fix(pt): use user seed in DpLoaderSet

deepmd/pt/entrypoints/main.py

@@ -109,7 +109,9 @@
         assert dist.is_nccl_available()
         dist.init_process_group(backend="nccl")
 
-    def prepare_trainer_input_single(model_params_single, data_dict_single, rank=0):
+    def prepare_trainer_input_single(
+        model_params_single, data_dict_single, rank=0, seed=None
+    ):
         training_dataset_params = data_dict_single["training_data"]
         validation_dataset_params = data_dict_single.get("validation_data", None)
         validation_systems = (
@@ -134,11 +136,14 @@
             stat_file_path_single = DPPath(stat_file_path_single, "a")
 
         # validation and training data
+        # avoid the same batch sequence among devices
+        rank_seed = (seed + rank) % (2**32) if seed is not None else None
         validation_data_single = (
             DpLoaderSet(
                 validation_systems,
                 validation_dataset_params["batch_size"],
                 model_params_single["type_map"],
+                seed=rank_seed,
             )
             if validation_systems
             else None
@@ -147,6 +152,7 @@
             training_systems,
             training_dataset_params["batch_size"],
             model_params_single["type_map"],
+            seed=rank_seed,
         )
         return (
             train_data_single,
@@ -155,6 +161,7 @@
         )
 
     rank = dist.get_rank() if dist.is_available() and dist.is_initialized() else 0
+    data_seed = config["training"].get("seed", None)
     if not multi_task:
         (
             train_data,
@@ -164,6 +171,7 @@
             config["model"],
             config["training"],
             rank=rank,
+            seed=data_seed,
         )
     else:
         train_data, validation_data, stat_file_path = {}, {}, {}
@@ -176,6 +184,7 @@
                 config["model"]["model_dict"][model_key],
                 config["training"]["data_dict"][model_key],
                 rank=rank,
+                seed=data_seed,
             )
 
     trainer = training.Trainer(

deepmd/pt/train/training.py

@@ -274,10 +274,6 @@ def get_lr(lr_params):
             self.opt_type, self.opt_param = get_opt_param(training_params)
 
         # Model
-        dp_random.seed(training_params["seed"])
-        if training_params["seed"] is not None:
-            torch.manual_seed(training_params["seed"])
-
         self.model = get_model_for_wrapper(model_params)
 
         # Loss
@@ -302,7 +298,6 @@ def get_lr(lr_params):
                 )
 
         # Data
-        dp_random.seed(training_params["seed"])
         if not self.multi_task:
             self.get_sample_func = single_model_stat(
                 self.model,

deepmd/pt/utils/dataloader.py

@@ -80,7 +80,8 @@
         seed=10,
         shuffle=True,
     ):
-        setup_seed(seed)
+        if seed is not None:
+            setup_seed(seed)
         if isinstance(systems, str):
             with h5py.File(systems) as file:
                 systems = [os.path.join(systems, item) for item in file.keys()]