ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-10 22:10:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

IQ4_K_R4 (#138)

Browse Source 
* iq4_k_r4: WIP

* iq4_k_r4: Zen4 and hopefully AVX2

On Zen4 we get PP-512(LLaMA-3.1-8B) = 232.6 t/s, up from 182.2 t/s
for iq4_k. Applying the extra shift costs a ~6 performance penalty.

* iq4_k_r4: AVX2

PP-512 = 227.60 t/s. The shifts are really costly.

* iq4_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 108 t/s, up from 58.2 t/s for iq4_k.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-12 16:04:20 +01:00
committed by GitHub
parent 66ade83e56
commit ce97b0325e
10 changed files with 443 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

114
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -4552,3 +4552,117 @@ void vec_dot_q2_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
GGML_UNUSED(by);
}
//
// ========================================= iq4_k_r4
//
void quantize_row_iq4_k_r4_ref(const float * x, block_iq4_k_r4 * y, int64_t k) {
quantize_iq4_k_r4(x, (void *)y, 4, k/4, nullptr);
}
void quantize_row_iq4_k_r4(const float * x, void * y, int64_t k) {
quantize_iq4_k_r4(x, y, 4, k/4, nullptr);
}
static void repack_iq4_k(int nrows, int n_per_row, const block_iq4_k * x, block_iq4_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_iq4_k * x4[4];
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].extra, 0, 8);
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] = x4[k][ibl].d;
auto extra = x4[k][ibl].extra;
for (int ib = 0; ib < QK_K/32; ++ib) {
if (extra & 1) y[ibl].extra[k+0] |= (1 << ib);
if (extra & 2) y[ibl].extra[k+4] |= (1 << ib);
extra >>= 2;
uint8_t sl1 = x4[k][ibl].scales_l[ib] & 0xf;
uint8_t sl2 = x4[k][ibl].scales_l[ib] >> 4;
uint8_t sh = x4[k][ibl].scales_h[ib/2] >> 4*(ib%2);
uint8_t sh1 = (sh >> 0) & 3;
uint8_t sh2 = (sh >> 2) & 3;
int i = 8*ib + k;
y[ibl].scales_l[i%32] |= (sl1 << 4*(i/32));
y[ibl].scales_h[i%16] |= (sh1 << 2*(i/16));
i += 4;
y[ibl].scales_l[i%32] |= (sl2 << 4*(i/32));
y[ibl].scales_h[i%16] |= (sh2 << 2*(i/16));
}
}
for (int ib = 0; ib < QK_K/32; ++ib) {
for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) {
y[ibl].qs[64*ib+4*k+i+ 0] = (x4[k][ibl].qs[16*ib+i+0] & 0xf) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0x0f) << 4); // 0....3 + 8...11 from each row
y[ibl].qs[64*ib+4*k+i+16] = (x4[k][ibl].qs[16*ib+i+0] >> 4) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0xf0)); // 16...19 + 24...27 from each row
y[ibl].qs[64*ib+4*k+i+32] = (x4[k][ibl].qs[16*ib+i+4] & 0xf) | ((x4[k][ibl].qs[16*ib+i+12] & 0x0f) << 4); // 4....7 + 12...15 from each row
y[ibl].qs[64*ib+4*k+i+48] = (x4[k][ibl].qs[16*ib+i+4] >> 4) | ((x4[k][ibl].qs[16*ib+i+12] & 0xf0)); // 20...23 + 28...31 from each row
}
}
}
x += 4*nblock;
y += nblock;
}
}
size_t quantize_iq4_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_IQ4_K, n_per_row);
std::vector<char> qtmp(4*row_size);
for (int row = 0; row < nrows; row += 4) {
quantize_iq4_k(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
repack_iq4_k(4, n_per_row, (const block_iq4_k *)qtmp.data(), (block_iq4_k_r4 *)qcur);
qcur += 4*row_size;
src += 4*n_per_row;
}
return nrows*row_size;
}
void dequantize_row_iq4_k_r4(const block_iq4_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]);
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 8*ib + k;
float dl1 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 3) << 4)) - 32);
is += 4;
float dl2 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 3) << 4)) - 32);
auto values1 = iq4k_values + (x[ibl].extra[k+0] & (1 << ib) ? 16 : 0);
auto values2 = iq4k_values + (x[ibl].extra[k+4] & (1 << ib) ? 16 : 0);
for (int i = 0; i < 4; ++i) {
y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+ 0] & 0xf];
y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+ 0] >> 4];
y4[k][QK_K*ibl+32*ib+i+16] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+16] & 0xf];
y4[k][QK_K*ibl+32*ib+i+24] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+16] >> 4];
y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+32] & 0xf];
y4[k][QK_K*ibl+32*ib+i+12] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+32] >> 4];
y4[k][QK_K*ibl+32*ib+i+20] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+48] & 0xf];
y4[k][QK_K*ibl+32*ib+i+28] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+48] >> 4];
}
}
}
}
}
void vec_dot_iq4_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_IQ4_K_R4, vx, 0, GGML_TYPE_Q8_K, 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);
}