[Bugfix] recompute dep filter param (#49010)

* recompute dep filter param * recompute dep for reshard
PaddlePaddle · Dec 14, 2022 · b9fad5d · b9fad5d
1 parent a8d139a
commit b9fad5d
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 0 deletions.
diff --git a/python/paddle/distributed/auto_parallel/reshard.py b/python/paddle/distributed/auto_parallel/reshard.py
@@ -2497,6 +2497,8 @@ def _reshard_output(self, block):
             "read",
             "write_to_array",
             "read_from_array",
+            "nop",
+            "depend",
         ]
         global _g_special_ops
         skip_ops += _g_special_ops

diff --git a/python/paddle/distributed/auto_parallel/utils.py b/python/paddle/distributed/auto_parallel/utils.py
@@ -2168,6 +2168,9 @@ def insert_dependencies_for_two_ops(
 
     def _select_best_depend_var(vars):
 
+        # parameter should not be dep var since it maybe partition in sharding pass
+        vars = [var for var in vars if not var.is_parameter]
+        assert len(vars) > 0
         vars_with_numels = [(var, get_var_numel(var)) for var in vars]
         vars_with_numels.sort(key=lambda x: x[1])