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
Tests: load the helpers ORPO rewriter constants are built with (#15296)
9b98bb9 Daniel Han committed 26 days ago
Backend CI: run the multi-rank gloo suites in their own Repo tests shard (#15291)
231a4aa Daniel Han committed 26 days ago
Studio: build native int8 image transformers per tensor on a cache miss (#15289)
df4b4a5 Daniel Han committed 26 days ago
Studio backend tests: forget MCP discovery state between tests (#15287)
83c1c62 Daniel Han committed 26 days ago
Studio: cache image native int8 builds in the local prequant cache (#15283)
48a9c8d Daniel Han committed 26 days ago
PP gloo parity test: size gemma3's grad bounds to its noise on AVX2 kernels (#15286)
4bdf6c4 Daniel Han committed 26 days ago
DPO: UNSLOTH_FUSED_PREFERENCE_LOSS on TRL 1.x DPOTrainer._compute_loss (#15277)
b6cde16 Daniel Han committed 27 days ago
Flex attention: fitting tiles for Gemma 3 on sm86 / sm89 (L4, A10, RTX 4090) instead of eager (#15271)
47aaa26 Daniel Han committed 27 days ago
PP gloo parity test: give gemma3 its own per-element grad tolerance (#15273)
8b050a7 Daniel Han committed 27 days ago
SDPA on cuDNN: bucket masked and prefill calls, keep decode off cuDNN where torch would pick it (#15268)
a2aca8d Daniel Han committed 27 days ago
Decode graph KV growth test: gate its CUDA skips on has_real_cuda() (#15270)
9d3f406 Daniel Han committed 27 days ago
Studio: give web search engines a connect timeout so a blocked network fails fast (#15143) (#15161)
d53a090 Debojyoti Banerjee committed 27 days ago
Fused preference loss test: take TRL's own trainer as the reference (#15269)
24015fc Daniel Han committed 27 days ago
Studio: allow editing a refreshed reply when the tab Date disagrees (#15087)
21088a4 Sourav Rajvi committed 27 days ago
Studio: let --no-reasoning-preserve and a custom chat template turn off the Qwen3.8 preserve_thinking default (#15170)
808515e Debojyoti Banerjee committed 27 days ago
Studio: stop refusing our own llama-server on Windows when psutil misreports its listener (#15022) (#15034)
5f558cb Debojyoti Banerjee committed 27 days ago
GRPO truncated-support gate test: lift the block that binds _support_kwargs (#15267)
b9ae4e7 Daniel Han committed 27 days ago
TP x FSDP2 parity test: keep AdamW out of its eps cliff on the lm_head copy (#15264)
ab7624b Daniel Han committed 27 days ago
Define _recover_vulkan_stall below the CPU fallback it calls (#15261)
5cc1e43 Daniel Han committed 27 days ago
Studio: keep a Settings-picked embedder visible in the Hub (#15256)
db1e8d9 Daniel Han committed 27 days ago
MoE dW accumulate test: compare in-place against grad + dW, check the exact 1 on CUDA only (#15254)
b8e8cd7 Daniel Han committed 27 days ago
Route PPO, RLOO, BCO, Online DPO, XPO, Nash-MD and GKD to the zoo fused linear kernels (#15244)
8fec282 Daniel Han committed 27 days ago
SFT: compute loss_type="dft" on TRL's fused LM head instead of silently training NLL (#15213)
2c81e9f Daniel Han committed 27 days ago
Refuse transformers / accelerate cp_size under Unsloth DPO and GRPO (#15229)
caf650e Daniel Han committed 27 days ago
Refuse accelerate config cp_size under Unsloth SFT, not only TrainingArguments (#15212)
af99ed7 Daniel Han committed 27 days ago
RoPE: constexpr the strides and head counts fixed per model (#15243)
4db3eae Daniel Han committed 27 days ago
Studio: show one composer button while a reply streams (#15248)
5e61155 Michael Han committed 27 days ago
Studio: make Keep multiple models loaded unload idle models before spilling a load to the CPU (#15246)
6112594 Michael Han committed 27 days ago
Decode graphs: grow the KV cache instead of reserving max_new_tokens, and capture on the model's GPU (#15237)
dc55427 Daniel Han committed 27 days ago
Compiled blocks: packed rows run varlen attention outside the graph (#14808)
e6f72a0 Daniel Han committed 27 days ago