ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-27 18:01:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

Q5_K_R4 (#132)

Browse Source 
* q5_k_r4: WIP

* q5_k_r4: Zen4 and AVX2

We get PP-512(LLaMA-3.1-8B) = 248.3 t/s on Zen4.
Q5_K_S has PP-512 = 190 t/s.

* q5_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 96.1 t/s.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-10 18:13:47 +01:00
committed by GitHub
parent 3a8795d422
commit e78e47b857
10 changed files with 454 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

119
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -4182,3 +4182,122 @@ void vec_dot_q6_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
}
//
// ========================================= q5_k_r4
//
void quantize_row_q5_k_r4_ref(const float * x, block_q5_k_r4 * y, int64_t k) {
quantize_q5_k_r4(x, (void *)y, 4, k/4, nullptr);
}
void quantize_row_q5_k_r4(const float * x, void * y, int64_t k) {
quantize_q5_k_r4(x, y, 4, k/4, nullptr);
}
namespace {
inline void convert_q5_k(const block_q5_K& x, uint8_t * L, uint8_t * Ld, uint8_t * Lm) {
for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
get_scale_min_k4(2*ib64+0, x.scales, Ld[2*ib64+0], Lm[2*ib64+0]);
get_scale_min_k4(2*ib64+1, x.scales, Ld[2*ib64+1], Lm[2*ib64+1]);
for (int j = 0; j < 32; ++j) {
L[64*ib64+j+ 0] = (x.qs[32*ib64+j] & 0xf) | (((x.qh[j] >> (2*ib64+0)) & 1) << 4);
L[64*ib64+j+32] = (x.qs[32*ib64+j] >> 4) | (((x.qh[j] >> (2*ib64+1)) & 1) << 4);
}
}
}
}
static void repack_q5_k(int nrows, int n_per_row, const block_q5_K * x, block_q5_k_r4 * y) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
const block_q5_K * x4[4];
uint8_t L[QK_K], Ld[QK_K/32], Lm[QK_K/32];
for (int row = 0; row < nrows; row += 4) {
for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
for (int ibl = 0; ibl < nblock; ++ibl) {
std::memset(y[ibl].scales_l, 0, QK_K/8);
std::memset(y[ibl].scales_h, 0, QK_K/16);
for (int k = 0; k < 4; ++k) {
y[ibl].d[k+0] = x4[k][ibl].d;
y[ibl].d[k+4] = x4[k][ibl].dmin;
convert_q5_k(x4[k][ibl], L, Ld, Lm);
for (int ib = 0; ib < QK_K/32; ++ib) {
y[ibl].scales_l[4*ib+k] = (Ld[ib] & 0xf) | ((Lm[ib] & 0xf) << 4);
uint8_t h = (Ld[ib] >> 4) | ((Lm[ib] >> 4) << 2);
y[ibl].scales_h[(4*ib+k)%16] |= (h << 4*((4*ib+k)/16));
for (int i = 0; i < 4; ++i) {
y[ibl].qs[64*ib+4*k+i+ 0] = (L[32*ib+i+ 0] & 0xf) | ((L[32*ib+i+ 8] & 0xf) << 4);
y[ibl].qs[64*ib+4*k+i+16] = (L[32*ib+i+16] & 0xf) | ((L[32*ib+i+24] & 0xf) << 4);
y[ibl].qs[64*ib+4*k+i+32] = (L[32*ib+i+ 4] & 0xf) | ((L[32*ib+i+12] & 0xf) << 4);
y[ibl].qs[64*ib+4*k+i+48] = (L[32*ib+i+20] & 0xf) | ((L[32*ib+i+28] & 0xf) << 4);
y[ibl].qh[16*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] >> 4) << 0) | ((L[32*ib+i+ 8] >> 4) << 1) | ((L[32*ib+i+ 4] >> 4) << 2) | ((L[32*ib+i+12] >> 4) << 3) |
((L[32*ib+i+16] >> 4) << 4) | ((L[32*ib+i+24] >> 4) << 5) | ((L[32*ib+i+20] >> 4) << 6) | ((L[32*ib+i+28] >> 4) << 7);
}
}
}
}
x += 4*nblock;
y += nblock;
}
}
size_t quantize_q5_k_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
char * qcur = (char *)dst;
auto row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
std::vector<char> qtmp(4*row_size);
for (int row = 0; row < nrows; row += 4) {
quantize_q5_K(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
repack_q5_k(4, n_per_row, (const block_q5_K *)qtmp.data(), (block_q5_k_r4 *)qcur);
qcur += 4*row_size;
src += 4*n_per_row;
}
return nrows*row_size;
}
void dequantize_row_q5_k_r4(const block_q5_k_r4 * x, float * y, int64_t k) {
auto n_per_row = k/4;
float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
int nblock = n_per_row/QK_K;
for (int ibl = 0; ibl < nblock; ++ibl) {
for (int k = 0; k < 4; ++k) {
const float d = GGML_FP16_TO_FP32(x[ibl].d[k+0]);
const float m = GGML_FP16_TO_FP32(x[ibl].d[k+4]);
auto ql = x[ibl].qs;
auto qh = x[ibl].qh;
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 4*ib + k;
float dl = d * ((x[ibl].scales_l[is] & 0xf) | (((x[ibl].scales_h[is%16] >> 4*(is/16)) & 0x03) << 4));
float ml = m * ((x[ibl].scales_l[is] >> 4) | (((x[ibl].scales_h[is%16] >> 4*(is/16)) & 0x0c) << 2));
for (int i = 0; i < 4; ++i) {
y4[k][QK_K*ibl+32*ib+i+ 0] = dl * ((ql[4*k+i+ 0] & 0xf) | ((qh[4*k+i] << 4) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+ 8] = dl * ((ql[4*k+i+ 0] >> 4) | ((qh[4*k+i] << 3) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+16] = dl * ((ql[4*k+i+16] & 0xf) | ((qh[4*k+i] >> 0) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+24] = dl * ((ql[4*k+i+16] >> 4) | ((qh[4*k+i] >> 1) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+ 4] = dl * ((ql[4*k+i+32] & 0xf) | ((qh[4*k+i] << 2) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+12] = dl * ((ql[4*k+i+32] >> 4) | ((qh[4*k+i] << 1) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+20] = dl * ((ql[4*k+i+48] & 0xf) | ((qh[4*k+i] >> 2) & 0x10)) - ml;
y4[k][QK_K*ibl+32*ib+i+28] = dl * ((ql[4*k+i+48] >> 4) | ((qh[4*k+i] >> 3) & 0x10)) - ml;
}
ql += 64;
qh += 16;
}
}
}
}
void vec_dot_q5_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
#if GGML_USE_IQK_MULMAT
if (iqk_mul_mat(1, 1, n, GGML_TYPE_Q5_K_R4, vx, 0, GGML_TYPE_Q8_K32, vy, 0, s, 0, 0, 1)) {
return;
}
#endif
GGML_ASSERT(n%QK4_NL == 0);
GGML_ASSERT(nrc == 1);
GGML_UNUSED(bs);
GGML_UNUSED(bx);
GGML_UNUSED(by);
}