move fsdp handling to accelerate (#23158)

* mixed precision support via accelerate

* fix issues

* fix for the sharded ddp case

* fix flax and tf failing tests

* `refactor the place to create `Accelerator` object

* move ddp prep to accelerate

* fix 😅

* resolving comments

* move fsdp handling to accelerate

* fixex

* fix saving

src/transformers/trainer.py

@@ -343,6 +343,12 @@ class Trainer:
         # create accelerator object
         self.accelerator = Accelerator()
 
+        # post accelerator creation setup
+        if getattr(self.accelerator.state, "fsdp_plugin", None) is not None:
+            fsdp_plugin = self.accelerator.state.fsdp_plugin
+            fsdp_plugin.limit_all_gathers = self.args.fsdp_config.get("limit_all_gathers", False)
+            fsdp_plugin.use_orig_params = self.args.fsdp_config.get("use_orig_params", False)
+
         # memory metrics - must set up as early as possible
         self._memory_tracker = TrainerMemoryTracker(self.args.skip_memory_metrics)
         self._memory_tracker.start()
@@ -464,7 +470,7 @@ class Trainer:
                 self.fsdp = ShardingStrategy.NO_SHARD
 
             self.backward_prefetch = BackwardPrefetch.BACKWARD_PRE
-            if "backward_prefetch" in self.args.fsdp_config and "backward_pos" in self.args.fsdp_config.get(
+            if "backward_prefetch" in self.args.fsdp_config and "backward_post" in self.args.fsdp_config.get(
                 "backward_prefetch", []
             ):
                 self.backward_prefetch = BackwardPrefetch.BACKWARD_POST
@@ -1479,63 +1485,7 @@ class Trainer:
                     cpu_offload=cpu_offload,
                 ).to(self.args.device)
         # Distributed training using PyTorch FSDP
-        elif self.fsdp is not None:
-            if not self.args.fsdp_config["xla"]:
-                # PyTorch FSDP!
-                from torch.distributed.fsdp.fully_sharded_data_parallel import CPUOffload, MixedPrecision
-                from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
-                from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
-
-                if FSDPOption.OFFLOAD in self.args.fsdp:
-                    cpu_offload = CPUOffload(offload_params=True)
-                else:
-                    cpu_offload = CPUOffload(offload_params=False)
-
-                auto_wrap_policy = None
-
-                if FSDPOption.AUTO_WRAP in self.args.fsdp:
-                    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:
-                        transformer_cls_to_wrap = set()
-                        for layer_class in self.args.fsdp_config["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
-                            transformer_layer_cls=transformer_cls_to_wrap,
-                        )
-                mixed_precision_policy = None
-                dtype = None
-                if self.args.fp16:
-                    dtype = torch.float16
-                elif self.args.bf16:
-                    dtype = torch.bfloat16
-                if dtype is not None:
-                    mixed_precision_policy = MixedPrecision(param_dtype=dtype, reduce_dtype=dtype, buffer_dtype=dtype)
-                if type(model) != FSDP:
-                    # XXX: Breaking the self.model convention but I see no way around it for now.
-                    signature = inspect.signature(FSDP.__init__).parameters.keys()
-                    kwargs = {}
-                    for arg in ["limit_all_gathers", "forward_prefetch", "backward_prefetch"]:
-                        if arg in signature:
-                            kwargs[arg] = getattr(self, arg)
-                    self.model = model = FSDP(
-                        model,
-                        sharding_strategy=self.fsdp,
-                        cpu_offload=cpu_offload,
-                        auto_wrap_policy=auto_wrap_policy,
-                        mixed_precision=mixed_precision_policy,
-                        device_id=self.args.device,
-                        **kwargs,
-                    )
-            else:
+        elif self.fsdp is not None and self.args.fsdp_config["xla"]:
             try:
                 from torch_xla.distributed.fsdp import XlaFullyShardedDataParallel as FSDP
                 from torch_xla.distributed.fsdp import checkpoint_module
@@ -1796,18 +1746,27 @@ class Trainer:
         if is_sagemaker_mp_enabled() and resume_from_checkpoint is not None:
             self._load_from_checkpoint(resume_from_checkpoint, model)
 
-        # for the rest of this function `model` is the outside model, whether it was wrapped or not
-        if model is not self.model:
-            self.model_wrapped = model
+        # as the model is wrapped, don't use `accelerator.prepare`
+        # this is for unhandled cases such as
+        # Fairscale Sharded DDP, FSDP-XLA, SageMaker MP/DP, DataParallel, IPEX
+        use_accelerator_prepare = True if model is self.model else False
 
         if delay_optimizer_creation:
             self.create_optimizer_and_scheduler(num_training_steps=max_steps)
 
         # prepare using `accelerator` prepare
+        if use_accelerator_prepare:
             model, self.optimizer, self.lr_scheduler = self.accelerator.prepare(
                 self.model, self.optimizer, self.lr_scheduler
             )
 
+        if getattr(self.accelerator.state, "fsdp_plugin", None) is not None:
+            self.model = model
+
+        # for the rest of this function `model` is the outside model, whether it was wrapped or not
+        if model is not self.model:
+            self.model_wrapped = model
+
         # Check if saved optimizer or scheduler states exist
         self._load_optimizer_and_scheduler(resume_from_checkpoint)
 
@@ -2894,7 +2853,11 @@ class Trainer:
             ShardedDDPOption.ZERO_DP_2 in self.args.sharded_ddp
             or ShardedDDPOption.ZERO_DP_3 in self.args.sharded_ddp
             or self.fsdp is not None
+            or getattr(self.accelerator.state, "fsdp_plugin", None) is not None
         ):
+            if getattr(self.accelerator.state, "fsdp_plugin", None) is not None:
+                self.accelerator.state.fsdp_plugin.save_model(self.accelerator, self.model, output_dir)
+            else:
                 state_dict = self.model.state_dict()
 
                 if self.args.should_save:

src/transformers/training_args.py

@@ -442,7 +442,7 @@ class TrainingArguments:
                     - `"backward_pre"` : Prefetches the next set of parameters before the current set of parameter's
                       gradient
                         computation.
-                    - `"backward_pos"` : This prefetches the next set of parameters after the current set of
+                    - `"backward_post"` : This prefetches the next set of parameters after the current set of
                       parameter’s
                         gradient computation.
                 - fsdp_forward_prefetch (`bool`, *optional*, defaults to `False`)
@@ -1504,6 +1504,32 @@ class TrainingArguments:
             if self.fsdp_config["xla_fsdp_grad_ckpt"]:
                 warnings.warn("`--xla_fsdp_grad_ckpt` is useful only when `--xla` is set to true.")
 
+        # accelerate integration for FSDP
+        if len(self.fsdp) > 0 and not self.fsdp_config["xla"]:
+            os.environ["ACCELERATE_USE_FSDP"] = "true"
+            from accelerate.utils.constants import (
+                FSDP_AUTO_WRAP_POLICY,
+                FSDP_SHARDING_STRATEGY,
+            )
+
+            for fsdp_option in self.fsdp:
+                if fsdp_option.upper() in FSDP_SHARDING_STRATEGY:
+                    # set environment variable for FSDP sharding strategy
+                    os.environ["FSDP_SHARDING_STRATEGY"] = str(FSDP_SHARDING_STRATEGY.index(fsdp_option.upper()) + 1)
+                elif fsdp_option == FSDPOption.OFFLOAD:
+                    os.environ["FSDP_OFFLOAD_PARAMS"] = "true"
+                elif fsdp_option == FSDPOption.AUTO_WRAP:
+                    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()
+
         if self.tpu_metrics_debug:
             warnings.warn(
                 "using `--tpu_metrics_debug` is deprecated and will be removed in version 5 of 🤗 Transformers. Use"