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 video: keep the stream encode hook out of a compiled VAE decode graph (#14822)
8cbcf05 Daniel Han committed 29 days ago
Studio images: keep the resident GGUF's base repo on Reapply and native reload (#14828)
505bbc2 Daniel Han committed 29 days ago
Studio MCP: drop a queued video's request and caller once it settles (#14825)
7aa1073 Daniel Han committed 29 days ago
Apply the repo formatter to 26 files merged without it (#14817)
dc80c04 Daniel Han committed 29 days ago
Studio tests: skip the two-card fast pin test on ROCm instead of erroring (#14618)
4cf6fcc Daniel Han committed 29 days ago
Studio tests: compare MLX snapshot paths as paths, not posix strings (#14809)
1809377 Daniel Han committed 29 days ago
FSDP2 NF4 saves: stage packed bytes on the CPU, restore QuantState on full DCP load (#14739)
59cd08b Daniel Han committed 29 days ago
Attention: let the SDPA GQA probe trace under torch.compile(dynamic=True) (#14779)
32650d2 Daniel Han committed 29 days ago
Fused LoRA: fall back to PEFT for variants, disabled or merged adapters, trainable bases and autocast dtype mismatches (#14806)
e228cbc Daniel Han committed 29 days ago
Studio diffusion: fast whole-component swap for low_vram offload (#14767)
4e2838e Daniel Han committed 29 days ago
Studio tests: pin the card name in the learned-peaks store fixture (#14804)
93a57f1 Daniel Han committed 29 days ago
Studio: cap Pillow below 13 for the MLX image engine (#14802)
1b598db Daniel Han committed 29 days ago
Studio MLX images: plan text encoder and dense bits from the Metal working set (#14793)
572a66b Daniel Han committed 29 days ago
Studio MLX images: convert Q4_K / Q5_K / Q6_K and Q5_0 / Q5_1 GGUFs (#14794)
5186735 Daniel Han committed 29 days ago
FP8 block decode: weight-only GEMV where it beats the block GEMM, fused small-M called directly (#14813)
9bc4713 Daniel Han committed 29 days ago
QAT: fake quantize Gemma embeddings on the float32 and Accelerate-hooked paths (#14702)
8e8eb3d Daniel Han committed 29 days ago
Studio MLX images: tile the VAE decode when the untiled one would not fit (#14796)
9a822f4 Daniel Han committed 29 days ago
MoE grouped GEMM: skip kernel-info logging work unless debug logging is on (#14787)
390b866 Daniel Han committed 1 month ago
Studio video: stream-encode Wan clips under the ROCm bf16 VAE decode (#14653)
d6677ab Daniel Han committed 1 month ago
Studio diffusion: seed dense fp16 denoisers on the default and max tiers from measured peaks (#14720)
7a4d599 Daniel Han committed 1 month ago
Studio diffusion: read ROCm from the wheel tag in every fast-path gate (#14654)
53719fc Daniel Han committed 1 month ago
Studio: resolve torchao's int8 GEMM home before the int8 fused install returns (#14771)
445599f Daniel Han committed 1 month ago
Studio Images: detect v-prediction SDXL fine-tunes from their own config, not their name (#14746)
6267ef5 Daniel Han committed 1 month ago
Studio diffusion: one host read per guided step in the CFG-batch and CFG-parallel argument checks (#14734)
cf3134d Daniel Han committed 1 month ago
MXFP4 packed Linear: bound the below-sm_80 weight decode by free memory (#14784)
4077567 Daniel Han committed 1 month ago
Fused preference loss: fall back when lm_head has module backward hooks (#14788)
2350ba8 Daniel Han committed 1 month ago
Gemma 4 vLLM fallback: check the engine's GPU, not the current one (#14778)
ad814c6 Daniel Han committed 1 month ago
MoE dW accumulation: fall back when autograd Node hooks can rewrite dW (#14759)
7de9c3b Daniel Han committed 1 month ago
GRPO routing replay: prime router formulas before compile and score the PEFT reference without replay (#14756)
956c4ef Daniel Han committed 1 month ago