def _process_batch():
dev_grad_batch, dev_events, job_event = queue.get()
dev_coalesced = []
# Coalesce the tensors on all devices and start a local reduction
for dev_id, grad_batch, event, stream in zip(device_ids, dev_grad_batch, dev_events, reduction_streams):
with torch.cuda.device(dev_id), torch.cuda.stream(stream):
stream.wait_event(event)
coalesced = _flatten_tensors(grad_batch)
dev_coalesced.append(coalesced)
# Wait for all copies to complete before starting the NCCL kernel
for stream in reduction_streams:
stream.synchronize()
nccl.reduce(dev_coalesced, root=device_ids[0], streams=nccl_streams)
# From now on we're only going to work on the first device (from device_ids)
grad_batch = dev_grad_batch[0]
coalesced = dev_coalesced[0]
reduce_stream = reduction_streams[0]
with torch.cuda.stream(reduce_stream):
reduce_stream.wait_stream(nccl_streams[0])
coalesced /= dist.get_world_size()
dist.all_reduce(coalesced, group=group_id)
for grad, reduced in zip(grad_batch, _unflatten_tensors(coalesced, grad_batch)):
grad.copy_(reduced)
job_event.set()
def _process_batch():
dev_grad_batch, dev_events, job_event = queue.get()
dev_coalesced = []
# Coalesce the tensors on all devices and start a local reduction
for dev_id, grad_batch, event, stream in zip(
device_ids, dev_grad_batch, dev_events, reduction_streams):
with torch.cuda.device(dev_id), torch.cuda.stream(stream):
stream.wait_event(event)
coalesced = _flatten_tensors(grad_batch)
dev_coalesced.append(coalesced)
# Wait for all copies to complete before starting the NCCL kernel
for stream in reduction_streams:
stream.synchronize()
nccl.reduce(dev_coalesced, root=0, streams=nccl_streams)
# From now on we're only going to work on the first device (from device_ids)
grad_batch = dev_grad_batch[0]
coalesced = dev_coalesced[0]
reduce_stream = reduction_streams[0]
with torch.cuda.stream(reduce_stream):
reduce_stream.wait_stream(nccl_streams[0])
coalesced /= dist.get_world_size()
dist.all_reduce(coalesced, group=group_id)
for grad, reduced in zip(
grad_batch, _unflatten_tensors(coalesced, grad_batch)):
grad.copy_(reduced)
job_event.set()
def _queue_reduction(self, bucket_idx):
grads_batch = self.buckets[bucket_idx]
grads_batch_coalesced = []
# coalesce the bucket
for dev_id, dev_grads_batch in zip(self.device_ids, grads_batch):
with torch.cuda.device(dev_id):
dev_grads_batch_coalesced = _flatten_dense_tensors(
dev_grads_batch)
grads_batch_coalesced.append(dev_grads_batch_coalesced)
# reduce to the first GPU in self.device_ids
if len(self.device_ids) > 1:
nccl.reduce(grads_batch_coalesced,
root=0,
streams=self.default_streams)
# divide by the number of processes here to reduce chances of overflow
grads_batch_coalesced[0] /= self.process_group.size()
# now work on the first gpu
reduction_work = self.process_group.allreduce(
[grads_batch_coalesced[0]], self.allreduce_opts)
self.reduction_works[bucket_idx] = reduction_work
self.buckets_coalesced[bucket_idx] = grads_batch_coalesced[0]
def reduce_add(inputs, destination=None):
"Reduces tensors from multiple GPUs and returns a result a specified device"
# TODO: try to find an input on another gpu, copy it,
# and accumulate into the copy
input_size = inputs[0].size()
for i, inp in enumerate(inputs):
assert inp.is_cuda, "reduce_add expects all inputs to be on GPUs"
if inp.size() != input_size:
got = 'x'.join(str(x) for x in inp.size())
expected = 'x'.join(str(x) for x in input_size)
raise ValueError("input {} has invalid size: got {}, but expected "
"{}".format(i, got, expected))
if destination is None:
destination = torch.cuda.current_device()
with torch.cuda.device(destination):
result = type(inp)(input_size).zero_()
if nccl.is_available(inputs) and inputs[0].get_device() == destination:
outputs = [result] + [t.new(t.size()) for t in inputs[1:]]
nccl.reduce(inputs, outputs)
return result
for inp in inputs:
input_correct_gpu = inp.cuda(result.get_device())
result.add_(input_correct_gpu)
return result
def test_reduce(self):
tensors = [torch.FloatTensor(128).uniform_() for i in range(nGPUs)]
expected = torch.FloatTensor(128).zero_()
for t in tensors:
expected.add_(t)
tensors = [tensors[i].cuda(i) for i in range(nGPUs)]
nccl.reduce(tensors)
self.assertEqual(tensors[0], expected)
def test_reduce(self, device, dtype):
tensors = [torch.zeros(128).uniform_().to(dtype=dtype) for i in range(nGPUs)]
expected = torch.zeros(128, dtype=dtype)
for t in tensors:
expected.add_(t)
tensors = [tensors[i].cuda(i) for i in range(nGPUs)]
nccl.reduce(tensors)
self.assertEqual(tensors[0], expected)
def test_reduce(self):
tensors = [torch.FloatTensor(128).uniform_() for i in range(nGPUs)]
expected = torch.FloatTensor(128).zero_()
for t in tensors:
expected.add_(t)
tensors = [tensors[i].cuda(i) for i in range(nGPUs)]
nccl.reduce(tensors)
self.assertEqual(tensors[0], expected)
def reduce_add(inputs, destination=None):
"""Sums tensors from multiple GPUs.
All inputs should have matching shapes, dtype, and layout. The output tensor
will be of the same shape, dtype, and layout.
Arguments:
inputs (Iterable[Tensor]): an iterable of tensors to add.
destination (int, optional): a device on which the output will be
placed (default: current device).
Returns:
A tensor containing an elementwise sum of all inputs, placed on the
:attr:`destination` device.
"""
destination = _get_device_index(destination, optional=True)
input_size = inputs[0].size()
root_index = None # index of input tensor that already is on the correct device
for i, inp in enumerate(inputs):
assert inp.device.type != "cpu", "reduce_add expects all inputs to be on GPUs"
if inp.get_device() == destination:
root_index = i
if inp.size() != input_size:
got = 'x'.join(str(x) for x in inp.size())
expected = 'x'.join(str(x) for x in input_size)
raise ValueError("input {} has invalid size: got {}, but expected "
"{}".format(i, got, expected))
if root_index is None:
raise RuntimeError(
"reduce_add expects destination to be on the same GPU with one of the tensors"
)
if len(inputs) == 1:
return inputs[0]
if nccl.is_available(inputs):
result = torch.empty_like(inputs[root_index])
nccl.reduce(inputs, output=result, root=root_index)
else:
destination_device = torch.device(inputs[root_index].device.type,
destination)
nonroot = [t for i, t in enumerate(inputs) if i != root_index]
# make a new tensor w/o clone
result = inputs[root_index] + nonroot[0].to(device=destination_device,
non_blocking=True)
for other in nonroot[1:]:
result.add_(other.to(device=destination_device, non_blocking=True))
return result
def _queue_reduction(self, bucket_idx):
grads_batch = self.buckets[bucket_idx]
grads_batch_coalesced = []
# coalesce the bucket
for dev_id, dev_grads_batch in zip(self.device_ids, grads_batch):
with torch.cuda.device(dev_id):
dev_grads_batch_coalesced = _flatten_dense_tensors(dev_grads_batch)
grads_batch_coalesced.append(dev_grads_batch_coalesced)
# reduce to the first GPU in self.device_ids
if len(self.device_ids) > 1:
nccl.reduce(grads_batch_coalesced, root=0, streams=self.default_streams)
# now work on the first gpu
reduction_work = c10d.all_reduce(grads_batch_coalesced[0], self.process_group)
self.reduction_works[bucket_idx] = reduction_work
self.buckets_coalesced[bucket_idx] = grads_batch_coalesced[0]
def test_collective_errors(self, device):
t = torch.rand(10).cuda(0)
with self.assertRaisesRegex(
TypeError, "Inputs should be a collection of tensors"):
nccl.all_reduce(t)
with self.assertRaisesRegex(
TypeError, "Inputs should be a collection of tensors"):
nccl.reduce(t)
with self.assertRaisesRegex(
TypeError, "Inputs should be a collection of tensors"):
nccl.broadcast(t)
with self.assertRaisesRegex(
TypeError, "Inputs should be a collection of tensors"):
nccl.all_gather(t, t)
with self.assertRaisesRegex(
TypeError, "Inputs should be a collection of tensors"):
nccl.reduce_scatter(t, t)
