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: size auto-quantised text encoders at the hosted file in the image picker (#14735)
f27fd4e Daniel Han committed 1 month ago
Studio diffusion: Automatic multi-GPU placement fixes and a placement decision in status (#14657)
94a1e3d Daniel Han committed 1 month ago
Studio Images: run Krea 2 GGUFs on stable-diffusion.cpp on CPU-only hosts (#14627)
e1b3124 Daniel Han committed 1 month ago
Studio: show a checkpoint's training loss when that step also evaluated (#13879)
299e905 Mohammad Hijjawi committed 1 month ago
MoE autotune cache: pick kernel configs for the operands' GPU (#14751)
08a5ae7 Daniel Han committed 1 month ago
RMSNorm kernels: do not specialize on the row count and row stride (#14786)
64c7a18 Daniel Han committed 1 month ago
FP8 block decode: fused act_quant and narrow-N GEMM for 1 to 16 tokens (#14792)
f232b9a Daniel Han committed 1 month ago
FP8 rowwise decode: one multi-row GEMV launch for 2 to 16 tokens (#14791)
444dcbf Daniel Han committed 1 month ago
tests: keep the fused-NVFP4 env in place after the split-reduction hook override (#14789)
dfc3419 Daniel Han committed 1 month ago
perf(studio): prefill MLX loads in the planned step while experts are on disk (#14466)
d0bdc7a Long Yixing committed 1 month ago
Studio: build the CFG-parallel replica outside inference mode and catch in-place edits of inference weights (#14626)
d5f0d8b Daniel Han committed 1 month ago
Studio diffusion: change baked LoRA subsets and weights without recompiling or recapturing (#14692)
38a3a34 Daniel Han committed 1 month ago
Studio Images: Hires fix and ESRGAN model upscalers (#14651)
45b095e Daniel Han committed 1 month ago
Two-GPU decode RoPE test: check the copy bitwise and the rounding at 2 ulp (#14752)
6394a86 Daniel Han committed 1 month ago
MoE LoRA targets: keep a requested router sibling when attention modules are also listed (#14745)
62a2ec2 Daniel Han committed 1 month ago
4-bit decode: no recompile from the eager setup, one output dtype under autocast, refresh the cached LoRA cast (#14754)
5503246 Daniel Han committed 1 month ago
Compiled blocks: packed GatedDeltaNet op, flex main kernel in training blocks, PrefixGrouper calls skip the block (0 graph breaks) (#14782)
3c16b36 Daniel Han committed 1 month ago
Studio NVFP4: keep ungated and gate-failing video builds out of Auto, check flashinfer against the gate record (#14659)
031a117 Daniel Han committed 1 month ago
Studio: clamp the top_k a model passes to search_knowledge_base (#14222)
772ac8b Roydon Sequeira committed 1 month ago
Studio sd.cpp: one CPU thread per usable physical core on Linux (#14694)
f2e45b5 Daniel Han committed 1 month ago
Block compile: let an enclosing torch.compile own wrapped decoder layers (#14704)
7e12c32 Daniel Han committed 1 month ago
Studio diffusion: arm Z-Image's prompt-length inputs for an int8 / fp8 compile (#14715)
725455e Daniel Han committed 1 month ago
FSDP2 QLoRA: keep 4-bit merge operands on one device in push_to_hub_merged (#14724)
ebdb28d Daniel Han committed 1 month ago
Studio diffusion: arm FLUX.1 prompt-length dynamic shapes so int8 does not recompile on a long prompt or new size (#14689)
6f0ca18 Daniel Han committed 1 month ago
Studio diffusion: replay a model-offloaded denoiser's CUDA graph across renders and pin torchao int8 host weights (#14666)
caa5574 Daniel Han committed 1 month ago
Studio diffusion: report compile and CUDA graph engagement in the image and video status (#14656)
3612651 Daniel Han committed 1 month ago
Studio sd.cpp: BF16 text encoders by default (Q8_0 / Q4_K_M changed pixels), opt-in lower precision (#14697)
08cd6a1 Daniel Han committed 1 month ago
Studio diffusion: give stored-fp8 denoisers their own planner label instead of dense (#14631)
38570e2 Daniel Han committed 1 month ago
Images: refuse Flux base / max shift values that build NaN sigmas (#14613)
ef990d2 Daniel Han committed 1 month ago