●Change8

July-2025

📦 unslothView on GitHub →
✨ 7 features🐛 18 fixes🔧 20 symbols

Summary

This release focuses heavily on stability, VRAM reduction (10-25% less), and broad model compatibility, including full fixes for Gemma 3N Vision and support for new models like Devstral 1.1 and MedGemma.

Migration Steps

  1. Update Unsloth using: `pip install --upgrade --force-reinstall --no-deps --no-cache-dir unsloth unsloth_zoo`
  2. For users of Gemma 3N, re-download all model checkpoints (Unsloth will handle this automatically).

✨ New Features

  • Gemma 3N Vision support is now functional.
  • Gemma 3 text and vision performance significantly improved on T4 hardware.
  • Achieved 10 to 25% reduction in VRAM consumption across all models.
  • Added support for Devstral 1.1 and MedGemma 27B, 4B models, including vision capabilities.
  • Enabled Unsloth support for AMD GPUs via vLLM.
  • Added Intel GPU support with vLLM integration.
  • Falcon H1 finetuning is now supported.

🐛 Bug Fixes

  • Fixed issues with Gemma 3 text and vision training losses (now 1 to 2 instead of 6 to 7).
  • Resolved downloads getting stuck between 90% and 95%.
  • Fixed compatibility issues for Qwen 2.5, Qwen 2, and GLM models.
  • Resolved GRPO compatibility with the latest main TRL branch.
  • Ensured compatibility with the latest main branches of TRL, PEFT, and Transformers.
  • Fixed an issue where transformers were being forcibly upgraded.
  • Fixed LoftQ implementation when used with FastBaseModel.
  • Fixed LoftQ None config issue for FastBaseModel.
  • Converted torch.bfloat16, torch.float16, etc., to vLLM valid dtypes.
  • Fixed quantized model parameter count method.
  • Fixed UnslothTrainingArguments not patching trl.Config correctly.
  • Fixed unnecessary warning when using transformers >= 4.53.0.
  • Silently skipped Falcon H1 import if transformers_version is less than 4.53.0.
  • Dynamically adjusted get_per_token_logps for TRL main upgrade.
  • Fixed casual mask implementation.
  • Explicitly checked for xformers existence before using it for attention.
  • Added fallback check for feed_forward if mlp does not exist in the layer module for Falcon H1.
  • Fixed device placement by moving inputs to the correct devices.

🔧 Affected Symbols

Gemma 3NGemma 3Qwen 2.5Qwen 2GLMGRPOTRLPEFTTransformersFalcon H1Devstral 1.1MedGemma 27BMedGemma 4BLoftQFastBaseModelUnslothTrainingArgumentstrl.Configtorch.bfloat16torch.float16get_per_token_logps