MSCCL++ Integration (#22734)

Co-authored-by: Caio Rocha <caiorocha@microsof.com>
Co-authored-by: empyreus <rjsouza1995@gmail.com>
This commit is contained in:
Caio Rocha
2026-06-08 21:13:13 -07:00
committed by GitHub
co-authored by Caio Rocha empyreus
parent 9c53031d2b
commit c2eae96c56
17 changed files with 397 additions and 1775 deletions
-14
View File
@@ -89,20 +89,6 @@ std::tuple<std::vector<int64_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta
void register_buffer(fptr_t _fa, const std::vector<fptr_t>& fake_ipc_ptrs);
void register_graph_buffers(
fptr_t _fa, const std::vector<std::vector<int64_t>>& handles, const std::vector<std::vector<int64_t>>& offsets);
// mscclpp
torch::Tensor mscclpp_generate_unique_id();
fptr_t mscclpp_init_context(
const torch::Tensor& unique_id,
const int64_t rank,
const int64_t world_size,
torch::Tensor& scratch,
torch::Tensor& put_buffer,
const int64_t nranks_per_node,
const std::vector<int64_t>& rank_to_node,
const std::vector<int64_t>& rank_to_ib,
const int64_t context_selection);
void mscclpp_allreduce(fptr_t _context, torch::Tensor& inp, torch::Tensor& out, int64_t nthreads, int64_t nblocks);
#endif
/*