●Change8

v0.1.36-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 1 breaking✨ 4 features🐛 13 fixes🔧 7 symbols

Summary

This release focuses heavily on stabilizing and fixing issues related to Gemma 4 training and quant formats, alongside significant stability and feature enhancements in Unsloth Studio, including the addition of speculative decoding.

⚠️ Breaking Changes

  • Users must re-download Gemma 4 GGUFs due to updated quant files, although the issues fixed were external to Unsloth.

Migration Steps

  1. Re-download Gemma 4 GGUFs to incorporate critical fixes related to CUDA buffer overlap, kv-cache, and token handling.
  2. Ensure Unsloth version is at least 2026.4.2 (or later, as 2026.4.4 is also mentioned) for optimal Gemma 4 support and fixes.

✨ New Features

  • Gemma 4 training and quant fixes implemented, resulting in stable training and performance improvements (~1.5x faster, ~60% less VRAM than FA2 setups).
  • Speculative decoding (ngram-mod) added to Unsloth Studio (on by default).
  • Ability to export and save Gemma 4 models enabled.
  • Unsloth Studio now allows non-LLM recipes to run and moves the Data tab first in executions.

🐛 Bug Fixes

  • Fixed exploding losses (300-400 down to 10-15) during Gemma 4 training when using gradient accumulation.
  • Fixed Index Error during inference for 26B and 31B Gemma 4 models when using standard transformers.
  • Resolved gibberish output when using `use_cache=False` for Gemma E2B and E4B models.
  • Fixed float16 audio -1e9 overflows.
  • Fixed Qwen3.5 and Gemma 4 training issues in Unsloth Studio.
  • Lazy-imported transformers in `model_config` to fix issues related to switching between transformers 4.x and 5.x versions.
  • Patched PEFT for `Gemma4ClippableLinear` in loader checkpoint path to fix export issues.
  • Fixed issues related to llama.cpp building.
  • Removed Gemma-4 from `FORCE_FLOAT32`.
  • Skipped redundant `HfFileSystem().glob()` calls in `loader.py`.
  • Fixed custom folder scan failure in Unsloth Studio when pointing directly at a model directory for GGUF variants.
  • Fixed empty chat threads on navigation and stabilized the new chat flow in Unsloth Studio.
  • Fixed Studio Colab button message fallback when proxy URL fails.

Affected Symbols

Gemma 4 training pipelineGemma 4 inference (26B, 31B)Gemma E2B/E4B with use_cache=Falsemodel_configGemma4ClippableLinearloader.pyis_vision_model()