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
Keep Mamba dt_proj out of 4-bit so QLoRA forward works (#15079)
995afcc Daniel Han committed 28 days ago
Native FSDP2 optimizer.pt resume: parameters the saved run never stepped (#15050)
925e1c0 Daniel Han committed 28 days ago
Studio: shorter Agent permissions description, even card padding (#15124)
4a868e2 Michael Han committed 28 days ago
pre-commit: format with ruff-unsloth instead of three passes around stock ruff (#14993)
ec6e383 Daniel Han committed 28 days ago
Unwrap compressed-tensors offload caches after the lazy first-forward decompress (#15095)
695ad64 Daniel Han committed 28 days ago
Wait for the sandbox pick to land instead of a fixed 5ms tick (#15120)
f31feed Daniel Han committed 28 days ago
Fused LoRA: plain Linear projections after merge_and_unload / unload no longer crash (#15119)
d8c28b7 Daniel Han committed 28 days ago
MXFP4 compressed-linear tests: fp16 where bf16 is unsupported (#15113)
724090f Daniel Han committed 28 days ago
Read the working folder picker labels from the component in project_folders (#15114)
6ca5884 Daniel Han committed 28 days ago
Studio diffusion: pack up-front group-offload pins into slabs (#15105)
84d7833 Daniel Han committed 28 days ago
Find a LAN address's actions in its row, not the text's parent (#15109)
54ddc69 Daniel Han committed 28 days ago
FP8 inductor fullgraph test: allow one ulp on GPUs where Inductor fuses the dequant (#15107)
e5c2cd6 Daniel Han committed 28 days ago
Studio: language and platform tabs for the API page code examples (#15084)
db76346 Michael Han committed 28 days ago
Studio: one red shade for every red button in light and dark (#15085)
09c38c8 Michael Han committed 28 days ago
Block compile: mark the batch dim dynamic so an uneven last batch does not recompile every layer (#15106)
7a0d86c Daniel Han committed 28 days ago
Studio UI polish: projects, API page, audio inputs, settings layout (#15080)
76e94d5 Michael Han committed 28 days ago
FP8 to NF4 load tests: compare in the load dtype on GPUs without bf16 (#15103)
91ec230 Daniel Han committed 28 days ago
Studio diffusion: keep the denoiser resident ahead of the text encoders (#15099)
35ec100 Daniel Han committed 28 days ago
Studio: contiguous k/v before SDPA on gfx1151 when they outgrow L2 (#15096)
2c2e044 Daniel Han committed 28 days ago
Studio: free the Qwen-Image-2.1 prefix KV cache before VAE decode (#15094)
91e3885 Daniel Han committed 28 days ago
Resume native FSDP2 optimizer checkpoints across torch AdamW and bnb 8-bit (#15092)
f6d79cf Daniel Han committed 28 days ago
GRPO trainer: follow zoo's torch < 2.12 gate for persistent TMA matmul (#15100)
9f68d57 Daniel Han committed 28 days ago
Load Mamba / FalconMamba with eager instead of a refused sdpa (#15076)
d4dab13 Daniel Han committed 28 days ago
FP8 tests: same automatic dequant route on sm89+, no Fp8Dequantize leak between files (#15078)
feddc0b Daniel Han committed 28 days ago
Studio diffusion: carry the next forward's first offload groups into the idle copy tail (#15088)
a31803a Daniel Han committed 28 days ago
Studio: stream an offloaded ControlNet through the group prefetcher (#15081)
1aba35c Daniel Han committed 28 days ago
Do not let the gap worker lock test deadlock its own worker (#15091)
f49cbdf Daniel Han committed 28 days ago
Bring two diffusion route test fakes up to the load id routes (#15090)
630996c Daniel Han committed 28 days ago
Full finetune pack-quantized compressed-tensors checkpoints (#15075)
0268444 Daniel Han committed 28 days ago