[trainer] fix: Add data.seed to config

docs/examples/config.rst

@@ -25,6 +25,7 @@ Data
      return_raw_chat: False
      return_full_prompt: False
      shuffle: True
+     seed: 42
      filter_overlong_prompts: False
      filter_overlong_prompts_workers: 1
      truncation: error
@@ -60,6 +61,8 @@ Data
   without applying chat template.
 - ``data.return_full_prompt``: Whether to return the full prompt with chat template
 - ``data.shuffle``: Whether to shuffle the data in the dataloader.
+- ``data.seed``: An integer seed to use when shuffling the data. If not set or set to
+  `null`, the data shuffling will not be seeded, resulting in a different data order on each run.
 - ``data.filter_overlong_prompts``: Default don't filter.
 - ``data.filter_overlong_prompts_workers``: For large-scale dataset, filtering
   overlong prompts could be timeconsuming. You cat set the ``filter_overlong_prompts_workers``

examples/split_placement/config/ppo_trainer_split.yaml

@@ -21,6 +21,7 @@ data:
   return_raw_chat: False
   return_full_prompt: False
   shuffle: True
+  seed: 42
 
 actor_rollout_ref:
   hybrid_engine: True

recipe/entropy/main_entropy.py

@@ -237,7 +237,9 @@ def create_rl_sampler(data_config, dataset):
     # use sampler for better ckpt resume
     if data_config.shuffle:
         train_dataloader_generator = torch.Generator()
-        train_dataloader_generator.manual_seed(data_config.get("seed", 1))
+        seed = data_config.get("seed")
+        if seed is not None:
+            train_dataloader_generator.manual_seed(seed)
         sampler = RandomSampler(data_source=dataset, generator=train_dataloader_generator)
     else:
         sampler = SequentialSampler(data_source=dataset)

recipe/prime/prime_ray_trainer.py

@@ -187,7 +187,9 @@ def _create_dataloader(self, *args, **kwargs):
         # use sampler for better ckpt resume
         if self.config.data.shuffle:
             train_dataloader_generator = torch.Generator()
-            train_dataloader_generator.manual_seed(self.config.data.get("seed", 1))
+            seed = self.config.data.get("seed")
+            if seed is not None:
+                train_dataloader_generator.manual_seed(seed)
             sampler = RandomSampler(data_source=self.train_dataset, generator=train_dataloader_generator)
         else:
             sampler = SequentialSampler(data_source=self.train_dataset)

tests/trainer/config/legacy_ppo_megatron_trainer.yaml

@@ -12,6 +12,7 @@ data:
   return_raw_chat: False
   return_full_prompt: False
   shuffle: True
+  seed: null # An integer seed to use when shuffling the data. If not set or set to `null`, the data shuffling will not be seeded, resulting in a different data order on each run.
   filter_overlong_prompts: False # for large-scale dataset, filtering overlong prompts could be timeconsuming. You cat set the filter_overlong_prompts_workers to use multiprocessing to speed up.
   filter_overlong_prompts_workers: 1
   truncation: error

tests/trainer/config/legacy_ppo_trainer.yaml

@@ -55,6 +55,10 @@ data:
   # Whether to shuffle the data in the dataloader.
   shuffle: True
 
+  # An integer seed to use when shuffling the data. If not set or set to
+  # `null`, the data shuffling will not be seeded, resulting in a different data order on each run.
+  seed: null
+
   # num dataloader workers
   dataloader_num_workers: 8
 

verl/trainer/config/_generated_ppo_megatron_trainer.yaml

@@ -286,6 +286,7 @@ data:
   return_raw_chat: false
   return_full_prompt: false
   shuffle: true
+  seed: null
   dataloader_num_workers: 8
   validation_shuffle: false
   filter_overlong_prompts: false

verl/trainer/config/_generated_ppo_trainer.yaml

@@ -282,6 +282,7 @@ data:
   return_raw_chat: false
   return_full_prompt: false
   shuffle: true
+  seed: null
   dataloader_num_workers: 8
   validation_shuffle: false
   filter_overlong_prompts: false

verl/trainer/config/data/legacy_data.yaml

@@ -48,6 +48,9 @@ return_full_prompt: False
 # Whether to shuffle the data in the dataloader.
 shuffle: True
 
+# Seed to use when shuffling the data
+seed: null
+
 # num dataloader workers
 dataloader_num_workers: 8
 

verl/trainer/main_ppo.py

@@ -403,7 +403,9 @@ def create_rl_sampler(data_config, dataset):
     # If shuffling is enabled in the data configuration, create a random sampler.
     elif data_config.shuffle:
         train_dataloader_generator = torch.Generator()
-        train_dataloader_generator.manual_seed(data_config.get("seed", 1))
+        seed = data_config.get("seed")
+        if seed is not None:
+            train_dataloader_generator.manual_seed(seed)
         sampler = RandomSampler(data_source=dataset, generator=train_dataloader_generator)
     else:
         # If shuffling is disabled, use a sequential sampler to iterate through the dataset in order.