ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 23:54:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq4_k: TG now works on CUDA

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-07-27 18:02:59 +03:00
parent 8a2d43813d
commit 41d20f6bb5
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cuda/vecdotq.cuh
View File

@@ -1253,10 +1253,10 @@ static __device__ __forceinline__ float vec_dot_iq4_k_q8_1(
const block_iq4_k * bq4 = (const block_iq4_k *) vbq + kbx;
const uint8_t * all_values = (const uint8_t *)iq4k_values;
// iqs is 0...7
const int ib32 = iqs;
// iqs is 0...28
const int ib32 = iqs/4;
// Why iqs/4 ?
const int32_t * q8 = (const int *)bq8_1[iqs/4].qs;
const int32_t * q8 = (const int *)bq8_1[ib32].qs;
const uint16_t * q4 = (const uint16_t *)bq4->qs + 8*ib32;
const uint16_t extra = bq4->extra >> 2*ib32;
int v1, v2;
@@ -1267,7 +1267,7 @@ static __device__ __forceinline__ float vec_dot_iq4_k_q8_1(
sumi1 = ggml_cuda_dp4a(v1, q8[j+0], sumi1);
sumi2 = ggml_cuda_dp4a(v2, q8[j+4], sumi2);
}
const float d = __half2float(bq4->d) * __low2float(bq8_1[iqs/4].ds);
const float d = __half2float(bq4->d) * __low2float(bq8_1[ib32].ds);
const uint8_t sh = bq4->scales_h[ib32/2] >> 4*(ib32%2);
const int ls1 = ((bq4->scales_l[ib32] & 0xf) | ((sh << 4) & 0x30)) - 32;
const int ls2 = ((bq4->scales_l[ib32] >> 4) | ((sh << 2) & 0x30)) - 32;