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: speculative decoding for the MLX inference backend (#13014)
6bd164f Long Yixing committed 1 month ago
Decide accepts_loss_kwargs from the loss head so gradient accumulation gets num_items_in_batch for every model (#12946)
99cee64 Daniel Han committed 1 month ago
Studio: install the llm-compressor-main shadow from a source archive instead of git (#13132)
5080782 Daniel Han committed 1 month ago
Studio: keep merged table cells in their columns when reading web pages (#13098)
cceb770 Nilay committed 1 month ago
MLX decision training: Clef rows and predict take images (#13127)
ba5239a Long Yixing committed 1 month ago
Studio: send Decision API images to the MLX engine for Clef (#13126)
e0e35a0 Long Yixing committed 1 month ago
fix(install): silence the missing AppArmor userns sysctl read (#13125)
f6fb4cd Jay Zhou committed 1 month ago
README: install PyTorch before Unsloth on Windows so uv does not resolve an old Unsloth (#13134)
512617e Daniel Han committed 1 month ago
Studio: pass the llama-server key through the env when Windows cannot open its file (#13133)
57a1756 Daniel Han committed 1 month ago
Studio: use the Vulkan llama.cpp bundle, not CPU, on Windows NVIDIA drivers below CUDA 12.4 (#13129)
236dc94 Daniel Han committed 1 month ago
Studio: fall back to the legacy stream when an older backend refuses a new chat field (#13130)
90e913e Daniel Han committed 1 month ago
Studio: repair orphan tool results in the chat template fallback (#11733)
e69ddda lorenzozane committed 1 month ago
Studio: avoid repeated Qwen-Image-2.1 VAE tile OOMs on ROCm (#13116)
4399ef8 oobabooga committed 1 month ago
Dequantize 4-bit nn.MultiheadAttention out_proj after load (#13090)
1c946d1 Daniel Han committed 1 month ago
Load Qwen2.5-Omni checkpoints that name Qwen2_5OmniModel (#13087)
9656b58 Daniel Han committed 1 month ago
SyntheticDataKit: take a port argument and move off 8000 when it is taken (#13072)
e4d2f4b Daniel Han committed 1 month ago
Right-pad encoder-decoder tokenizers and load SeamlessM4T without auto_model (#13066)
c2a24a1 Daniel Han committed 1 month ago
Make precomputed inputs_embeds train under gradient checkpointing (#13088)
8cae9f8 Daniel Han committed 1 month ago
Point a 4-bit load that does not fit the GPU at offload_layers = "auto" (#13052)
0bafe1d Daniel Han committed 1 month ago
Repair RoPE for plain transformers loads after an Unsloth load (reward models in Online DPO / GRPO) (#13055)
113d5ef Daniel Han committed 1 month ago
Fix TRL PPOTrainer under Unsloth: rollout crash, sampling filters and left padding (#13108)
eac50ba Daniel Han committed 1 month ago
Fix PPOTrainer with an Unsloth policy: wrapper GC toggles, policy generate, RoPE on value / reward models (#13067)
bf0a464 Daniel Han committed 1 month ago
Fix LLaVA image token count when processor_config lacks num_additional_image_tokens (#13091)
60d2f0c Daniel Han committed 1 month ago
Keep paged bitsandbytes optimizer state paged on checkpoint resume (#13092)
0f14817 Daniel Han committed 1 month ago
Keep slow tokenizers transformers cannot convert instead of crashing the load (#13073)
fa8e70e Daniel Han committed 1 month ago