Page first direct IO kernel (#10060)

Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
This commit is contained in:
huangtingwei
2025-09-10 13:35:34 +08:00
committed by GitHub
parent 737d73ed5b
commit 5be8c2f7f7
5 changed files with 358 additions and 2 deletions

View File

@@ -331,6 +331,14 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
"transfer_kv_direct(Tensor[] src_layers, Tensor[] dst_layers, Tensor src_indices, Tensor dst_indices, int "
"page_size) -> ()");
m.impl("transfer_kv_direct", torch::kCUDA, &transfer_kv_direct);
m.def(
"transfer_kv_per_layer_direct_pf_lf(Tensor[] src_ptrs, Tensor[] dst_ptrs, Tensor src_indices, "
"Tensor dst_indices, int layer_id, int page_size)->() ");
m.impl("transfer_kv_per_layer_direct_pf_lf", torch::kCUDA, &transfer_kv_per_layer_direct_pf_lf);
m.def(
"transfer_kv_all_layer_direct_lf_pf(Tensor[] src_ptrs, Tensor[] dst_ptrs, Tensor src_indices, "
"Tensor dst_indices, int page_size) ->() ");
m.impl("transfer_kv_all_layer_direct_lf_pf", torch::kCUDA, &transfer_kv_all_layer_direct_lf_pf);
/*
* From csrc/memory