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: fused fp16 Wan block kernels on T4 (Wan2.2-5B 11.7 to 10.0 s per step, within base's spread) (#12531)
0f14c9e Daniel Han committed 1 month ago
Studio: make the fused int8 kernels bit-exact on torchao 0.17 and round int32 to bf16 twice like PyTorch (#12633)
335fb3d Daniel Han committed 1 month ago
Studio: keep dark ink visible in transparent source images (#12576)
0cf8c57 Nilay committed 1 month ago
Studio int8 GEMM: stop the device probe from reserving 0.7 to 1.4 GB of spill memory (#12525)
331cfda Daniel Han committed 1 month ago
Studio: pin inductor's dynamic_scale_rblock off so compiled renders match across servers (#12588)
b4e29c2 Daniel Han committed 1 month ago
Studio: keep a load's diffusers / peft import off the post-warm import (#12572)
132af56 Daniel Han committed 1 month ago
Studio: decode a resident Wan2.2-TI2V-5B VAE untiled when it fits (#12603)
037c518 Daniel Han committed 1 month ago
Studio: move torchao int8 weights back to the GPU after an oversized request streams pinned groups (#12617)
b9df859 Daniel Han committed 1 month ago
Keep the load-time gradient checkpointing mode in for_training (#12629)
2557cba Daniel Han committed 1 month ago
studio: allow full access for the installation owner (#12604)
dd3a704 Maheswar Kumar committed 1 month ago
Studio small-host route: prefetch the streamed text encoder and fuse the int8 dequant (T4 FLUX.1-schnell 5.19 to 4.49 s per step, pixel-identical) (#12532)
0c6ef28 Daniel Han committed 1 month ago
Point five mapper rows at their upstream repo instead of an unpublished unsloth 16bit name (#12587)
5ae9c3b Leo Borcherding committed 1 month ago
Studio: stream a torchao denoiser from an unpinned host copy instead of leaving it on the GPU (#12551)
a71db3c oobabooga committed 1 month ago
Windows ROCm: use math attention where the fused SDPA kernels fail (#12639)
ad13de2 Daniel Han committed 1 month ago
Studio: stop the Load Model panel flagging Auto context loads as over VRAM (#12555)
58c9cd6 oobabooga committed 1 month ago
Unsloth Studio / Desktop: stop spam-clicking sidebar rows from queueing a navigation per click (#12544)
8254e67 Leo Borcherding committed 1 month ago
Studio: keep earlier tool calls in chat history on safetensors and MLX models (#12574)
faaca35 Nilay committed 1 month ago
Studio: use ComfyUI's default settings for FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2 and HunyuanVideo-1.5 (#12616)
e8f30ca Daniel Han committed 1 month ago
Studio: renew the chat-run lease while long prefill is still advancing (#12619)
0b67c51 Daniel Han committed 1 month ago
Studio: MiniMax-H3 GGUF uses the BF16 cuBLAS matmul path by default on sm80+ (1.2 to 2.3x per step, closer to an F32 reference) (#12527)
e640be3 Daniel Han committed 1 month ago
Let agent desktop apps use the model Unsloth is serving (#12582)
dd5131f Nilay committed 1 month ago
Studio: keep the prompt queue's more menu open while queued prompts send (#12546)
813d8af Leo Borcherding committed 1 month ago
Unsloth Studio / Desktop: keep release-notes table links from splitting mid-word in the update popup (#12543)
9d90475 Leo Borcherding committed 1 month ago
Studio: keep the VRAM figure visible on the Run preview hardware row (#12545)
1fb1c9a Leo Borcherding committed 1 month ago
Let API requests that ask for a JSON reply still call their tools (#12578)
a0136ac Nilay committed 1 month ago
fix(studio): WSL2 Windows localhost hint in startup banner (#11187) (#12621)
a550c89 Daniel Han committed 1 month ago
Studio: use the reply on screen when turning chats into training data (#12584)
1db4f70 Nilay committed 1 month ago
Studio: read text files saved in older Windows encodings correctly (#12581)
aab08e5 Nilay committed 1 month ago
Chat UI: find the Full access consent dialog by its slots, not its wording (#12640)
93769f5 Daniel Han committed 1 month ago