Merge branch 'saaketh/remove_deepspeed' of https://github.com/mosaicm…

…l/composer into saaketh/remove_deepspeed

merge type beat

composer/callbacks/checkpoint_saver.py

@@ -513,7 +513,7 @@ def _save_checkpoint(self, state: State, logger: Logger):
                         logger=logger,
                     )
             else:
-                remote_file_name = self.remote_file_name.format(state,).lstrip('/')
+                remote_file_name = self.remote_file_name.format(state).lstrip('/')
 
             log.debug(f'Uploading checkpoint to {remote_file_name}')
             try:
@@ -533,7 +533,7 @@ def _save_checkpoint(self, state: State, logger: Logger):
 
             # symlinks stay the same with sharded checkpointing
             if self.latest_remote_file_name is not None:
-                symlink_name = self.latest_remote_file_name.format(state,).lstrip('/') + '.symlink'
+                symlink_name = self.latest_remote_file_name.format(state).lstrip('/') + '.symlink'
 
                 # create and upload a symlink file
                 symlink_filename = os.path.join(

composer/trainer/trainer.py

@@ -1233,7 +1233,7 @@ def __init__(
                 **parallelism_config_args,
             ) if len(parallelism_config_args) > 0 else None
         if parallelism_config is not None or dist.get_world_size() > 1:
-            # FSDP requries torch.distributed to be initialized, even if the world size is 1
+            # FSDP requires torch.distributed to be initialized, even if the world size is 1
             # And torch.distributed is always required for multi-rank training
             dist.initialize_dist(device, dist_timeout)
         if parallelism_config is not None: