vllm/csrc/attention.cpp
2023-10-16 00:59:57 -07:00

43 lines
1.1 KiB
C++

#include <torch/extension.h>
#include <c10/util/Optional.h>
void paged_attention_v1(
torch::Tensor& out,
torch::Tensor& query,
torch::Tensor& key_cache,
torch::Tensor& value_cache,
torch::Tensor& head_mapping,
float scale,
torch::Tensor& block_tables,
torch::Tensor& context_lens,
int block_size,
int max_context_len,
const c10::optional<torch::Tensor>& alibi_slopes);
void paged_attention_v2(
torch::Tensor& out,
torch::Tensor& exp_sums,
torch::Tensor& max_logits,
torch::Tensor& tmp_out,
torch::Tensor& query,
torch::Tensor& key_cache,
torch::Tensor& value_cache,
torch::Tensor& head_mapping,
float scale,
torch::Tensor& block_tables,
torch::Tensor& context_lens,
int block_size,
int max_context_len,
const c10::optional<torch::Tensor>& alibi_slopes);
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def(
"paged_attention_v1",
&paged_attention_v1,
"Compute the attention between an input query and the cached keys/values using PagedAttention.");
m.def(
"paged_attention_v2",
&paged_attention_v2,
"PagedAttention V2.");
}