Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Commits on main
Studio: Audio Edit step numbers as text, standard Transcribe icon (#15484)
9d184f3 Michael Han committed 25 days ago
Studio: API examples behind one tab row, an access group, API page fixes (#15467)
9fb0221 Michael Han committed 25 days ago
Studio: solid primary buttons on update notifications (#15470)
6b95f80 Michael Han committed 25 days ago
Studio: Library chat, project and section cards keep one size (#15471)
9e09d19 Michael Han committed 25 days ago
Studio: top-align the settings rows the shared row components do not cover (#15468)
7ff6f3e Michael Han committed 25 days ago
Studio: Run settings divider, preset picker fill and a clearer Always load (#15472)
ceaad59 Michael Han committed 25 days ago
Studio: a switch per permission mode in Settings > Sandbox (#15469)
fab26f4 Michael Han committed 25 days ago
Studio: centre the chat column under classic scroll bars in WebKit (#15392)
8d0a5aa Michael Han committed 25 days ago
Keep DeepseekV4 per-label rope bases when rope_scaling is replaced on transformers 5.21 (#15425)
388e0ad Daniel Han committed 25 days ago
Studio video: pre-seed the denoiser compile's float specialisations on a returning load (#15406)
e417bfd Daniel Han committed 25 days ago
Studio tests: queue each FIFO admission waiter before starting the next (#15457)
92287a4 Daniel Han committed 25 days ago
Tests: fix the two spoof-safety lint failures on main (#15454)
a4e651e Daniel Han committed 25 days ago
Attention: packed fp32 rerouted to xformers keeps an fp32 output without autocast (#15410)
f001b5b Daniel Han committed 25 days ago
Native FSDP2 fp16 scaler: agree on overflow over ranks without an explicit unscale_, and give fused optimizers one 0 / 1 flag over all devices (#15416)
291a4c7 Daniel Han committed 25 days ago
gpt-oss: read config.transformers_weights before the standard filenames when probing the expert layout (#15386)
029a596 Daniel Han committed 25 days ago
FSDP2: gather before clearing the post-forward mesh in native_fsdp2_unsharded (#15426)
dbe9846 Daniel Han committed 25 days ago
Tensor parallel: average fp16 dp gradients in fp32 so finite gradients cannot sum to inf (#15415)
91bbf46 Daniel Han committed 25 days ago
Tensor parallel: generate, adapter / tokenizer saves, merged and GGUF exports, and reload cells under UNSLOTH_TP_SIZE (#15368)
78b2f82 Daniel Han committed 25 days ago
FSDP2 bitsandbytes: no per-parameter device sync, bounded GPU memory in the optimizer save (#15384)
8abe47a Daniel Han committed 25 days ago
loader: do not restore an fp8 quantization_config over dequantized weights (#15364)
c3a2907 Daniel Han committed 25 days ago
Keep SFT loss_type="dft" on the DFT objective when label smoothing is set (#15396)
2debcf3 Daniel Han committed 25 days ago
Studio backend tests: close a scoped MCP session a test leaves cached (#15407)
0ea0473 Daniel Han committed 26 days ago
FSDP2 / DDP: only keep torchao weights FSDP2 cannot shard out of fully_shard and the DDP broadcast (#15373)
f93c871 Daniel Han committed 26 days ago
TP x FSDP2 parity test: per-arch step-norm tolerance for phi3's fused shards (#15400)
c4d7842 Daniel Han committed 26 days ago
Studio video: refuse a NaN render on the native MiniMax-H3 route (#15394)
d5126dd Daniel Han committed 26 days ago
Multi-account Playwright: retry the login page once when the app's own logout redirect aborts it (#15398)
5fd8d9f Daniel Han committed 26 days ago
TP Trainer parity tests: keep eps 1e-4 to the UNSLOTH_TP_SIZE runs, native Trainer runs back on the default (#15397)
f63cedd Daniel Han committed 26 days ago
Studio sd.cpp: verify a new bundle starts before it replaces the installed one; Windows ROCm without the HIP SDK uses Vulkan (#15371)
7d9f122 Daniel Han committed 26 days ago
Studio diffusion: pre-seed the background compile's float specialisations on a returning load (#15393)
c4d100a Daniel Han committed 26 days ago