Skip to content

[Runtime][KVCache] AttentionWithFusedQKV and RoPE mode #16456

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tqchen merged 1 commit into from
Jan 24, 2024
Merged

[Runtime][KVCache] AttentionWithFusedQKV and RoPE mode #16456

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions src/runtime/relax_vm/kv_cache.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,6 +137,19 @@ class AttentionKVCache : public Object {
virtual void Attention(int64_t layer_id, NDArray q_data, NDArray k_data, NDArray v_data,
Optional<NDArray> mask, NDArray o_data) = 0;

/*!
* \brief Compute attention with Q/K/V data which are concatenated along
* the head dimension.
* \param layer_id The model layer where the attention compute happens.
* \param qkv_data The input Q/K/V data, in layout
* `(total_length, num_qo_heads + 2 * num_kv_heads, head_dim)`.
* \param mask The input mask data, in layout `(total_sqr_length)`.
* \param o_data The output O data, in layout `(total_length, num_qo_heads, head_dim)`.
* \sa AttentionKVCache::Attention
*/
virtual void AttentionWithFusedQKV(int64_t layer_id, NDArray qkv_data, Optional<NDArray> mask,
NDArray o_data) = 0;

/************** Debug Helpers **************/

/*!
Expand Down
Loading