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: Qwen-Image-2.1 eager renders before the deferred compile now match ComfyUI (#15908)
dbf73e8 Daniel Han committed 23 days ago
Studio: Images icon for Reference, file search icon for Chat with files (#15806)
926a8f9 Michael Han committed 23 days ago
Studio: put a just-created API token in the Agents command and API examples (#15890)
3db27a2 Michael Han committed 23 days ago
Studio: lead the Live monitor tiles with the reading, unrounded (#15897)
c8ccc5b Michael Han committed 23 days ago
Studio tests: measure the int8 / fp8 dequant and nibble-swap peaks in a fresh process (#15912)
84d45b3 Daniel Han committed 23 days ago
Studio tests: keep GGUF loads off the host's Metal budget unless a test asks (#15910)
da8ec32 Daniel Han committed 23 days ago
Packed attention: varlen probes under fullgraph + dynamic=True compile (#15891)
0430356 Daniel Han committed 23 days ago
Studio: show the real error when a Responses stream fails (#15816)
9b3376d Nilay committed 23 days ago
Studio: keep refusing a symlinked spill where there is no O_NOFOLLOW (Windows) (#15887)
cf80abd Daniel Han committed 23 days ago
GRPO: take the per-step host syncs out of the loss path (#15855)
b80df64 Daniel Han committed 23 days ago
Studio tests: skip rag_conn where this Python's sqlite3 cannot load extensions (#15862)
9a9d5f5 Daniel Han committed 23 days ago
Desktop: re-enable GPU compositing on NVIDIA X11 with driver 610 or newer (#15546)
c3879b1 oobabooga committed 23 days ago
Backend CI: split the fsdp2 Repo tests shard in two (#15851)
06314b9 Daniel Han committed 23 days ago
Studio: skip the Keep resident page-lock when the llama-server --help probe is inconclusive (#15833)
1b3ad97 Daniel Han committed 23 days ago
Tests: FP8 routing tests pin the NVIDIA path so they pass on ROCm runners (#15838)
338616d Daniel Han committed 23 days ago
Block CUDA graphs: only release the cudagraph pool for tree functions the block backend compiled (#15777)
dda083b Daniel Han committed 23 days ago
Studio diffusion: load the int8 GEMM operands through TMA descriptors on sm120 (#15829)
6cdce47 Daniel Han committed 23 days ago
Mamba LoRA: apply a trained dt_proj / x_proj / out_proj adapter in eval and cached decode (#15730)
802c589 Daniel Han committed 23 days ago
Studio: emit no legacy load flag when the llama-server --help probe is inconclusive (#15828)
987731e Daniel Han committed 23 days ago
Studio tests: drain the background sandbox mover between tests (#15830)
36d3c10 Daniel Han committed 23 days ago
CI: turn update checks off in every job that drives a browser (#15815)
266b278 Daniel Han committed 23 days ago
FP8 scaled_mm backend: send 17-64 rows to torch._scaled_mm when the 16-bit weight spills L2 (sm89) (#15791)
01c94d3 Daniel Han committed 23 days ago
checkpoint_skip_layers="auto": skip the planning collective under DDP.join() (#15779)
a14f61a Daniel Han committed 23 days ago
Studio native cache write: recapture when an offload moves weights before the live-read swap (#15776)
6c7febb Daniel Han committed 23 days ago
Studio: keep MiniMax-H3's streamed top-level group on the card between denoise steps (#15796)
c4c785c Daniel Han committed 23 days ago
Studio: test that a batched fine-tune stays in its own account (#15825)
53bd212 Daniel Han committed 23 days ago
Studio: raise the startup bundle budget by the Trick or treat row's growth (#15822)
3db711a Daniel Han committed 23 days ago