This commit is contained in:
Kawrakow 2026-06-14 07:42:39 +00:00
parent 670a3f6f5b
commit c73bfbe9ce

View File

@ -1173,13 +1173,17 @@ struct FlashQKV {
S += expf(s - fms.M[j]);
}
}
GGML_ASSERT(S > 0);
if (S > 0) {
auto norm = F16::set1(1/S);
//auto norm = F16::set1(fms.S[j] > 0 ? 1/fms.S[j] : 0.f);
for (int i = 0; i < D/F16::block_size; ++i) {
auto r = F16::load(R + F16::block_size*i);
F16::store(qkv + F16::block_size*i, F16::mul(norm, r));
}
} else {
for (int i = 0; i < D/F16::block_size; ++i) {
F16::store(qkv + F16::block_size*i, F16::zero());
}
}
}
template <typename FMS>