Fix: align Qwen2.5-VL inference rope index with training by passing s… (#41153)

Fix: align Qwen2.5-VL inference rope index with training by passing second_per_grid_ts

Author: Xqle

src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -1558,6 +1558,7 @@ def prepare_inputs_for_generation(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
                     video_grid_thw=video_grid_thw,
+                    second_per_grid_ts=second_per_grid_ts,
                     attention_mask=attention_mask,
                 )
                 self.model.rope_deltas = rope_deltas

src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py

@@ -814,6 +814,7 @@ def prepare_inputs_for_generation(
                     model_inputs.get("input_ids", None),
                     image_grid_thw=image_grid_thw,
                     video_grid_thw=video_grid_thw,
+                    second_per_grid_ts=second_per_grid_ts,
                     attention_mask=attention_mask,
                 )
                 self.model.rope_deltas = rope_deltas