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 diffusion: drop stale pinned host layout entries after a dtype change in low_vram offload (#14857)
d61d157 Daniel Han committed 28 days ago
Native FSDP2 float16: start the loss scale where one GPU's scaled backward fits (#15033)
e472a56 Daniel Han committed 28 days ago
Backend CI: split the 3.13 shards a-f / g-r instead of a-k / l-r (#15061)
cff73aa Daniel Han committed 28 days ago
mxfp4_packed_matmul: exact gradients for grad-requiring inputs (#15060)
b60f1a1 Daniel Han committed 28 days ago
Attention: keep training SDPA off cuDNN attention on SM100 (#15019)
5b258ae Daniel Han committed 28 days ago
FSDP2 native: fewer collectives under gradient accumulation and checkpoint recompute (#15023)
96f60da Daniel Han committed 28 days ago
Decode LoRA cast cache: refresh after fused optimizer steps (#15052)
5c38c96 Daniel Han committed 28 days ago
FSDP2 optimizer_cls test: follow #15008, which keeps bitsandbytes Adam / AdamW on local shards (#15049)
20f3b3a Daniel Han committed 28 days ago
Block compile: keep a caller's flex BACKEND instead of adding FORCE_USE_FLEX_ATTENTION (#15044)
d3e3edc Daniel Han committed 28 days ago
Studio: return a video render's pinned host memory to the driver at render end (#15036)
74288f4 Daniel Han committed 28 days ago
Unsloth Studio / Desktop: remember the text encoder and VAE files per image model (#14845)
711332a Leo Borcherding committed 28 days ago
Unsloth Studio / Desktop: image generation ignores the VRAM Budget setting when planning offload (#14816)
38c5e3a Leo Borcherding committed 28 days ago
FSDP2: keep bitsandbytes 8-bit AdamW state by stepping the local shards (#15008)
f674295 Daniel Han committed 28 days ago
Name the encoding on two more diffusion.py reads in test_diffusion_zimage_cfg_split (#15040)
66dfb16 Daniel Han committed 28 days ago
Studio: route Apple Silicon image requests to the MLX engine by default, diffusers MPS for the rest (#14883)
1492952 Daniel Han committed 28 days ago
FSDP2 QLoRA: reuse the planned NF4 dequant launch for all-gathered weights (#15011)
60db9ad Daniel Han committed 28 days ago
Studio diffusion: correct the fused RMSNorm accuracy notes (#15035)
89d6805 Daniel Han committed 28 days ago
Settings tabs smoke: hold the shortcut recorder box before pressing Backspace (#15032)
132a208 Daniel Han committed 28 days ago
Keep frozen load_in_fp8 weights out of FSDP2 sharding and the DDP broadcast (#15001)
888c51b Daniel Han committed 28 days ago
Opt-in torchao float8 training for full finetuning (#15007)
7b70f09 Daniel Han committed 28 days ago
Studio diffusion: record CUDA graphs thread-locally and keep other threads' syncs out of a recording (#14964)
9dfd045 Daniel Han committed 28 days ago
Studio Images: Detailer (face fix) pass with SAM 3 and inpaint (#14616)
edd6136 Daniel Han committed 28 days ago
Studio: keep the Z-Image CFG split on when group offload leaves every group resident (#14955)
8785d7b Daniel Han committed 28 days ago
Scan baseline: re-review the eight unsloth-zoo sites 2026.10.8 moved (#15031)
5bd30d0 Daniel Han committed 28 days ago
SFT: keep the router aux loss off for single-expert configs (Jamba) (#15010)
f48b996 Daniel Han committed 28 days ago
Studio Images: decide VAE tiling at the decode and stream resident groups there, so offloaded edits decode untiled (#15006)
b0c7f68 Daniel Han committed 28 days ago
Compiled blocks: keep the xformers stand-in SDPA off cuDNN attention (#15009)
217fa4c Daniel Han committed 28 days ago
Block compile: router_logits capture no longer recompiles every layer (#14998)
012192a Daniel Han committed 28 days ago
Native FSDP2 float16: skip the step when every rank overflows under fused AdamW (#14996)
96e1778 Daniel Han committed 28 days ago
Native FSDP2: faster native_clip_grad_norm_ that follows torch clip_grad_norm_ dtypes (#14973)
9df97d9 Daniel Han committed 28 days ago