q5_1: 125 t/s -> 253 t/s

This commit is contained in:
Iwan Kawrakow
2025-06-18 11:43:09 +03:00
parent 8d393fccf5
commit c6048c478a
2 changed files with 3 additions and 2 deletions

View File

@@ -1837,7 +1837,7 @@ bool iqk_convert_legacy_quants_q8_r8(int type, int n, const void * vx, size_t bx
case GGML_TYPE_Q4_0 : iqk_convert_qX_q80_r8<block_q4_0, Q4_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q4_1 : iqk_convert_qX_1_q8_1_r8<block_q4_1, Q4_1_Dequantizer>(n, vx, bx, vy, nrc_x); break;
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>(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_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;

View File

@@ -261,6 +261,7 @@ struct MulMat {
case GGML_TYPE_Q4_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q4_1 : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;
case GGML_TYPE_Q5_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q5_1 : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;
case GGML_TYPE_Q6_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q8_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
@@ -412,7 +413,7 @@ bool iqk_convert_repack(int typeA, int n, const void * vx, size_t bx, void * vy,
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
//case GGML_TYPE_Q5_1:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q6_0:
case GGML_TYPE_Q8_0:
case GGML_TYPE_IQ4_NL: