ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 10:21:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

IQ3_S_R4 (#162)

Browse Source 
* iq3_s_r4: WIP

* iq3_s_r4: Zen4

* iq3_s_r4: slightly better Zen4

* iq3_s_r4: AVX2

* iq3_s_r4: NEON

* iq3_s_r4: rearrange quants

* iq3_s_r4: rearranged quants - AVX2

* iq3_s_r4: rearranged quants - NEON

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-23 14:34:23 +01:00
committed by GitHub
parent 1a0a35dcd1
commit 167479e027
10 changed files with 394 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -5696,6 +5696,109 @@ void vec_dot_iq3_xxs_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_
GGML_UNUSED(by);
}
//
// ========================================= iq3_s_r4
//
void quantize_row_iq3_s_r4_ref(const float * x, block_iq3_s_r4 * y, int64_t k) {
quantize_iq3_s_r4(x, (void *)y, 4, k/4, nullptr);
}
void quantize_row_iq3_s_r4(const float * x, void * y, int64_t k) {
quantize_iq3_s_r4(x, y, 4, k/4, nullptr);
}
static void repack_iq3_s(int nrows, int n_per_row, const block_iq3_s * x, block_iq3_s_r4 * y) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
const block_iq3_s * 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].scales, 0, QK_K/16);
std::memset(y[ibl].signs, 0, QK_K/2);
std::memset(y[ibl].qh, 0, QK_K/8);
for (int k = 0; k < 4; ++k) {
y[ibl].d[k] = x4[k][ibl].d;
for (int ib = 0; ib < QK_K/64; ++ib) {
int j = 8*ib + k;
y[ibl].scales[(j+0)%16] |= ((x4[k][ibl].scales[ib] & 0xf) << 4*((j+0)/16));
y[ibl].scales[(j+4)%16] |= ((x4[k][ibl].scales[ib] >> 4) << 4*((j+4)/16));
}
for (int ib = 0; ib < QK_K/32; ++ib) {
y[ibl].qh[4*ib+k] = x4[k][ibl].qh[ib]; // leave ot like this?
for (int i = 0; i < 4; ++i) {
y[ibl].qs[32*ib+k+8*i+0] = x4[k][ibl].qs[8*ib+i+0];
y[ibl].qs[32*ib+k+8*i+4] = x4[k][ibl].qs[8*ib+i+4];
}
for (int i = 0; i < 4; ++i) {
y[ibl].signs[16*ib+4*k+i] = (((x4[k][ibl].signs[4*ib+0] >> i) & 1) << 0) | (((x4[k][ibl].signs[4*ib+0] >> (4+i)) & 1) << 1) |
(((x4[k][ibl].signs[4*ib+1] >> i) & 1) << 2) | (((x4[k][ibl].signs[4*ib+1] >> (4+i)) & 1) << 3) |
(((x4[k][ibl].signs[4*ib+2] >> i) & 1) << 4) | (((x4[k][ibl].signs[4*ib+2] >> (4+i)) & 1) << 5) |
(((x4[k][ibl].signs[4*ib+3] >> i) & 1) << 6) | (((x4[k][ibl].signs[4*ib+3] >> (4+i)) & 1) << 7);
}
}
}
}
x += 4*nblock;
y += nblock;
}
}
size_t quantize_iq3_s_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_IQ3_S, n_per_row);
std::vector<char> qtmp(4*row_size);
for (int row = 0; row < nrows; row += 4) {
quantize_iq3_s(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
repack_iq3_s(4, n_per_row, (const block_iq3_s *)qtmp.data(), (block_iq3_s_r4 *)qcur);
qcur += 4*row_size;
src += 4*n_per_row;
}
return nrows*row_size;
}
void dequantize_row_iq3_s_r4(const block_iq3_s_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 l = 4*ib + k;
float dl = d * (1 + 2*((x[ibl].scales[l%16] >> 4*(l/16)) & 0xf));
for (int i = 0; i < 4; ++i) {
auto grid1 = (const uint8_t *)(iq3s_grid + x[ibl].qs[32*ib+k+8*i+0] + ((x[ibl].qh[4*ib+k] << (8-i)) & 0x100));
auto grid2 = (const uint8_t *)(iq3s_grid + x[ibl].qs[32*ib+k+8*i+4] + ((x[ibl].qh[4*ib+k] << (4-i)) & 0x100));
for (int j = 0; j < 4; ++j) {
y4[k][QK_K*ibl+32*ib+4*i+ 0+j] = dl * grid1[j] * (x[ibl].signs[16*ib+4*k+j] & (1 << (i+0)) ? -1 : 1);
y4[k][QK_K*ibl+32*ib+4*i+16+j] = dl * grid2[j] * (x[ibl].signs[16*ib+4*k+j] & (1 << (i+4)) ? -1 : 1);
}
}
}
}
}
}
void vec_dot_iq3_s_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_IQ3_S_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);
}
//================================================
void iqk_repack_tensor(struct ggml_tensor * tensor) {
constexpr int kChunk = 8;
if (!tensor) return;
@@ -5711,6 +5814,11 @@ void iqk_repack_tensor(struct ggml_tensor * tensor) {
{ GGML_TYPE_IQ4_KS, { GGML_TYPE_IQ4_KS_R4, 4, (Repack::repack_func)repack_iq4_ks} },
{ GGML_TYPE_IQ4_NL, { GGML_TYPE_IQ4_NL_R4, 4, (Repack::repack_func)repack_iq4_nl} },
{ GGML_TYPE_IQ2_BN, { GGML_TYPE_IQ2_BN_R4, 4, (Repack::repack_func)repack_iq2_bn} },
{ GGML_TYPE_IQ2_XXS,{ GGML_TYPE_IQ2_XXS_R4,4, (Repack::repack_func)repack_iq2_xxs} },
{ GGML_TYPE_IQ2_XS, { GGML_TYPE_IQ2_XS_R4, 4, (Repack::repack_func)repack_iq2_xs} },
{ GGML_TYPE_IQ2_S, { GGML_TYPE_IQ2_S_R4, 4, (Repack::repack_func)repack_iq2_s} },
{ GGML_TYPE_IQ3_XXS,{ GGML_TYPE_IQ3_XXS_R4,4, (Repack::repack_func)repack_iq3_xxs} },
{ GGML_TYPE_IQ3_S, { GGML_TYPE_IQ3_S_R4, 4, (Repack::repack_func)repack_iq3_s} },
{ GGML_TYPE_Q2_K, { GGML_TYPE_Q2_K_R4, 4, (Repack::repack_func)repack_q2_k} },
{ GGML_TYPE_Q3_K, { GGML_TYPE_Q3_K_R4, 4, (Repack::repack_func)repack_q3_k} },
{ GGML_TYPE_Q4_K, { GGML_TYPE_Q4_K_R4, 4, (Repack::repack_func)repack_q4_k} },