Update min_dalle/min_dalle.py

min_dalle/min_dalle.py

@@ -251,6 +251,9 @@ class MinDalle:
                     prev_tokens=image_tokens[i],
                     token_index=token_indices[[i]]
                 )
+                
+                del attention_state
+                del image_tokens
              
 
             with torch.cuda.amp.autocast(dtype=torch.float16):
@@ -260,8 +263,8 @@ class MinDalle:
                         is_seamless=is_seamless,
                         is_verbose=is_verbose
                     )
-            del attention_state
-            del image_tokens
+                    
+