ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-27 01:29:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix FA bug on AVX2 (#364)

Browse Source 
* Fix FA bug on AVX2

* Also this was wrong

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-05-02 07:09:09 +02:00
committed by GitHub
parent d37add8b39
commit 1ea1df4b2d
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -17120,11 +17120,12 @@ void compute_helper_q(KHelper& kh, VHelper& vh, int nq1, int nk1, int stride_q,
vh.reset_block();
block_q8_0_r8 q8r8[Dk/QK8_0 * k_step/8];
HelperQ80R8<Dk, k_step> khr8((const char *)q8r8, Dk/QK8_0*sizeof(block_q8_0));
HelperQ80<Dk, QK8_0>::convert(q_step, stride_q, q, q8);
auto q8r = (typename HelperQ80R8<Dk, k_step>::block_q8 *)qptr;
HelperQ80<Dk, QK8_0>::convert(q_step, stride_q, q, q8r);
auto mr = mask;
for (int k1 = 0; k1 < nk1/k_step; ++k1) {
HelperQ80R8<Dk, k_step>::repack(k_step, kh.data, kh.stride, q8r8);
KQHelper::mul_mask_kq(khr8, stride_m, q8, mr, fms);
HelperQ80R8<Dk, k_step>::repack(k_step, kh.block, kh.stride, q8r8);
KQHelper::mul_mask_kq(khr8, stride_m, q8r, mr, fms);
fqkv.accumulate_qkv(vh, fms);
kh.next_block();
vh.next_block();
@@ -17236,7 +17237,8 @@ struct FlashAttn {
std::is_same_v<KHelper, HelperQ8KV<Dk, k_step>> ||
std::is_same_v<KHelper, HelperQ8KVR8<Dk, k_step>>) {
constexpr size_t kMaxOnStackSize = 576;
auto q_size = q_step*(Dk/KHelper::block_size_q)*sizeof(typename KHelper::block_q8);
//auto q_size = q_step*(Dk/KHelper::block_size_q)*sizeof(typename KHelper::block_q8);
auto q_size = q_step*(Dk/QK8_2*sizeof(block_q8_2));
q_size = GGML_PAD(q_size, 64);
if (q_size > kMaxOnStackSize) {
auto qptr = get_q_storage(q_size);