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: refuse playlists at the entry of the remaining upload decoders (#14755)
9e8c2de Daniel Han committed 27 days ago
Studio installer (AMD, Linux): CPU torch for a lone gfx90c instead of kernel-less ROCm wheels (#15002)
e3f2044 Leo Borcherding committed 27 days ago
Studio drafter: split TRL packed rows at seq_lengths, reattach a Hub full fine-tune's MTP head on export (#14911)
40290cf Daniel Han committed 27 days ago
Studio: keep enhancing prompts from an already downloaded FLUX.3 upsampler quant (#14836)
3af50bd Daniel Han committed 27 days ago
llama.cpp installer: keep Qualcomm Adreno on the CPU prebuilt unless Vulkan is requested (#14709)
084d397 Daniel Han committed 27 days ago
adamw_flash: keep the unquantized bf16 Adam variance in fp32 so it can decay (#14749)
04f21d3 Daniel Han committed 27 days ago
Studio: API keys cannot approve host-path tool calls out of the sandbox (#14757)
ec0895c Daniel Han committed 27 days ago
Tests: skip the QAT MX and FP8 block GEMM GPU cases on the recorded real CUDA, not the spoofable probe (#14785)
c993bf3 Daniel Han committed 27 days ago
Studio: an offline audio.cpp listing checks the account grant on the disk pool (#14829)
6eeb690 Daniel Han committed 27 days ago
Skip the torchao blockwise GEMM CUDA case on a real CUDA probe (#15211)
6375906 Daniel Han committed 27 days ago
Packed metadata: keep the tensor path under torch.compile, fall back to pageable copies (#15199)
18380d9 Daniel Han committed 27 days ago
Name the encoding on the tensor parallel index reads; give the PPO unwrap test _decode_graphs (#15209)
65e2e9f Daniel Han committed 27 days ago
Studio diffusion: an int8 prequant read-back hits the compile bundle the runtime quantize saved (#15192)
c9fab47 Daniel Han committed 27 days ago
FP8 block GEMM: route empty inputs to the Triton kernel (#15205)
3b2773c Daniel Han committed 27 days ago
Studio: scheduled tasks that run a prompt on a schedule and save each answer as a new chat (#13871)
ba5c3b2 alkinun committed 27 days ago
Clef: keep fp16 loss scaling on the float32 path (fixes Gemma 3 not training on T4) (#14482)
7302874 Indrajit swain committed 27 days ago
Flex attention: tuned backward tiles on A100 for head_dim 256 (Gemma 3 LoRA 1.9-2.7x faster) (#15193)
0705fa5 Daniel Han committed 27 days ago
Gemma 3: keep flex attention only where its head_dim 256 template fits the card (#14670)
137a093 Daniel Han committed 27 days ago
Skip the norm warps CUDA case on a real CUDA probe (#15194)
f639676 Daniel Han committed 27 days ago
Tensor parallelism v2: sharded loading, data parallel x tensor parallel, and the Trainer (#15174)
957adb3 Daniel Han committed 27 days ago
Fused Qwen3 QK norm + RoPE: on by default on NVIDIA CUDA (#15189)
28ace86 Daniel Han committed 27 days ago
Decode graphs on by default for RL trainer generation, released on return to training (#15188)
cc1b758 Daniel Han committed 27 days ago
SFT: pick padded batches + block compile over padding-free when micro batches are host bound (#15185)
6988ffc Daniel Han committed 27 days ago
Studio diffusion: run an eager native int8 W8A8 linear as one compiled function (bit-identical, 1.33x streamed) (#15183)
34304b5 Daniel Han committed 27 days ago
Bump the pip group across 1 directory with 2 updates (#15053)
abfe8dc dependabot[bot] committed 27 days ago
Studio video: pin offloaded weights at their size, not a power of two (#15179)
494a857 Daniel Han committed 27 days ago
Give the single-row decode harness the real _decode_graphs helpers (#15182)
bf9eaf3 Daniel Han committed 27 days ago