def allreduce_grad(self, model):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params(model)
itemsize = 4
n_elems_total = sum(param.grad.size for param in params)
n_elems_per_node = int(math.ceil(n_elems_total / self.inter_size))
n_bytes_per_node = n_elems_per_node * itemsize
n_bytes_buffer = n_bytes_per_node * self.inter_size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
_memory_utility.pack_params(params, itemsize, 'grad',
self.gpu_buffer_a)
# Intra-node reduce
self.intra_nccl_comm.reduce(self.gpu_buffer_a.ptr(),
self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, nccl.NCCL_SUM, 0,
stream.ptr)
# Inter-node allreduce
if self.intra_rank == 0:
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size, self.gpu_buffer_a,
self.gpu_buffer_b, n_bytes_buffer, n_elems_per_node,
n_bytes_per_node, stream)
# Intra-node bcast
self.intra_nccl_comm.bcast(self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, 0, stream.ptr)
_memory_utility.unpack_params(params, itemsize, 'grad',
self.gpu_buffer_b)
def multi_node_mean_grad(self, model, zero_fill=False):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model, zero_fill)
itemsize = 4
n_elems_total = _memory_utility.count_grad_elements(params, zero_fill)
n_elems_per_node_2d = int(math.ceil(n_elems_total / self.size))
n_elems_per_node_1d = n_elems_per_node_2d * self.inter_size
n_bytes_per_node_1d = n_elems_per_node_1d * itemsize
n_bytes_per_node_2d = n_elems_per_node_2d * itemsize
n_bytes_buffer = n_bytes_per_node_2d * self.size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
allreduce_grad_dtype = np.float32
_memory_utility.pack_params(params, 'grad', self.gpu_buffer_a,
allreduce_grad_dtype, zero_fill)
if chainer.is_debug():
stream.synchronize()
array_a = self.gpu_buffer_a.array(n_elems_total)
array_b = self.gpu_buffer_b.array(n_elems_total)
self._check_ready_to_allreduce(array_a, array_b)
# Intra-node reduce-scatter (1st dimension)
self.intra_nccl_comm.reduceScatter(self.gpu_buffer_a.ptr(),
self.gpu_buffer_b.ptr(),
n_elems_per_node_1d,
nccl.NCCL_FLOAT, nccl.NCCL_SUM,
stream.ptr)
# Inter-node allreduce (2nd dimension)
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size, self.gpu_buffer_a,
self.gpu_buffer_b, n_bytes_per_node_1d, n_elems_per_node_2d,
n_bytes_per_node_2d, stream)
# Intra-node allgather (1st dimension)
self.intra_nccl_comm.allGather(self.gpu_buffer_b.ptr(),
self.gpu_buffer_a.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT,
stream.ptr)
if chainer.is_debug():
stream.synchronize()
self._ensure_all_finite(self.gpu_buffer_a.array(n_elems_total))
_memory_utility.unpack_params(params, 'grad', self.gpu_buffer_a,
allreduce_grad_dtype, zero_fill)
def allreduce_grad(self, model):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model)
itemsize = 4
n_elems_total = sum(param.grad.size for param in params)
n_elems_per_node_2d = int(math.ceil(n_elems_total / self.size))
n_elems_per_node_1d = n_elems_per_node_2d * self.inter_size
n_bytes_per_node_1d = n_elems_per_node_1d * itemsize
n_bytes_per_node_2d = n_elems_per_node_2d * itemsize
n_bytes_buffer = n_bytes_per_node_2d * self.size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
allreduce_grad_dtype = np.float32
_memory_utility.pack_params(
params, 'grad', self.gpu_buffer_a, allreduce_grad_dtype)
if chainer.is_debug():
stream.synchronize()
array_a = self.gpu_buffer_a.array(n_elems_total)
array_b = self.gpu_buffer_b.array(n_elems_total)
self.check_ready_to_allreduce(array_a, array_b)
# Intra-node reduce-scatter (1st dimension)
self.intra_nccl_comm.reduceScatter(
self.gpu_buffer_a.ptr(), self.gpu_buffer_b.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT, nccl.NCCL_SUM, stream.ptr)
# Inter-node allreduce (2nd dimension)
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size,
self.gpu_buffer_a, self.gpu_buffer_b,
n_bytes_per_node_1d, n_elems_per_node_2d,
n_bytes_per_node_2d, stream)
# Intra-node allgather (1st dimension)
self.intra_nccl_comm.allGather(
self.gpu_buffer_b.ptr(), self.gpu_buffer_a.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT, stream.ptr)
if chainer.is_debug():
stream.synchronize()
self.ensure_all_finite(self.gpu_buffer_a.array(n_elems_total))
_memory_utility.unpack_params(
params, 'grad', self.gpu_buffer_a, allreduce_grad_dtype)
def allreduce_grad(self, model, zero_fill=False):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model, zero_fill)
itemsize = 4
n_elems_total = _memory_utility.count_grad_elements(params, zero_fill)
n_elems_per_node = int(math.ceil(n_elems_total / self.inter_size))
n_bytes_per_node = n_elems_per_node * itemsize
n_bytes_buffer = n_bytes_per_node * self.inter_size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
allreduce_grad_dtype = np.float32
_memory_utility.pack_params(params, 'grad', self.gpu_buffer_a,
allreduce_grad_dtype, zero_fill, stream)
if chainer.is_debug():
stream.synchronize()
array_a = self.gpu_buffer_a.array(n_elems_total)
array_b = self.gpu_buffer_b.array(n_elems_total)
self.check_ready_to_allreduce(array_a, array_b)
# Intra-node reduce
self.intra_nccl_comm.reduce(self.gpu_buffer_a.ptr(),
self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, nccl.NCCL_SUM, 0,
stream.ptr)
# Inter-node allreduce
if self.intra_rank == 0:
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size, self.gpu_buffer_a,
self.gpu_buffer_b, n_bytes_buffer, n_elems_per_node,
n_bytes_per_node, stream)
# Intra-node bcast
self.intra_nccl_comm.bcast(self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, 0, stream.ptr)
if chainer.is_debug():
stream.synchronize()
self.ensure_all_finite(self.gpu_buffer_b.array(n_elems_total))
_memory_utility.unpack_params(params, 'grad', self.gpu_buffer_b,
allreduce_grad_dtype, zero_fill, stream)
def allreduce_grad(self, model):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model)
itemsize = 4
n_elems_total = sum(param.grad.size for param in params)
n_elems_per_node_2d = int(math.ceil(n_elems_total / self.size))
n_elems_per_node_1d = n_elems_per_node_2d * self.inter_size
n_bytes_per_node_1d = n_elems_per_node_1d * itemsize
n_bytes_per_node_2d = n_elems_per_node_2d * itemsize
n_bytes_buffer = n_bytes_per_node_2d * self.size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
allreduce_grad_dtype = np.float32
_memory_utility.pack_params(params, 'grad', self.gpu_buffer_a,
allreduce_grad_dtype)
# Intra-node reduce-scatter (1st dimension)
self.intra_nccl_comm.reduceScatter(self.gpu_buffer_a.ptr(),
self.gpu_buffer_b.ptr(),
n_elems_per_node_1d,
nccl.NCCL_FLOAT, nccl.NCCL_SUM,
stream.ptr)
# Inter-node allreduce (2nd dimension)
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size, self.gpu_buffer_a,
self.gpu_buffer_b, n_bytes_per_node_1d, n_elems_per_node_2d,
n_bytes_per_node_2d, stream)
# Intra-node allgather (1st dimension)
self.intra_nccl_comm.allGather(self.gpu_buffer_b.ptr(),
self.gpu_buffer_a.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT,
stream.ptr)
_memory_utility.unpack_params(params, 'grad', self.gpu_buffer_a,
allreduce_grad_dtype)
def allreduce_grad(self, model):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model)
itemsize = 4
n_elems_total = sum(param.grad.size for param in params)
n_elems_per_node_2d = int(math.ceil(n_elems_total / self.size))
n_elems_per_node_1d = n_elems_per_node_2d * self.inter_size
n_bytes_per_node_1d = n_elems_per_node_1d * itemsize
n_bytes_per_node_2d = n_elems_per_node_2d * itemsize
n_bytes_buffer = n_bytes_per_node_2d * self.size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
_memory_utility.pack_params(
params, itemsize, 'grad', self.gpu_buffer_a)
# Intra-node reduce-scatter (1st dimension)
self.intra_nccl_comm.reduceScatter(
self.gpu_buffer_a.ptr(), self.gpu_buffer_b.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT, nccl.NCCL_SUM, stream.ptr)
# Inter-node allreduce (2nd dimension)
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size,
self.gpu_buffer_a, self.gpu_buffer_b,
n_bytes_per_node_1d, n_elems_per_node_2d,
n_bytes_per_node_2d, stream)
# Intra-node allgather (1st dimension)
self.intra_nccl_comm.allGather(
self.gpu_buffer_b.ptr(), self.gpu_buffer_a.ptr(),
n_elems_per_node_1d, nccl.NCCL_FLOAT, stream.ptr)
_memory_utility.unpack_params(
params, itemsize, 'grad', self.gpu_buffer_a)
def allreduce_grad(self, model):
self._init_comms()
stream = chainer.cuda.Stream.null
params = _memory_utility.extract_params_set_grad(model)
itemsize = 4
n_elems_total = sum(param.grad.size for param in params)
n_elems_per_node = int(math.ceil(n_elems_total / self.inter_size))
n_bytes_per_node = n_elems_per_node * itemsize
n_bytes_buffer = n_bytes_per_node * self.inter_size
self.gpu_buffer_a.assign(n_bytes_buffer)
self.gpu_buffer_b.assign(n_bytes_buffer)
_memory_utility.pack_params(
params, itemsize, 'grad', self.gpu_buffer_a)
# Intra-node reduce
self.intra_nccl_comm.reduce(
self.gpu_buffer_a.ptr(), self.gpu_buffer_b.ptr(), n_elems_total,
nccl.NCCL_FLOAT, nccl.NCCL_SUM, 0, stream.ptr)
# Inter-node allreduce
if self.intra_rank == 0:
_communication_utility.inter_allreduce_gpu(
self.inter_mpi_comm, self.size,
self.gpu_buffer_a, self.gpu_buffer_b,
n_bytes_buffer, n_elems_per_node, n_bytes_per_node, stream)
# Intra-node bcast
self.intra_nccl_comm.bcast(
self.gpu_buffer_b.ptr(), n_elems_total, nccl.NCCL_FLOAT, 0,
stream.ptr)
_memory_utility.unpack_params(
params, itemsize, 'grad', self.gpu_buffer_b)
