Update modeling_minicpmv.py

modeling_minicpmv.py

@@ -297,25 +297,24 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
         vector_reshaped = pad_embedding_vector.view(1, 1, 4096)
         
         for place, tensor in enumerate(batch):
-            to_add = []
-            for pl in range(0, max_x - tensor.shape[1]):
-                to_add.append(0)
-            bound = 0
-            
-            for pl in range(max_x - tensor.shape[1], max_x):
-                if bound == 0:
-                    bound = pl
-            
-                #print("here")
-                tensor = torch.cat((vector_reshaped, tensor), dim=1)
-                to_add.append(1)
-            # tensor = tensor.to(self.device)
-            # if bound != 0:
-            #     print(tensor[0,bound-1], " BOUND")
-            #     print(tensor[0,bound], " OTHERSIDE")
-            #     print(pad_embedding_vector, " VECTOR")
+            # Calculate how much padding is needed on the left
+            padding_needed = max_x - tensor.shape[1]
+        
+            # Create the list for the attention mask, marking the padded tokens
+            to_add = [0] * padding_needed + [1] * tensor.shape[1]
+        
+            # Create the padding tensor of the correct size
+            padding_tensor = vector_reshaped.expand(tensor.shape[0], padding_needed, tensor.shape[2])
+        
+            # Concatenate the padding tensor to the left of the original tensor
+            tensor = torch.cat((padding_tensor, tensor), dim=1)
+        
             print(tensor.shape, "UPDATED_SHAPE")
+            
+            # Update the batch with the padded tensor
             batch[place] = tensor
+            
+            # Append the attention mask for this tensor
             attention_mask.append(to_add)
         
         attention_mask = torch.tensor(attention_mask)