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
Context parallel (Ulysses) for DPO and GRPO training (#15041)
1d3e5f5 Daniel Han committed 28 days ago
Give the embedder unload race test room for a slow release (#15082)
ad90a00 Daniel Han committed 28 days ago
Read diffusion.py as UTF-8 in the NVFP4 act-quant test (#15086)
ef44420 Daniel Han committed 28 days ago
Studio: keep host memory unpinned on integrated GPUs (#15077)
cfc593f Daniel Han committed 28 days ago
Studio: multistage companion and CFG replica fit checks do not count CUDA graph pool cache as free (#15069)
8262da2 Daniel Han committed 28 days ago
Studio: name image and video loads so an unload ends only its own (#14827)
c7f088c Daniel Han committed 28 days ago
Studio: run the image-load gates before the MLX router's diffusers fallback (#15066)
0b81657 Daniel Han committed 28 days ago
Studio Images: run Detailer inpaints eager and uncounted (#15063)
b276691 Daniel Han committed 28 days ago
Studio diffusion/video: share the Hub reachability verdict across cached-load calls (#15059)
d497ab2 Daniel Han committed 28 days ago
Studio NVFP4: tag the compile cache in both switch states so a legacy cache never matches (#15055)
2b0c60e Daniel Han committed 28 days ago
Studio: CFG batch, H3 budget and stream residency fit checks do not count CUDA graph pool cache as free (#15067)
3049b61 Daniel Han committed 28 days ago
Forget the last NVML inventory between backend tests (#15072)
a4f015b Daniel Han committed 28 days ago
Native FSDP2: sync initial trainable weights across ranks; shard FastSentenceTransformer and FastDiffusionModel (#15012)
c52ef02 Daniel Han committed 28 days ago
FastDiffusionModel: record an explicit float32 request so T4 training stays float32 (#15062)
12e8658 Daniel Han committed 28 days ago
Native FSDP2: sharded resume reads for weights and an opt-in sharded optimizer checkpoint (#15042)
75a2887 Daniel Han committed 28 days ago
Studio diffusion: do not demote a resident transformer to whole-module offload for the 2048 worst case (#15065)
38011fd Daniel Han committed 28 days ago
Studio: leave rows with a null caption out of vision training (#14726)
c18042b Vineeth Sai Varikuntla committed 28 days ago
FP8 dequant under compile: keep the torch path for bf16 on cards without native bf16 (#15071)
ea73d37 Daniel Han committed 28 days ago
MoE per-expert dW: count pipelined shared memory, fit tiles that overflow the card (#15070)
5f3d662 Daniel Han committed 28 days ago
Studio MLX: retry a refused coexist diffusers fallback with MLX unloaded (#15056)
5950335 Daniel Han committed 28 days ago
Studio: untiled VAE fit checks do not count CUDA graph pool cache as free (#15064)
09e30cf Daniel Han committed 28 days ago
Studio: fp32 VAE upcast frees each 16-bit device tensor as it is cast (#15058)
f8743f8 Daniel Han committed 28 days ago
Studio: split guided Z-Image on the background-compile recording render (#15057)
c785486 Daniel Han committed 28 days ago
MOSS-Image: build the float64 timestep sinusoid on the CPU off CUDA (#14929)
8737871 Daniel Han committed 28 days ago
Studio diffusion: one bounded Hub probe per process when the Hub is unreachable (#15039)
5b2bc06 Daniel Han committed 28 days ago
Studio diffusion: key the CFG-batch verdict on the weight quant, not just the module layout (#15047)
a1d9411 Daniel Han committed 28 days ago
Train on CPU when no GPU is found, instead of refusing to import (#14717)
e89baf3 Daniel Han committed 28 days ago
Studio diffusion: keep the VRAM Budget reserve free in the decode-time residency release (#15048)
6792704 Daniel Han committed 28 days ago
Studio sd.cpp: use BF16 cuBLAS for Q8_0 only when every GGUF the process loads is Q8_0 (#15054)
74b6995 Daniel Han committed 28 days ago