Add the interface to use SHARP to FSDP strategy (#8202)

Signed-off-by: Sangkug Lym <slym@nvidia.com> Signed-off-by: stevehuang52 <heh@nvidia.com>
NVIDIA · Jan 31, 2024 · cc78cb8 · cc78cb8
1 parent 069773c
commit cc78cb8
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 0 deletions.
diff --git a/nemo/collections/nlp/parts/megatron_trainer_builder.py b/nemo/collections/nlp/parts/megatron_trainer_builder.py
@@ -69,6 +69,7 @@ def _training_strategy(self) -> Union[NLPDDPStrategy, NLPFSDPStrategy]:
                 sharded_checkpoint=sharded_checkpoint,
                 precision=self.cfg.trainer.precision,
                 nccl_communicator_config_path=self.cfg.model.get('nccl_communicator_config_path', None),
+                sharp=self.cfg.model.get('sharp', False),
             )
 
         return NLPDDPStrategy(

diff --git a/nemo/collections/nlp/parts/nlp_overrides.py b/nemo/collections/nlp/parts/nlp_overrides.py
@@ -517,6 +517,7 @@ def __init__(
         sharded_checkpoint: bool = False,
         precision: Union[int, str] = 'bf16-mixed',
         nccl_communicator_config_path: Optional[str] = None,
+        sharp: bool = False,
         **kwargs: Union[Any, Dict[str, Any]],
     ) -> None:
         if not HAVE_APEX:
@@ -561,6 +562,7 @@ def __init__(
         )
 
         self.nccl_communicator_config_path = nccl_communicator_config_path
+        self.sharp = sharp
         super().__init__(**kwargs)
 
     def _set_mixed_precision_recipe(