Skip to content

[OPTIMIZATION] Optimizes the single_query_cached_kv_attention kernel #420

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
merged 3 commits into from
Aug 4, 2023
Merged

[OPTIMIZATION] Optimizes the single_query_cached_kv_attention kernel #420

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1f256ee
[OPTIMIZATION] Optimizes the
naed90 Jul 10, 2023
b27f396
Rename NUM_THREAD_GROUPS_LOWER_BOUND
naed90 Aug 4, 2023
455038f
Assumes NUM_THREADS % THREAD_GROUP_SIZE == 0
naed90 Aug 4, 2023
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
11 changes: 7 additions & 4 deletions csrc/attention/attention_kernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,6 +83,8 @@ __global__ void single_query_cached_kv_attention_kernel(
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride) {
constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
constexpr int NUM_THREAD_GROUPS = NUM_THREADS / THREAD_GROUP_SIZE; // Note: This assumes THREAD_GROUP_SIZE divides NUM_THREADS
assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
constexpr int NUM_TOKENS_PER_THREAD_GROUP = (BLOCK_SIZE + WARP_SIZE - 1) / WARP_SIZE;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int thread_idx = threadIdx.x;
Expand Down Expand Up @@ -116,12 +118,13 @@ __global__ void single_query_cached_kv_attention_kernel(
// th vectors of the query, and so on.
// NOTE(woosuk): Because q is split from a qkv tensor, it may not be contiguous.
const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
Q_vec q_vecs[NUM_VECS_PER_THREAD];
__shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
#pragma unroll
for (int i = 0; i < NUM_VECS_PER_THREAD; i++) {
for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD; i += NUM_THREAD_GROUPS) {
const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
q_vecs[i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
q_vecs[thread_group_offset][i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
}
__syncthreads(); // TODO(naed90): possible speedup if this is replaced with a memory wall right before we use q_vecs

// Memory planning.
extern __shared__ char shared_mem[];
Expand Down Expand Up @@ -169,7 +172,7 @@ __global__ void single_query_cached_kv_attention_kernel(

// Compute dot product.
// This includes a reduction across the threads in the same thread group.
float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs, k_vecs);
float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs[thread_group_offset], k_vecs);
// Add the ALiBi bias if slopes are given.
qk += (alibi_slope != 0) ? alibi_slope * (token_idx - context_len) : 0;

Expand Down