add util for ram efficient loading of model when using fsdp

docs/source/en/internal/trainer_utils.md

@@ -32,6 +32,8 @@ Most of those are only useful if you are studying the code of the Trainer in the
 
 [[autodoc]] torch_distributed_zero_first
 
+[[autodoc]] load_pretrained_model_only_on_rank0
+
 ## Callbacks internals
 
 [[autodoc]] trainer_callback.CallbackHandler

src/transformers/__init__.py

@@ -3059,7 +3059,10 @@
     _import_structure["sagemaker"] = []
     _import_structure["time_series_utils"] = []
     _import_structure["trainer"] = ["Trainer"]
-    _import_structure["trainer_pt_utils"] = ["torch_distributed_zero_first"]
+    _import_structure["trainer_pt_utils"] = [
+        "load_pretrained_model_only_on_rank0",
+        "torch_distributed_zero_first",
+    ]
     _import_structure["trainer_seq2seq"] = ["Seq2SeqTrainer"]
 
 # TensorFlow-backed objects
@@ -6598,7 +6601,7 @@
 
         # Trainer
         from .trainer import Trainer
-        from .trainer_pt_utils import torch_distributed_zero_first
+        from .trainer_pt_utils import load_pretrained_model_only_on_rank0, torch_distributed_zero_first
         from .trainer_seq2seq import Seq2SeqTrainer
 
     # TensorFlow

src/transformers/modeling_utils.py

@@ -105,6 +105,14 @@
 _init_weights = True
 
 
+def is_fsdp_enabled():
+    return os.environ["ACCELERATE_USE_FSDP"]
+
+
+def is_fsdp_enabled_and_dist_rank_0():
+    return is_fsdp_enabled() and torch.distributed.is_initialized() and torch.distributed.get_rank() == 0
+
+
 if is_sagemaker_mp_enabled():
     import smdistributed.modelparallel.torch as smp
     from smdistributed.modelparallel import __version__ as SMP_VERSION
@@ -457,7 +465,11 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
             )
         return safe_load_file(checkpoint_file)
     try:
-        if is_deepspeed_zero3_enabled() and torch.distributed.is_initialized() and torch.distributed.get_rank() > 0:
+        if (
+            (is_deepspeed_zero3_enabled() or is_fsdp_enabled())
+            and torch.distributed.is_initialized()
+            and torch.distributed.get_rank() > 0
+        ):
             map_location = "meta"
         else:
             map_location = "cpu"
@@ -541,7 +553,7 @@ def load(module: nn.Module, state_dict, prefix=""):
                     with deepspeed.zero.GatheredParameters(params_to_gather, modifier_rank=0):
                         if torch.distributed.get_rank() == 0:
                             module._load_from_state_dict(*args)
-            else:
+            elif not is_fsdp_enabled() or is_fsdp_enabled_and_dist_rank_0():
                 module._load_from_state_dict(*args)
 
         for name, child in module._modules.items():
@@ -1481,7 +1493,7 @@ def _get_resized_embeddings(
             with deepspeed.zero.GatheredParameters(old_embeddings.weight, modifier_rank=0):
                 if torch.distributed.get_rank() == 0:
                     new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
-        else:
+        elif not is_fsdp_enabled() or is_fsdp_enabled_and_dist_rank_0():
             new_embeddings.weight.data[:n, :] = old_embeddings.weight.data[:n, :]
 
         return new_embeddings
@@ -1565,7 +1577,7 @@ def _get_resized_lm_head(
                     # Copy bias weights to new lm head
                     if has_new_lm_head_bias:
                         new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
-        else:
+        elif not is_fsdp_enabled() or is_fsdp_enabled_and_dist_rank_0():
             # Copy old lm head weights to new lm head
             if not transposed:
                 new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]
@@ -2193,6 +2205,9 @@ def from_pretrained(
         commit_hash = kwargs.pop("_commit_hash", None)
         variant = kwargs.pop("variant", None)
 
+        if is_fsdp_enabled():
+            low_cpu_mem_usage = True
+
         if use_auth_token is not None:
             warnings.warn(
                 "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers.", FutureWarning
@@ -3265,23 +3280,24 @@ def _find_mismatched_keys(
                 )
 
                 if low_cpu_mem_usage:
-                    new_error_msgs, offload_index, state_dict_index = _load_state_dict_into_meta_model(
-                        model_to_load,
-                        state_dict,
-                        loaded_keys,
-                        start_prefix,
-                        expected_keys,
-                        device_map=device_map,
-                        offload_folder=offload_folder,
-                        offload_index=offload_index,
-                        state_dict_folder=state_dict_folder,
-                        state_dict_index=state_dict_index,
-                        dtype=dtype,
-                        is_quantized=is_quantized,
-                        is_safetensors=is_safetensors,
-                        keep_in_fp32_modules=keep_in_fp32_modules,
-                    )
-                    error_msgs += new_error_msgs
+                    if not is_fsdp_enabled() or is_fsdp_enabled_and_dist_rank_0():
+                        new_error_msgs, offload_index, state_dict_index = _load_state_dict_into_meta_model(
+                            model_to_load,
+                            state_dict,
+                            loaded_keys,
+                            start_prefix,
+                            expected_keys,
+                            device_map=device_map,
+                            offload_folder=offload_folder,
+                            offload_index=offload_index,
+                            state_dict_folder=state_dict_folder,
+                            state_dict_index=state_dict_index,
+                            dtype=dtype,
+                            is_quantized=is_quantized,
+                            is_safetensors=is_safetensors,
+                            keep_in_fp32_modules=keep_in_fp32_modules,
+                        )
+                        error_msgs += new_error_msgs
                 else:
                     error_msgs += _load_state_dict_into_model(model_to_load, state_dict, start_prefix)
 

src/transformers/trainer.py

@@ -461,10 +461,6 @@ def __init__(
             ):
                 self.backward_prefetch = BackwardPrefetch.BACKWARD_POST
 
-            self.forward_prefetch = False
-            if self.args.fsdp_config.get("forward_prefect", False):
-                self.forward_prefetch = True
-
             self.limit_all_gathers = False
             if self.args.fsdp_config.get("limit_all_gathers", False):
                 self.limit_all_gathers = True
@@ -1379,12 +1375,12 @@ def _wrap_model(self, model, training=True, dataloader=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
+                "transformer_layer_cls_to_wrap", default_transformer_cls_names_to_wrap
             )
 
-            if self.args.fsdp_config["fsdp_min_num_params"] > 0:
+            if self.args.fsdp_config["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"]
+                    size_based_auto_wrap_policy, min_num_params=self.args.fsdp_config["min_num_params"]
                 )
             elif fsdp_transformer_layer_cls_to_wrap is not None:
                 transformer_cls_to_wrap = set()
@@ -3825,7 +3821,6 @@ def create_accelerator_and_postprocess(self):
             fsdp_plugin.limit_all_gathers = self.args.fsdp_config.get(
                 "limit_all_gathers", fsdp_plugin.limit_all_gathers
             )
-            fsdp_plugin.use_orig_params = self.args.fsdp_config.get("use_orig_params", fsdp_plugin.use_orig_params)
 
         if self.is_deepspeed_enabled:
             if getattr(self.args, "hf_deepspeed_config", None) is None:

src/transformers/trainer_pt_utils.py

@@ -1125,3 +1125,16 @@ def smp_nested_concat(tensor):
         # It doesn't seem possible to check here if `tensor` is a StepOutput because StepOutput lives in `smp.step`
         # which is also the name of the decorator so Python is confused.
         return tensor.concat().detach().cpu()
+
+
+def load_pretrained_model_only_on_rank0(model_cls, config_cls, model_name_or_path):
+    from accelerate.state import PartialState
+
+    state = PartialState()
+    if state.is_main_process:
+        model = model_cls.from_pretrained(model_name_or_path, return_dict=True)
+    else:
+        with torch.device("meta"):
+            config = config_cls.from_pretrained(model_name_or_path)
+            model = model_cls.from_config(config)
+    return model

src/transformers/training_args.py

@@ -458,10 +458,17 @@ class TrainingArguments:
                     FSDP's forward prefetch mode (useful only when `fsdp` field is passed).
                      If `"True"`, then FSDP explicitly prefetches the next upcoming all-gather while executing in the
                      forward pass.
-                - limit_all_gathers (`bool`, *optional*, defaults to `False`)
+                - fsdp_limit_all_gathers (`bool`, *optional*, defaults to `False`)
                     FSDP's limit_all_gathers (useful only when `fsdp` field is passed).
                      If `"True"`, FSDP explicitly synchronizes the CPU thread to prevent too many in-flight
                      all-gathers.
+                - fsdp_use_orig_params (`bool`, *optional*, defaults to `False`)
+                    If `"True"`, allows non-uniform `requires_grad` during init, which means support for interspersed frozen and trainable paramteres.
+                    Useful in cases such as parameter-efficient fine-tuning.
+                    Please refer this [blog](https://dev-discuss.pytorch.org/t/rethinking-pytorch-fully-sharded-data-parallel-fsdp-from-first-principles/1019
+                - fsdp_sync_module_states (`bool`, *optional*, defaults to `False`)
+                    If `"True"`, each individually wrapped FSDP unit will broadcast module parameters from rank 0
+                    to ensure they are the same across all ranks after initialization
                 - xla (`bool`, *optional*, defaults to `False`):
                     Whether to use PyTorch/XLA Fully Sharded Data Parallel Training. This is an experimental feature
                     and its API may evolve in the future.
@@ -1511,40 +1518,44 @@ def __post_init__(self):
             self.fsdp_config = {}
 
         if isinstance(self.fsdp_config, str):
+            if len(self.fsdp) == 0:
+                warnings.warn("`--fsdp_config` is useful only when `--fsdp` is specified.")
             with io.open(self.fsdp_config, "r", encoding="utf-8") as f:
                 self.fsdp_config = json.load(f)
+                for k, v in self.fsdp_config.items():
+                    if k.startswith("fsdp_"):
+                        self.fsdp_config[k.replace("fsdp", "")] = v
+                        del self.fsdp_config[k]
 
         if self.fsdp_min_num_params > 0:
             warnings.warn("using `--fsdp_min_num_params` is deprecated. Use fsdp_config instead ", FutureWarning)
 
         self.fsdp_config["fsdp_min_num_params"] = max(
-            self.fsdp_config.get("fsdp_min_num_params", 0), self.fsdp_min_num_params
+            self.fsdp_config.get("min_num_params", 0), self.fsdp_min_num_params
         )
 
-        # if fsdp_config["fsdp_transformer_layer_cls_to_wrap"] is specified as a string, convert it to a list with a single object
-        if isinstance(self.fsdp_config.get("fsdp_transformer_layer_cls_to_wrap", None), str):
-            self.fsdp_config["fsdp_transformer_layer_cls_to_wrap"] = [
-                self.fsdp_config["fsdp_transformer_layer_cls_to_wrap"]
-            ]
+        # if fsdp_config["transformer_layer_cls_to_wrap"] is specified as a string, convert it to a list with a single object
+        if isinstance(self.fsdp_config.get("transformer_layer_cls_to_wrap", None), str):
+            self.fsdp_config["transformer_layer_cls_to_wrap"] = [self.fsdp_config["transformer_layer_cls_to_wrap"]]
 
         if self.fsdp_transformer_layer_cls_to_wrap is not None:
             warnings.warn(
                 "using `--fsdp_transformer_layer_cls_to_wrap` is deprecated. Use fsdp_config instead ", FutureWarning
             )
-            self.fsdp_config["fsdp_transformer_layer_cls_to_wrap"] = self.fsdp_config.get(
-                "fsdp_transformer_layer_cls_to_wrap", []
+            self.fsdp_config["transformer_layer_cls_to_wrap"] = self.fsdp_config.get(
+                "transformer_layer_cls_to_wrap", []
             ) + [self.fsdp_transformer_layer_cls_to_wrap]
 
-        if len(self.fsdp) == 0 and self.fsdp_config["fsdp_min_num_params"] > 0:
+        if len(self.fsdp) == 0 and self.fsdp_config["min_num_params"] > 0:
             warnings.warn("`--fsdp_min_num_params` is useful only when `--fsdp` is specified.")
 
-        if len(self.fsdp) == 0 and self.fsdp_config.get("fsdp_transformer_layer_cls_to_wrap", None) is not None:
+        if len(self.fsdp) == 0 and self.fsdp_config.get("transformer_layer_cls_to_wrap", None) is not None:
             warnings.warn("`--fsdp_transformer_layer_cls_to_wrap` is useful only when `--fsdp` is specified.")
 
         if (
             len(self.fsdp) > 0
-            and self.fsdp_config["fsdp_min_num_params"] > 0
-            and self.fsdp_config.get("fsdp_transformer_layer_cls_to_wrap", None) is not None
+            and self.fsdp_config["min_num_params"] > 0
+            and self.fsdp_config.get("transformer_layer_cls_to_wrap", None) is not None
         ):
             raise ValueError(
                 "`--fsdp_min_num_params` and `--fsdp_transformer_layer_cls_to_wrap` are mutually exclusive."
@@ -1574,23 +1585,29 @@ def __post_init__(self):
                 FSDP_SHARDING_STRATEGY,
             )
 
+            prefix = "FSDP_"
             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)
+                    os.environ[f"{prefix}SHARDING_STRATEGY"] = str(
+                        FSDP_SHARDING_STRATEGY.index(fsdp_option.upper()) + 1
+                    )
                 elif fsdp_option == FSDPOption.OFFLOAD:
-                    os.environ["FSDP_OFFLOAD_PARAMS"] = "true"
+                    os.environ[f"{prefix}OFFLOAD_PARAMS"] = "true"
                 elif fsdp_option == FSDPOption.AUTO_WRAP:
-                    os.environ["FSDP_AUTO_WRAP_POLICY"] = FSDP_AUTO_WRAP_POLICY[0]
+                    os.environ[f"{prefix}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]
+                        os.environ[f"{prefix}MIN_NUM_PARAMS"] = str(self.fsdp_config["min_num_params"])
+                        os.environ[f"{prefix}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[f"{prefix}TRANSFORMER_CLS_TO_WRAP"] = ",".join(
+                            self.fsdp_config["transformer_layer_cls_to_wrap"]
                         )
             prefetch_policy = self.fsdp_config.get("fsdp_backward_prefetch", "NO_PREFETCH")
-            os.environ["FSDP_BACKWARD_PREFETCH"] = prefetch_policy.upper()
+            os.environ[f"{prefix}BACKWARD_PREFETCH"] = prefetch_policy.upper()
+            os.environ[f"{prefix}FORWARD_PREFETCH"] = self.fsdp_config.get("forward_prefect", "false")
+            os.environ[f"{prefix}SYNC_MODULE_STATES"] = self.fsdp_config.get("sync_module_states", "false")
+            os.environ[f"{prefix}USE_ORIG_PARAMS"] = self.fsdp_config.get("use_orig_params", "false")
 
         if self.tpu_metrics_debug:
             warnings.warn(

src/transformers/utils/dummy_pt_objects.py

@@ -8493,6 +8493,10 @@ def __init__(self, *args, **kwargs):
         requires_backends(self, ["torch"])
 
 
+def load_pretrained_model_only_on_rank0(*args, **kwargs):
+    requires_backends(load_pretrained_model_only_on_rank0, ["torch"])
+
+
 def torch_distributed_zero_first(*args, **kwargs):
     requires_backends(torch_distributed_zero_first, ["torch"])