[Feature request] Add a new contain API for cache senario
Lifann opened this issue · 1 comments
Lifann commented
We use pure-device hkv as a device cache for feature ids which has high probability to be used in next few steps training/inference. It's critical to have a contain API for performance if there are multiple level of caches, in different throughput, interact with each other:
void contain(const size_type n, const key_type* keys, // (n)
bool* founds, // (n)
score_type* scores = nullptr, // (n)
cudaStream_t stream = 0) const;