per-sample media budgets in processor; per-sample MRoPE position ids for batched vision inputs

modeling_moss_vl.py

@@ -1732,6 +1732,27 @@ class MossVLModel(MossVLPreTrainedModel):
             rope_deltas: (batch_size,) - position offset due to vision tokens
         """
         batch_size, max_vision_seq_len, _ = cross_attention_states.shape
+        if batch_size > 1:
+            # Match frames to text markers within each sample. Blank images and
+            # right-truncated media may have no corresponding marker; flattening
+            # across the batch would assign their metadata to the next sample.
+            # Keep the original single-sample computation, including in-place
+            # text position updates and generation's per-sample rope_deltas.
+            results = [
+                self.compute_vision_position_ids(
+                    input_ids[i:i + 1],
+                    position_ids[:, i:i + 1],
+                    [vision_token_info[i]],
+                    cross_attention_states[i:i + 1],
+                    None if attention_mask is None else attention_mask[i:i + 1],
+                )
+                for i in range(batch_size)
+            ]
+            return (
+                torch.cat([result[0] for result in results], dim=1),
+                position_ids,
+                torch.cat([result[2] for result in results], dim=0),
+            )
         device = cross_attention_states.device
         image_token_id = self.config.image_token_id
         merge_size = self.visual.spatial_merge_size