●Change8

v0.1.0-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 2 breaking✨ 11 features🐛 18 fixes🔧 5 symbols

Summary

This release introduces Unsloth Studio (Beta), a new web UI for training and running LLMs, alongside significant feature additions like Mixtral support, ROCM compatibility, and various stability and security fixes across the platform.

⚠️ Breaking Changes

  • Removed Blackwell flex attention disable workaround from studio. This might affect users running on specific Blackwell hardware configurations if they relied on the previous workaround.
  • Applied removal of `use_reentrant` to all TRL trainer configs. Users relying on this specific parameter in custom TRL setups might need to adjust their configurations.

Migration Steps

  1. For MacOS/Linux/WSL users: Ensure `cmake` is installed. If not, run `brew install cmake`.
  2. For MacOS/Linux/WSL users: Install `uv` using `curl -LsSf https://astral.sh/uv/install.sh | sh`.
  3. For Windows users: Install Python 3.13 using `winget install -e --id Python.Python.3.13`.
  4. For Windows users: Install `uv` using `winget install --id=astral-sh.uv -e`.
  5. All users should use `uv venv unsloth_studio --python 3.13` to create environments.
  6. All users should install Unsloth via `uv pip install unsloth --torch-backend=auto`.
  7. Run `unsloth studio setup` to initialize the Studio environment.
  8. Launch Studio using `unsloth studio -H 0.0.0.0 -p 8888`.
  9. For Docker users, use the provided `docker run` command, ensuring to set a password for Jupyter.

✨ New Features

  • Launch of Unsloth Studio (Beta), a new open-source web UI for training and running LLMs.
  • Support for Sequence Classification.
  • VLMs support for GRPO (Generalized Reinforcement Learning from Human Feedback).
  • ROCM support added.
  • Mixtral model support added.
  • Initial refactoring of Attention mechanisms.
  • Ability for users to upload evaluation datasets in the Studio.
  • Improvements to AI Assist, including updated default model, output parsing, logging, and dataset mapping UX.
  • Studio now supports per-model inference defaults, GGUF slider fix, and reasoning toggle.
  • Studio UI improvements: training progress display, CUDA lib path handling, and fix for dataset_num_proc.
  • Studio UI updates: Creative/Precise presets updated, and showing "Off" for disabled samplers (e.g., repetition penalty = 1).

🐛 Bug Fixes

  • Fixed respecting `llm_int8_skip_modules` for VLM.
  • Prevented AI-assist model config RCE via untrusted Hugging Face repos.
  • Disabled remote code execution in seed inspect dataset loads.
  • Fixed installation of data-designer plugin to be non-editable for Colab compatibility.
  • Patched VLM trainer to resize images correctly.
  • Fixed Compare Mode Deadlock, Cancel Event Poisoning & IPC Optimization in Studio.
  • Fixed GGUF inference in Studio regarding reasoning tokens, max_tokens, server flags, and GPU allocation.
  • Chat only uses GGUF for Mac devices.
  • Fixed Studio frontend build producing empty Tailwind CSS.
  • Fixed setup.sh crash on Mac with empty gitignore array.
  • Fixed Ctrl+C not terminating backend on Linux.
  • Fixed VLM GRPO matmul shape mismatch in `_get_per_token_logps_and_entropies`.
  • Resolved CUDA toolkit mismatch on multi-CUDA Windows systems.
  • Added Qwen3.5 version gate in loader dispatch.
  • Fixed xformers Blackwell guard with broader coverage and documentation.
  • Fixed stale GGUF metadata, updated helper model, and improved authentication in Studio.
  • Fixed slow cancellation of GGUF generation in Studio.
  • Fixed streaming issues in Studio (implied by truncated note).

Affected Symbols

TRL trainer configsllm_int8_skip_modules_get_per_token_logps_and_entropiesdata-designer pluginsetup.sh