t # This is a combination of 2 commits.

allreduce overlap with wgrad gemm

change custom delay to dummy add

megatron/arguments.py

@@ -462,6 +462,10 @@ def _add_training_args(parser):
     group.add_argument('--dataloader-type', type=str, default=None,
                        choices=['single', 'cyclic'],
                        help='Single pass vs multiple pass data loader')
+    group.add_argument('--async-tensor-parallel-allreduce', action='store_true',
+                       help='Enable asynchronous excution of tensor-parallel allreduce '
+                       'with other GPU operators',
+                       dest='async_tensor_parallel_allreduce')
     return parser
 
 

megatron/initialize.py

@@ -176,11 +176,22 @@ def _initialize_distributed():
             else:
                 args.local_rank = device
             torch.cuda.set_device(device)
+        # Increase cuda stream priority of NCCL ops when overlapping with other ops
+        if (args.async_tensor_parallel_allreduce and 
+                args.tensor_model_parallel_size > 1):
+            from torch._C._distributed_c10d import ProcessGroupNCCL
+
+            pg_options = ProcessGroupNCCL.Options()
+            pg_options.is_high_priority_stream = True
+            pg_options._timeout = timedelta(days=7)
+        else:
+            pg_options = None
         # Call the init process
         torch.distributed.init_process_group(
             backend=args.distributed_backend,
             world_size=args.world_size, rank=args.rank,
-            timeout=timedelta(days=7))
+            timeout=timedelta(days=7),
+            pg_options=pg_options)
 
     # Set the tensor model-parallel, pipeline model-parallel, and
     # data-parallel communicators.

megatron/mpu/layers.py

@@ -27,6 +27,7 @@ from torch.nn.parameter import Parameter
 
 from .initialize import get_tensor_model_parallel_rank
 from .initialize import get_tensor_model_parallel_world_size
+from .initialize import get_tensor_model_parallel_group
 from .mappings import copy_to_tensor_model_parallel_region
 from .mappings import gather_from_tensor_model_parallel_region
 from .mappings import reduce_from_tensor_model_parallel_region
@@ -198,6 +199,37 @@ class VocabParallelEmbedding(torch.nn.Module):
         return output
 
 
+class ColumnParallelLinearFunction(torch.autograd.Function):
+    """
+    Column-parallel linear layer execution with asynchronous all-reduce
+    execution in backprop.
+    """
+    @staticmethod
+    def forward(ctx, input, weight, bias, use_bias):
+        ctx.save_for_backward(input, weight)
+        ctx.use_bias = use_bias
+        output = torch.matmul(input, weight.t())
+        if use_bias:
+            output = output + bias
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, weight = ctx.saved_tensors
+        use_bias = ctx.use_bias
+        grad_input = grad_output.matmul(weight)
+        # Asyncronous all-reduce
+        handle = torch.distributed.all_reduce(
+                grad_input, group=get_tensor_model_parallel_group(), async_op=True)
+        # Delay the start of weight gradient computation shortly (3us) to have
+        # all-reduce scheduled first and have GPU resources allocated
+        _ = torch.empty(1, device=grad_output.device) + 1
+        grad_weight = grad_output.t().matmul(input)
+        grad_bias = grad_output.sum(dim=0) if use_bias else None
+        handle.wait()
+        return grad_input, grad_weight, grad_bias, None
+
+
 class ColumnParallelLinear(torch.nn.Module):
     """Linear layer with column parallelism.
 
@@ -272,16 +304,29 @@ class ColumnParallelLinear(torch.nn.Module):
                 self.bias.zero_()
         else:
             self.register_parameter('bias', None)
+        self.async_tensor_parallel_allreduce = (args.async_tensor_parallel_allreduce
+                                                and world_size > 1)
 
 
 
     def forward(self, input_):
-        # Set up backprop all-reduce.
-        input_parallel = copy_to_tensor_model_parallel_region(input_)
-        # Matrix multiply.
-
         bias = self.bias if not self.skip_bias_add else None
-        output_parallel = F.linear(input_parallel, self.weight, bias)
+
+        if self.async_tensor_parallel_allreduce:
+            input_shape = input_.shape
+            input_ = input_.view(input_shape[0] * input_shape[1],input_shape[2])
+            # Maxtrix multiply with asynchronouse tensor-parallel all-reduce execution
+            output_parallel = ColumnParallelLinearFunction.apply(
+                    input_, self.weight, bias, bias is not None)
+            output_parallel = output_parallel.view(
+                    input_shape[0], input_shape[1], output_parallel.shape[1])
+        else:
+            # Set up backprop all-reduce.
+            input_parallel = copy_to_tensor_model_parallel_region(input_)
+
+            # Matrix multiply.
+            output_parallel = F.linear(input_parallel, self.weight, bias)
+
         if self.gather_output:
             # All-gather across the partitions.
             output = gather_from_tensor_model_parallel_region(output_parallel)