[Breaking] Switch from rabit to the collective communicator (#8257)
* Switch from rabit to the collective communicator * fix size_t specialization * really fix size_t * try again * add include * more include * fix lint errors * remove rabit includes * fix pylint error * return dict from communicator context * fix communicator shutdown * fix dask test * reset communicator mocklist * fix distributed tests * do not save device communicator * fix jvm gpu tests * add python test for federated communicator * Update gputreeshap submodule Co-authored-by: Hyunsu Philip Cho <chohyu01@cs.washington.edu>
This commit is contained in:
@@ -24,6 +24,10 @@ class NcclDeviceCommunicator : public DeviceCommunicator {
|
||||
int32_t const rank = communicator_->GetRank();
|
||||
int32_t const world = communicator_->GetWorldSize();
|
||||
|
||||
if (world == 1) {
|
||||
return;
|
||||
}
|
||||
|
||||
std::vector<uint64_t> uuids(world * kUuidLength, 0);
|
||||
auto s_uuid = xgboost::common::Span<uint64_t>{uuids.data(), uuids.size()};
|
||||
auto s_this_uuid = s_uuid.subspan(rank * kUuidLength, kUuidLength);
|
||||
@@ -52,8 +56,15 @@ class NcclDeviceCommunicator : public DeviceCommunicator {
|
||||
}
|
||||
|
||||
~NcclDeviceCommunicator() override {
|
||||
dh::safe_cuda(cudaStreamDestroy(cuda_stream_));
|
||||
ncclCommDestroy(nccl_comm_);
|
||||
if (communicator_->GetWorldSize() == 1) {
|
||||
return;
|
||||
}
|
||||
if (cuda_stream_) {
|
||||
dh::safe_cuda(cudaStreamDestroy(cuda_stream_));
|
||||
}
|
||||
if (nccl_comm_) {
|
||||
dh::safe_nccl(ncclCommDestroy(nccl_comm_));
|
||||
}
|
||||
if (xgboost::ConsoleLogger::ShouldLog(xgboost::ConsoleLogger::LV::kDebug)) {
|
||||
LOG(CONSOLE) << "======== NCCL Statistics========";
|
||||
LOG(CONSOLE) << "AllReduce calls: " << allreduce_calls_;
|
||||
@@ -61,16 +72,28 @@ class NcclDeviceCommunicator : public DeviceCommunicator {
|
||||
}
|
||||
}
|
||||
|
||||
void AllReduceSum(double *send_receive_buffer, int count) override {
|
||||
dh::safe_cuda(cudaSetDevice(device_ordinal_));
|
||||
dh::safe_nccl(ncclAllReduce(send_receive_buffer, send_receive_buffer, count, ncclDouble,
|
||||
ncclSum, nccl_comm_, cuda_stream_));
|
||||
allreduce_bytes_ += count * sizeof(double);
|
||||
allreduce_calls_ += 1;
|
||||
void AllReduceSum(float *send_receive_buffer, size_t count) override {
|
||||
DoAllReduceSum<ncclFloat>(send_receive_buffer, count);
|
||||
}
|
||||
|
||||
void AllReduceSum(double *send_receive_buffer, size_t count) override {
|
||||
DoAllReduceSum<ncclDouble>(send_receive_buffer, count);
|
||||
}
|
||||
|
||||
void AllReduceSum(int64_t *send_receive_buffer, size_t count) override {
|
||||
DoAllReduceSum<ncclInt64>(send_receive_buffer, count);
|
||||
}
|
||||
|
||||
void AllReduceSum(uint64_t *send_receive_buffer, size_t count) override {
|
||||
DoAllReduceSum<ncclUint64>(send_receive_buffer, count);
|
||||
}
|
||||
|
||||
void AllGatherV(void const *send_buffer, size_t length_bytes, std::vector<std::size_t> *segments,
|
||||
dh::caching_device_vector<char> *receive_buffer) override {
|
||||
if (communicator_->GetWorldSize() == 1) {
|
||||
return;
|
||||
}
|
||||
|
||||
dh::safe_cuda(cudaSetDevice(device_ordinal_));
|
||||
int const world_size = communicator_->GetWorldSize();
|
||||
int const rank = communicator_->GetRank();
|
||||
@@ -95,6 +118,9 @@ class NcclDeviceCommunicator : public DeviceCommunicator {
|
||||
}
|
||||
|
||||
void Synchronize() override {
|
||||
if (communicator_->GetWorldSize() == 1) {
|
||||
return;
|
||||
}
|
||||
dh::safe_cuda(cudaSetDevice(device_ordinal_));
|
||||
dh::safe_cuda(cudaStreamSynchronize(cuda_stream_));
|
||||
}
|
||||
@@ -136,6 +162,19 @@ class NcclDeviceCommunicator : public DeviceCommunicator {
|
||||
return id;
|
||||
}
|
||||
|
||||
template <ncclDataType_t data_type, typename T>
|
||||
void DoAllReduceSum(T *send_receive_buffer, size_t count) {
|
||||
if (communicator_->GetWorldSize() == 1) {
|
||||
return;
|
||||
}
|
||||
|
||||
dh::safe_cuda(cudaSetDevice(device_ordinal_));
|
||||
dh::safe_nccl(ncclAllReduce(send_receive_buffer, send_receive_buffer, count, data_type, ncclSum,
|
||||
nccl_comm_, cuda_stream_));
|
||||
allreduce_bytes_ += count * sizeof(T);
|
||||
allreduce_calls_ += 1;
|
||||
}
|
||||
|
||||
int const device_ordinal_;
|
||||
Communicator *communicator_;
|
||||
ncclComm_t nccl_comm_{};
|
||||
|
||||
Reference in New Issue
Block a user