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 diffusion: launch the VAE and MOSS Triton kernels on the input's own card (#17099)
f4668b4 Daniel Han committed 11 days ago
Studio image postprocess: convert to uint8 in row strips with one float32 copy each (#17091)
a7ff181 Daniel Han committed 11 days ago
Studio: the load prefetch skips chunks already in the page cache and budgets its read buffers (#17112)
de1568a Daniel Han committed 12 days ago
Studio int8: fused MLPs and v2 Linears quantize their input with the probed one-pass kernel (#17114)
86a41b0 Daniel Han committed 12 days ago
Studio diffusion: pair Qwen-Image-2.1's guidance branches per streamed block (#17111)
c953832 Daniel Han committed 12 days ago
Studio: judge the compiled ROCm W8A8 linear per card, not once per process (#17078)
0854cbe Daniel Han committed 12 days ago
Studio: queued image generations wait on their own pool, not the default executor (#17108)
8f8b34c Daniel Han committed 12 days ago
Studio video: fall back to the tiled VAE decode when the Linux / macOS CPU allocator refuses (#17075)
ed670cc Daniel Han committed 12 days ago
Studio: keep the LoRA fuse host backup off integrated GPU pins and inside the host RAM reserve (#17086)
ffeb8f7 Daniel Han committed 12 days ago
Studio MiniMax-H3: keep only the conditioner hidden state H3 reads (about 2 GiB less at 4096 tokens) (#17095)
90f5bcf Daniel Han committed 12 days ago
Studio: run the int8 / fp8 Triton widen on the operand's GPU (#17133)
8723ad5 Daniel Han committed 12 days ago
Studio: a denoiser-only LoRA change keeps the cached text embeddings (#17105)
4db4ac0 Daniel Han committed 12 days ago
Studio diffusion: keep single-file int8 weights as int8 on T4 / pre-Ampere NVIDIA and Intel XPU under offload (#17079)
3967e81 Daniel Han committed 12 days ago
Studio: per-tool MCP switches at the end of Settings > Agents (#15629)
decd94a Etherl committed 12 days ago
Context parallel: Ulysses for DeepSeek (V2 / V3 MLA, V3.2 indexer, V4 sink attention) (#15882)
66ee929 Immanuel Sun committed 12 days ago
Studio text encoders: keep the fp8 storage across a layerwise-cast forward instead of re-casting it (#17103)
14f6b6d Daniel Han committed 12 days ago
Studio LTX VAE: keep only the open tile band on the card during tiled decode (#17060)
faf7150 Daniel Han committed 12 days ago
Studio diffusion: FLUX.3 modulation heads project once per sample; offloaded Base edits reuse the conditioning streams (#17059)
3194a83 Daniel Han committed 12 days ago
Studio LTX-2.5: keep a fixed negative prompt in the encoder cache (refresh entries on a hit) (#17138)
33a1f1b Daniel Han committed 12 days ago
Studio MiniMax-H3: normalise a reference video one VAE chunk at a time (#17096)
870041c Daniel Han committed 12 days ago
Studio: re-merge only the layers of the LoRA adapters whose weight changed (#17102)
1b1d919 Daniel Han committed 12 days ago
Studio GGUF: load IQ1_S / IQ1_M / IQ2_XXS / IQ2_XS / IQ2_S / IQ3_XXS / IQ3_S on the diffusers path (#17071)
66eb0c2 Daniel Han committed 12 days ago
Studio diffusion: expandable VAE decode keeps a runtime allocator opt-in and restores the exact settings (#17076)
85f2252 Daniel Han committed 12 days ago
Studio MLX convert: pack 3 / 5 / 6-bit codes from Fortran-order or strided arrays (#17082)
750d728 Daniel Han committed 12 days ago
Studio: let peft 0.19 attach a LoRA to torchao int8 / fp8 weights of quantised single files (#17087)
f46fbab Daniel Han committed 12 days ago
RoPE: rotate Q and K in one launch each way under torch.compile (#17109)
315fd89 Daniel Han committed 12 days ago
Studio CLI tests: keep the --secure re-exec tests off the cloudflared download (#17129)
028cae9 Daniel Han committed 12 days ago
Studio: bit-identical one-kernel RoPE for FLUX.3 eager attention (L4: -12% s/step) (#17136)
e3eb43d Daniel Han committed 12 days ago
Studio: run FLUX.3 in float32 on fp16-only GPUs (T4 renders were black) (#17124)
c6fb555 Daniel Han committed 12 days ago