CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-12 10:47:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: support fattn sinks (#15126)

Browse Source
This commit is contained in:
Jeff Bolz
2025-08-07 15:44:20 -05:00
committed by GitHub
parent a0552c8bee
commit c4f53563df
6 changed files with 148 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
View File

@@ -329,6 +329,27 @@ void main() {
return;
}
if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
float ms = 1.0f;
float vs = 1.0f;
if (sink > Mf[r]) {
ms = exp(Mf[r] - sink);
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
Of[r][d] *= ACC_TYPE(ms);
}
} else {
vs = exp(sink - Mf[r]);
}
Lf[r] = Lf[r]*ms + vs;
}
}
float Lfrcp[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Lfrcp[r] = 1.0 / Lf[r];