torch distributed: add support for user-specified parameter synchroni…

…zation

returnn/torch/distributed.py

@@ -3,20 +3,72 @@
 """
 
 from __future__ import annotations
-from typing import Optional, Any, Dict
+from abc import abstractmethod, ABC
+import logging
+import numpy
 import os
 import socket
-import logging
+from typing import Callable, Optional, Any, Dict, Type
 
 import torch
 from torch.nn.parallel import DistributedDataParallel
 
-from returnn.config import Config
-from returnn.util.basic import CollectionReadCheckCovered
+from returnn.util.basic import CollectionReadCheckCovered, OptionalNotImplementedError
 
 _logger = logging.getLogger("returnn.torch.distributed")
 
 
+class ParamSynchronizer(ABC):
+    """
+    Custom parameter synchronization primitive.
+
+    Contains a callback that is called after every train step to synchronize model parameters
+    across processes/nodes.
+    """
+
+    @abstractmethod
+    def __init__(self, *, rank: int, size: int, local_rank: int, local_size: int, **kwargs):
+        """
+        `__init__` called after the default global process group is created.
+        Can be used to initialize any additional custom process (sub)groups.
+
+        Note this function is passed a randomly named kwarg on every invocation to ensure forwards compatbility.
+
+        :param rank: global rank of the current process across all nodes
+        :param size: global world size across all nodes
+        :param local_rank: local rank of the current process on the current node
+        :param local_rank: local world size on the current node
+        """
+        super().__init__()
+
+    def make_distributed_model(self, *, module: torch.nn.Module, **kwargs) -> DistributedDataParallel:
+        """
+        Creates an associated `DistributedDataParallel` for the given module for gradient synchronization.
+
+        This function can be left unimplemented if no gradient synchronization is done.
+
+        Note this function is passed a randomly named kwarg on every invocation to ensure forwards compatbility.
+        """
+        raise OptionalNotImplementedError
+
+    @abstractmethod
+    def step(self, *, module: torch.nn.Module, train_step_idx: int, **kwargs):
+        """
+        Parameter synchronization callback called after every train step.
+
+        Note this function is passed a randomly named kwarg on every invocation to ensure forwards compatbility.
+
+        :param module: the NN being trained
+        :param train_step_idx: the current train step
+        :param kwargs: any additional kwargs.
+        """
+        raise NotImplementedError
+
+    def __call__(self, *args, **kwargs):
+        """forwards to :func:``step``"""
+        return self.step(*args, **kwargs)
+
+
 class DistributedContext:
     """
     This class setups some helper functions for torch distributed training
@@ -26,6 +78,9 @@ def __init__(self, options: Dict[str, Any]):
         import torch.distributed as dist
 
         self._opts = CollectionReadCheckCovered(options)
+        # Only used to generate forwards compatibility ensuring random kwargs, therefore
+        # the seed is not important
+        self._rng = numpy.random.default_rng()
 
         # when no backend is specified, both gloo and nccl backends will be created
         # the gloo backend will be used for collectives with CPU tensors and
@@ -42,8 +97,11 @@ def __init__(self, options: Dict[str, Any]):
             % (socket.gethostname(), os.getpid(), self._rank, self._size, self._local_rank, self._local_size)
         )
 
+        self._custom_sync_class: Optional[Type[ParamSynchronizer]] = self._opts.get("synchronizer", None)
+        self._custom_sync: Optional[Callable] = None
         self._reduce_type = self._opts.get("reduce_type", "grad")
         self._param_sync_step: Optional[int] = self._opts.get("param_sync_step", None)
+
         if self._reduce_type == "param":
             assert isinstance(self._param_sync_step, int) and self._param_sync_step > 0, (
                 f"reduce_type param: param_sync_step must be a positive int,"
@@ -52,6 +110,23 @@ def __init__(self, options: Dict[str, Any]):
             _logger.info(f"reduce_type param: param_sync_step {self._param_sync_step}")
         elif self._reduce_type == "grad":
             _logger.info("reduce_type grad")
+        elif self._reduce_type == "custom":
+            if issubclass(self._custom_sync_class, ParamSynchronizer):
+                self._custom_sync = self._custom_sync_class(
+                    rank=self._rank,
+                    size=self._size,
+                    local_rank=self._local_rank,
+                    local_size=self._local_size,
+                    **{f"fwd_compatible_random_kwarg_{self._rng.integers(0, 100)}": None},
+                )
+            elif isinstance(self._custom_sync_class, Callable):
+                self._custom_sync = self._custom_sync_class
+            else:
+                raise ValueError(
+                    f"synchronizer must either be a callable or a class inheriting from {ParamSynchronizer.__name__}"
+                )
+
+            _logger.info(f"reduce_type custom: {type(self._custom_sync)}")
         else:
             raise ValueError(f"invalid reduce_type {self._reduce_type!r}")
 
@@ -70,6 +145,8 @@ def _check_no_unknown_opts(self):
             self._opts.get("options")
         if self._reduce_type == "param":
             self._opts.get("sync_on_cpu")
+        if self._reduce_type == "custom":
+            self._opts.get("synchronizer")
 
         self._opts.assert_all_read()
 
@@ -102,7 +179,24 @@ def maybe_make_distributed_module(self, module: torch.nn.Module) -> Optional[Dis
         """
         if self._reduce_type == "param":
             return None
-        assert self._reduce_type == "grad"
+        assert self._reduce_type in ["custom", "grad"]
+
+        if self._reduce_type == "custom":
+            assert isinstance(self._custom_sync, (ParamSynchronizer, Callable))
+
+            if isinstance(self._custom_sync, ParamSynchronizer):
+                try:
+                    return self._custom_sync.make_distributed_model(
+                        module=module, **{f"fwd_compatible_random_kwarg_{self._rng.integers(0, 100)}": None}
+                    )
+                except OptionalNotImplementedError:
+                    pass
+            else:
+                # callable short form does not have support for DistributedDataParallel
+                pass
+
+            return None
+
         cls = self._opts.get("class", DistributedDataParallel)
         if cls is not DistributedDataParallel:
             _logger.warning(f"Using custom class {cls} instead of DistributedDataParallel, might be unsupported.")
@@ -115,7 +209,14 @@ def maybe_make_distributed_module(self, module: torch.nn.Module) -> Optional[Dis
 
     def step_after_param_update(self, *, module: torch.nn.Module, epoch_step_idx: int):
         """one train step"""
-        if self._reduce_type == "param" and ((epoch_step_idx % self._param_sync_step) == (self._param_sync_step - 1)):
+        if self._reduce_type == "custom":
+            with torch.no_grad():  # TODO: do we want this for all syncers?
+                self._custom_sync(
+                    module=module,
+                    train_step_idx=epoch_step_idx,
+                    **{f"fwd_compatible_random_kwarg_{self._rng.integers(0, 100)}": None},
+                )
+        elif self._reduce_type == "param" and ((epoch_step_idx % self._param_sync_step) == (self._param_sync_step - 1)):
             _sync_params_avg(module=module, sync_on_cpu=self._opts.get("sync_on_cpu", False))
 
 
@@ -155,7 +256,7 @@ def _sync_params_avg(*, module: torch.nn.Module, sync_on_cpu: bool = False):
 
     if sync_on_cpu:
         for param in module.parameters():
-            # Separately move each param to CPU (instead of the whole module), to safe CPU memory.
+            # Separately move each param to CPU (instead of the whole module), to save CPU memory.
             param_cpu = param.to(torch.device("cpu"))
             # On CPU, we are likely using Gloo, and Gloo does not support AVG
             dist.all_reduce(param_cpu.data, op=dist.ReduceOp.SUM)
@@ -166,12 +267,11 @@ def _sync_params_avg(*, module: torch.nn.Module, sync_on_cpu: bool = False):
     if dist.get_backend() == "gloo":
         # Gloo does not support AVG
         reduce_op = dist.ReduceOp.SUM
+    elif hasattr(dist.ReduceOp, "AVG"):
+        reduce_op = dist.ReduceOp.AVG
     else:
-        if hasattr(dist.ReduceOp, "AVG"):
-            reduce_op = dist.ReduceOp.AVG
-        else:
-            # Older PyTorch versions do not have ReduceOp.AVG.
-            reduce_op = dist.ReduceOp.SUM
+        # Older PyTorch versions do not have ReduceOp.AVG.
+        reduce_op = dist.ReduceOp.SUM
 
     for param in module.parameters():
         dist.all_reduce(param.data, op=reduce_op)