## unslothai/unsloth — v0.1.526-beta…v0.1.527-beta

_80 commits._

### Features
- **Studio: add MiniMax H3 video generation (#7989)** (495ba21)
- **Resolve latest mlx-vlm and route its new processors (#8251)** (c7ffee7)
- **add folder picking and drag-drop for paired image and caption datasets (#8216)** (c2c04a4)

### Fixes
- **remove the unused images train dialog and fix media page responsiveness (#8215)** (cf28871)
- **Fix Studio loaded-model GGUF quant label (#8238)** (1357ee8)
- **fix(studio): verify TLS against the OS trust store at runtime (corporate TLS-inspection proxies) (#8108)** (16fac7b)
- **fix(studio): wire llama's OpenMP runtime into slim whisper installs (#8230)** (954468a)
- **Studio: fix Windows desktop updates failing integrity checks (#8185)** (968d2e5)
- **fix(studio): reserve a bounded OS headroom on unified ROCm APUs, not a flat 20% (#7960)** (7aa0217)
- **fix(studio): release the inference subprocess's VRAM when its last model unloads (#8221)** (cdeea9c)

### Backend
- **Versioning** (154fa22)
- **Install the Linux CUDA sd-cli, and bump the pin two releases forward (#8208)** (ba1faf0)
- **Make the diffusion checkpoint-resume tests describe the host they run on (#8273)** (f16769c)
- **Studio: keep the overlay stack off a maximised Live monitor (#8190)** (fbf5eba)
- **Turn off embedding offload when the model is dispatched across devices (#8272)** (e0d3b1d)
- **Say when attention has no fused kernel, and log the request behind a failed video (#8233)** (7494373)
- **Studio: LTX-2 video LoRA training from still images (#8196)** (f02e49f)
- **Studio: size FLUX.2-klein 9B LoRA training from its own weights, not the 4B family number (#8263)** (5a8e4ef)
- **Studio: make the image memory plan resolution-aware and stop reserving the text encoder during sampling (#8224)** (f55d736)
- **Studio: let image generation be stopped (#8219)** (f9656fd)
- **Assert the image family's pipeline class in the training preflight, before the GPU teardown (#8236)** (bcfc626)
- **Restore SFT loss_type='nll' default over TRL 1.7.0 chunked_nll (#8261)** (43b3cd1)
- **Studio: stop leaking child processes on an abnormal exit (#8170)** (c6ad59a)
- **Studio: grey the sidebar dot a finished chat leaves (#8258)** (3a6f395)
- **Stop the CPU fake-train tests from spoofing torch for the whole session (#8259)** (29fa2f7)
- **Desktop: keep probing the announced backend port instead of discarding it (#8211)** (eb892d6)
- **Studio: budget a pre-cast fp8 text encoder at the size it actually loads (#8218)** (f12ab6b)
- **Studio: name the quant a diffusion GGUF pick actually loaded (#8241)** (f21b04c)
- **Studio: realign the image precision contracts with the fail-closed refusal (#8257)** (d9c15f6)
- **Studio: align Images and settings layouts (#8164)** (2ca477d)
- **Carry the sliding window into SDPA, and treat a zero window as no window (#8253)** (7088df8)
- **Do not let a broken TensorFlow install break the Unsloth import (#8124)** (57339b4)
- **Keep an explicitly requested float32 model in float32 without bfloat16 (#7867)** (85a036a)
- **Report the network error when the offline retry finds nothing cached (#8121)** (f0ef75e)
- **Diffusion: allow fp8 on Qwen-Image again, and split the training deny (#8018)** (8e55860)
- **GRPO: wrap the model that owns the head when falling back to hidden states (#8231)** (4a02eb6)
- **Make the desktop release contract tests fail when the contract breaks (#8228)** (0cd73cf)
- **Reject an incompatible GGUF and base pair before the download, not after (#8158)** (1d11616)
- **studio: infer splits for metadata-free cached datasets (#8144)** (eda3b26)
- **Diffusion: source every image family's companions from unsloth mirrors (#8030)** (b8f09bc)
- **Name a torchvision whose compiled ops do not match torch (#8128)** (2536edd)
- **Make install.ps1 work with the user's PowerShell profile loaded (#8161)** (a151ac8)
- **Fetch the llama.cpp validation model only when the smoke test reads it (#8248)** (82bf2e6)
- **Desktop: confirm quit during training  (#7790)** (4c6aa87)
- **Do not scan vendored packages that refuse to import on our floor (#8252)** (d458029)
- **Stop padding-free SFT from tripping the TRL >= 1.0.0 max_length guard (#7951)** (1fe8c7a)
- **Resume diffusion training from a stop-and-save checkpoint (#8163)** (4e7755b)
- **Studio: only clean up GGUF artifacts the export owns (#7940)** (e8fcbd6)
- **Backfill the conversion symbols peft imports and transformers 5 dropped (#8051)** (f48eeb2)
- **Format the release-permissions test so pre-commit.ci stops failing on main (#8247)** (f7b3116)
- **Format the release permissions test to the kwarg-spacing hook (#8245)** (4bd9a2b)
- **Let the floating monitor paint over the notification stack (#8199)** (b62a16d)
- **Stop the idle-unload TTL test racing a loaded runner (#8175)** (1a23d56)
- **Let the pipeline gate survive a diffusers whose lazy submodule cannot import (#8227)** (0d9e6bd)
- **Repo tests: pin the post-publish VirusTotal contract instead of the old pre-flight one (#8240)** (b844e55)
- **Tell users to upgrade the dependency, not transformers, when a --no-deps install leaves a floor unmet (#8212)** (f652f84)
- **Studio: stream oversized diffusion components on low-VRAM GPUs (#8141)** (a08234d)
- **Revert "Say when xFormers is installed but its kernels cannot load (#8157)" (#8237)** (ee76b5f)
- **Say when xFormers is installed but its kernels cannot load (#8157)** (72aec0f)
- **Pin the training precision, video preset and image metadata contracts (#8162)** (5d461d7)
- **Report the precision actually in use, and refuse an explicit one we cannot honor (#8165)** (39147e4)
- **Install a CUDA-matched xFormers on Windows instead of whatever PyPI serves (#8156)** (e219b46)
- **Stop a finished download reporting 0 B and offering Retry (#8160)** (68408d0)
- **Studio: preserve training config values across YAML load and save (#7942)** (113a2cf)
- **Desktop: stop accidental reloads in the Windows WebView (#7945)** (4532df2)
- **Studio: surface an MLX completion-masking miss as a warning, not a status line (#8133)** (6fd4e9a)
- **Studio: only dodge an overlay box that reaches the corner stack (#8210)** (da6aeae)
- **Repin six UI contracts on what they guard, not how it was spelled (#8197)** (37f43eb)
- **Studio: stop MTP forcing llama-server to a single parallel slot (#8172)** (dc49ff6)
- **tests: build the hub-error fixtures without calling __init__ (huggingface_hub 1.x) (#8207)** (b4b780b)
- **Desktop release: run the VirusTotal scan after publishing, in its own job (#8194)** (b8dac29)
- **tests: make `pytest tests` collect cleanly and drop the deleted QVQ registry mirror (#8206)** (68e5c3f)
- **Desktop release: start the publish job alongside the build matrix (#8193)** (ae59a9c)
- **Desktop release: build the three platform legs in parallel (#8191)** (f9dc723)
- **studio: serve the deep research event stream over post so proxies stop buffering it (#8205)** (2a56df0)
- **Studio: eliminate Cargo and Clippy warnings (#8192)** (161a301)
- **Bump install.sh / install.ps1 pins to unsloth>=2026.8.9 and unsloth-zoo>=2026.8.6 (#8189)** (68cf69d)
- **Studio: make the gallery video link absolute so the desktop app can play and save it (#8173)** (b65cae5)
- **Desktop: show a closing overlay while the backend shuts down (#8151)** (68eefd3)

### Tests
- **test(studio): stop the backend suite depending on a live Hub (#8174)** (a6914ed)

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