EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[CPU] remove process_group from inputs of shm_allreduce and shm_allgather (#7486)

Browse Source
This commit is contained in:
Chunyuan WU
2025-07-01 12:54:11 +08:00
committed by GitHub
parent ff2e9c9479
commit 6005eceee3
2 changed files with 9 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
sgl-kernel/csrc/cpu/interface.cpp
View File

@@ -47,71 +47,26 @@ void initialize(int64_t size, int64_t rank) {
}
}
void shm_allreduce(
torch::Tensor& data, c10::intrusive_ptr<c10d::ProcessGroup> process_group, c10::intrusive_ptr<c10d::ReduceOp> op) {
void shm_allreduce(torch::Tensor& data, int64_t op) {
RECORD_FUNCTION("sgl-kernel::shm_allreduce", std::vector<c10::IValue>({data}));
TORCH_CHECK(op == c10d::ReduceOp::SUM, "Only torch.distributed.ReduceOp.SUM is supported");
auto numel = data.numel();
int data_size = 0;
bool data_type_fallback = false;
switch (data.scalar_type()) {
case c10::ScalarType::BFloat16:
data_size = numel * 2;
break;
case c10::ScalarType::Float:
data_size = numel * 4;
break;
default:
data_type_fallback = true;
}
if (data_type_fallback || !all_ranks_local_p) {
// Fallback to torch distributed allreduce
std::vector<torch::Tensor> tensors = {data};
process_group->allreduce(tensors)->wait();
} else {
all_reduce_outer_loop(data, numel, data_size);
}
int data_size = numel * data.element_size();
all_reduce_outer_loop(data, numel, data_size);
return;
}
torch::Tensor shm_allgather(torch::Tensor& data, c10::intrusive_ptr<c10d::ProcessGroup> process_group, int64_t dim) {
torch::Tensor shm_allgather(torch::Tensor& data, int64_t dim) {
RECORD_FUNCTION("sgl-kernel::shm_allgather", std::vector<c10::IValue>({data}));
auto numel = data.numel();
int data_size = 0;
bool data_type_fallback = false;
switch (data.scalar_type()) {
case c10::ScalarType::BFloat16:
data_size = numel * 2;
break;
case c10::ScalarType::Float:
data_size = numel * 4;
break;
default:
data_type_fallback = true;
}
int data_size = numel * data.element_size();
if (dim < 0) {
dim += data.dim();
}
if (data_type_fallback || !all_ranks_local_p) {
// Fallback to torch distributed allreduce
std::vector<std::vector<torch::Tensor>> output_tensors(1);
auto world_size = process_group->getSize();
for (int i = 0; i < world_size; i++) {
output_tensors[0].push_back(torch::empty_like(data));
}
std::vector<torch::Tensor> input_tensors = {data};
process_group->allgather(output_tensors, input_tensors)->wait();
return torch::cat(output_tensors[0], dim).contiguous();
}
std::vector<int64_t> result_shape = data.sizes().vec();
result_shape[dim] *= world_size;
torch::Tensor result_tensor = torch::empty(result_shape, data.options());