happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix SHMEM_STAGING indexing

This commit is contained in:
Ruben Ortlam 2026-03-20 14:32:40 +01:00
parent 6d61237467
commit 82afe402cb
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
View File

@ -407,9 +407,11 @@ void main() {
#if defined(DATA_A_Q4_0)
if (d_per_step < 8) {
[[unroll]] for (uint32_t d = 0; d < d_per_step; d++) {
uint sub = (d_block + d) % 4;
uint pos = d_tid * (HSK_per_thread / 4) + d_block + d;
uint sub = pos % 4;
uint shift = ((pos % 8) >= 4) ? 4 : 0;
uint vui = kblocksh[buf_ib].qs[sub];
uint shift = ((d_block + d) >= 4) ? 4 : 0;
k_quants[d] = int32_t((vui >> shift) & 0x0F0F0F0F);
}
} else {
@ -422,7 +424,7 @@ void main() {
}
#elif defined(DATA_A_Q8_0)
[[unroll]] for (uint32_t d = 0; d < d_per_step; d++) {
k_quants[d] = kblocksh[buf_ib].qs[d_block % 8 + d];
k_quants[d] = kblocksh[buf_ib].qs[(d_tid * (HSK_per_thread / 4) + d_block) % 8 + d];
}
#endif
} else {