Merge pull request #13104 from jacquesqiao/fix-sparse-grad-merge

fix sparse grad merge on pserver

Author: jacquesqiao

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

@@ -438,7 +438,7 @@ def transpiler_test_impl(self):
         # 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)

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -1390,13 +1390,11 @@ def _append_pserver_grad_merge_ops(self, optimize_block,
                 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,