## unslothai/unsloth — v0.1.701-beta…v0.1.702-beta

_116 commits._

### Features
- **Studio: add image and video generation presets (#8390)** (4ca3836)
- **Studio: detect real GGUF imatrix support, and give the local export its credential (#8603)** (0c6c1b9)
- **Unsloth Studio: add ChatGPT subscription chat with Codex tools (#8511)** (ef8daca)
- **Add files via upload** (f20d465)
- **Add files via upload** (fd5dadb)

### Fixes
- **Fix gallery and sidebar menu interactions (#8582)** (06c90e4)
- **Fix three frontend contract tests that main is currently red on (#8703)** (b2158c2)
- **Keep the #8577 AMD peer guards message-only, and fix the table drift they exposed (#8689)** (99bcfd3)
- **Say ROCm does not cover RDNA 1 instead of advising a fix that cannot work (#8577)** (d5a2160)
- **fix(audio): install the audio decode shim before the training worker loads a dataset (#8436)** (97841d8)
- **fix(audio): do not read an unreadable tokenizer_config.json as "not an audio model" (#8625)** (e10ed80)
- **fix(amd): gate ROCm GPU selection and crash recovery on the build's arch coverage (#7670)** (5a3e9fc)
- **tests: do not read a host refusal as an H3 reference-load regression (#8638)** (620bd47)
- **Fix the desktop-auth routes stub so the health test runs again (#8590)** (75d983d)
- **Images: fix the img2img VAE dtype crash and make the Resolution control bound Transform (#8583)** (6d37709)
- **Fix Apple Silicon M4+ CPU frequency reported as MHz instead of GHz (#8571)** (94cc7ee)
- **Fix Cloudflare documentation link (#8572)** (51fac01)

### Backend
- **Versioning** (662fead)
- **Desktop: keep "Run Unsloth at login" when something deletes the Run value (#8707)** (4dab08b)
- **Windows: stop depending on the generated unsloth.exe console script (#8592)** (1b48147)
- **studio: honour the gpu selection for image and video loads (#8645)** (e7ea3cb)
- **Studio: run the local tool loop against every capable external provider (#8665)** (0e158a8)
- **Apply the kwarg-spacing formatter to the #8677 system poll test (#8699)** (c160215)
- **Model hub: show the Meta mark for Meta's own orgs (#8691)** (e93eece)
- **Reduce antivirus false positives in the desktop installers (#8586)** (5a5bf64)
- **Windows: stop the oversize guard tests from emptying os.environ (#8696)** (79fbf94)
- **Studio: stop the /api/system poll from pinning a CUDA/HIP primary context (#8677)** (44a05f3)
- **Windows: start the backend from a usable folder on login autostart (#8575)** (715535d)
- **Studio: set DYLD_LIBRARY_PATH for llama-server on macOS, and classify macOS startup failures (#8574)** (ce3f5c9)
- **Studio: keep running when the main window closes (#8675)** (e95fd2b)
- **studio: harden the launcher-refresh installer fetch (#8542)** (18dbedb)
- **Studio: import Open WebUI chat exports (#8643)** (42c8671)
- **Studio: give GGML_CUDA_ENABLE_UNIFIED_MEMORY a real off switch (#8651) (#8680)** (7822b7c)
- **Restore the #8335 WMI guard anchor after the RDNA1 wrapper (#8684)** (9ae9b2c)
- **tests: require the export pin fallback to cover a half-built unsloth_zoo (#8685)** (fd27f7f)
- **Studio: run the transformer-quant smoke probe in a child so planning a download costs no VRAM (#8671)** (0e70c4c)
- **Studio: three follow-ups to the Deep Research main-thread work (#8633)** (8993f33)
- **Attach long pastes as a text file in Chat (#8472)** (faf3343)
- **Studio: square off the MiniMax H3 mode dialog (#8659)** (3ddf4a2)
- **Update README.md** (72e7257)
- **Stop the GRPO hidden-states wrapper paying for logits it discards (#8576)** (baa96a7)
- **Do not let a CUDA-mismatched torchaudio take the whole import with it (#8496)** (b86944e)
- **Stop the idle-unload tests racing a fixed wall-clock window (#8674)** (06355d4)
- **Studio: stop a streaming research run re-rendering the whole chat (#8634)** (b979529)
- **Match the export pin test to the widened exception handler (#8673)** (665a004)
- **Make the VRAM budget fraction tunable (#8589)** (98cefae)
- **tests: follow the remote-connection contract through its refactor (#8467)** (620b9fa)
- **Studio: optimize startup by deferring fine-tuning actions (#8624)** (710beb1)
- **Free the intermediate 16-bit merge when the GGUF quants will not fit (#8500)** (71656a0)
- **studio: close the delete-vs-load races around the H3 companion repos (#8657)** (b4702b2)
- **Keep the drafterless retry intact when the arch gate narrows the argv (#8667)** (f4a7458)
- **Ask whether a device is present before asking what it can do (#8653)** (4bff729)
- **Studio: allow max output overrides for custom providers (#8512)** (5e184f5)
- **Update issue templates** (08508fd)
- **Stop the APU unified-memory tests inheriting the shell's GPU mask (#8662)** (794b16e)
- **tests: pin the remote GGUF compute reserve (#8660)** (965f7b4)
- **Studio: switch llama.cpp backends from the UI (#8520)** (5426a78)
- **Surface an untyped Responses error frame instead of skipping it (#8650)** (ab9f23b)
- **Read the Responses event type from the SSE event field (#8608)** (b337630)
- **Let a remote GGUF estimate be priced without its compute reserve (#8641)** (d4139d6)
- **Studio: validate legacy sd binary discovery (#8560)** (22cb12e)
- **Studio: contain RAG embedder torch allocation crashes (#8609)** (ec9db28)
- **Desktop: flush gated tool approval cards immediately (#8628)** (2c058bf)
- **Detect the Radeon AI PRO R9700 (gfx1201): it carries neither 9070 nor 9080, so name inference found nothing (#8573)** (8502cf8)
- **Studio: confirm before clearing the video gallery (#8354)** (a4c77ed)
- **Refuse an image / video GGUF before launching llama-server, and open it on its own page (#8584)** (a3ed5b5)
- **Write a readable traceback under each JSON log record (#8585)** (e005b1e)
- **studio: stop the MiniMax-H3 refusal telling users to delete /usr/bin (#8569)** (fe37ed0)
- **studio: reduce main-thread work in Deep Research and tooltips (#8525)** (a14c86e)
- **launch embedding ggufs with --embedding (#8524)** (2c1d09c)
- **Studio: stop telling the model it is sandboxed under Full access (#8562)** (dac7b51)
- **Studio: defer optional GPU startup work (#8564)** (3c590a5)
- **Studio: require managed backend for linked folders (#8536)** (a1e96ec)
- **Harden the workflow-trigger lint: scan .yaml, and host it outside the workflow it audits (#8545)** (90a6a23)
- **Studio: keep the extras install working under a hardened uv.toml / pip.conf (#8579)** (88262f5)
- **Studio: two ways past the stdio MCP UI-session gate (#8551)** (6d4a9ff)
- **Studio: drop the speculative drafter under Auto when only the model fits in VRAM (#8435)** (a7e073e)
- **Studio: validate external provider base URLs before proxying (#8549)** (5eca9fa)
- **security: the network check could not see httpx2 (#8565)** (4a53e80)
- **Pin the ROCm-on-WSL bootstrap to immutable refs (#8540)** (e468965)
- **Studio: honor a request's enable_tools: false instead of overriding it (#8547)** (098a6a0)
- **tests: drop --single-process from the Chromium launch args (#8563)** (9c46cc4)
- **studio: fail closed on HF commit-operation uploads in the sandbox gate (#8544)** (efbc909)
- **Update README.md** (4caf91c)
- **Studio: only a UI session may define a local (stdio) MCP command (#8550)** (7dfee7e)
- **Studio: linear-time tool signal scanning in the safetensors and healer paths (#8494)** (9a3a30e)
- **Studio: finish the backend CI cleanup #8506 started (#8554)** (b01c8b9)
- **Studio: verify the flash-attn import after installing it (#8465)** (947b4bb)
- **security: lockfile audit must block non-registry sources and missing integrity by default (#8541)** (f5c64f4)
- **Pin sha256 for triton-xpu 3.6.0 wheels in intelgputorch210 (#8543)** (f27c6fc)
- **Studio: drop a duplicate import that breaks the frontend build (#8553)** (bdd8a8f)
- **Studio: drop the duplicated HubModelPicker import in model-selector (#8534)** (993e3e4)
- **Clear the four main CI reds blocking every open PR (#8506)** (a23951b)
- **Install torchao in Backend CI, and stop one test's allowlist answer leaking into the rest (#8486)** (4f520e0)
- **Studio: read ?model= from the diffusion page's own route match, not the root one (#8260)** (7bbcc8d)
- **Studio backend performance: five superlinear paths in the routes and data layer (#8499)** (87cd048)
- **Studio: cut backend start time and stop blocking the event loop (#8498)** (d7e88a8)
- **Studio: stop rescanning the whole answer on every streamed token (#8538)** (0e47051)
- **Studio: give the tool-call strip one owner and one scan order (#8427)** (48cf5ff)
- **Studio: stop the memory guards trusting an over-reported free VRAM on Windows ROCm (#8482)** (b53de1e)
- **Studio: report host VRAM usage when no single GPU's usage can be attributed (#8481)** (26ef15b)
- **Route spoofed Strix Halo GPUs to the AMD per-gfx index (#8480)** (5714530)
- **Studio: classify a moved or mixed model folder from the checkpoint, not from directory order (#8475)** (671bc1b)
- **Studio: name a connected model the provider dropped instead of its raw id (#8470)** (dfae53b)
- **Drop socket reads that arrive after an h11 connection is closed (#8469)** (c63df2d)
- **Studio: keep the compiled cache when a sibling backend is live (#8457)** (1ef30b9)
- **Studio: stop building test scratch paths inside a macOS sensitive root (#8485)** (e2a8043)
- **Studio: remove obsolete onboarding and model code (#8453)** (bdb4e33)
- **Read macOS zombie status from sysctl, the call that answers (#8493)** (2ca3f66)
- **Update README.md** (4acc36b)
- **Update README.md** (8fd61b5)
- **Update README.md** (4916481)
- **Update README.md** (d96eef5)
- **Bump install.sh / install.ps1 pin to unsloth>=2026.8.15 (#8491)** (14c6ce3)

### Chore
- **ci: stop asserting torch on the Intel Mac clean-machine leg (#8693)** (29806d3)
- **CI: run the Studio desktop unit tests on macOS (#8487)** (e91ab9a)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/unslothai/unsloth?utm_source=github-action)._