Automated PR - 2026-04-13

2026-04-13 14:29:35 +00:00
parent 59ca828d5a
commit d887bbd1e0
29 changed files with 463 additions and 184 deletions
@@ -125,7 +125,8 @@ class TrainingStrategy(ABC):
            audio_pred: Audio prediction from the transformer model (None for video-only)
            inputs: The prepared model inputs containing targets and masks
        Returns:
-            Scalar loss tensor
+            Per-element loss tensor of shape [B,]. The trainer reduces to a scalar
+            before backward(). Returning unreduced loss enables per-sigma-bucket tracking.
        """

    def get_checkpoint_metadata(self) -> dict[str, Any]: