Implement block copy kernel to optimize beam search (#32)

This commit is contained in:
Woosuk Kwon
2023-04-07 17:45:07 -07:00
committed by GitHub
parent a490aafa36
commit 0f40557af6
6 changed files with 154 additions and 48 deletions

View File

@@ -9,8 +9,8 @@ void swap_blocks(
const std::map<int64_t, int64_t>& block_mapping);
void copy_blocks(
torch::Tensor& src,
torch::Tensor& dst,
std::vector<torch::Tensor>& key_caches,
std::vector<torch::Tensor>& value_caches,
const std::map<int64_t, std::vector<int64_t>>& block_mapping);
void reshape_and_cache(