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 small-host INT8: armed layers compile once (eager-only call counters, follow-up to #17167) (#17193)
4acaa6c Daniel Han committed 11 days ago
Studio: a streamed GGUF denoiser's top-level weights re-point to their host copy instead of copying back every step (FLUX.3 Q8_0 T4: -11% per image) (#17222)
61babeb Daniel Han committed 11 days ago
Tool activity gate: wait for each card to open before reading it, not a fixed sleep (#17234)
43f8a88 Daniel Han committed 11 days ago
Studio Qwen-Image-2.1: encode a condition image again after load_state_dict(assign=True) on its VAE (#17197)
8429c9a Daniel Han committed 11 days ago
Studio FLUX.3: keep reference-image modulation at one row per span in the joint blocks (#17181)
f2bb9c2 Daniel Han committed 11 days ago
Studio FLUX.3: price float32 compute on fp16-only GPUs from the measured activation table (x2), so a GGUF Q4 denoiser stays resident on a T4 (#17219)
3977ee2 Daniel Han committed 11 days ago
Studio FLUX.3 on fp16-only GPUs: an explicit INT8 / FP8 quant reads the bf16 checkpoint as stored instead of widening the transformer to float32 (#17228)
184cdb8 Daniel Han committed 11 days ago
Studio FLUX.3: run attention in float16 on fp16-only GPUs (T4: -15% per image GGUF, -34% INT8) (#17223)
cc0c4a0 Daniel Han committed 11 days ago
Studio: give each rotated native int8 Linear its own ConvRot buffer (FLUX.3 INT8 on T4 failed its first render) (#17217)
6511d95 Daniel Han committed 11 days ago
Studio FLUX.3 manual cast: keep LoRA adapters at the fp32 compute dtype and out of the bf16 fold (#17201)
1185624 Daniel Han committed 11 days ago
Studio LoRA fuse: charge the host RAM guard what the pinned backup really holds (#17191)
4eeffee Daniel Han committed 11 days ago
Studio diffusion: keep Qwen-Image-2.1 CFG pairing through its first-render compile (#17177)
10f638a Daniel Han committed 11 days ago
Studio sd.cpp: pin the prebuilt resynced with upstream master-1050 (#17148)
e430c01 Daniel Han committed 11 days ago
Studio: install FlashInfer 0.7.3 prebuilt kernels for diffusion NVFP4 (#17215)
0a41a7a Daniel Han committed 11 days ago
Studio CUDA graphs: the sm_89 fp8 stream-K guard reads every split card, not only the primary (follow-up to #17166) (#17194)
64ccbc2 Daniel Han committed 11 days ago
Studio FLUX.3 / Qwen-Image-2.1: prompt encodes keep only the hidden states they read (#17182)
cd95bc8 Daniel Han committed 11 days ago
Studio diffusion: keep a streamed denoiser's blocks resident for the denoise phase, not the whole render (A100 8 GiB Qwen-Image-2.1 -18%, FLUX.3 -21 to -24%) (#17207)
0bfc894 Daniel Han committed 11 days ago
Studio compile cache: torch 2.7 keeps reused graphs in saved bundles (follow-up to #17110) (#17196)
50c236c Daniel Han committed 11 days ago
Studio MiniMax-H3: keep the conditioner read capture and tile decode traceable under torch.compile (#17095, #17097) (#17190)
0c946b4 Daniel Han committed 11 days ago
Studio LTX-2.5: auto text-encoder precision loads the hosted int8 ConvRot encoder (#17179)
e865fa2 Daniel Han committed 11 days ago
Studio: run Detailer inpaints on the image-generation pool (#17188)
185abcf Daniel Han committed 11 days ago
Studio: keep segment CUDA graphs on unprotected NVFP4 steps (#17216)
6bda1b7 Daniel Han committed 11 days ago
Studio MiniMax-H3: keep the whole denoiser forward compilable with the render text cache (follow-up to #17203) (#17226)
83fd83f Daniel Han committed 11 days ago
Compiled RoPE: keep per-tensor launches when only Q or K needs a gradient (#17205)
2179383 Daniel Han committed 11 days ago
Studio Qwen-Image-2.1: one GEMM for the dense gate + value and Q / K / V projections (#17214)
8504a4f Daniel Han committed 11 days ago
Studio: decode the FLUX.2 / FLUX.3 VAE in row strips (36-43% less decode VRAM, same speed) (#17202)
3fdd122 Daniel Han committed 11 days ago
Studio diffusion: size CPU / MPS VAE decode tiles from free host memory (CPU 512 px decode 2.1x faster) (#17186)
3be02a6 Daniel Han committed 11 days ago
Studio: int8 ConvRot by default for FLUX.2 klein and FLUX.3 klein (LPIPS 0.10 to 0.04) (#17209)
4a45dbd Daniel Han committed 11 days ago
Studio GGUF: dequantise bit-identical to llama.cpp, rounded once to the compute dtype (#17204)
b67e563 Daniel Han committed 11 days ago