Whether a Sample is Involved in Either Understanding Loss Calculation or Generation Loss Calculation

Hello, although the packed batch contains samples from different categories (e.g., t2i_pretrain, vlm_sft, unified_edit), the attention mask separates these samples. Therefore, each sample undergoes either understanding loss calculation or generation loss calculation, but not both.

No source files, tests, or entry points are named. Start by locating the packed-batch attention-mask and understanding/generation-loss paths; done would require resolving whether each sample contributes to only one loss, but the issue does not define a code change or acceptance test.

Written by the indexing model from the issue text.