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: read GPU metadata from the CUDA driver API, not torch.cuda.get_device_properties (#17120)
2f8467f Daniel Han committed 12 days ago
Studio: run single-file fp8 on torchao under group offload, and keep Qwen-Image-2.1's timestep table in fp32 (#17127)
83e55c4 Daniel Han committed 12 days ago
Studio tests: import the new test helpers top-level when a sibling imports the module that way (#17139)
2fcd260 Daniel Han committed 12 days ago
Studio diffusion tests: plan as Linux does whatever the host can compile (#17135)
b7dc64f Daniel Han committed 12 days ago
Studio tests: keep the multi-GB fake checkpoints sparse on Windows (#17119)
a4323ea Daniel Han committed 12 days ago
Backend CI: retry the ruff-unsloth wheel install on a release-asset 504 (#17130)
9c2ad15 Daniel Han committed 12 days ago
Studio sd.cpp: pick the Q8_0 BF16 cuBLAS route from the diffusion GGUF, not its companions (1.57x on A100) (#17098)
ba22bb2 Daniel Han committed 12 days ago
Context parallel: Ulysses for MiniMax (M1 Lightning hybrid, M2) (#15881)
fa82bcc Immanuel Sun committed 12 days ago
Studio: AUTO W8A8 on ROCm RDNA keeps bf16 when the int8 plan still offloads (#17104)
dd2c39d Daniel Han committed 12 days ago
Studio sd.cpp: pass the Windows core count to MiniMax-H3 when sd.cpp would halve it (#17089)
4439f87 Daniel Han committed 12 days ago
Studio: keep nvfp4 / mxfp8 layers packed when no native runtime runs them (#17067)
285946a Daniel Han committed 12 days ago
Studio: widen stored fp8 / int8 weights without full float32 temporaries (#17068)
57d5fff Daniel Han committed 12 days ago
Studio MiniMax-H3: run int8 / fp8 denoisers where torchao's runtime does not (Ampere fp8, ROCm, MPS) (#17064)
1dcfcf2 Daniel Han committed 12 days ago
Studio tests: skip pre-quant tests without torchao and pin the Mac memory reading (#17117)
39f0ea3 Daniel Han committed 12 days ago
Studio tests: keep two Linux-host assumptions off the Windows and macOS legs (#17107)
8a11e11 Daniel Han committed 12 days ago
Studio: offloaded native int8 denoisers release their safetensors file once pinned (LTX-2.5 holds ~23 GiB less host mapping) (#17065)
3c1fbf2 Daniel Han committed 12 days ago
Context parallel: register the Qwen3.5 hybrid cp x native FSDP2 cell, add coverage tests (#17039)
6244269 Daniel Han committed 12 days ago
Studio diffusion: Automatic CFG-parallel no longer refuses an idle card over its own CUDA context (#17057)
49453f3 Daniel Han committed 12 days ago
Tests: skip the compiled half of the fp8 codes layer test on Windows without MSVC (#17073)
a1c6047 Daniel Han committed 12 days ago
GRPO: score a hooked output head through its own forward (#17050)
a174823 Daniel Han committed 12 days ago
Context parallel: Ulysses for Kimi-Linear (KDA + MLA), and fix Kimi-Linear MoE gate backward (#15880)
b61e6ee Immanuel Sun committed 12 days ago
PPO / RLOO fused log-probs: fall back under torch global module hooks (#17043)
0148e62 Daniel Han committed 12 days ago
Studio diffusion: decode image VAEs into one expandable allocator segment (SDXL holds 1.15 GiB less VRAM) (#17045)
570619d Daniel Han committed 12 days ago
ORPO / CPO / DPO / KTO chunked head: fall back under torch's global module hooks (#15259)
22a024f Shubhankar Tripathy committed 12 days ago
Tensor parallel sharded loading: allow init_lora_weights = "orthogonal" (#15218)
c92009b Shubhankar Tripathy committed 12 days ago
TP x FSDP2 save: one writer when only tp-mesh DTensors remain (#15257)
984888b Shubhankar Tripathy committed 12 days ago
Studio tests: drain deferred work before the auth refresh fixture restores window (#17055)
3b0a483 Daniel Han committed 12 days ago
Context parallel: Unsloth's fast Mistral, Granite and Cohere attention, and DPO / GRPO on fast Qwen3 (#16893)
534aa39 Daniel Han committed 13 days ago
Refuse fp8 training layers under tensor parallel instead of silently training in 16 bit (#16881)
fab4676 Daniel Han committed 13 days ago