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
Pipeline parallel: per-stage sharded loading (load models larger than one GPU) (#15656)
bcd329b Daniel Han committed 20 days ago
Keep tensor subclass types through the fused LoRA accumulates (#16249)
4b9a0eb Daniel Han committed 21 days ago
Run the dynamic-rank LoRA compile test on hosts without Triton (#16256)
d8efd15 Daniel Han committed 21 days ago
Give the WMI deadline check headroom for Windows PowerShell job startup (#16257)
48c57aa Daniel Han committed 21 days ago
Pipeline parallel: every torch.distributed.pipelining schedule (Interleaved1F1B, LoopedBFS, zero bubble, ZBV, DualPipeV) (#15655)
323696d Daniel Han committed 21 days ago
Tests: compare the disabled 4-bit fused LoRA forward with dequantize + linear, not bitsandbytes' fused GEMM (#16252)
1d6ff58 Daniel Han committed 21 days ago
LoRA: add the rank-r term inside the base GEMM so small updates are not rounded away (#16071)
91989b1 Daniel Han committed 21 days ago
Vision GC offload: each model gets its own non-reentrant offloader (#16194)
f863657 Daniel Han committed 21 days ago
Tests: follow the tiled MLP hooks from #16125 (encoding, loader stubs, trainer logger) (#16196)
9dc1055 Daniel Han committed 21 days ago
Tiled MLP: one flag or one line for any model, on the new sequence tiling engine (#16125)
673a44b Daniel Han committed 21 days ago
Studio Images: SDXL-Turbo opens at its native 512x512 canvas (#16168)
377738b Daniel Han committed 21 days ago
Dependencies: the studio extra carries usearch like studio.txt (#16176)
82ffe01 Daniel Han committed 21 days ago
Tests: the skipped-menu linearity check times each size as the best of five (#16170)
19a5770 Daniel Han committed 21 days ago
Studio: HNSW (usearch) tier for gigantic RAG scopes, kept consistent with SQLite (#16145)
5c1dc91 Daniel Han committed 21 days ago
Studio MLX estimate: price attention scores where mlx really holds them, at the width it holds them (#16021)
3ec3350 Daniel Han committed 21 days ago
Studio: MOSS-Image negative-branch layer skip follows the reference workflows per checkpoint (#16127)
b1e0b00 Daniel Han committed 21 days ago
Dependencies: pin pymupdf 1.28.2 on Windows ARM64, the only release with a win_arm64 wheel (#16160)
9448864 Daniel Han committed 21 days ago
Studio: download files over 50GB from ModelScope over HTTPS (#15989) (#15994)
3b9ad4d Indrajit swain committed 21 days ago
Tests: fused LoRA TP parity compares O and dX at each tensor fp32 scale (#16144)
4d28d72 Daniel Han committed 21 days ago
Studio: load LFM3 custom-code models (route by saving transformers version, fix generate, chat and export) (#16096)
87a4d57 Daniel Han committed 21 days ago
Studio: warn when a Windows GPU load leaves the system short of commit (#16046)
d3b48f3 Debojyoti Banerjee committed 21 days ago
Studio Windows AMD: pair triton-windows with torch after the dependency stack (#16074)
a8c95bc Daniel Han committed 21 days ago
Flex wrapper: send short no-grad queries to the main kernel where flex_decoding cannot compile (#16092)
fae3823 Daniel Han committed 21 days ago
Tests: the titlebar smoke treats Vite outdated pre-bundle 504 as a reload (#16130)
febcd3e Daniel Han committed 21 days ago
Tests: the supplied-component twins check runs where diffusers is not installed (#16126)
8c74da0 Daniel Han committed 21 days ago
Studio MLX estimate: ask the quantized-attention route about a cache longer than the chunk (#15970)
245b9d0 Debojyoti Banerjee committed 21 days ago
Don't warn about loss_type = "nll" on a TRL whose SFTConfig predates it (#16042)
3cf4d45 Debojyoti Banerjee committed 21 days ago
Bump install.sh / install.ps1 pins to unsloth>=2026.10.11 (#16131)
c233ae3 Daniel Han committed 21 days ago