@@ -24,7 +24,7 @@ from super_gradients.common.factories.callbacks_factory import CallbacksFactory
 from super_gradients.common.data_types.enum import MultiGPUMode, StrictLoad, EvaluationType
 from super_gradients.training.models.all_architectures import ARCHITECTURES
 from super_gradients.common.decorators.factory_decorator import resolve_param
-from super_gradients.common.environment import env_helpers
+from super_gradients.common.environment import ddp_utils
 from super_gradients.common.abstractions.abstract_logger import get_logger, mute_current_process
 from super_gradients.common.factories.list_factory import ListFactory
 from super_gradients.common.factories.losses_factory import LossesFactory
@@ -81,7 +81,6 @@ from super_gradients.training.utils.callbacks import (
     ContextSgMethods,
     LRCallbackBase,
 )
-from super_gradients.common.environment import environment_config
 from super_gradients.training.utils import HpmStruct
 from super_gradients.training.utils.hydra_utils import load_experiment_cfg, add_params_to_cfg
 from omegaconf import OmegaConf
@@ -1433,7 +1432,7 @@ class Trainer:
                         logger.warning("\n[WARNING] - Tried running on multiple GPU but only a single GPU is available\n")
                 else:
                     if requested_multi_gpu == MultiGPUMode.AUTO:
-                        if env_helpers.is_distributed():
+                        if ddp_utils.is_distributed():
                             requested_multi_gpu = MultiGPUMode.DISTRIBUTED_DATA_PARALLEL
                         else:
                             requested_multi_gpu = MultiGPUMode.DATA_PARALLEL
@@ -1456,7 +1455,7 @@ class Trainer:
         batch you specify times the number of GPUs. In the literature there are several "best practices" to set
         learning rates and schedules for large batch sizes.
         """
-        local_rank = environment_config.DDP_LOCAL_RANK
+        local_rank = ddp_utils.DDP_LOCAL_RANK
         if local_rank > 0:
             mute_current_process()
 

@@ -11,9 +11,10 @@ from torch.distributed.elastic.multiprocessing.errors import record
 from torch.distributed.launcher.api import LaunchConfig, elastic_launch
 
 from super_gradients.common.data_types.enum import MultiGPUMode
-from super_gradients.common.environment.env_helpers import find_free_port, is_distributed
+from super_gradients.common.environment.argparse_utils import EXTRA_ARGS
+from super_gradients.common.environment.ddp_utils import find_free_port, is_distributed
 from super_gradients.common.abstractions.abstract_logger import get_logger
-from super_gradients.common.environment import environment_config
+
 
 logger = get_logger(__name__)
 
@@ -238,7 +239,7 @@ def restart_script_with_ddp(num_gpus: int = None):
         metrics_cfg={},
     )
 
-    elastic_launch(config=config, entrypoint=sys.executable)(*sys.argv, *environment_config.EXTRA_ARGS)
+    elastic_launch(config=config, entrypoint=sys.executable)(*sys.argv, *EXTRA_ARGS)
 
     # The code below should actually never be reached as the process will be in a loop inside elastic_launch until any subprocess crashes.
     sys.exit("Main process finished")