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
Train the float8 FSDP2 reference in a fresh process like the sharded ranks (#15250)
c9e0db6 Daniel Han committed 27 days ago
Studio: read diffusion.py as utf-8 in the encoder pin structural test (#15242)
fabc696 Daniel Han committed 27 days ago
Pipeline parallel x native FSDP2 data parallel x tensor parallel for dense decoders (#15204)
85234fe Daniel Han committed 27 days ago
Studio diffusion: price a released pipeline from its bf16 table while its cache is incomplete (#15236)
4533c46 Daniel Han committed 27 days ago
Studio: offer, size and pin every MIG GPU instance CUDA enumerates on R570+ drivers (#14981)
674f7f4 Daniel Han committed 27 days ago
MoE dW: fall back to the legacy kernel when a per-expert tile overflows and no smaller tile is known (#15238)
2b3a1b3 Daniel Han committed 27 days ago
Studio diffusion: read a pipeline's .fp16 weight files on fp16 loads (SDXL reads 6.5 GB, not 12.9 GB) (#15195)
3042f52 Daniel Han committed 27 days ago
Inductor readback barrier: key the compile cache when a user pass is already set (#15208)
7ebf5e4 Daniel Han committed 27 days ago
Cross entropy: keep the loss and label gradient of confident tokens (#15140)
80fabbe Ming Liu committed 27 days ago
Studio: keep loading a downloaded retired FLUX.3 full copy from disk (#14835)
d72ad8d Daniel Han committed 27 days ago
Route ORPO, CPO, KTO, DPO and TRL 1.15's fused head to the zoo fused linear kernel (#14185)
e6a58b7 Daniel Han committed 27 days ago
Drop the duplicate has_real_cuda import in the torchao route test (#15239)
a4a6350 Daniel Han committed 27 days ago
Block CUDA graphs: copy parameter gradients out of the graph pool (#15207)
f9ee5b8 Daniel Han committed 27 days ago
Mixed NVIDIA hosts: bf16 and compile follow the current device; check where the weights landed after loading (#14622)
8d5200d Daniel Han committed 27 days ago
Native FSDP2: write sub-mesh saves from the mesh's own first rank (#14761)
8e64f1a Daniel Han committed 27 days ago
Flex tuned backward: leave caller tiles and small sparse blocks to Inductor (#15228)
99288b8 Daniel Han committed 27 days ago
Studio diffusion: streamed offload copies outside inference mode, so each compiled block traces once (#15223)
7047e55 Daniel Han committed 27 days ago
Native FSDP2: keep accelerate's FSDP2 for UNSLOTH_FSDP2=1 on XPU / NPU / other non-CUDA installs (#14760)
40ef4b9 Daniel Han committed 27 days ago
Native FSDP2: restore the exact reshard-after-forward setting after generate (#14762)
5c1e04e Daniel Han committed 27 days ago
Context parallel gather: correct gradients under torch.compile, and a guard for in-place edits of returned tensors (#15231)
327dc5d Daniel Han committed 27 days ago
RoPE cache: a cold build inside a compiled block traces again (no is_inference_mode_enabled under compile) (#14728)
1fe53a8 Daniel Han committed 27 days ago
Fused LoRA: run PEFT for a partially merged block whose input needs no grad (#14748)
42591ca Daniel Han committed 27 days ago
Installer: Vulkan llama.cpp for ROCm GPUs without a bundle or hipcc; keep Strix wheels below ROCm 7.13 (#14662)
8b0427e Daniel Han committed 27 days ago
Studio: load GGUF on CPU when the auto-selected Vulkan driver hangs or has no device (#14521)
21b0577 Daniel Han committed 27 days ago
FSDP2: restore the checkpoint's 4-bit QuantState on a full DCP load (#14747)
c066804 Daniel Han committed 27 days ago
Stamp no bitsandbytes config on a dequantized fp8 load, and read a quantization_config object before stamping (#14803)
f0c87b1 Daniel Han committed 27 days ago
Studio diffusion: write the runtime int8 / fp8 prequant cache after the load, off the first image (#15202)
da0542b Daniel Han committed 27 days ago
Studio: keep a local folder's own text encoder when the text encoder setting is unset (#14738)
e3bba82 Daniel Han committed 27 days ago
Studio: read MiniMax-H3 card VRAM on ROCm without a torch context (#14732)
97a4208 Daniel Han committed 27 days ago
FSDP2 merged_4bit: keep the base bias on the GPU when lora_bias merges into it (#14801)
01b05ba Daniel Han committed 27 days ago