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: bf16 Wan VAE decode on ROCm (gfx11 / gfx12) (#12681)
6295d6c Daniel Han committed 1 month ago
Block Swap: Stream frozen transformer blocks from host RAM so a dense model can train past the card's VRAM (#11832)
7911672 Leo Borcherding committed 1 month ago
Studio: fused int8 GEMM for block-streamed image denoisers (#12677)
29bcc23 Daniel Han committed 1 month ago
Studio: async block prefetch for MiniMax-H3's streamed denoiser (#12683)
8e11ba1 Daniel Han committed 1 month ago
Studio: deterministic FLUX.1, Z-Image and Qwen-Image renders on sm80, sm89 and sm120 (#12674)
1170b03 Daniel Han committed 1 month ago
docs(save): correct save_method values in docstrings (16bit/4bit -> merged_16bit/merged_4bit) (#10394)
100a5cf 陈志谦 committed 1 month ago
fix(studio): discover On Device GGUFs from disk before Hub (#12460)
e4e875b Lee Jackson committed 1 month ago
Studio: run fp16 on ROCm GPUs without native bf16 (RDNA2 and older, Vega) (#12676)
8216fa9 Daniel Han committed 1 month ago
Studio: deterministically load explicitly mentioned skills before generation (#12538)
9d74ab7 Wasim Yousef Said committed 1 month ago
Studio: per-model automatic step skip at each model's measured step count (1.44 to 1.81x, MiniMax-H3 on max) (#12652)
341f5ba Daniel Han committed 1 month ago
Studio: channels_last for 3D-conv VAEs and untiled A14B decode (#12672)
729cc3f Daniel Han committed 1 month ago
Studio: honour an explicit flash attention request on ROCm (gfx1151 8 to 12% faster per step) (#12667)
52f0d16 Daniel Han committed 1 month ago
Studio: do not reserve the OS share twice on Linux ROCm APUs (#12675)
24af86a Daniel Han committed 1 month ago
Studio: faster first start for image loads (VAE kernel prebuild, early quant probe, one FLUX.1 single-block graph) (#12666)
ec80623 Daniel Han committed 1 month ago
Apply the DoRA magnitude in fast_linear_forward (#12642)
93434b1 Vineeth Sai Varikuntla committed 1 month ago
Studio: give the MiniMax-H3 sd.cpp engine a CUDA-12 cuDNN for fused attention on sm80+ Linux (#12662)
858afa7 Daniel Han committed 1 month ago
Studio: read pre-quantized diffusion checkpoints from safetensors on torchao 0.17, 0.18 and main (#12645)
0f239dd Daniel Han committed 1 month ago
Installers: resolve studio from the venv, never from the caller's directory (#12682)
4a6ec2e Daniel Han committed 1 month ago
Studio (AMD, Windows): pin the multi-arch ROCm torch to rocm7.14.0, whose fused attention works (#12670)
e54e4fa Daniel Han committed 1 month ago
fix(studio): clarify image preview loading and allow retry (#12450)
5d4b7f1 Lee Jackson committed 1 month ago
Studio frontend: bump hono, ip-address, qs, dompurify and @babel/core overrides for security advisories (#12679)
93e8299 Daniel Han committed 1 month ago
Studio: set the ROCm AOTriton opt-in from the inference package so every entry point gets fused attention (#12668)
a0af6f1 Daniel Han committed 1 month ago
Studio: pin inductor's reduction configs for LTX-2 compiles so renders match across servers (#12651)
0584134 Daniel Han committed 1 month ago
Studio: stream offloaded denoiser blocks without host waits (B200 FLUX.1 16 GB 0.59x, L4 0.90x per step) (#12664)
3d0a0e6 Daniel Han committed 1 month ago
Studio: load the hosted FP8 checkpoints for FLUX.1-dev, FLUX.2-klein-4B, FLUX.2-dev and Qwen-Image-2512, and seed Krea 2 (#12649)
c03a5e0 Daniel Han committed 1 month ago
Studio: SageAttention 2 from the kernels hub, and FlashAttention 4 dependencies that load on a fresh install (#12654)
766a7ca Daniel Han committed 1 month ago
Studio: never fail, slow down or render noise on an explicit SageAttention or FlashAttention 4 request (#12641)
23308cd Daniel Han committed 1 month ago