unslothai/unsloth July-2025

Less VRAM + bug fixes

on GitHub

More VRAM reduction, faster & bug fixes

Please update Unsloth! pip install --upgrade --force-reinstall --no-deps --no-cache-dir unsloth unsloth_zoo

1. Gemma 3N Vision now works and is fixed! Please re-download all model checkpoints (Unsloth will auto do it) Try Kaggle Notebook! There is also a challenge with a prize pool of $100,000!
2. Gemma 3 text and vision are all fixed for T4, and is much faster. Losses of 6 to 7 are now fixed - it should be 1 to 2.
3. 10 to 25% less VRAM consumption for all models. Also faster compiling and less errors. Unsloth is now more stable!
4. Downloads stuck at 90% to 95% fixed!
5. Qwen 2.5, Qwen 2, GLM all fixed as well.
6. GRPO now works with latest main TRL
7. Main TRL, PEFT, Transformers all work
8. Forced upgrading transformers is now fixed.
9. Falcon H1 finetuning should work great! Notebooks incoming
10. Devstral 1.1 and MedGemma 27B, 4B support with vision
11. Many many many more bug fixes - this release of Unsloth should be much more stable and error tolerant!

Please update Unsloth! pip install --upgrade --force-reinstall --no-deps --no-cache-dir unsloth unsloth_zoo

What's Changed

• Gemma 3N by @danielhanchen in #2809
• Add instructions for installing unsloth on RTX 5090 by @jeromeku in #2812
• Add falcon h1 by @dhiaEddineRhaiem in #2650
• Granite4 support by @mmathew23 in #2799
• import undefined transformers_version for falcon model by @mmathew23 in #2822
• Fix LoftQ with FastBaseModel by @mehmetoguzderin in #2826
• Create stale.yml by @danielhanchen in #2832
• Create stale.yml by @danielhanchen in #2836
• Added conda/mamba section to blackwell installation readme by @rolandtannous in #2817
• Gemma 3N bug fixes by @danielhanchen in #2842
• Fix loftq None config for FastBaseModel by @mmathew23 in #2848
• Convert torch.bfloat16, torch.float16, etc. to vLLM valid dtypes by @rishabh135 in #2811
• [Feature] enable unsloth on amd gpu by @billishyahao in #2520
• Fix Gemma 3N by @danielhanchen in #2854
• fix quantized model parameter count method by @rolandtannous in #2855
• Update CSM for faster inference (no compile) by @mmathew23 in #2865
• Fix UnslothTrainingArguments not patching trl.Config properly by @Erland366 in #2873
• Fix unnecessary warning for transformers >= 4.53.0 by @mmathew23 in #2867
• Update README.md by @danielhanchen in #2885
• Many bug fixes by @danielhanchen in #2908
• silenty skip falcon h1 import if transformers_version < 4.53.0 by @mmathew23 in #2912
• Dynamically adjust get_per_token_logps [trl main upgrade] by @Datta0 in #2911
• [Intel] add intel gpu with vllm support by @leizhenyuan in #2903
• [bugs] fix for casual mask by @leizhenyuan in #2868
• Explicitly check if xformers exists for attention by @Datta0 in #2889
• Falcon H1: if mlp doesn't exist in layer module check for feed_forward by @mmathew23 in #2913
• Move inputs to right devices. by @Datta0 in #2919
• Many bug fixes by @danielhanchen in #2927

New Contributors

• @dhiaEddineRhaiem made their first contribution in #2650
• @mehmetoguzderin made their first contribution in #2826
• @rishabh135 made their first contribution in #2811
• @billishyahao made their first contribution in #2520

Full Changelog: June-2025...July-2025