diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index 1d9ab44ed447468fb8383c52747d14970ae27ced..b85501ef6b80d1f5004aa0dd08c3123d3bda48a5 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -438,7 +438,7 @@ class TestLocalLookupTable(TestDistLookupTableBase): # 2 optimize for table adam # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num self.assertEqual([op.type for op in pserver1.blocks[2].ops], - ["sum", "adam", "scale", "scale"]) + ["sum", "scale", "adam", "scale", "scale"]) trainer, _ = self.get_trainer() self.assertEqual(len(trainer.blocks), 1) diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index a6266a7b0c9ac40eac7b2823fc7ddf38f55357a9..bddeb6617c1743de946b3c5b4b0a465d85f35ce3 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -1390,13 +1390,11 @@ class DistributeTranspiler(object): inputs={"X": vars2merge}, outputs={"Out": merged_var}, attrs={"use_mkldnn": False}) - # TODO(panyx0718): What if it's SELECTED_ROWS. - if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: - optimize_block.append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainer_num)}) + optimize_block.append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainer_num)}) return merged_var def _append_pserver_ops(self, optimize_block, opt_op, endpoint,