ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'origin/main' into r5945

Browse source 
# Conflicts:
#	install.sh
This commit is contained in:
Daniel Han 2026-07-19 13:21:33 +00:00
commit 661f73bf50
331 changed files with 7932 additions and 1763 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.gitattributes vendored
View file

@ -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.

21
.github/scripts/agent-guides-drive.sh vendored
View file

@ -36,6 +36,23 @@ AGENT="${2:?usage: agent-guides-drive.sh <mode> <agent>}"
# Determinism (seed/temp) is applied at the server level by
# serve-unsloth-run.sh --extra; agents inherit it through the API.
TIMEOUT="${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 )' "$raw" || true)"
# The launch command is the last non-export, non-status line. start.py
# prints "Studio <url> · model <id>" and "Updated ..." status lines first.
CONNECT_CMD="$(grep -vE '^(export |unset |Studio |Updated |Disabled |Warning|Loading)' "$raw" \
# prints "Unsloth <url> · model <id>" and "Updated ..." status lines first.
CONNECT_CMD="$(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"

2
.github/scripts/assert-llama-loads.sh vendored
View file

@ -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.

2
.github/scripts/assert-prompt-cache.sh vendored
View file

@ -31,7 +31,7 @@
# (llama_cpp.py:337-340). So default: ~/.unsloth/studio/logs/llama-server/.
#
# <P> 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-<STUDIO_PORT>`
# glob would never match). We pick the newest llama-*.log instead.
#

4
.github/scripts/hf-download-with-retry.sh vendored
View file

@ -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

4
.github/workflows/lint-ci.yml vendored
View file

@ -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

16
.github/workflows/local-agent-guides-ci.yml vendored
View file

@ -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

12
.github/workflows/mlx-ci.yml vendored
View file

@ -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"

8
.github/workflows/release-desktop.yml vendored
View file

@ -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

30
.github/workflows/security-audit.yml vendored
View file

@ -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

14
.github/workflows/studio-api-smoke.yml vendored
View file

@ -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

2
.github/workflows/studio-backend-ci.yml vendored
View file

@ -64,7 +64,7 @@ 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

2
.github/workflows/studio-export-capability-ci.yml vendored
View file

@ -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:

4
.github/workflows/studio-frontend-ci.yml vendored
View file

@ -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

56
.github/workflows/studio-inference-smoke.yml vendored
View file

@ -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 <think> 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

4
.github/workflows/studio-load-orchestrator-ci.yml vendored
View file

@ -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:

10
.github/workflows/studio-mac-api-smoke.yml vendored
View file

@ -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

58
.github/workflows/studio-mac-inference-smoke.yml vendored
View file

@ -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 <think> 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

4
.github/workflows/studio-mac-install-matrix.yml vendored
View file

@ -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.

18
.github/workflows/studio-mac-ui-smoke.yml vendored
View file

@ -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

16
.github/workflows/studio-mac-update-smoke.yml vendored
View file

@ -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

2
.github/workflows/studio-tauri-smoke.yml vendored
View file

@ -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:

34
.github/workflows/studio-ui-smoke.yml vendored
View file

@ -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

12
.github/workflows/studio-update-smoke.yml vendored
View file

@ -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

16
.github/workflows/studio-windows-api-smoke.yml vendored
View file

@ -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

90
.github/workflows/studio-windows-inference-smoke.yml vendored
View file

@ -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()

32
.github/workflows/studio-windows-ui-smoke.yml vendored
View file

@ -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

26
.github/workflows/studio-windows-update-smoke.yml vendored
View file

@ -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-<tag>-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

10
.github/workflows/wheel-smoke.yml vendored
View file

@ -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()

16
README.md
View file

@ -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):

8
build.sh
View file

@ -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"

36
install.ps1
View file

@ -191,7 +191,7 @@ 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 = $envOverride
@ -771,7 +771,7 @@ function Find-FreeLaunchPort {
return `$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:`$existingPort"
@ -787,7 +787,7 @@ try {
`$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
@ -1453,7 +1453,7 @@ 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 (
@ -1464,7 +1464,7 @@ 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..."
@ -1483,7 +1483,7 @@ 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 {
@ -1513,7 +1513,7 @@ 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 = $true
@ -1532,7 +1532,7 @@ exit 0
substep "$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 $VenvDir -PathType Container) {
@ -1541,7 +1541,7 @@ exit 0
# ── Helper: run amd-smi without triggering a UAC elevation prompt ──
# amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing
# DiskPart UAC prompt mid-install (Studio backend amd.py hits the same).
# DiskPart UAC prompt mid-install (Unsloth backend amd.py hits the same).
# __COMPAT_LAYER=RunAsInvoker forces it (and helpers it spawns) to run
# un-elevated; on failure the WMI name -> gfx fallback still resolves the arch.
function Invoke-AmdSmiNoElevate {
@ -1668,7 +1668,7 @@ exit 0
function Test-HipinfoIsVenvInternal {
param([AllowNull()][string]$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 }
@ -1678,7 +1678,7 @@ 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) {
@ -1957,7 +1957,7 @@ 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"
@ -2724,8 +2724,8 @@ 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
@ -2927,7 +2927,7 @@ 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")
}
@ -3038,7 +3038,7 @@ 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 }
@ -3056,7 +3056,7 @@ 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
@ -3121,7 +3121,7 @@ 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)

198
install.sh
View file

@ -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) && {
@ -1374,7 +1376,7 @@ 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.)" "$C_WARN"
fi
fi
@ -1442,7 +1444,7 @@ if [ "$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 ""
@ -1679,7 +1681,7 @@ _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.
@ -1829,11 +1831,13 @@ 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 $STUDIO_HOME/unsloth_studio that lacks Unsloth sentinels.
# Accept the in-VENV ownership marker so partial-install retries are
# not blocked. Sentinels must be regular files: -f follows symlinks
# to files (the legitimate ln -s shim shape) but rejects directories
@ -1846,6 +1850,12 @@ if [ -x "$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 "$VENV_DIR"
@ -1854,7 +1864,7 @@ elif [ "$_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 "$STUDIO_HOME/.venv/bin/python" -c "import sys; print(sys.executable)" >/dev/null 2>&1; then
@ -1911,7 +1921,7 @@ if [ ! -x "$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
@ -2198,6 +2208,68 @@ _torch_flavor_tag() {
esac
}
# Whether release base $1 (X.Y[.Z...]) falls inside constraint window $2
# ("torch>=A.B[.C],<D.E.F"). Compares at major.minor granularity, which is exact
# for the windows this script uses (ceilings are always X.Y.0); a non-.0 ceiling
# would only make this conservative (excludes the whole ceiling minor). Anything
# unparseable answers "no" so the caller fails toward the supported range.
_torch_release_in_window() {
_trw_con="$2"
case "$_trw_con" in
"torch>="*",<"*) ;;
*) 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 ($1): cuXXX / cpu / rocm (rocmX.Y and gfx* ->
# rocm). Empty on an unknown leaf (odd mirror) so the repair safely no-ops.
_expected_torch_flavor_tag() {
@ -2346,7 +2418,7 @@ _pick_radeon_wheel() {
# the installer -- always returns 0. Runs the idempotent helper (ROCm 7.2 +
# librocdxg), then sources the env it persisted so detection finds the GPU.
# Export the ROCm-on-WSL env into this process and persist it to /etc/profile.d
# so non-login Studio/llama launches inherit it. Idempotent (writes only when
# so non-login Unsloth/llama launches inherit it. Idempotent (writes only when
# the drop-in is missing); no-op without librocdxg, so never fires off WSL.
# /etc/profile.d is root-owned -- sudo-tee when not root, else ROCm vanishes
# after this shell on a non-root reinstall. Best-effort either way.
@ -2391,7 +2463,7 @@ _maybe_bootstrap_rocm_wsl() {
rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9]/ && !/generic/{found=1} END{exit !found}'; then
# rocminfo may work only via the transient env _ensure_rocm_probe_env
# just set, which dies with the installer. Persist the drop-in so login
# shells (Studio, llama.cpp) inherit it -- else a reinstall over an
# shells (Unsloth, llama.cpp) inherit it -- else a reinstall over an
# existing /opt/rocm (uninstall keeps ROCm but drops it) loses the GPU.
_persist_rocm_wsl_dropin
return 0
@ -2413,7 +2485,7 @@ _maybe_bootstrap_rocm_wsl() {
# shellcheck disable=SC1091
. /etc/profile.d/unsloth-rocm-wsl.sh || true
else
# librocdxg present but the env drop-in is gone (e.g. a Studio
# librocdxg present but the env drop-in is gone (e.g. an Unsloth
# uninstall removed it while keeping shared ROCm). Restore the env.
_persist_rocm_wsl_dropin
fi
@ -2489,12 +2561,32 @@ case "$_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 $_PREV_TORCH_VER -- keeping it (set UNSLOTH_TORCH_UPGRADE=1 to get the newest release)"
fi
fi
# Auto-detect GPU for AMD ROCm based
# get_torch_index_url must have chosen */rocm*
# (gfx in rocminfo or amd-smi list). Then require rocminfo "Marketing Name:.*Radeon".
@ -2716,6 +2808,43 @@ esac
# ── Install unsloth directly into the venv (no activation needed) ──
tauri_log "STEP" "Installing PyTorch"
_VENV_PY="$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
@ -2740,9 +2869,13 @@ 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)..."
@ -2924,8 +3057,20 @@ 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 "$TORCH_INDEX_URL"
fi
fi
# AMD ROCm: install bitsandbytes (once, after torch, for all ROCm paths).
# Gate on SKIP_TORCH=false so a user running with --no-torch on a ROCm
@ -2938,9 +3083,10 @@ elif [ -n "$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 [ "$SKIP_TORCH" = true ]; then
# No-torch: install unsloth + unsloth-zoo with --no-deps, then
# runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
@ -2964,6 +3110,7 @@ elif [ -n "$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
@ -2981,8 +3128,11 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
"unsloth @ git+https://github.com/unslothai/unsloth@${UNSLOTH_INSTALL_REF}" unsloth-zoo
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 "$_UNSLOTH_TORCH_OVERRIDES"
_UNSLOTH_TORCH_OVERRIDES=""
# aarch64 + NVIDIA (DGX Spark / GB10 / N1X): unsloth's x86_64-oriented cuXXX
# extras break 4-bit QLoRA, but aarch64 manylinux wheels work (verified on
# sm_121 via PTX JIT). Best-effort: no wheel keeps 16-bit LoRA / full finetuning.
@ -3072,7 +3222,7 @@ if [ "$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=""
@ -3266,7 +3416,7 @@ 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

2
scripts/install_rocm_wsl_strixhalo.sh
View file

@ -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 <<EOF

4
scripts/lockfile_supply_chain_audit.py
View file

@ -2,7 +2,7 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
"""Lockfile supply-chain audit for the Studio frontend and Tauri shell.
"""Lockfile supply-chain audit for the Unsloth frontend and Tauri shell.
Runs BEFORE `npm ci` / `cargo fetch` in CI. Refuses to proceed when a
lockfile contains patterns indicating supply-chain injection (npm
@ -294,7 +294,7 @@ CARGO_REGISTRY_SOURCE = "registry+https://github.com/rust-lang/crates.io-index"
# Cargo non-registry source allowlist: `(crate_name, exact_source_string)`.
# Both must match verbatim; bumping the pinned SHA forces a re-review.
# Studio's Tauri shell pulls `fix-path-env` from git because it is not
# Unsloth's Tauri shell pulls `fix-path-env` from git because it is not
# published to crates.io; commit c4c45d5 was reviewed when it landed.
CARGO_SOURCE_ALLOWLIST: tuple[tuple[str, str], ...] = (
(

2
scripts/scan_npm_packages.py
View file

@ -62,7 +62,7 @@ REPO_ROOT = Path(__file__).resolve().parents[1]
# Hard caps (deliberately conservative; npm tarballs in this repo are
# all well under these limits, so a packaging spike is noticeable).
# ─────────────────────────────────────────────────────────────────────
# Caps calibrated against the real Studio frontend transitive closure:
# Caps calibrated against the real Unsloth frontend transitive closure:
# - typescript.js is 9.1 MB (TS compiler bundled into one file)
# - mermaid 11.x dist/mermaid.js.map is ~12 MB (sourcemap)
# - lightningcss-linux-x64-{gnu,musl}.node is 10 MB

4
scripts/scan_packages_baseline.json
View file

@ -95,8 +95,8 @@
"file": "fastapi/routing.py",
"check": "C2 polling/beaconing loop detected",
"severity": "CRITICAL",
"evidence": "L586: while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
"evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
"evidence": "L587: while True: sha256:06c2c7f15d73bf192e5e3272c5ff5fcaeff7f6774fef5f4eca6ef473ae50e2b3",
"evidence_hash": "57acd497f404c203e4450d0580ad85aa8a33406e8d64ad06fbac6cf47d97b24d"
},
{
"package": "fastmcp-slim",

20
scripts/stamp_studio_release.py
View file

@ -2,7 +2,7 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Stamp and verify display-only Studio release metadata for builds."""
"""Stamp and verify display-only Unsloth release metadata for builds."""
from __future__ import annotations
@ -50,7 +50,7 @@ MAX_VERSION_LENGTH = 64
PLACEHOLDER = """# 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
@ -145,7 +145,7 @@ def build_info_source(version: str | None) -> 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

14
scripts/uninstall.ps1
View file

@ -103,7 +103,7 @@ 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:
# <root>\share\studio.conf, <root>\unsloth_studio\.unsloth-studio-owned,
# or <root>\bin\unsloth.exe.
function _IsStudioRoot {
@ -184,7 +184,7 @@ 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.
@ -227,7 +227,7 @@ 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 }
@ -243,8 +243,8 @@ function Uninstall-UnslothStudio {
return $false
}
# Stop a Studio backend whose port is recorded in <DataDir>\studio.port.
# Only kills if the listening PID's exe path is under a known Studio root.
# Stop an Unsloth backend whose port is recorded in <DataDir>\studio.port.
# Only kills if the listening PID's exe path is under a known Unsloth root.
function _StopByPortFile {
param([string]$PortFile, [string[]]$KnownRoots)
if (-not (Test-Path -LiteralPath $PortFile -PathType Leaf)) { return }
@ -392,7 +392,7 @@ 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
@ -497,7 +497,7 @@ 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.

14
scripts/uninstall.sh
View file

@ -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"
@ -281,7 +281,7 @@ _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..."

12
studio/MCP.md
View file

@ -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

2
studio/Unsloth_Studio_Colab.ipynb
View file

@ -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)"
]
},
{

2
studio/backend/assets/chat_templates/gemma-4-edge.jinja
View file

@ -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<channel|>" block on enable_thinking=false is
NOT emitted. Google ships a distinct template for E2B/E4B (google/gemma-4-E2B-it,

2
studio/backend/assets/chat_templates/gemma-4.jinja
View file

@ -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.
-#}

2
studio/backend/auth/authentication.py
View file

@ -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))

10
studio/backend/auth/bootstrap_timeout.py
View file

@ -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 "

6
studio/backend/auth/storage.py
View file

@ -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

6
studio/backend/auth/terminal_prompt.py
View file

@ -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

8
studio/backend/cloudflare_tunnel.py
View file

@ -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://<ip>:<port> is often unreachable (https-vs-http, blocked ports,
closed security groups); a cloudflared quick tunnel gives a free
https://*.trycloudflare.com URL that works anywhere, with no account or domain.
Best-effort throughout: any failure collapses to "no URL" and Studio keeps
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()

12
studio/backend/colab.py
View file

@ -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;">
<img src="https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/unsloth-gem.png"
height="48" style="display:block;">
Shareable Studio Link is Ready!
Shareable Unsloth Link is Ready!
</h2>
<a href="{cloudflare_url}" onclick="var w=window.open(this.href,'_blank');if(!w){{return true;}}return false;"
style="display: inline-flex; align-items: center; gap: 10px; padding: 14px 28px;
@ -223,7 +223,7 @@ def _shareable_link_html(cloudflare_url: str) -> str:
def _show_and_embed(port: int, *, cloudflare_url: "str | None" = None):
"""Render the Studio header + iframe for *port*, with a shareable-link card above
"""Render the Unsloth header + iframe for *port*, with a shareable-link card above
when *cloudflare_url* is set. Falls back to serve_kernel_port_as_iframe."""
url = get_colab_url(port)
logger.info(f"🌐 Unsloth Studio URL: {url}")
@ -281,7 +281,7 @@ def start(port: int = 8888, *, cloudflare: bool = False):
Args:
port: Port to bind/serve on.
cloudflare: Opt in to a shareable Cloudflare HTTPS link reachable from any
device (default OFF). It exposes Studio's login page beyond Colab, so it
device (default OFF). It exposes Unsloth's login page beyond Colab, so it
stays an explicit opt-in; the default shows only the in-tab proxy iframe.
Usage:
@ -292,10 +292,10 @@ def start(port: int = 8888, *, cloudflare: bool = False):
logger.info("🦥 Starting Unsloth Studio...")
# Fast path: Studio already running (cell re-run). Re-launching would collide on
# Fast path: Unsloth already running (cell re-run). Re-launching would collide on
# the port, so just re-show the link and iframe.
if _is_studio_healthy(port):
logger.info(f" Studio is already running on port {port} — reusing existing server.")
logger.info(f" Unsloth is already running on port {port} — reusing existing server.")
# try/finally: tear the tunnel down even if interrupted mid-start/render.
try:
cf_url = start_cloudflare_tunnel(port) if cloudflare else None

6
studio/backend/core/data_recipe/jobs/parse.py
View file

@ -133,7 +133,7 @@ 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."),
),
)
@ -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."),
),
)

2
studio/backend/core/data_recipe/local_callable_validators.py
View file

@ -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())

4
studio/backend/core/data_recipe/service.py
View file

@ -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)

2
studio/backend/core/inference/__init__.py
View file

@ -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.
"""

2
studio/backend/core/inference/anthropic_compat.py
View file

@ -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

2
studio/backend/core/inference/chat_template_helpers.py
View file

@ -150,7 +150,7 @@ def _split_partial_marker(text: str, marker: str) -> tuple[str, str]:
class ReasoningChannelNormalizer:
"""Incrementally convert one native reasoning channel to ``<think>``.
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.

4
studio/backend/core/inference/chat_templates.py
View file

@ -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.
"""

14
studio/backend/core/inference/external_provider.py
View file

@ -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:

795
studio/backend/core/inference/llama_cpp.py
View file

File diff suppressed because it is too large Load diff

2
studio/backend/core/inference/llama_keepwarm.py
View file

@ -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
)

65
studio/backend/core/inference/llama_server_args.py
View file

@ -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] = []

2
studio/backend/core/inference/llama_stats.py
View file

@ -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.
"""

10
studio/backend/core/inference/local_model_resolver.py
View file

@ -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)

2
studio/backend/core/inference/mcp_client.py
View file

@ -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.

93
studio/backend/core/inference/mlx_inference.py
View file

@ -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 <think> 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 <think> 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

17
studio/backend/core/inference/orchestrator.py
View file

@ -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,

4
studio/backend/core/inference/passthrough_healing.py
View file

@ -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 (``<tool_call>{...}</tool_call>``,
Gemma ``<|tool_call>...``, ``<function=...>`` 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``.

8
studio/backend/core/inference/pricing.py
View file

@ -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):

5
studio/backend/core/inference/providers.py
View file

@ -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,
},

2
studio/backend/core/inference/safetensors_agentic.py
View file

@ -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,

2
studio/backend/core/inference/sandbox_site/sitecustomize.py
View file

@ -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.

2
studio/backend/core/inference/tool_loop_controller.py
View file

@ -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.

14
studio/backend/core/inference/tools.py
View file

@ -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/<pid>/environ unreadable to its children.
"""Make the Unsloth process's /proc/<pid>/environ unreadable to its children.
Stripping the child env is not enough on Linux: a bypassed same-UID child
can read /proc/<getppid()>/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/<parent>/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/<parent>/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."
)

31
studio/backend/core/inference/worker.py
View file

@ -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,

2
studio/backend/core/rag/captioner.py
View file

@ -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

25
studio/backend/core/rag/config.py
View file

@ -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

2
studio/backend/core/rag/embed_llama_server.py
View file

@ -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).
"""

2
studio/backend/core/rag/embeddings.py
View file

@ -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"}

2
studio/backend/core/training/resume.py
View file

@ -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)

2
studio/backend/core/training/trainer.py
View file

@ -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

6
studio/backend/core/training/training.py
View file

@ -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

12
studio/backend/core/training/worker.py
View file

@ -1126,7 +1126,7 @@ _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
@ -1283,7 +1283,7 @@ _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(
(
@ -1335,7 +1335,7 @@ 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] = []
@ -1938,7 +1938,7 @@ 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)
@ -2147,7 +2147,7 @@ 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)
@ -2853,7 +2853,7 @@ 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:

2
studio/backend/hub/services/download_lifecycle.py
View file

@ -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

40
studio/backend/hub/services/models/cache_inventory.py
View file

@ -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,

2
studio/backend/hub/services/models/folder_browser.py
View file

@ -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

17
studio/backend/hub/services/models/local_inventory.py
View file

@ -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),

2
studio/backend/hub/services/models/ollama.py
View file

@ -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 ``<ollama_dir>/.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 ``<ollama_dir>/.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:

338
studio/backend/hub/tests/test_model_services.py
View file

@ -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)

2
studio/backend/hub/utils/state_dir.py
View file

@ -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:

28
studio/backend/main.py
View file

@ -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

26
studio/backend/mcp_server.py
View file

@ -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

160
studio/backend/models/inference.py
View file

@ -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

2
studio/backend/models/training.py
View file

@ -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."
),
)

4
studio/backend/plugins/data-designer-github-repo-seed/README.md
View file

@ -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 .

2
studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py
View file

@ -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).

2
studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
View file

@ -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

2
studio/backend/requirements/extras-no-deps.txt
View file

@ -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

2
studio/backend/requirements/no-torch-runtime.txt
View file

@ -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

2
studio/backend/requirements/single-env/constraints.txt
View file

@ -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

4
studio/backend/requirements/studio.txt
View file

@ -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

Some files were not shown because too many files have changed in this diff Show more