Adds timeout argument to training_args to avoid socket timeouts in DDP (

huggingface#18562)

* chore(training_args): Adds support for timeout argument.

* fix(training_args): Passes make style through changes.

* fix(training_args): Removes wrong docstring sentence.

* fix(training_args): Fixes timeout not being JSON serializable.

* fix(training_args_sm): Also updates timeout to timeout_delta.

* fix(training_args): Fixes PR according to suggestions.

src/transformers/sagemaker/training_args_sm.py

@@ -92,7 +92,7 @@ def _setup_devices(self) -> "torch.device":
         elif is_sagemaker_dp_enabled():
             import smdistributed.dataparallel.torch.torch_smddp  # noqa: F401
 
-            torch.distributed.init_process_group(backend="smddp")
+            torch.distributed.init_process_group(backend="smddp", timeout=self.ddp_timeout_delta)
             self.local_rank = int(os.getenv("SMDATAPARALLEL_LOCAL_RANK"))
             device = torch.device("cuda", self.local_rank)
             self._n_gpu = 1
@@ -111,7 +111,7 @@ def _setup_devices(self) -> "torch.device":
             # Here, we'll use torch.distributed.
             # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
             if not torch.distributed.is_initialized():
-                torch.distributed.init_process_group(backend="nccl")
+                torch.distributed.init_process_group(backend="nccl", timeout=self.ddp_timeout_delta)
             device = torch.device("cuda", self.local_rank)
             self._n_gpu = 1
 

src/transformers/training_args.py

@@ -18,6 +18,7 @@
 import os
 import warnings
 from dataclasses import asdict, dataclass, field
+from datetime import timedelta
 from enum import Enum
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
@@ -481,6 +482,11 @@ class TrainingArguments:
             are also available. See the [Ray documentation](
             https://docs.ray.io/en/latest/tune/api_docs/analysis.html#ray.tune.ExperimentAnalysis.get_best_trial) for
             more options.
+        ddp_timeout (`int`, *optional*, defaults to 1800):
+            The timeout for `torch.distributed.init_process_group` calls, used to avoid GPU socket timeouts when
+            performing slow operations in distributed runnings. Please refer the [PyTorch documentation]
+            (https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group) for more
+            information.
         use_mps_device (`bool`, *optional*, defaults to `False`):
             Whether to use Apple Silicon chip based `mps` device.
     """
@@ -971,6 +977,12 @@ class TrainingArguments:
             )
         },
     )
+    ddp_timeout: Optional[int] = field(
+        default=1800,
+        metadata={
+            "help": "Overrides the default timeout for distributed training (value should be given in seconds)."
+        },
+    )
 
     def __post_init__(self):
         # Handle --use_env option in torch.distributed.launch (local_rank not passed as an arg then).
@@ -1291,6 +1303,13 @@ def eval_batch_size(self) -> int:
         eval_batch_size = per_device_batch_size * max(1, self.n_gpu)
         return eval_batch_size
 
+    @property
+    def ddp_timeout_delta(self) -> timedelta:
+        """
+        The actual timeout for torch.distributed.init_process_group since it expects a timedelta variable.
+        """
+        return timedelta(seconds=self.ddp_timeout)
+
     @cached_property
     @torch_required
     def _setup_devices(self) -> "torch.device":
@@ -1358,7 +1377,9 @@ def _setup_devices(self) -> "torch.device":
                         f"num_cpu_threads_per_process unset, we set it at {num_cpu_threads_per_process} to improve oob"
                         " performance."
                     )
-                torch.distributed.init_process_group(backend=self.xpu_backend, rank=rank, world_size=size)
+                torch.distributed.init_process_group(
+                    backend=self.xpu_backend, rank=rank, world_size=size, timeout=self.ddp_timeout_delta
+                )
         elif is_torch_tpu_available():
             device = xm.xla_device()
             self._n_gpu = 0
@@ -1369,7 +1390,7 @@ def _setup_devices(self) -> "torch.device":
         elif is_sagemaker_dp_enabled():
             import smdistributed.dataparallel.torch.torch_smddp  # noqa: F401
 
-            dist.init_process_group(backend="smddp")
+            dist.init_process_group(backend="smddp", timeout=self.ddp_timeout_delta)
             self.local_rank = int(os.getenv("SMDATAPARALLEL_LOCAL_RANK"))
             device = torch.device("cuda", self.local_rank)
             self._n_gpu = 1
@@ -1431,7 +1452,7 @@ def _setup_devices(self) -> "torch.device":
             # Here, we'll use torch.distributed.
             # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
             if not torch.distributed.is_initialized():
-                torch.distributed.init_process_group(backend="nccl")
+                torch.distributed.init_process_group(backend="nccl", timeout=self.ddp_timeout_delta)
             device = torch.device("cuda", self.local_rank)
             self._n_gpu = 1