Update modeling_minicpmv.py

modeling_minicpmv.py

@@ -272,7 +272,16 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
             counter += 1
             batch.append(input_embeds)
             
-        batch = torch.cat(batch, dim=0)
+        # batch = torch.cat(batch, dim=0)
+        # pad_sequence(embeddings_list, batch_first=True, padding_value=0.0)
+        max_length = max(tensor.shape[1] for tensor in batch)
+
+        # Step 2: Automatically pad each tensor to have the same length (L) in the last dimension
+        padded_tensors = [torch.nn.functional.pad(tensor, (0, max_length - tensor.shape[1])) for tensor in batch]
+        
+        # Step 3: Stack the padded tensors into a single batch
+        batch = torch.cat(padded_tensors, dim=0)
+        
         # output_ids = self._decode(input_embeds, tokenizer, **kwargs)
         if stream:
             kwargs.pop("decode_text")