This release fixes the attention implementation of a supervised run over Gemma 4 26B-A4B.
Changes
A supervised run over google/gemma-4-26B-A4B-it on two or more H200 GPUs that names no config.model.attentionImplementation trains with flex_attention, as the other Gemma 4 models do. Previously, it trained with sdpa, and its packing start check could fail the run.
