[AutoParallel] keep input order (#48963)

* [AutoParallel] keep input order

* rm annotation

python/paddle/distributed/auto_parallel/dist_saver.py

@@ -195,17 +195,16 @@ class DistributedSaver:
             used_inputs += op.input_arg_names
             used_outputs += op.output_arg_names
 
-        dist_feed_vars_names = list(set(feed_vars_names) & set(used_inputs))
-        dist_fetch_vars_names = list(set(fetch_vars_names) & set(used_outputs))
+        for idx, var_name in enumerate(feed_vars_names):
+            if var_name not in used_inputs:
+                feed_vars_names.pop(idx)
+        for idx, var_name in enumerate(fetch_vars_names):
+            if var_name not in used_outputs:
+                fetch_vars_names.pop(idx)
 
-        dist_feed_vars = [
-            global_block.vars[name] for name in dist_feed_vars_names
-        ]
-        dist_fetch_vars = [
-            global_block.vars[name] for name in dist_fetch_vars_names
-        ]
+        dist_feed_vars = [global_block.vars[name] for name in feed_vars_names]
+        dist_fetch_vars = [global_block.vars[name] for name in fetch_vars_names]
 
-        # NOTE: `paddle.static.save_inference_model` does not support subblock.
         dist_filename = filename + "_dist" + str(rank_id)
         dist_path = os.path.join(dirname, dist_filename)
         paddle.static.save_inference_model(