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 Hub: list every image and video model under Recommended (#14421)
444f632 Michael Han committed 1 month ago
gpt-oss: load_in_4bit from a BF16 checkpoint no longer loads random experts and router (#14400)
723a43b Daniel Han committed 1 month ago
RoPE: multi-head tiles, Q and K in one launch, no clones (2-2.7x per call) (#14417)
20e8179 Daniel Han committed 1 month ago
Tests: run the precision template in one namespace, keep the block-compile packing test off leaked env (#14425)
b93e4dc Daniel Han committed 1 month ago
Studio no-torch runtime: pin certifi 2026.10.19, the floor fastmcp-slim now requires (#14418)
77e6eaf Daniel Han committed 1 month ago
DDP: gradient_as_bucket_view for full finetuning, opt-in bf16 / fp16 gradient all-reduce (#14399)
a3395cf Daniel Han committed 1 month ago
FP8 byte dequant: pass enable_fp_fusion at the launch (#14416)
d568d16 Daniel Han committed 1 month ago
GRPO under DeepSpeed ZeRO-3: score with the gathered head, no_grad instead of inference_mode (#14395)
b824412 Daniel Han committed 1 month ago
Studio video stream encode: a failed final flush falls back instead of hanging the render (#14413)
731cb16 Daniel Han committed 1 month ago
Studio: show the retrieved passages behind a document search (#13987)
9185659 Jonathan Ye committed 1 month ago
QAT test: spoof CUDA in the unpickling child on GPU-less runners (#14415)
f095e78 Daniel Han committed 1 month ago
Tests: read checked-in sources as UTF-8 in three new test files (#14414)
b26ee31 Daniel Han committed 1 month ago
Fuse the decode residual add into RMSNorm (one Triton launch per norm) (#14382)
8cb1d11 Daniel Han committed 1 month ago
DeepSpeed ZeRO-2 / ZeRO-3: stage detection at load, MoE and fused LoRA leaves, precision resync, reentrant checkpointing (#14394)
b098d77 Daniel Han committed 1 month ago
Studio: run MiniMax-H3 PDD checkpoints with their 32-head output bank, refuse the ones that cannot run (#14410)
6901dd3 Daniel Han committed 1 month ago
FastLanguageModel LoRA / QLoRA: compile each decoder block as one region (#14398)
e24afff Daniel Han committed 1 month ago
Studio: pick ROCm flash attention by a whole-block timing probe instead of a model table (#13792)
3c14a2b Daniel Han committed 1 month ago
Native FSDP2: keep blocks that share parameters in the root unit (Zamba2) (#14384)
dc47ce1 Daniel Han committed 1 month ago
Block compile: find decoder layers structurally and wrap mixed stacks per class (#14397)
c663d49 Daniel Han committed 1 month ago
int4 packed: trace the training dequantize inside torch.compile (#14385)
66fb355 Daniel Han committed 1 month ago
FP8 block dequant below sm89: decode e4m3fn bytes in one Triton pass (#14408)
98010f6 Daniel Han committed 1 month ago
Studio: offload-first retry for an OOM VAE decode, fuse GroupNorm+SiLU on uncovered VAEs by structure (#13802)
c57c21f Daniel Han committed 1 month ago
Studio: fix First Block Cache on FLUX.2 (dual-stream head, single-stream tail) (#14401)
4c5e4bd Daniel Han committed 1 month ago
Studio: retry json_object as json_schema when a custom provider (LM Studio) refuses it (#11813)
d6641ec Saravana Priyan committed 1 month ago
Refuse load_in_8bit with fast_inference instead of silently loading vLLM in 16-bit (#13019)
c2f9ce1 Daniel Han committed 1 month ago
Studio: prefer discrete GPUs over unified-memory APUs in automatic GGUF placement; clamp Vulkan free to total (#14372)
4279fe4 Daniel Han committed 1 month ago
Block compile: keep the stock forward under DeepSpeed ZeRO-3 (#14396)
57c2a94 Daniel Han committed 1 month ago
Raise the startup bundle budget for the Rosetta chat-only messages (#14374) (#14407)
598cfae Daniel Han committed 1 month ago