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: refuse Laya questions whose instructions would be cut (#15602)
71ff9e5 Nilay committed 24 days ago
Keep a dataset's thinking text when training Gemma 4 E2B and E4B (#15569)
1b0b58b Nilay committed 24 days ago
install-kernels: install mamba_ssm on sm70 with Triton 3.4+, matching unsloth_zoo (#15557)
9b442f4 Daniel Han committed 24 days ago
Condense module docstrings, keeping the guarantees they state (#15594)
cb678f5 Daniel Han committed 24 days ago
Studio: count the cards a CUDA + Vulkan llama.cpp reaches only through Vulkan (#15466)
d4fa58c Debojyoti Banerjee committed 24 days ago
Distributed tests: take every MASTER_PORT from the shared rendezvous helper (#15679)
3c63f72 Daniel Han committed 24 days ago
Fast decode: follow a model moved after loading (GRPO on a non-default GPU raised cuda:1 and cuda:0) (#15677)
7f3c5a1 Daniel Han committed 24 days ago
Studio: Scheduled task dialog padding, pill name field, dark New task button (#15647)
4611107 Michael Han committed 24 days ago
Studio: restore the reload icons replaced in #15232 (#15648)
774b56c Michael Han committed 24 days ago
Studio: expanded composer fills the chat area, net of attachments (#15650)
a674707 Michael Han committed 24 days ago
Studio: recheck that the user is still away before an automatic task swaps models (#15436)
3c4802b Daniel Han committed 24 days ago
checkpoint_skip_layers="auto" under torchrun: every rank keeps the smallest planned count (#15545)
570b9b8 Daniel Han committed 24 days ago
Context parallel and async GRPO: recognise gated fast forwards (#15559)
68fad07 Daniel Han committed 24 days ago
Flash impl: warn when an explicit fa3 / fa4 request falls back on the device or torch (#15554)
d8cd1e7 Daniel Han committed 24 days ago
Block CUDA graphs: run MoE layers and flex attention blocks compiled without graphs (#15553)
f442e80 Daniel Han committed 24 days ago
Tensor parallel float16: unscale DTensor and plain grads apart in the native ShardedGradScaler (#15551)
def3364 Daniel Han committed 24 days ago
Fix the FSDP2 QLoRA non-float parameter guard on Python 3.9 (#14774)
f0c879a Daniel Han committed 24 days ago
Run each FlashAdamW update on its parameter's CUDA device (#14773)
9d719c0 Daniel Han committed 24 days ago
Keep LoRA adapters on Mamba x_proj / dt_proj / out_proj when the fused training kernel is available (#15539)
986558e Daniel Han committed 24 days ago
Studio: rounded browser tabs and hovers, Recents above Suggested once visited (#15649)
c4d2ae2 Michael Han committed 24 days ago
Studio: Export page in the Train style, Train polish, hide Q4NX without an AMD NPU (#15640)
e8b9273 Michael Han committed 24 days ago
Tests: fail grouped tests by name when --dist loadgroup stops grouping (pytest 9.2) (#15638)
e54ae4f Daniel Han committed 24 days ago
Studio: audio field titles stop focusing their field, quieter Optional tags (#15651)
224faa2 Michael Han committed 24 days ago
Studio: Library cards stay apart at every zoom, one size, 5 per row (#15643)
e4c96e5 Michael Han committed 24 days ago
Studio: Qwen-Image-Edit streamed-tier activation seed keeps more of the DiT resident on tight cards (#15645)
602eda4 Daniel Han committed 24 days ago
fix(studio): reject standalone GGUF adapters before model teardown (#15421)
bb7ac2b AAFI MALEK committed 24 days ago
Import hoist verifier: a local import re-pointed after its definition moved is not a clash (#15622)
e7bf2e9 Daniel Han committed 24 days ago