[ShardedDDP] Sync buffers + small cleanup (#112)

- adding the buffer broadcast option
- minor cleanup in shardedDDP

benchmarks/oss.py

@@ -62,7 +62,7 @@ def train(
 ):
     assert not use_sdp or (use_sdp and use_oss), "ShardedDataParallel requires OSS"
     # DDP
-    dist_init(rank, world_size, backend)
+    dist_init(rank=rank, world_size=world_size, backend=backend)
 
     # Setup
     torch.cuda.set_device(rank)
@@ -81,7 +81,11 @@ def train(
 
     if use_sdp:
         ddp = ShardedDataParallel(
-            module=model, optimizer=OPTIM, optimizer_params={"lr": 1e-4, "momentum": 0.9}, world_size=world_size,
+            module=model,
+            optimizer=OPTIM,
+            optimizer_params={"lr": 1e-4, "momentum": 0.9},
+            world_size=world_size,
+            broadcast_buffers=False,
         )
         ddp.train()
         optimizer = ddp.optimizer

fairscale/nn/data_parallel/sharded_ddp.py

@@ -25,7 +25,7 @@ class ShardedDataParallel(nn.Module):
     """Implements distributed data parallel training with optimizer state sharding.
 
     A simplified version of :class:`torch.nn.parallel.DistributedDataParallel`.
-    This version uses a c10d process group for communication and does not
+    This version uses a c10d process group for communication and optionally
     broadcast buffers.
 
     Args:
@@ -33,6 +33,8 @@ class ShardedDataParallel(nn.Module):
         optimizer (~torch.optim.Optimizer): optimizer to be used for training
         optimizer_params(Dict): extra parameters for the optimizer
         world_size (int): number of parallel workers
+        broadcast_buffers (bool): flag that enables syncing (broadcasting) buffers of
+        the module at beginning of the forward function. (default: ``True``)
         process_group (optional): the c10d process group to be used for
             distributed gradient reduction. If None, the default WORLD process group
             will be used.
@@ -47,6 +49,7 @@ class ShardedDataParallel(nn.Module):
         optimizer: Type[torch.optim.Optimizer],
         optimizer_params: Dict[str, Any],
         world_size: int,
+        broadcast_buffers: bool,
         process_group: Any = None,
         buffer_size: int = 2 ** 28,
     ):
@@ -56,6 +59,8 @@ class ShardedDataParallel(nn.Module):
         self.world_size = world_size
         self.process_group = process_group if process_group is not None else dist.group.WORLD
         self.rank = dist.get_rank(self.process_group)
+        self.broadcast_buffers = broadcast_buffers
+        self.authoritative_rank = 0
 
         # Never use a bigger buffer than the number of model params
         self.buffer_size = min(buffer_size, sum(p.numel() for p in self.module.parameters()))
@@ -71,7 +76,7 @@ class ShardedDataParallel(nn.Module):
         # Build the sharded optimizer
         self.sharded_optimizer = OSS(self.module.parameters(), optim=optimizer, group=process_group, **optimizer_params)
 
-        # sanity checks
+        # Sanity checks
         assert len(self.sharded_optimizer.param_to_rank) == len(
             list(self.module.parameters())
         ), "number of params do not match"
@@ -109,6 +114,9 @@ class ShardedDataParallel(nn.Module):
                 raise RuntimeError("OssDdp requires explicit reduction, must call OssDdp.reduce")
             if not self.accumulate_grads:
                 self.need_reduction = True
+            if self.broadcast_buffers and len(list(self.module.buffers())) > 0:
+                self._sync_buffers()
+
         return self.module(*inputs, **kwargs)
 
     def reduce(self) -> None:
@@ -118,52 +126,35 @@ class ShardedDataParallel(nn.Module):
         """
         assert self.module.training, "Cannot call reduce in eval"
 
-        def reduce_params(params: List[Parameter], params_rank: int) -> None:
-            """ Helper to reduce a list of params that should fix in the buffer. """
+        def reduce_grads(params: List[Parameter], params_rank: int) -> None:
+            """ Helper to reduce a list of params that should fit in the buffer.
+            NOTE: All param gradients are assumed to exist"""
             assert self.buffer is not None
+
+            # Fill in the packed IO buffer
             buffer: Tensor = cast(Tensor, self.buffer)
-            nonzero_buffer = False
             if len(params) > 1:
                 offset = 0
                 for p in params:
                     sz = p.numel()
-                    if p.grad is not None:
-                        # The type error could have been fixed in later
-                        # version of pytorch. Same elsewhere.
                     buffer[offset : offset + sz].copy_(p.grad.data.view(-1))  # type: ignore
-                        nonzero_buffer = True
-                    else:
-                        buffer[offset : offset + sz].zero_()
                     offset += sz
             else:
                 # we only have a single grad to reduce
-                p = params[0]
-                if p.grad is not None:
-                    buffer = p.grad.data
-                    nonzero_buffer = True
-                elif p.numel() <= self.buffer.numel():
-                    buffer = buffer[: p.numel()]
-                    buffer.zero_()
-                else:
-                    buffer = torch.zeros_like(p)
+                buffer = params[0].grad.data  # type: ignore
 
-            if nonzero_buffer:
+            # Reduce
             buffer.div_(self.world_size)  # type: ignore
+            dist.reduce(tensor=buffer, dst=params_rank, group=self.process_group)  # type: ignore
 
-            dist.reduce(buffer, params_rank, group=self.process_group)  # type: ignore
-
+            # Copy reduced grads back into their original place, or free corresponding memory
             if params_rank == self.rank:
-                # copy reduced grads back into their original place
                 offset = 0
                 for p in params:
                     sz = p.numel()
-                    if p.grad is not None:
                     p.grad.data.copy_(buffer[offset : offset + sz].view_as(p))  # type: ignore
-                    else:
-                        p.grad = buffer[offset : offset + sz].view_as(p).clone()
                     offset += sz
             else:
-                # wipe the grads
                 for p in params:
                     p.grad = None
 
@@ -195,7 +186,7 @@ class ShardedDataParallel(nn.Module):
                     if sz > self.buffer.numel():
                         # reduce big params directly
                         assert param_rank is not None
-                        reduce_params([param], cast(int, param_rank))
+                        reduce_grads([param], cast(int, param_rank))
                     else:
                         # smaller params are packed together from the same device
                         # and same rank.
@@ -203,7 +194,7 @@ class ShardedDataParallel(nn.Module):
                             last_param_rank is not None and last_param_rank != param_rank
                         ):
                             assert last_param_rank is not None
-                            reduce_params(buffered_params, cast(int, last_param_rank))
+                            reduce_grads(buffered_params, cast(int, last_param_rank))
                             offset = 0
                             buffered_params.clear()
                         buffered_params.append(cast(Parameter, param))
@@ -211,6 +202,21 @@ class ShardedDataParallel(nn.Module):
 
                 if len(buffered_params) > 0:
                     assert param_rank is not None
-                    reduce_params(buffered_params, cast(int, param_rank))
+                    reduce_grads(buffered_params, cast(int, param_rank))
 
         reduction_fn()
+
+    def _sync_buffers(self) -> None:
+        """
+        Sync all the param buffers in between ranks.
+        TODO: Could be worth bucketing ?
+        """
+        _ = list(
+            map(
+                lambda x: x.wait(),
+                map(
+                    lambda x: dist.broadcast(x, self.authoritative_rank, self.process_group, async_op=True),
+                    self.module.buffers(),
+                ),
+            )
+        )

tests/nn/data_parallel/test_sharded_ddp.py

@@ -37,12 +37,20 @@ def run_one_step(rank, world_size, backend, device, temp_file_name):
     if device == torch.device("cuda"):
         torch.cuda.set_device(rank)
 
+    # Any model works. Add one different buffer per rank
     model = Sequential(Linear(2, 3), Linear(3, 4)).to(device)
+    model.register_buffer("test_buffer", torch.ones((1)) * rank)
+    model.to(device)
 
     ddp = ShardedDataParallel(
-        module=model, optimizer=torch.optim.SGD, optimizer_params={"lr": 0.1, "momentum": 0.99}, world_size=world_size
+        module=model,
+        optimizer=torch.optim.SGD,
+        optimizer_params={"lr": 0.01, "momentum": 0.99},
+        world_size=world_size,
+        broadcast_buffers=True,
     )
     optimizer = ddp.optimizer
+    model = ddp.module
 
     input_tensor = torch.rand((64, 2)).to(device)
     output = ddp(input_tensor).abs().sum() / input_tensor.numel()
@@ -58,10 +66,9 @@ def run_one_step(rank, world_size, backend, device, temp_file_name):
             if param.requires_grad:
                 assert param.grad.abs().sum().item() > 0.0, "The reduce step should have populated all the gradients"
 
-    # Check that the optimization process makes sense (ie. loss goes down for the same data)
-    optimizer.step()
-    new_eval = ddp(input_tensor).abs().sum() / input_tensor.numel()
-    # assert new_eval.item() < output.item()
+    # Check that all the buffers are in sync (authoritative rank is 0, its buffer is 0)
+    for b in model.buffers():
+        assert b.cpu().item() == 0.0
 
 
 def run_test(backend, device, world_size=2):
@@ -76,7 +83,7 @@ def run_eval_mode(_unused):
     )
     model = Sequential(Linear(2, 3), Linear(3, 4))
     optimizer_params = {"lr": 0.1, "momentum": 0.99}
-    ddp = ShardedDataParallel(model, torch.optim.SGD, optimizer_params, 1)
+    ddp = ShardedDataParallel(model, torch.optim.SGD, optimizer_params, 1, broadcast_buffers=False)
     optimizer = ddp.optimizer
 
     ddp.eval()