Modeling fix (quic#605)

Signed-off-by: Mohit Soni <mohisoni@qti.qualcom.com>
Co-authored-by: Mohit Soni <mohisoni@qti.qualcom.com>
Signed-off-by: abhishek-singh591 <sabhis@qti.qualcomm.com>

Author: 2 people

QEfficient/transformers/models/modeling_auto.py

@@ -1412,6 +1412,8 @@ def kv_offload_generate(
                 if x.startswith("past_") or x.endswith("_RetainedState")
             ]
         )
+        if not_mllama:
+            lang_session.skip_buffers(vision_outputs.keys())
 
         # Get first token
         lang_inputs["input_ids"] = outputs["logits"].argmax(2)

QEfficient/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -953,6 +953,7 @@ def smart_resize(
         grid_height = grid_h * grid_w
         grid_width = patch_size * patch_size * temporal_patch_size * channel
         vision_size = grid_height // 4
+        vision_size = vision_size * num_frames
         grid_height = grid_height * batch_size
 
         vision = [