Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 05:48:47 +01:00
Code Issues Packages Projects Releases Wiki Activity

metal : improve dequantize precision to match CPU (#4836)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-01-09 19:37:08 +02:00 committed by GitHub
parent 9a818f7c42
commit 4dccb38d9a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
ggml-metal.metal
View File

@ -3841,8 +3841,8 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4]; uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
: (scale_2&kmask2) | ((scale_1&kmask1) << 4); : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h); float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
const half ml = 4.h * dl; const float ml = 4.f * dl;
il = (il/2) & 3; il = (il/2) & 3;
const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
@ -3909,7 +3909,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
uint8_t ul = 1 << (il/2); uint8_t ul = 1 << (il/2);
il = il & 3; il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales); const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.h; const float d = il < 2 ? xb->d : xb->d / 16.f;
const float min = xb->dmin; const float min = xb->dmin;
const float dl = d * sc[0]; const float dl = d * sc[0];
const float ml = min * sc[1]; const float ml = min * sc[1];
@ -3942,17 +3942,17 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
#if QK_K == 256 #if QK_K == 256
ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1); ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
qh = qh + 32*(il/8) + 16*(il&1); qh = qh + 32*(il/8) + 16*(il&1);
half sc = scales[(il%2) + 2 * ((il/2))]; float sc = scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3; il = (il/2) & 3;
#else #else
ql = ql + 16 * (il&1); ql = ql + 16 * (il&1);
half sc = scales[il]; float sc = scales[il];
#endif #endif
const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F;
const half coef = il>1 ? 1.f/16.h : 1.h; const float coef = il>1 ? 1.f/16.f : 1.f;
const half ml = d_all * sc * 32.h; const float ml = d_all * sc * 32.f;
const half dl = d_all * sc * coef; const float dl = d_all * sc * coef;
for (int i = 0; i < 16; ++i) { for (int i = 0; i < 16; ++i) {
const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2)) const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
: ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4)); : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));