vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fixed bug in new metal Q2_K implementation

Browse source
This commit is contained in:
Iwan Kawrakow 2023-07-21 10:21:19 +03:00
parent 417546c8c3
commit bdf3b6e0d7
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-metal.metal
View file

@ -1246,7 +1246,7 @@ kernel void kernel_mul_mat_q2_K_f32(
yl[i+24] = y4[i+96]; sumy[3] += yl[i+24]; yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
} }
device const uint8_t * sc = (device const uint8_t *)x[ib].scales + is; device const uint8_t * sc = (device const uint8_t *)x[ib].scales + 8*im + is;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir; device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
device const half * dh = &x[ib].d; device const half * dh = &x[ib].d;