EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 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

28
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
View File

@@ -248,6 +248,34 @@ void main() {
// resize L by using smear/reduce
coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> S;
coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Mr;
// resize M by using smear/reduce
coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
// O, Ldiag, Mr all have the same type so all element locations match
[[unroll]] for (uint32_t i = 0; i < Ldiag.length(); ++i) {
ACC_TYPE sink = S[i];
ACC_TYPE ms = ACC_TYPE(1.0f);
ACC_TYPE vs = ACC_TYPE(1.0f);
if (sink > Mr[i]) {
ms = exp(Mr[i] - sink);
O[i] *= ms;
} else {
vs = exp(sink - Mr[i]);
}
Ldiag[i] = Ldiag[i]*ms + vs;
}
}
[[unroll]]
for (int k = 0; k < Ldiag.length(); ++k) {
Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];