From 3f54b4978672cc236bf38e707b24f036cb37f3f7 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Wed, 11 Jun 2025 15:01:34 +0300 Subject: Faster iq1_s GEMM via repacking to Q8_0_R8 (#517) TG is slightly faster too - 24.4 vs 23.1 t/s on the Ryzen-5975WX Co-authored-by: Iwan Kawrakow --- ggml/src/ggml.c | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'ggml/src/ggml.c') diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index bf5a8519..e4e954a1 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1205,7 +1205,11 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_iq1_s, .from_float_ref = (ggml_from_float_t)quantize_row_iq1_s_ref, .vec_dot = ggml_vec_dot_iq1_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, }, -- cgit v1.2.3