fsdp fixes and enhancements (#24980)

* fix fsdp prepare to remove the warnings and fix excess memory usage

* Update training_args.py

* parity for FSDP+XLA

* Update trainer.py

docs/source/en/main_classes/trainer.md

@@ -441,7 +441,7 @@ as the model saving with FSDP activated is only available with recent fixes.
 - Remaining FSDP config is passed via `--fsdp_config <path_to_fsdp_config.json>`. It is either a location of
   FSDP json config file (e.g., `fsdp_config.json`) or an already loaded json file as `dict`. 
   - If auto wrapping is enabled, you can either use transformer based auto wrap policy or size based auto wrap policy.
-    - For transformer based auto wrap policy, please specify `fsdp_transformer_layer_cls_to_wrap` in the config file. 
+    - For transformer based auto wrap policy, it is recommended to specify `fsdp_transformer_layer_cls_to_wrap` in the config file. If not specified, the default value is `model._no_split_modules` when available.
       This specifies the list of transformer layer class name (case-sensitive) to wrap ,e.g, [`BertLayer`], [`GPTJBlock`], [`T5Block`] ....
       This is important because submodules that share weights (e.g., embedding layer) should not end up in different FSDP wrapped units.
       Using this policy, wrapping happens for each block containing Multi-Head Attention followed by couple of MLP layers. 
@@ -482,7 +482,7 @@ Pass `--fsdp "full shard"` along with following changes to be made in `--fsdp_co
   This setting can only be used when the xla flag is set to true, and an auto wrapping policy is specified through
   `fsdp_min_num_params` or `fsdp_transformer_layer_cls_to_wrap`. 
 - You can either use transformer based auto wrap policy or size based auto wrap policy.
-  - For transformer based auto wrap policy, please specify `fsdp_transformer_layer_cls_to_wrap` in the config file. 
+  - For transformer based auto wrap policy, it is recommended to specify `fsdp_transformer_layer_cls_to_wrap` in the config file. If not specified, the default value is `model._no_split_modules` when available.
     This specifies the list of transformer layer class name (case-sensitive) to wrap ,e.g, [`BertLayer`], [`GPTJBlock`], [`T5Block`] ....
     This is important because submodules that share weights (e.g., embedding layer) should not end up in different FSDP wrapped units.
     Using this policy, wrapping happens for each block containing Multi-Head Attention followed by couple of MLP layers. 

src/transformers/trainer.py

@@ -1377,18 +1377,24 @@ def _wrap_model(self, model, training=True, dataloader=None):
                 raise ImportError("Missing XLA FSDP related module; please make sure to use torch-xla >= 2.0.")
             auto_wrap_policy = None
             auto_wrapper_callable = None
+            default_transformer_cls_names_to_wrap = getattr(model, "_no_split_modules", None)
+            fsdp_transformer_layer_cls_to_wrap = self.args.fsdp_config.get(
+                "fsdp_transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap
+            )
+
             if self.args.fsdp_config["fsdp_min_num_params"] > 0:
                 auto_wrap_policy = functools.partial(
                     size_based_auto_wrap_policy, min_num_params=self.args.fsdp_config["fsdp_min_num_params"]
                 )
-            elif self.args.fsdp_config.get("fsdp_transformer_layer_cls_to_wrap", None) is not None:
+            elif fsdp_transformer_layer_cls_to_wrap is not None:
                 transformer_cls_to_wrap = set()
-                for layer_class in self.args.fsdp_config["fsdp_transformer_layer_cls_to_wrap"]:
+                for layer_class in fsdp_transformer_layer_cls_to_wrap:
                     transformer_cls = get_module_class_from_name(model, layer_class)
                     if transformer_cls is None:
                         raise Exception("Could not find the transformer layer class to wrap in the model.")
                     else:
                         transformer_cls_to_wrap.add(transformer_cls)
+
                 auto_wrap_policy = functools.partial(
                     transformer_auto_wrap_policy,
                     # Transformer layer class to wrap
@@ -1600,6 +1606,7 @@ def _inner_training_loop(
             and self.sharded_ddp != ShardedDDPOption.SIMPLE
             or is_sagemaker_mp_enabled()
             or self.fsdp is not None
+            or self.is_fsdp_enabled
         )
 
         # We need to reset the scheduler, as its parameters may be different on subsequent calls
@@ -1631,6 +1638,8 @@ def _inner_training_loop(
         use_accelerator_prepare = True if model is self.model else False
 
         if delay_optimizer_creation:
+            if use_accelerator_prepare:
+                self.model = self.accelerator.prepare(self.model)
             self.create_optimizer_and_scheduler(num_training_steps=max_steps)
 
         # prepare using `accelerator` prepare

src/transformers/training_args.py

@@ -1567,14 +1567,14 @@ def __post_init__(self):
                 elif fsdp_option == FSDPOption.OFFLOAD:
                     os.environ["FSDP_OFFLOAD_PARAMS"] = "true"
                 elif fsdp_option == FSDPOption.AUTO_WRAP:
+                    os.environ["FSDP_AUTO_WRAP_POLICY"] = FSDP_AUTO_WRAP_POLICY[0]
                     if self.fsdp_config["fsdp_min_num_params"] > 0:
                         os.environ["FSDP_MIN_NUM_PARAMS"] = str(self.fsdp_config["fsdp_min_num_params"])
                         os.environ["FSDP_AUTO_WRAP_POLICY"] = FSDP_AUTO_WRAP_POLICY[1]
                     elif self.fsdp_config.get("fsdp_transformer_layer_cls_to_wrap", None) is not None:
                         os.environ["FSDP_TRANSFORMER_CLS_TO_WRAP"] = ",".join(
                             self.fsdp_config["fsdp_transformer_layer_cls_to_wrap"]
                         )
-                        os.environ["FSDP_AUTO_WRAP_POLICY"] = FSDP_AUTO_WRAP_POLICY[0]
             prefetch_policy = self.fsdp_config.get("fsdp_backward_prefetch", "NO_PREFETCH")
             os.environ["FSDP_BACKWARD_PREFETCH"] = prefetch_policy.upper()