diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2025-02-09 09:14:52 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2025-02-09 09:14:52 +0200 |
| commit | 33390c4b74fa52875d6028c5c9aaf84f17288c25 (patch) | |
| tree | 97578e59530316c4b0637590d28e7e46ac8bb892 /ggml/src/ggml.c | |
| parent | 6d7b58eade37e45e3d8286a2353658047539d2b2 (diff) | |
Use Q8_K_128 for IQ1_S_R4 and IQ1_M_R4 matrix multiplications (#194)
* iq1_s_r4: Use Q8_K_128 instead of Q8_1_X4 for gemm (AVX2/Zen4)
* iq1_m_r4: Use Q8_K_128 instead of Q8_1_X4 for gemm (AVX2/Zen4)
* iq1_s_r4: Use Q8_K_128 instead of Q8_1_X4 for gemm (Neon)
* iq1_m_r4: Use Q8_K_128 instead of Q8_0_X4 for gemm (Neon)
* Simdify q8_K128 quantization also on Neon
* Cleanup
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml.c')
| -rw-r--r-- | ggml/src/ggml.c | 13 |
1 files changed, 11 insertions, 2 deletions
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index b19fb006..e07dd547 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1192,7 +1192,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_iq1_s_r4, .from_float_ref = (ggml_from_float_t)quantize_row_iq1_s_r4_ref, .vec_dot = vec_dot_iq1_s_r4_q8_k, - .vec_dot_type = GGML_TYPE_Q8_1_X4, + .vec_dot_type = GGML_TYPE_Q8_K128, .nrows = 1, .row_meta_size = 2, }, @@ -1218,7 +1218,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_iq1_m_r4, .from_float_ref = (ggml_from_float_t)quantize_row_iq1_m_r4_ref, .vec_dot = vec_dot_iq1_m_r4_q8_k, - .vec_dot_type = GGML_TYPE_Q8_0_X4, + .vec_dot_type = GGML_TYPE_Q8_K128, .nrows = 1, .row_meta_size = 2, }, @@ -1354,6 +1354,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_q8_K64, .row_meta_size = 0, }, + [GGML_TYPE_Q8_K128] = { + .type_name = "q8_K128", + .blck_size = 128, + .type_size = sizeof(block_q8_K128), + .is_quantized = true, + .from_float = quantize_row_q8_K128, + .row_meta_size = 0, + }, [GGML_TYPE_Q8_K16] = { .type_name = "q8_K16", .blck_size = 64, @@ -16161,6 +16169,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ1_M_R4: case GGML_TYPE_Q8_K: case GGML_TYPE_Q8_K64: + case GGML_TYPE_Q8_K128: case GGML_TYPE_Q8_K16: case GGML_TYPE_Q8_K32: case GGML_TYPE_Q4_0_4_4: |