Skip to content

Shut off NCCL when using only 1 device #1678

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
elliottslaughter wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Shut off NCCL when using only 1 device #1678

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 35 additions & 12 deletions lib/kernels/src/kernels/optimizer_kernels.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,18 +23,41 @@ void sgd_update_task(device_stream_t const &stream,
sgd_v_ptr = sgd_v.value().get_float_ptr();
}

gpu_sgd_nccl_update_task(
/*stream=*/stream.require_gpu(),
/*lr=*/lr,
/*momentum=*/momentum,
/*nesterov=*/nesterov,
/*weight_decay=*/weight_decay,
/*handle=*/handle.require_for_gpu(),
/*weight_grad_ptr=*/weight_grad.get_float_ptr(),
/*size=*/
get_num_elements(weight_grad.shape.dims).int_from_positive_int(),
/*weight_ptr=*/weight.get_float_ptr(),
/*sgd_v_ptr=*/sgd_v_ptr);
PerDeviceFFHandle const &gpu_handle = handle.require_for_gpu();
#ifdef FF_USE_NCCL
bool have_nccl_comm = (gpu_handle.ncclComm != nullptr);
#else
bool have_nccl_comm = false;
#endif

if (have_nccl_comm) {
gpu_sgd_nccl_update_task(
/*stream=*/stream.require_gpu(),
/*lr=*/lr,
/*momentum=*/momentum,
/*nesterov=*/nesterov,
/*weight_decay=*/weight_decay,
/*handle=*/gpu_handle,
/*weight_grad_ptr=*/weight_grad.get_float_ptr(),
/*size=*/
get_num_elements(weight_grad.shape.dims).int_from_positive_int(),
/*weight_ptr=*/weight.get_float_ptr(),
/*sgd_v_ptr=*/sgd_v_ptr);
} else {
// No communicator (single-rank), so the gradient allreduce is a no-op.
gpu_sgd_ps_update_task(
/*stream=*/stream.require_gpu(),
/*lr=*/lr,
/*momentum=*/momentum,
/*nesterov=*/nesterov,
/*weight_decay=*/weight_decay,
/*weight_grad_ptr=*/weight_grad.get_float_ptr(),
/*size=*/
get_num_elements(weight_grad.shape.dims).int_from_positive_int(),
/*num_replicas=*/num_replicas,
/*weight_ptr=*/weight.get_float_ptr(),
/*sgd_v_ptr=*/sgd_v_ptr);
}
} else {
ASSERT(stream.is_cpu());
ASSERT(handle.is_for_cpu());
Expand Down
18 changes: 13 additions & 5 deletions lib/kernels/src/managed_per_device_ff_handle.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,10 +18,16 @@ ManagedPerDeviceFFHandle::ManagedPerDeviceFFHandle(
checkCUDA(cudaMalloc(&this->handle->workSpace, this->handle->workSpaceSize));

#ifdef FF_USE_NCCL
ncclUniqueId ncclId;
checkNCCL(ncclGetUniqueId(&ncclId));
checkNCCL(ncclCommInitRank(
&handle->ncclComm, num_ranks, ncclId, my_rank)); // todo generalize
// A single-rank communicator can only ever perform no-op collectives, but
// still costs ~500MiB of device memory, so skip creating it entirely.
if (num_ranks == 1) {
this->handle->ncclComm = nullptr;
} else {
ncclUniqueId ncclId;
checkNCCL(ncclGetUniqueId(&ncclId));
checkNCCL(ncclCommInitRank(
&handle->ncclComm, num_ranks, ncclId, my_rank)); // todo generalize
}
#endif
}

Expand All @@ -41,7 +47,9 @@ ManagedPerDeviceFFHandle::~ManagedPerDeviceFFHandle() {
checkCUBLAS(cublasDestroy(this->handle->blas));
checkCUDA(cudaFree(this->handle->workSpace));
#ifdef FF_USE_NCCL
checkNCCL(ncclCommDestroy(this->handle->ncclComm));
if (this->handle->ncclComm != nullptr) {
checkNCCL(ncclCommDestroy(this->handle->ncclComm));
}
#endif
delete this->handle;
}
Expand Down
Loading