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
Give the startup transfer budget headroom again after #14524 set it to its own build (#14571)
cbdb0be Daniel Han committed 1 month ago
Studio: pin the master-982 stable-diffusion.cpp prebuilt so MOSS-Image GGUFs load (#14565)
5ab80ff Daniel Han committed 1 month ago
Studio Images: per-generation Shift for flow-matching models and SDXL textual inversion embeddings (#14563)
38d2124 Daniel Han committed 1 month ago
Studio: show the AMD NPU by default again, with a clearer setup notice (#14524)
3f22b3d Daniel Han committed 1 month ago
Take the least of three RssAnon peaks in the block dequant host RAM check test (#14561)
94135ef Daniel Han committed 1 month ago
Studio: Z-Image CFG split under static step skip and CFG-parallel (#14551)
6e63ceb Daniel Han committed 1 month ago
MoE autotune: run the kernel tuning eagerly when called from a compiled block (#14559)
76d91cf Daniel Han committed 1 month ago
Count the context ceiling's per-device reserve gate from #14492 in the compute buffer wiring test (#14560)
35b4cf4 Daniel Han committed 1 month ago
Studio Images: paste generation parameters, inpaint mask blur and only-masked area, seed buttons (#14508)
94a8bfa Daniel Han committed 1 month ago
Studio: discrete-first GGUF placement keeps explicit device picks, matching context ceiling; in-process Rosetta check (#14492)
d22feec Daniel Han committed 1 month ago
Studio: FLUX.3 upsampler pin follows torch's vendor, single-GPU hosts launch as before, non-blocking Vulkan probe (#14491)
3f74ad1 Daniel Han committed 1 month ago
Load the real loader_utils helpers into the precision conflict test harness (#14545)
ce1ab5b Daniel Han committed 1 month ago
XPU bf16 on the current device, keep explicit bf16 on XPU, MLX full finetuning keeps its default dtype (#14493)
a7a37d3 Daniel Han committed 1 month ago
Studio: refuse a single-file diffusion load that would exhaust host RAM instead of crashing (#14510)
1032b5f Daniel Han committed 1 month ago
Format the grouped GEMM single-flight wrapper so pre-commit.ci passes again (#14544)
d8d88af Daniel Han committed 1 month ago
Keep two Backend CI tests valid on uv 0.13 and on runners with no accelerator (#14535)
a0425b4 Daniel Han committed 1 month ago
Studio: Z-Image guided steps run at the guidance-off batch so turning CFG on recompiles nothing (#14541)
b6980ac Daniel Han committed 1 month ago
Wait for the Unsloth-managed folder label in the project folders browser smoke (#14542)
ee98841 Daniel Han committed 1 month ago
Studio diffusion: compile group-offloaded blocks below their hooks on the deferred profile and for planned-offload video (#14528)
3649767 Daniel Han committed 1 month ago
Jetson / Rosetta follow-ups: an explicit GPU index beats the Jetson CPU record, JetPack 7 off the CPU note (#14494)
ce27c36 Daniel Han committed 1 month ago
MoE grouped GEMM: share autotuners across identical cards; fix the TMA test on one-GPU hosts and macOS (#14490)
ec80925 Daniel Han committed 1 month ago
NVFP4 fused dX: skip split K on sm89 when the output already fills the GPU (#14536)
0b124ea Daniel Han committed 1 month ago
Cast float32 activations into the float16 projections of Falcon-H1 / Nemotron-H fp16 loads (#14532)
c4b4a35 Daniel Han committed 1 month ago
Load FP8 checkpoints dequantized to 16-bit on GPUs below sm89 (#14531)
b7e66d7 Daniel Han committed 1 month ago
Studio diffusion: place bf16 and GGUF / int8 packed denoisers from their learned peak (#13860)
bf2725e Daniel Han committed 1 month ago
Dequantize fine-grained fp8 checkpoints on a 16bit load in FastLanguageModel, and save them as 16bit (#14516)
795f8d0 Daniel Han committed 1 month ago
Studio diffusion: key learned peaks on the unsloth mirror a fresh install downloads (#13847)
fc91656 Daniel Han committed 1 month ago
QAT: qat_scheme = "nvfp4", and strip NVFP4 / MX QAT linears before saving (#14512)
68de924 Daniel Han committed 1 month ago