[Core][Distributed] Refactor ipc buffer init in CustomAllreduce (#10030)

Signed-off-by: Hanzhi Zhou <hanzhi713@gmail.com>
This commit is contained in:
Hanzhi Zhou
2024-11-06 23:50:47 -08:00
committed by GitHub
parent d7263a1bb8
commit 6192e9b8fe
9 changed files with 218 additions and 260 deletions
+9 -13
View File
@@ -199,20 +199,16 @@ void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
#ifndef USE_ROCM
using fptr_t = int64_t;
fptr_t init_custom_ar(torch::Tensor& meta, torch::Tensor& rank_data,
const std::vector<std::string>& handles,
const std::vector<int64_t>& offsets, int64_t rank,
bool full_nvlink);
void all_reduce_reg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out);
void all_reduce_unreg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& reg_buffer,
torch::Tensor& out);
fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
torch::Tensor& rank_data, int64_t rank, bool full_nvlink);
void all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
fptr_t reg_buffer, int64_t reg_buffer_sz_bytes);
void dispose(fptr_t _fa);
int64_t meta_size();
void register_buffer(fptr_t _fa, torch::Tensor& t,
const std::vector<std::string>& handles,
const std::vector<int64_t>& offsets);
std::tuple<torch::Tensor, std::vector<int64_t>> get_graph_buffer_ipc_meta(
fptr_t _fa);
void register_graph_buffers(fptr_t _fa, const std::vector<std::string>& handles,
void register_buffer(fptr_t _fa, const std::vector<int64_t>& fake_ipc_ptrs);
std::tuple<std::vector<int64_t>, std::vector<int64_t>>
get_graph_buffer_ipc_meta(fptr_t _fa);
void register_graph_buffers(fptr_t _fa,
const std::vector<std::vector<int64_t>>& handles,
const std::vector<std::vector<int64_t>>& offsets);
#endif