else () if not launch: _print_env(env, command, unset_env = unset_env, wsl_env_bridge = wsl_env_bridge) @@ -1306,7 +1306,7 @@ def write_openclaw_config( ) return before = json.dumps(config, sort_keys = True) - # Studio is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path). + # Unsloth is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path). provider_model = {"id": model["id"], "name": model["id"]} window = model.get("context_length") or model.get("max_context_length") if window: @@ -1570,14 +1570,14 @@ def write_pi_config(base: str, key: str, model: dict, path: Path) -> None: return before = json.dumps(config, sort_keys = True) # Pi reads custom providers from ~/.pi/agent/models.json (HOME-relocated for the - # session). Studio is a generic OpenAI-compatible /v1 endpoint, and the key lives + # session). Unsloth is a generic OpenAI-compatible /v1 endpoint, and the key lives # in the config rather than the env (matching openclaw/opencode). provider_model = {"id": model["id"]} window = model.get("context_length") or model.get("max_context_length") if window: window = int(window) # An unspecified model defaults to contextWindow 128000 / maxTokens 16384, - # far larger than a small Studio context, so Pi compacts too late and overflows + # far larger than a small Unsloth context, so Pi compacts too late and overflows # the server. Pin the real window and a sane output cap (mirrors OpenCode). provider_model["contextWindow"] = window provider_model["maxTokens"] = min(window // 4, 8192) @@ -1606,7 +1606,7 @@ def claude( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point Claude Code at the running Studio server and start it.""" + """Point Claude Code at the running Unsloth server and start it.""" base, key, entry = _connect( api_key, model, @@ -1690,7 +1690,7 @@ def codex( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point OpenAI Codex at the running Studio server and start it.""" + """Point OpenAI Codex at the running Unsloth server and start it.""" base, key, entry = _connect( api_key, model, @@ -1734,7 +1734,7 @@ def openclaw( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point OpenClaw at the running Studio server and start it.""" + """Point OpenClaw at the running Unsloth server and start it.""" base, key, entry = _connect( api_key, model, @@ -1791,7 +1791,7 @@ def opencode( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point OpenCode at the running Studio server and start it.""" + """Point OpenCode at the running Unsloth server and start it.""" base, key, entry = _connect( api_key, model, @@ -1835,7 +1835,7 @@ def opencode( # setting them in the highest-priority inline overlay neutralizes any user allowlist # or denylist for the launch. It is session-only: it lives in OPENCODE_CONFIG_CONTENT # for this invocation and never touches the user's config files, so their normal - # `opencode` is unchanged; only this session is limited to the Studio provider. + # `opencode` is unchanged; only this session is limited to the Unsloth provider. # small_model is opencode's separate model for lightweight tasks; pin it to the # session model too, or a user/project small_model on another (now filtered) # provider would resolve a not-found error mid-session. The session serves one @@ -1869,7 +1869,7 @@ def hermes( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point Hermes (Nous Research) at the running Studio server and start it.""" + """Point Hermes (Nous Research) at the running Unsloth server and start it.""" native_args = [*_yolo_command_flags("hermes", yolo), *ctx.args] command = ["hermes", *_hermes_resume_oneshot_args(native_args)] base, key, entry = _connect( @@ -1902,7 +1902,7 @@ def pi( yolo: bool = _YOLO_OPTION, persist: bool = _PERSIST_OPTION, ): - """Point Pi (coding agent) at the running Studio server and start it.""" + """Point Pi (coding agent) at the running Unsloth server and start it.""" base, key, entry = _connect( api_key, model, diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py index 09355bd454..f2f41fc583 100644 --- a/unsloth_cli/commands/studio.py +++ b/unsloth_cli/commands/studio.py @@ -40,7 +40,7 @@ def _enable_verbose_access_logs() -> None: # UNSLOTH_STUDIO_HOME wins when both env vars are set. def _looks_like_installer_managed_studio_home(candidate: Path) -> bool: """Sentinel check (studio.conf or bin shim) so a dev venv named - unsloth_studio is not misidentified as a custom Studio root. + unsloth_studio is not misidentified as a custom Unsloth root. """ shim_name = "unsloth.exe" if platform.system() == "Windows" else "unsloth" return (candidate / "share" / "studio.conf").is_file() or ( @@ -212,7 +212,7 @@ def _find_run_py() -> Optional[Path]: run_py = _PACKAGE_ROOT / "studio" / "backend" / "run.py" if run_py.is_file(): return run_py - # 2. Studio venv's site-packages (Linux + Windows layouts) + # 2. Unsloth venv's site-packages (Linux + Windows layouts) for pattern in ( "lib/python*/site-packages/studio/backend/run.py", "Lib/site-packages/studio/backend/run.py", @@ -273,7 +273,7 @@ def _find_setup_script() -> Optional[Path]: s = _PACKAGE_ROOT / "studio" / name if s.is_file(): return s - # 2. Studio venv's site-packages + # 2. Unsloth venv's site-packages for pattern in ( f"lib/python*/site-packages/studio/{name}", f"Lib/site-packages/studio/{name}", @@ -641,7 +641,7 @@ def _create_desktop_secret_in_cli() -> str: def _should_prompt_password_change( *, cloudflare: Optional[bool], host: str, secure: bool, api_only: bool ) -> bool: - """Whether this launch will expose Studio through the Cloudflare tunnel. + """Whether this launch will expose Unsloth through the Cloudflare tunnel. CLI mirror of run.py's _cloudflare_tunnel_should_start, minus the Colab case (Colab launches never come through this CLI path). --secure implies @@ -747,7 +747,7 @@ def _apply_supplied_password_before_launch(supplied_password: "str | None") -> N conn = _connect_auth_db() except (OSError, sqlite3.Error) as exc: typer.echo( - f"Error: --password could not open the Studio auth database ({exc}); not starting.", + f"Error: --password could not open the Unsloth auth database ({exc}); not starting.", err = True, ) raise typer.Exit(1) @@ -767,7 +767,7 @@ def _apply_supplied_password_before_launch(supplied_password: "str | None") -> N raise typer.Exit(1) if not row[2]: typer.echo( - "Error: a Studio admin password is already set; --password only sets " + "Error: an Unsloth admin password is already set; --password only sets " "the initial password. Run `unsloth studio reset-password` first " "(or change it in the UI).", err = True, @@ -790,7 +790,7 @@ def _apply_supplied_password_before_launch(supplied_password: "str | None") -> N # Any DB failure fails closed (typer.Exit is not caught here, so the # deliberate Exit(1) branches above propagate unchanged). typer.echo( - f"Error: --password could not update the Studio auth database ({exc}); not starting.", + f"Error: --password could not update the Unsloth auth database ({exc}); not starting.", err = True, ) raise typer.Exit(1) @@ -813,9 +813,9 @@ def _strip_seeded_bootstrap_password_or_exit(*, context: str) -> None: bootstrap_file.unlink(missing_ok = True) except OSError as exc: typer.echo( - "Error: refusing to publish Studio on a public Cloudflare URL: " + "Error: refusing to publish Unsloth on a public Cloudflare URL: " f"could not remove the seeded bootstrap password file ({exc}), so an " - f"older Studio child could still serve the default credential ({context}). " + f"older Unsloth child could still serve the default credential ({context}). " "Delete it manually or change the admin password (run `unsloth studio` " "locally with a terminal attached, or `unsloth studio reset-password`), " "then retry.", @@ -851,7 +851,7 @@ def _require_servable_frontend_or_exit( return frontend typer.echo( "Error: --frontend points at a directory with no index.html, so a " - "public Studio launch would have no login page to change the seeded " + "public Unsloth launch would have no login page to change the seeded " "admin password. Point --frontend at a built dist, rebuild it (re-run " "install.sh), or use --api-only.", err = True, @@ -862,7 +862,7 @@ def _require_servable_frontend_or_exit( if resolved is not None: return resolved typer.echo( - "Error: the Studio frontend is not built, so a public launch would have " + "Error: the Unsloth frontend is not built, so a public launch would have " "no login page to change the seeded admin password. Build it (re-run " "install.sh), pass --frontend PATH to a built dist, or use --api-only.", err = True, @@ -892,8 +892,8 @@ def _validate_inproc_backend_before_strip( _load_run_module() except Exception as exc: typer.echo( - f"Error: the Studio backend could not be loaded ({exc}); refusing to " - "expose Studio publicly before it is confirmed runnable. Re-run: " + f"Error: the Unsloth backend could not be loaded ({exc}); refusing to " + "expose Unsloth publicly before it is confirmed runnable. Re-run: " "unsloth studio setup", err = True, ) @@ -902,7 +902,7 @@ def _validate_inproc_backend_before_strip( def _tunnel_binary_confirmed_unavailable() -> bool: """True only if cloudflared is provably unavailable (found nowhere on PATH or - in the Studio cache AND the download failed), so the tunnel cannot start. + in the Unsloth cache AND the download failed), so the tunnel cannot start. Used on the --secure path (loopback bind, so the tunnel is the ONLY public exposure) to skip stripping the seeded recovery password before a public URL @@ -921,7 +921,7 @@ def _tunnel_binary_confirmed_unavailable() -> bool: if not tunnel_py.is_file(): return False # ensure_cloudflared() lazily imports utils.paths.storage_roots to resolve the - # Studio bin cache. The outer CLI hasn't added studio/backend to sys.path yet, + # Unsloth bin cache. The outer CLI hasn't added studio/backend to sys.path yet, # so that import would fail and return None (a false "unavailable" that wrongly # refuses --secure). Add the backend dir so the cache path resolves as in the child. added_backend_path = False @@ -946,7 +946,7 @@ def _tunnel_binary_confirmed_unavailable() -> bool: def _child_self_suppresses(*, in_studio_venv: bool, child_run_py: Optional[Path]) -> bool: - """True when the child that will serve Studio is provably THIS install's + """True when the child that will serve Unsloth is provably THIS install's backend, whose pre-bind gate sets app.state.suppress_bootstrap_injection and so never serves the seeded credential publicly -- even with .bootstrap_password on disk. The parent-side strip is then unnecessary and can be skipped to avoid @@ -1002,8 +1002,8 @@ def _enforce_password_change_before_exposure( # Refuse rather than risk a child serving the default login; a transient # lock clears on retry. typer.echo( - "Error: refusing to publish Studio on a public Cloudflare URL: could " - f"not open the Studio auth database ({exc}) to confirm the admin " + "Error: refusing to publish Unsloth on a public Cloudflare URL: could " + f"not open the Unsloth auth database ({exc}) to confirm the admin " "password was changed. Retry (a transient database lock clears), or " "change the password first (run `unsloth studio` locally with a " "terminal attached, or `unsloth studio reset-password`).", @@ -1028,8 +1028,8 @@ def _enforce_password_change_before_exposure( except OSError: pass typer.echo( - "Error: refusing to publish Studio on a public Cloudflare URL: could " - f"not initialize the admin account ({exc}), so a re-exec'd Studio " + "Error: refusing to publish Unsloth on a public Cloudflare URL: could " + f"not initialize the admin account ({exc}), so a re-exec'd Unsloth " "child could regenerate and serve a default credential. Retry (a " "transient database lock clears), or change the password first (run " "`unsloth studio` locally with a terminal attached, or `unsloth " @@ -1053,7 +1053,7 @@ def _enforce_password_change_before_exposure( # regenerate; we just couldn't read must_change back. Strip the seeded # file so nothing serves it, failing closed if the strip itself fails. typer.echo( - f"Warning: could not read the Studio admin state back ({exc}); " + f"Warning: could not read the Unsloth admin state back ({exc}); " "removing the seeded bootstrap password before public exposure.", err = True, ) @@ -1066,7 +1066,7 @@ def _enforce_password_change_before_exposure( # the launch: it never arms for api-only, and TIMEOUT=0 disables it. if api_only or not _bootstrap_deadline_active(): typer.echo( - "Error: refusing to publish Studio on a public Cloudflare " + "Error: refusing to publish Unsloth on a public Cloudflare " "URL: the default admin password was never changed, no " "terminal is attached to change it here, and the bootstrap " "shutdown deadline does not apply to this launch (api-only, " @@ -1085,12 +1085,12 @@ def _enforce_password_change_before_exposure( # fails). Keep the file for LOCAL recovery; must_change stays set # and the deadline arms. typer.echo( - "Warning: Studio is being exposed publicly while the admin " + "Warning: Unsloth is being exposed publicly while the admin " "account still uses its auto-generated bootstrap password. The " "login page forces a change and the credential is never served " "on the public page. Set a new password by running `unsloth " "studio` locally with a terminal attached, or `unsloth studio " - "reset-password`; Studio shuts down after ~1h if the password " + "reset-password`; Unsloth shuts down after ~1h if the password " "stays unchanged (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT).", err = True, ) @@ -1104,7 +1104,7 @@ def _enforce_password_change_before_exposure( # uncertainty.) if secure and _tunnel_binary_confirmed_unavailable(): typer.echo( - "Error: refusing to expose Studio: the Cloudflare tunnel binary " + "Error: refusing to expose Unsloth: the Cloudflare tunnel binary " "(cloudflared) is unavailable and could not be downloaded, so no " "public URL can start. The seeded bootstrap password is preserved " "for recovery; fix connectivity and retry, or change the password " @@ -1121,11 +1121,11 @@ def _enforce_password_change_before_exposure( # forces a change and the timer still arms; only the on-disk copy goes. _strip_seeded_bootstrap_password_or_exit(context = "no terminal to change it") typer.echo( - "Warning: Studio is being exposed publicly while the admin account " + "Warning: Unsloth is being exposed publicly while the admin account " "still uses its auto-generated bootstrap password. The seeded password " "file has been removed so it is not served on the public page. Set a new " "password by running `unsloth studio` locally with a terminal attached, " - "or `unsloth studio reset-password`; Studio shuts down after ~1h if the " + "or `unsloth studio reset-password`; Unsloth shuts down after ~1h if the " "password stays unchanged (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT).", err = True, ) @@ -1146,7 +1146,7 @@ def _enforce_password_change_before_exposure( new_password = _password_prompt.prompt_new_password(_is_current_password) except (KeyboardInterrupt, EOFError): typer.echo( - "\nError: password change aborted; refusing to expose Studio " + "\nError: password change aborted; refusing to expose Unsloth " "with the default admin password. Re-run and set a password, " "or launch without --secure/--cloudflare.", err = True, @@ -1247,7 +1247,7 @@ def studio_default( cloudflare: Optional[bool] = typer.Option( None, "--cloudflare/--no-cloudflare", - help = "Expose Studio on a PUBLIC internet URL via a free Cloudflare HTTPS " + help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS " "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; " "pass --cloudflare to enable it (--secure implies it). --no-cloudflare forces " "it off but does not change a raw wildcard bind.", @@ -1404,7 +1404,7 @@ def studio_default( studio_python = _studio_venv_python() run_py = _find_run_py() if not (studio_python and run_py): - typer.echo("Studio not set up. Run install.sh first.") + typer.echo("Unsloth Studio not set up. Run install.sh first.") raise typer.Exit(1) # A public UI launch must have a servable login page BEFORE the gate can # strip the seeded .bootstrap_password, or the child has no way to change @@ -1510,7 +1510,7 @@ def studio_default( rc = proc.wait() if rc != 0: typer.echo( - f"\nError: Studio server exited unexpectedly (code {rc}).", + f"\nError: Unsloth server exited unexpectedly (code {rc}).", err = True, ) typer.echo( @@ -1522,7 +1522,7 @@ def studio_default( else: os.execvp(str(studio_python), args) else: - typer.echo("Studio not set up. Run install.sh first.") + typer.echo("Unsloth Studio not set up. Run install.sh first.") raise typer.Exit(1) run_mod = _load_run_module() @@ -1733,7 +1733,7 @@ def run( cloudflare: Optional[bool] = typer.Option( None, "--cloudflare/--no-cloudflare", - help = "Expose Studio on a PUBLIC internet URL via a free Cloudflare HTTPS " + help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS " "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; " "pass --cloudflare to enable it (--secure implies it). --no-cloudflare forces " "it off but does not change a raw wildcard bind.", @@ -1769,16 +1769,16 @@ def run( "process list and shell history. Rotate later with `unsloth studio reset-password`.", ), ): - """Start Studio, load a model, print an API key -- one-liner server. + """Start Unsloth, load a model, print an API key -- one-liner server. - Unknown flags pass through to llama-server (GGUF only). Studio + Unknown flags pass through to llama-server (GGUF only). Unsloth rejects managed flags with HTTP 400: model identity, network (--host/--port/--path/--api-prefix/--reuse-port), auth/TLS (--api-key/--ssl-*), single-model UI (--ui/--models-*/--webui), and parallel slots (use --parallel above). Full denylist in studio/backend/core/inference/llama_server_args.py. Other knobs (-c, -ngl, --jinja, --flash-attn, -t, ...) pass through and - last-wins-override Studio's auto-set value. + last-wins-override Unsloth's auto-set value. Example: unsloth studio run --model unsloth/Qwen3-1.7B-GGUF --gguf-variant UD-Q4_K_XL @@ -1792,7 +1792,7 @@ def run( # Set before any re-exec so the in-venv server inherits it via the env. # `run --verbose` used to pass through to llama-server (its own -v); keep - # that by forwarding --log-verbose so we add Studio logs without dropping it. + # that by forwarding --log-verbose so we add Unsloth logs without dropping it. if verbose: _enable_verbose_access_logs() if not any(a in ("--verbose", "-v", "--log-verbose") for a in extra_llama_args): @@ -1878,14 +1878,14 @@ def run( if not in_studio_venv: studio_python = _studio_venv_python() if not studio_python: - typer.echo("Studio not set up. Run install.sh first.") + typer.echo("Unsloth Studio not set up. Run install.sh first.") raise typer.Exit(1) # Re-exec via the studio venv's `unsloth` console-script. studio_bin = studio_python.parent / "unsloth" if not studio_bin.is_file(): - typer.echo("Studio venv missing 'unsloth' entry point. Re-run: unsloth studio setup") + typer.echo("Unsloth venv missing 'unsloth' entry point. Re-run: unsloth studio setup") raise typer.Exit(1) - # `run` serves the same Studio UI (unless --api-only); a public launch must + # `run` serves the same Unsloth UI (unless --api-only); a public launch must # have a servable login page BEFORE the gate strips the seeded password, or # the child has no way to change it. Validate here and forward the resolved # dist so a shadowed child that can't self-resolve one still serves it. @@ -2216,7 +2216,7 @@ def stop(): import signal as _signal if not _PID_FILE.is_file(): - typer.echo("No running Studio server found (no PID file).") + typer.echo("No running Unsloth server found (no PID file).") raise typer.Exit(0) pid_text = _PID_FILE.read_text().strip() @@ -2229,7 +2229,7 @@ def stop(): # Check if still alive (os.kill(pid, 0) is invalid on Windows -- see _pid_alive). if not _pid_alive(pid): - typer.echo(f"Studio server (PID {pid}) is not running. Cleaning up stale PID file.") + typer.echo(f"Unsloth server (PID {pid}) is not running. Cleaning up stale PID file.") _PID_FILE.unlink(missing_ok = True) raise typer.Exit(0) @@ -2239,13 +2239,13 @@ def stop(): subprocess.run(["taskkill", "/PID", str(pid), "/F"], check = True) else: os.kill(pid, _signal.SIGTERM) - typer.echo(f"Sent shutdown signal to Studio server (PID {pid}).") + typer.echo(f"Sent shutdown signal to Unsloth server (PID {pid}).") except ProcessLookupError: - typer.echo(f"Studio server (PID {pid}) already exited.") + typer.echo(f"Unsloth server (PID {pid}) already exited.") _PID_FILE.unlink(missing_ok = True) raise typer.Exit(0) except Exception as e: - typer.echo(f"Failed to stop Studio server (PID {pid}): {e}", err = True) + typer.echo(f"Failed to stop Unsloth server (PID {pid}): {e}", err = True) raise typer.Exit(1) # Wait briefly for the process to exit and clean up. @@ -2253,10 +2253,10 @@ def stop(): time.sleep(0.5) if not _pid_alive(pid): _PID_FILE.unlink(missing_ok = True) - typer.echo("Studio server stopped.") + typer.echo("Unsloth server stopped.") raise typer.Exit(0) - typer.echo("Studio server is shutting down (may take a few seconds).") + typer.echo("Unsloth server is shutting down (may take a few seconds).") # ── unsloth studio setup / update ───────────────────────────────────── @@ -2471,7 +2471,7 @@ def setup( help = "Full pip/build output during setup for troubleshooting.", ), ): - """Run Studio setup (called by install.ps1 / install.sh).""" + """Run Unsloth setup (called by install.ps1 / install.sh).""" _run_setup_script(verbose = verbose) @@ -2632,10 +2632,10 @@ def provision_desktop_auth(): @studio_app.command("reset-password") def reset_password(): - """Reset the Studio admin password. + """Reset the Unsloth admin password. Deletes the auth database so that a fresh admin account with a new - random password is created on the next server start. The Studio + random password is created on the next server start. The Unsloth server must be restarted after running this command. """ auth_dir = STUDIO_HOME / "auth" @@ -2647,7 +2647,7 @@ def reset_password(): had_db = db_file.exists() # Delete auth.db FIRST and prove it is gone before touching the seeded - # credential files. If it cannot be removed (a running Studio or Windows + # credential files. If it cannot be removed (a running Unsloth or Windows # holds it open, or a read-only auth dir), abort with the credential files # untouched: deleting them while an un-resettable DB (must_change_password=1) # survives would lock a forgotten-password reset out of any recovery @@ -2657,7 +2657,7 @@ def reset_password(): except OSError as exc: typer.echo( f"Error: could not delete the auth database ({exc}). Stop any running " - "Studio and retry; no credential files were changed.", + "Unsloth and retry; no credential files were changed.", err = True, ) raise typer.Exit(1) @@ -2679,7 +2679,7 @@ def reset_password(): except OSError as exc: typer.echo( f"Error: could not remove or clear {path.name} ({exc}); delete " - "it manually before restarting Studio or the old password may " + "it manually before restarting Unsloth or the old password may " "be reused.", err = True, ) diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py index 56633408fb..ae6f8dcfd4 100644 --- a/unsloth_cli/tests/test_inference_chat.py +++ b/unsloth_cli/tests/test_inference_chat.py @@ -372,7 +372,7 @@ def test_find_studio_server_none_when_not_running(monkeypatch): def test_find_studio_server_prefers_ipv4_loopback_for_localhost(monkeypatch): - # localhost resolving ::1-first must not hide a Studio bound to 127.0.0.1: + # localhost resolving ::1-first must not hide an Unsloth bound to 127.0.0.1: # discovery tries each loopback address and returns the one that answers. import socket import urllib.request diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py index 7e76465144..2405ba0480 100644 --- a/unsloth_cli/tests/test_start.py +++ b/unsloth_cli/tests/test_start.py @@ -648,7 +648,7 @@ def test_resolve_model_attaches_to_loaded_catalog_hit_without_reload(monkeypatch def test_resolve_model_remote_studio_does_not_casefold_attach(monkeypatch): - # Against a remote Studio the local existence probe cannot see server-side paths, + # Against a remote Unsloth the local existence probe cannot see server-side paths, # so a case-variant loaded id must NOT attach without a load: it could be a distinct # server-side path on a case-sensitive host. The load endpoint resolves the request. calls = [] @@ -738,7 +738,7 @@ def test_no_launch_output_is_parseable(fake_studio): result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"]) assert result.exit_code == 0, result.output lines = [ln for ln in result.output.splitlines() if ln.strip()] - skip = ("export ", "unset ", "Studio ", "Updated ", "Disabled ", "Warning", "Loading") + skip = ("export ", "unset ", "Unsloth ", "Updated ", "Disabled ", "Warning", "Loading") body = [ln for ln in lines if not ln.startswith(skip)] assert "codex --oss --profile unsloth_api" in body[-1] assert any(ln.startswith("export CODEX_HOME=") for ln in lines) @@ -767,7 +767,7 @@ def test_no_launch_last_line_is_self_contained(fake_studio, tmp_path): def test_no_launch_claude_last_line_blanks_conflicting_auth(fake_studio): # The unset vars must be neutralized inline too, or a partial copy would send the - # user's own ANTHROPIC_API_KEY to the Studio base. + # user's own ANTHROPIC_API_KEY to the Unsloth base. result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"]) assert result.exit_code == 0, result.output last = [ln for ln in result.output.splitlines() if ln.strip()][-1] @@ -814,7 +814,7 @@ def test_https_loopback_never_auto_serves(fake_studio, monkeypatch): ) result = CliRunner().invoke(start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"]) assert result.exit_code == 1 - assert "No running Studio server" in result.output + assert "No running Unsloth server" in result.output assert started["called"] is False @@ -967,7 +967,7 @@ def test_connect_model_flag_forwards_load_options(fake_studio): def test_connect_model_flag_matches_canonical_id(fake_studio, monkeypatch): - # Studio registers a loaded model under a canonical id (resolved identifier + # Unsloth registers a loaded model under a canonical id (resolved identifier # / casing) that can differ from the path we passed. The agent must connect # to that model, not silently fall through to the first loaded one. requested = "Unsloth/Qwen3.5-35B-A3B" @@ -1024,7 +1024,7 @@ def test_connect_model_bare_id_matches_loaded_without_reload(fake_studio): def test_connect_model_variant_suffix_defers_to_server_dedup(fake_studio): # `--model repo:QUANT` splits into a VALID load payload (bare repo + gguf_variant), - # never the `:`-suffixed repo id Studio rejects. The variant knob defers to + # never the `:`-suffixed repo id Unsloth rejects. The variant knob defers to # /api/inference/load, whose already-loaded dedup answers without reloading when the # active variant+settings match -- so a second session running the same command # attaches without evicting the first, while a genuinely different quant reloads. @@ -1116,7 +1116,7 @@ def test_connect_no_model_loaded_errors(fake_studio, monkeypatch): def test_connect_requested_model_not_loaded_fails(fake_studio, monkeypatch): - # Studio never surfaces the requested model; fail loudly rather than + # Unsloth never surfaces the requested model; fail loudly rather than # silently connecting to whatever else happens to be loaded. inner = start._http_json @@ -1187,7 +1187,7 @@ def test_connect_nonloopback_explicit_key_is_allowed(fake_studio, monkeypatch): def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatch): - # A key saved for a remote (non-loopback) Studio is replayed on keyless runs; + # A key saved for a remote (non-loopback) Unsloth is replayed on keyless runs; # auto-minting stays blocked for non-loopback. remote = "http://studio.example:8888" monkeypatch.setattr(start, "find_studio_server", lambda: remote) @@ -1201,7 +1201,7 @@ def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatc def test_connect_studio_server_errors_on_explicit_remote(monkeypatch): - # A user who pointed UNSLOTH_STUDIO_URL at a remote Studio should get an + # A user who pointed UNSLOTH_STUDIO_URL at a remote Unsloth should get an # error, not a silent local model load (which they did not ask for). import typer @@ -1242,7 +1242,7 @@ def test_connect_unverified_loopback_without_cached_key_refuses_to_mint( def test_connect_replays_saved_key_without_identity_check(fake_studio, tmp_path, monkeypatch): - # A "saved" key (e.g. for an SSH-tunnelled Studio the handshake can't match) + # A "saved" key (e.g. for an SSH-tunnelled Unsloth the handshake can't match) # replays on keyless runs without the handshake, scoped to its own base. cache = tmp_path / "agent_api_key.json" cache.write_text(json.dumps({"servers": {BASE: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}})) @@ -1358,7 +1358,7 @@ def _serve_redirect(target): def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch): - # A squatter could 302 /api/auth/identity to the real Studio and relay its + # A squatter could 302 /api/auth/identity to the real Unsloth and relay its # proof; redirects must be refused so the squatter's base isn't accepted. import unsloth_cli._inference as inference @@ -1384,7 +1384,7 @@ def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch): def test_verify_studio_identity_rejects_relayed_proof(tmp_path, monkeypatch): - # A squatter that proxies the nonce to the real Studio on another port gets a + # A squatter that proxies the nonce to the real Unsloth on another port gets a # proof bound to *that* port; the client expects one bound to the port it # connected to, so the relayed proof is rejected. import unsloth_cli._inference as inference @@ -1435,7 +1435,7 @@ def test_connect_no_studio_errors(fake_studio, monkeypatch): monkeypatch.setattr(start, "find_studio_server", lambda: None) result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"]) assert result.exit_code == 1 - assert "No running Studio server" in result.output + assert "No running Unsloth server" in result.output @pytest.fixture(autouse = True) @@ -1564,7 +1564,7 @@ def test_no_serve_preserves_error(fake_studio, monkeypatch): start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-serve"] ) assert result.exit_code == 1 - assert "No running Studio server" in result.output + assert "No running Unsloth server" in result.output assert started["called"] is False @@ -1578,7 +1578,7 @@ def test_no_launch_never_serves(fake_studio, monkeypatch): start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-launch"] ) assert result.exit_code == 1 - assert "No running Studio server" in result.output + assert "No running Unsloth server" in result.output assert started["called"] is False @@ -1871,7 +1871,7 @@ def _opencode_inline_config(output: str) -> dict: def test_opencode_inline_scopes_session_to_studio_provider(fake_studio): # opencode filters even config-defined providers through enabled/disabled_providers, # and a model pin does not bypass that gate. The inline overlay (session-only, highest - # layer, arrays replace) allowlists our provider and clears the denylist so the Studio + # layer, arrays replace) allowlists our provider and clears the denylist so the Unsloth # model always loads regardless of the user's config, without reading or editing it. result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"]) assert result.exit_code == 0, result.output @@ -2657,7 +2657,7 @@ def test_agent_api_key_auto_started_rejected_env_key_falls_back(fake_studio, tmp def test_agent_api_key_auto_started_accepted_key_is_honored(fake_studio, tmp_path): - # An explicit key the fresh server accepts (e.g. persisted in this Studio + # An explicit key the fresh server accepts (e.g. persisted in this Unsloth # home's auth db across restarts) keeps working exactly as before. key = start._agent_api_key(BASE, "sk-unsloth-deadbeefdeadbeef", auto_started = True) assert key == "sk-unsloth-deadbeefdeadbeef" @@ -2937,11 +2937,11 @@ def test_hermes_resume_oneshot_rejects_usage_file(monkeypatch, usage_arg): def test_native_resume_flag_passes_through_unchanged(fake_studio, monkeypatch): # The persistence flag is --persist, NOT --resume, so an agent's own # `--resume ` (e.g. `unsloth start claude --resume `) still flows - # through to the agent verbatim and is not swallowed as a Studio option. + # through to the agent verbatim and is not swallowed as an Unsloth option. monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude") monkeypatch.setattr(start, "_claude_flags", lambda: []) captured = _capture_launch(monkeypatch, ["claude", "--resume", "some-session-guid"]) assert captured["command"][-2:] == ["--resume", "some-session-guid"] - # Studio never auto-appends its own resume token when the user drives resume. + # Unsloth never auto-appends its own resume token when the user drives resume. assert captured["command"].count("--resume") == 1 assert "--continue" not in captured["command"] diff --git a/unsloth_cli/tests/test_studio_cloudflare_flag.py b/unsloth_cli/tests/test_studio_cloudflare_flag.py index fb57d7aaf4..7287737f75 100644 --- a/unsloth_cli/tests/test_studio_cloudflare_flag.py +++ b/unsloth_cli/tests/test_studio_cloudflare_flag.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 -"""Tests for the `--cloudflare/--no-cloudflare` Studio flag. +"""Tests for the `--cloudflare/--no-cloudflare` Unsloth flag. Pins the typer Option (tri-state, default off / None) on both `unsloth studio` and `unsloth studio run`, and that the chosen polarity reaches the re-exec'd diff --git a/unsloth_cli/tests/test_studio_password_prompt.py b/unsloth_cli/tests/test_studio_password_prompt.py index 7bbdfe3703..6e9a2c1d52 100644 --- a/unsloth_cli/tests/test_studio_password_prompt.py +++ b/unsloth_cli/tests/test_studio_password_prompt.py @@ -885,7 +885,7 @@ def test_run_non_tty_deletes_bootstrap_password_file(monkeypatch, tmp_path): def test_run_missing_frontend_exits_before_stripping_bootstrap(monkeypatch, tmp_path): # Regression (item B / reviewer finding 4): `unsloth studio run` serves the - # same Studio UI and strips the seeded password on a headless public launch, + # same Unsloth UI and strips the seeded password on a headless public launch, # so a missing frontend dist must abort BEFORE the strip -- the same lockout # guard as `unsloth studio`, not just `studio run`'s model-load residual. import typer as _typer @@ -1104,7 +1104,7 @@ def test_cli_update_password_truncates_locked_bootstrap_after_change(monkeypatch def test_reset_password_fails_closed_when_db_cannot_be_deleted(monkeypatch, tmp_path): - # If auth.db cannot be removed (running Studio / Windows lock, read-only dir), + # If auth.db cannot be removed (running Unsloth / Windows lock, read-only dir), # reset must abort BEFORE touching the credential files -- deleting them while # an un-resettable must_change_password=1 DB survives would lock a # forgotten-password reset out with no recovery credential. diff --git a/unsloth_cli/tests/test_studio_secure_flag.py b/unsloth_cli/tests/test_studio_secure_flag.py index 5e5895309c..2a67aad95a 100644 --- a/unsloth_cli/tests/test_studio_secure_flag.py +++ b/unsloth_cli/tests/test_studio_secure_flag.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 -"""Tests for the `--secure/--no-secure` Studio flag: option registration, +"""Tests for the `--secure/--no-secure` Unsloth flag: option registration, re-exec/run_server forwarding, the forced 127.0.0.1 bind, and rejection alongside --no-cloudflare or before a subcommand. Modeled on test_studio_cloudflare_flag.py.""" diff --git a/unsloth_cli/tests/test_studio_verbose_flag.py b/unsloth_cli/tests/test_studio_verbose_flag.py index 4af32fd4a2..20468b5f02 100644 --- a/unsloth_cli/tests/test_studio_verbose_flag.py +++ b/unsloth_cli/tests/test_studio_verbose_flag.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 -"""Tests for the `--verbose/-v` Studio flag: option registration on both the +"""Tests for the `--verbose/-v` Unsloth flag: option registration on both the plain callback and the `run` subcommand, re-exec forwarding, the access-log env override, and rejection before a subcommand. Modeled on test_studio_secure_flag.py.""" @@ -123,7 +123,7 @@ def test_run_without_verbose_leaves_env_unset(monkeypatch): def test_run_verbose_preserves_llama_server_verbosity(monkeypatch): - # Studio consumes --verbose but still forwards llama-server's own verbosity. + # Unsloth consumes --verbose but still forwards llama-server's own verbosity. monkeypatch.delenv(_DEDUP, raising = False) monkeypatch.delenv(_POLL, raising = False) captured = _invoke_run(monkeypatch, _BASE + ["--verbose"])