Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: support flash attention GQA/split_k with small batches (#18938)

Browse Source
This commit is contained in:
Jeff Bolz
2026-01-21 17:43:43 +01:00
committed by GitHub
parent 067b8d7af3
commit 33f890e579
7 changed files with 74 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+7 -6
View File
@@ -90,7 +90,7 @@ void main() {
barrier();
}
uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02+iq3*p.nb03) / 4;
[[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSK / 4);
@@ -141,9 +141,9 @@ void main() {
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
#endif
uint32_t m_offset = 0;
uint32_t m_offset = gqa_iq1*KV;
if (p.nem2 != 1 || p.nem3 != 1) {
m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
}
[[dont_unroll]]
@@ -370,7 +370,8 @@ void main() {
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
@@ -382,7 +383,7 @@ void main() {
}
}
o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -428,7 +429,7 @@ void main() {
}
}
uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
if (p.gqa_ratio > 1) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {