fix

Author: HydrogenSulfate

deepmd/pd/train/training.py

@@ -783,7 +783,7 @@ def step(_step_id, task_key="Default") -> None:
                         loss.backward()
 
                 # gradient accumulation
-                if _step_id % self.acc_freq == 0:
+                if (_step_id + 1) % self.acc_freq == 0:
                     # fuse + allreduce manually before optimization if use DDP + no_sync
                     # details in https://github.com/PaddlePaddle/Paddle/issues/48898#issuecomment-1343838622
                     if self.world_size > 1:

source/tests/pd/test_training.py

@@ -150,9 +150,7 @@ def setUp(self) -> None:
         self.config["model"] = deepcopy(model_se_e2_a)
         self.config["training"]["numb_steps"] = 1
         self.config["training"]["save_freq"] = 1
-        # import paddle
         enable_prim(True)
-        # assert paddle.framework.core._is_eager_prim_enabled()
 
     def tearDown(self) -> None:
         DPTrainTest.tearDown(self)
@@ -170,9 +168,7 @@ def setUp(self) -> None:
         self.config["training"]["numb_steps"] = 1
         self.config["training"]["save_freq"] = 1
         self.config["training"]["acc_freq"] = 4
-        # import paddle
         enable_prim(True)
-        # assert paddle.framework.core._is_eager_prim_enabled()
 
     def tearDown(self) -> None:
         DPTrainTest.tearDown(self)