ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-21 15:09:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_s: use q8_k_r8

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-06-15 18:29:12 +03:00
parent a2f5c251fb
commit 6760096deb
3 changed files with 61 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml.c
View File

@@ -1158,11 +1158,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.from_float = quantize_row_iq3_s,
.from_float_ref = (ggml_from_float_t)quantize_row_iq3_s_ref,
.vec_dot = ggml_vec_dot_iq3_s_q8_K,
#ifdef __AVX2__
.vec_dot_type = GGML_TYPE_Q8_2_X4,
#else
.vec_dot_type = GGML_TYPE_Q8_K,
#endif
.nrows = 1,
.row_meta_size = 0,
},

77
ggml/src/iqk/iqk_gemm_iquants.cpp
View File

@@ -2554,6 +2554,49 @@ void iqk_convert_iq3_xxs_q8_0_r8(int n, const void * vx, size_t bx, void * vy, i
}
}
void iqk_convert_iq3_s_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
int nb = n/QK_K;
const block_iq3_s * x8[8];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
int16_t ls[16];
SignHelper sh;
IndexHelperIQ3S helper;
uint32_t block[8];
__m256i values[8];
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq3_s *)((const char *)vx + (ix + k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
float d = GGML_FP16_TO_FP32(x8[k][i].d);
auto qs = x8[k][i].qs;
auto qh = x8[k][i].qh;
auto signs = (const uint16_t *)x8[k][i].signs;
helper.make2(qs+ 0, qh+0, values+0);
helper.make2(qs+16, qh+2, values+2);
sh.sign_4_values(signs+0, values+0);
helper.make2(qs+32, qh+4, values+4);
helper.make2(qs+48, qh+6, values+6);
sh.sign_4_values(signs+8, values+4);
for (int ib32 = 0; ib32 < 8; ++ib32) {
ls[2*ib32 + 0] = (2*((x8[k][i].scales[ib32/2] >> 4*(ib32%2)) & 0xf) + 1);
ls[2*ib32 + 1] = ls[2*ib32 + 0];
}
float dnew = convert_to_q8_k_r8(k, 127, values, ls, block, y[i].qs);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
}
y += nb;
}
}
void iqk_convert_iq3_s_q8_0_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
@@ -2661,22 +2704,22 @@ bool iqk_set_kernels_iquants(int ne00, int typeA, int typeB, std::array<mul_mat_
// return false;
//}
if (ggml_type(typeA) == GGML_TYPE_IQ3_S) {
if (ggml_type(typeB) == GGML_TYPE_Q8_2_X4) {
//IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_K_q8_2_IQ_N, DequantizerIQ3S, kernels);
kernels[0] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 1, 8>;
kernels[1] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 2, 8>;
kernels[2] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 3, 8>;
kernels[3] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 4, 8>;
kernels[4] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 5, 8>;
kernels[5] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 6, 8>;
kernels[6] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 7, 8>;
kernels[7] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 8, 8>;
func16 = nullptr;
return true;
}
return false;
}
//if (ggml_type(typeA) == GGML_TYPE_IQ3_S) {
// if (ggml_type(typeB) == GGML_TYPE_Q8_2_X4) {
// //IQK_SET_MUL_MAT_FUNCTIONS_T(mul_mat_qX_K_q8_2_IQ_N, DequantizerIQ3S, kernels);
// kernels[0] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 1, 8>;
// kernels[1] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 2, 8>;
// kernels[2] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 3, 8>;
// kernels[3] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 4, 8>;
// kernels[4] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 5, 8>;
// kernels[5] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 6, 8>;
// kernels[6] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 7, 8>;
// kernels[7] = mul_mat_qX_K_q8_2_IQ_N<DequantizerIQ3S, 8, 8>;
// func16 = nullptr;
// return true;
// }
// return false;
//}
if (ggml_type(typeB) != GGML_TYPE_Q8_K) {
return false;
@@ -2739,7 +2782,7 @@ bool iqk_convert_iquants_q80_r8(int type, int n, const void * vx, size_t bx, voi
case GGML_TYPE_IQ2_XS : iqk_convert_iq2_xs_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ2_S : iqk_convert_iq2_s_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ3_XXS: iqk_convert_iq3_xxs_q8_k_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ3_S : iqk_convert_iq3_s_q8_0_r8 (n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ3_S : iqk_convert_iq3_s_q8_k_r8 (n, vx, bx, vy, nrc_x); break;
default: return false;
}
return true;

2
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -243,7 +243,7 @@ struct MulMat {
case GGML_TYPE_IQ2_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_S : return nrc_y >= 16 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ3_S : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ3_S : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ1_S : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q3_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_Q4_K : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;