Update modeling_minicpmv.py

modeling_minicpmv.py

@@ -288,16 +288,24 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
 
         # Step 2: Automatically pad each tensor to have the same length (L) in the last dimension
         attention_mask = []
+        
+        embedding_layer = self.get_input_embeddings()
+
+        # Retrieve the embedding vector for pad_token_id
+        pad_embedding_vector = embedding_layer.weight[0]
+        vector_reshaped = pad_embedding_vector.view(1, 1, 10)
+        
         for tensor in batch:
             to_add = []
             for pl in range(tensor.shape[1]):
                 to_add.append(1)
             for pl in range(tensor.shape[1], max_x):
+                tensor = torch.cat((tensor, vector_reshaped), dim=1)
                 to_add.append(0)
             attention_mask.append(to_add)
         attention_mask = torch.tensor(attention_mask)
         
-        padded_tensors = [torch.nn.functional.pad(tensor, (0, 0, 0, max_x - tensor.shape[1]), padding_value=0.0) for tensor in batch]
+        # padded_tensors = [torch.nn.functional.pad(tensor, (0, 0, 0, max_x - tensor.shape[1]), padding_value=0.0) for tensor in batch]
         
         # Step 3: Stack the padded tensors into a single batch
         for stuff in batch: