ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-05 19:40:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

This does it for iq4_nl on Zen4, but FA does not work

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-23 19:45:59 +03:00
parent 8e30a22c80
commit 8b68a21aea
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/iqk/iqk_gemm_legacy_quants.cpp
View File

@@ -605,6 +605,14 @@ struct IQ4_NL_Dequantizer {
}
};
struct IQ4_NL0_Dequantizer {
Dequantizer4bit b4;
const __m256i values = load_iq4k_values_256();
inline __m256i dequant(const block_iq4_nl * x) const {
return _mm256_shuffle_epi8(values, b4.dequant(x->qs));
}
};
//=============================
static inline __m128i load_unsigned_mxfp4_values_128() {
static const uint8_t kvalues_mxfp4_unsigned[16] = {12, 13, 14, 15, 16, 18, 20, 24, 12, 11, 10, 9, 8, 6, 4, 0};
@@ -1945,7 +1953,7 @@ bool iqk_convert_legacy_quants_q8_r8(int type, int n, const void * vx, size_t bx
case GGML_TYPE_Q5_0 : iqk_convert_qX_q80_r8<block_q5_0, Q5_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q5_1 : iqk_convert_qX_1_q8_1_r8<block_q5_1, Q5_1_Dequantizer<block_q5_1>>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q6_0 : iqk_convert_qX_q80_r8<block_q6_0, Q6_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_NL: iqk_convert_qX_q80_r8<block_iq4_nl, IQ4_NL_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_NL: iqk_convert_qX_q80_r8<block_iq4_nl, IQ4_NL0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q8_0 : iqk_convert_q80_q80_r8(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_MXFP4 : iqk_convert_qX_q80_r8<block_mxfp4, MXFP40_Dequantizer>(n, vx, bx, vy, nrc_x); break;
default: return false;