[fix] don't import ProcessGroup eagerly (#1074)

* [fix] don't import ProcessGroup eagerly

- move the import into typing since it is only used for type checking
- fixes #1057

* more fixes

* one more

* tested at least
Co-authored-by: Min Xu <min.xu.public@gmail.com>

fairscale/internal/parallel.py

@@ -7,13 +7,16 @@
 
 from enum import Enum
 import sys
-from typing import List, Optional, Sequence
+from typing import TYPE_CHECKING, List, Optional, Sequence
 
 import torch
 import torch.distributed as dist
-from torch.distributed import ProcessGroup
 import torch.nn.functional as F
 
+if TYPE_CHECKING:
+    # See comments in FSDP code for reason of this import.
+    from torch.distributed import ProcessGroup
+
 
 def chunk_and_pad(tensor: torch.Tensor, num_chunks: int) -> List[torch.Tensor]:
     """Chunk a given Tensor into num_chunks parts and add any necessary padding."""
@@ -27,7 +30,7 @@ def chunk_and_pad(tensor: torch.Tensor, num_chunks: int) -> List[torch.Tensor]:
     return chunks
 
 
-def validate_process_group(device: torch.device, process_group: ProcessGroup) -> None:
+def validate_process_group(device: torch.device, process_group: "ProcessGroup") -> None:
     """Do a quick test in case user called FSDP without calling torch.cuda.set_device()
     correctly. This can easily happen in cpu_offload case where the model resides on
     the CPU.
@@ -67,7 +70,7 @@ class ProcessGroupName(str, Enum):
 
 def get_process_group_cached(
     name: ProcessGroupName = ProcessGroupName.default, ranks: Optional[Sequence[int]] = None
-) -> ProcessGroup:
+) -> "ProcessGroup":
     """
     Singleton PyTorch distributed group cache. Inspired by the code from fairseq.
 

fairscale/internal/reduce_scatter_bucketer.py

@@ -5,12 +5,14 @@
 
 import functools
 import os
-from typing import Callable, Dict, List, Optional, Tuple
+from typing import TYPE_CHECKING, Callable, Dict, List, Optional, Tuple
 
 import torch
 from torch import Tensor
 import torch.distributed as dist
-from torch.distributed import ProcessGroup
+
+if TYPE_CHECKING:
+    from torch.distributed import ProcessGroup
 
 # TODO: Remove the toggle-enable_nccl_base_collectives when github open issue #801 is resolved.
 if os.getenv("ENABLE_NCCL_BASE_COLLECTIVES", "1") == "0":
@@ -20,7 +22,7 @@ else:
 
 
 class Bucket:
-    def __init__(self, data: Tensor, group: ProcessGroup):
+    def __init__(self, data: Tensor, group: "ProcessGroup"):
         self.data = data
         self.group = group
         self.offset = 0
@@ -99,13 +101,13 @@ class ReduceScatterBucketer:
 
     def __init__(self, bucket_cap_mb: int = 25):
         self.bucket_cap_mb = bucket_cap_mb
-        self.buckets: Dict[Tuple[torch.dtype, torch.device, ProcessGroup], Bucket] = {}
+        self.buckets: Dict[Tuple[torch.dtype, torch.device, "ProcessGroup"], Bucket] = {}
 
     @torch.no_grad()
     def reduce_scatter_async(
         self,
         input_list: List[Tensor],
-        group: ProcessGroup,
+        group: "ProcessGroup",
         callback_fn: Optional[Callable] = None,
     ) -> None:
         """
@@ -186,7 +188,7 @@ class ReduceScatterBucketer:
         bucket_size = self.bucket_cap_mb * MB / element_size
         return int(bucket_size // num_shards)
 
-    def _get_bucket(self, tensor: Tensor, group: ProcessGroup) -> Bucket:
+    def _get_bucket(self, tensor: Tensor, group: "ProcessGroup") -> Bucket:
         # TODO (Min): the `group` used here in the key is the object hash, not the content
         #     hash. That means if FSDP instances are initialized with different process groups,
         #     even when the group members are in fact the same, we end up creating different

fairscale/nn/__init__.py

@@ -5,11 +5,18 @@
 
 from typing import List
 
+import torch.distributed as dist
+
 from .checkpoint import checkpoint_wrapper
-from .data_parallel import FullyShardedDataParallel, ShardedDataParallel
+from .data_parallel import FullyShardedDataParallel
+
+if dist.is_available():
+    # Prevent import failure if dist is not available. #1057
+    from .data_parallel import ShardedDataParallel
+    from .moe import MOELayer, Top2Gate
+    from .pipe import Pipe, PipeRPCWrapper
+
 from .misc import FlattenParamsWrapper
-from .moe import MOELayer, Top2Gate
-from .pipe import Pipe, PipeRPCWrapper
 from .wrap import auto_wrap, config_auto_wrap_policy, default_auto_wrap_policy, enable_wrap, wrap
 
 __all__: List[str] = []

fairscale/nn/data_parallel/__init__.py

@@ -5,6 +5,8 @@
 
 from typing import List
 
+import torch.distributed as dist
+
 from .fully_sharded_data_parallel import (
     FullyShardedDataParallel,
     OffloadConfig,
@@ -12,6 +14,9 @@ from .fully_sharded_data_parallel import (
     auto_wrap_bn,
     no_pre_load_state_dict_hook,
 )
-from .sharded_ddp import ShardedDataParallel
+
+if dist.is_available():
+    # Prevent import failure if dist is not available. #1057
+    from .sharded_ddp import ShardedDataParallel
 
 __all__: List[str] = []

fairscale/nn/data_parallel/fully_sharded_data_parallel.py

@@ -35,7 +35,6 @@ from typing import (
 import torch
 from torch.autograd import Variable
 import torch.distributed as dist
-from torch.distributed import ProcessGroup
 import torch.nn as nn
 import torch.nn.functional as F
 from torch.nn.parameter import Parameter
@@ -58,6 +57,12 @@ from . import fsdp_optim_utils as ou
 
 if TYPE_CHECKING:
     from collections import OrderedDict  # noqa: F401
+
+    # See #1057. On some platform, torch.distributed may not have ProcessGroup
+    # So we only import it during type checking, which is not done on default
+    # import and only done by developer (doing it on supported platforms I presume).
+    from torch.distributed import ProcessGroup
+
 # TODO: Remove the toggle here when github open issue #801 is resolved.
 if os.getenv("ENABLE_NCCL_BASE_COLLECTIVES", "1") == "0":
     enable_nccl_base_collectives = False
@@ -308,7 +313,7 @@ class FullyShardedDataParallel(nn.Module):
     def __init__(
         self,
         module: nn.Module,
-        process_group: Optional[ProcessGroup] = None,
+        process_group: Optional["ProcessGroup"] = None,
         # The type for the process_group_reduce_scatter only can be either ProcessGroup or ProcessGroupName
         process_group_reduce_scatter: Any = ProcessGroupName.reduce_scatter,
         reshard_after_forward: bool = True,
@@ -352,6 +357,9 @@ class FullyShardedDataParallel(nn.Module):
             self.process_group_reduce_scatter = get_process_group_cached(ProcessGroupName.reduce_scatter)
         else:
             # If a specific process group is passed in, the reduce_scatter will use the passed in process group.
+            # Delay the import here since this type may not be available on certain platforms.
+            from torch.distributed import ProcessGroup
+
             if isinstance(process_group_reduce_scatter, ProcessGroup):
                 self.process_group_reduce_scatter = process_group_reduce_scatter
             else:
@@ -2648,7 +2656,7 @@ def _unpad(shard: torch.Tensor, pad: int) -> torch.Tensor:
 def auto_wrap_bn(
     module: nn.Module,
     single_rank_pg: bool = False,
-    process_group: Optional[ProcessGroup] = None,
+    process_group: Optional["ProcessGroup"] = None,
     fsdp_config: Optional[Dict[str, Any]] = None,
     wrap_it: bool = True,
     assert_on_collision: bool = True,