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
DeepSeek-V4.1 (transformers port): load, place and train it with LoRA / QLoRA (#17625)
3221ed0 Daniel Han committed 7 days ago
Studio: keep legacy function calls and their results on chat import (#17153)
0650065 DimitriChrysafis committed 7 days ago
Studio: read more chat export shapes without losing turns (#16971)
3fc613e DimitriChrysafis committed 7 days ago
Block compile: drop a caller's strict mark_dynamic so Qwen3.8-Flash-Next layers stay compiled (#17609)
f1c5502 Daniel Han committed 7 days ago
Strict Dynamo regions: turn suppress_errors off only where the flag exists (#17492)
e60d3d5 Daniel Han committed 7 days ago
PEFT MoE snapshot: add transformers' Dots3NoteTextModel alias (#17606)
d9827dc Daniel Han committed 7 days ago
Bound the batched multi-LoRA loss memory: fused Triton cross entropy on GPU, chunked loss elsewhere (#16655)
480fd26 Daniel Han committed 7 days ago
LoRA decode: return the caller's out tensor so batched decode keeps one mm per projection (#16588)
562d802 Daniel Han committed 7 days ago
Studio: keep tagged edit text lossless on a no-op save (#17157)
8f99876 DimitriChrysafis committed 7 days ago
Build multi-LoRA expert groups when the batch is set (#16651)
3de707f Daniel Han committed 7 days ago
Fix tensor parallel adapter save after the MoE expert LoRA factor swap (#16654)
f2fa80c Daniel Han committed 7 days ago
Tests: skip the case-kept clip sidecar test on a case-insensitive filesystem (#17602)
e0ef452 Daniel Han committed 7 days ago
Tests: check the MLX dataset slice through the shared slice_dataset_rows helper (#17601)
75f55aa Daniel Han committed 7 days ago
Studio: decode UTF-16 without BOM and repair .doc pieces split mid-surrogate (#16969)
bb377fe DimitriChrysafis committed 7 days ago
Normalize save_method before the merge so a respelled merged_16bit still writes config.json (#16933)
3696d7d Vineeth Sai Varikuntla committed 7 days ago
Triton kernels: GPU guard against mid-generate recompiles, comment every tl.constexpr (#16575)
2c94592 Daniel Han committed 7 days ago
Keep the user override when the torch config guard restores a deleted entry (#16652)
1c43803 Daniel Han committed 7 days ago
Fix SDXL checkpoint selection for catalog downloads and Run (#17582)
74d2c1c Lee Jackson committed 7 days ago
get_chat_template: write the chosen system message into the Ollama Modelfile (#16863)
5249097 Mohammad Hijjawi committed 7 days ago
Multi-LoRA: keep the base output dtype on the PyTorch path under autocast (#16649)
b13d8b0 Daniel Han committed 7 days ago
Studio: keep diffusion captions that hold U+2028, U+2029 or U+0085 (#17436)
5541af5 Vineeth Sai Varikuntla committed 7 days ago
Studio: keep imported system messages as system through save and load (#17052)
a7ee424 DimitriChrysafis committed 7 days ago
Ollama Phi-4 template: write roles as text, not <|user|> pseudo tokens (#17285)
b2784e3 Jae Birdsall committed 7 days ago
Legacy CLI: accept true/false spellings for --use_gradient_checkpointing (#17115)
a190904 Vineeth Sai Varikuntla committed 7 days ago
Studio: send the OpenCode Go session header on the older /zen/go base URL (#15981)
287844f Daniel Han committed 7 days ago
check_dataset: measure the text SFT trains on (#17047)
5bec89f Shubhankar Tripathy committed 7 days ago
Studio: honor Claude Code's thinking-off setting on /v1/messages (#17557)
53eca64 oobabooga committed 7 days ago