●Change8

May-2025

📦 unslothView on GitHub →
✨ 5 features🐛 12 fixes🔧 5 symbols

Summary

This release introduces official support for Qwen3 models, including fine-tuning capabilities for the 30B MoE variant. Numerous bug fixes address compatibility issues, quantization errors, and improve overall stability.

Migration Steps

  1. Update Unsloth via `pip install --upgrade --force-reinstall unsloth unsloth_zoo`.

✨ New Features

  • Added support for Qwen3 models, including 14B and 4B variants.
  • The 30B MoE model ("unsloth/Qwen3-30B-A3B") is now fine-tunable in Unsloth.
  • Introduced support for full finetuning via `full_finetuning = True` parameter in `FastModel.from_pretrained`.
  • Added support for custom `auto_model` for wider model compatibility (e.g., Whisper, Bert).
  • Added support for passing `attn_implementation` argument for attention configuration.

🐛 Bug Fixes

  • Fixed issue where `load_in_4bit = True` conflicted with `fast_inference = True`.
  • Resolved issue where `unsloth_fast_generate` model was not defined.
  • Ensured `trust_remote_code` propagates correctly down to `unsloth_compile_transformers`.
  • Improved error reporting by showing `peft_error` when applicable.
  • Fixed generation prompt error message handling.
  • Fixed configuration issue related to `config.torch_dtype` in `LlamaModel_fast_forward_inference`.
  • Updated support for new FFT 8bit functionality.
  • Fixed a `NameError` by adding a missing `importlib` import in utilities.
  • Added QLoRA Train and Merge16bit Test functionality.
  • Fixed compatibility issues with Transformers version 4.45.
  • Fixed saving 4bit quantization for Vision-Language Models (VLM).
  • Applied various fixes specific to Qwen3 inference and Qwen3 QK norm.

Affected Symbols

FastModelunsloth_compile_transformersunsloth_fast_generateLlamaModel_fast_forward_inferencemapper.py