ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 03:11:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

q8_KV_r8: don't use nrc_y = 16 on Zen4

Browse Source 
This is faster - 350 t/s. Why?
Much better than the 290 t/s we had before, but still slower
than the 370 t/s for q8_k_r8.
This commit is contained in:
Iwan Kawrakow
2025-02-18 16:39:45 +02:00
parent 2b9526c8b6
commit e08e292bea
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -9367,9 +9367,9 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) {
mm.funcs[5] = mul_mat_q8_KV_r8_q8_KV<6>; mm.funcs[5] = mul_mat_q8_KV_r8_q8_KV<6>;
mm.funcs[6] = mul_mat_q8_KV_r8_q8_KV<7>; mm.funcs[6] = mul_mat_q8_KV_r8_q8_KV<7>;
mm.funcs[7] = mul_mat_q8_KV_r8_q8_KV<8>; mm.funcs[7] = mul_mat_q8_KV_r8_q8_KV<8>;
#ifdef HAVE_FANCY_SIMD //#ifdef HAVE_FANCY_SIMD
mm.func16 = mul_mat_q8_KV_r8_q8_KV<16>; // mm.func16 = mul_mat_q8_KV_r8_q8_KV<16>;
#endif //#endif
expected_typeB = GGML_TYPE_Q8_KV; expected_typeB = GGML_TYPE_Q8_KV;
break; break;
case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ4_K_R4: