From 43e65a672a98d931998559785b58f1e980e87f54 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Sun, 8 Dec 2024 15:27:13 +0100 Subject: Faster IQ4_XS_R4 on Zen4 (#128) * Faster iq4_xs_r4 on Zen4 The trick is to simply prepare the Q8 block sums for blocks of 32 as floats. This brings PP-512 up to 254.6 t/s from 224 t/s. * Fix broken matrix x vector product on Zen4 --------- Co-authored-by: Iwan Kawrakow --- ggml/src/ggml.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) (limited to 'ggml/src/ggml.c') diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 69bb6d88..974e42b2 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1124,6 +1124,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_q8_K16, .row_meta_size = 20, }, + [GGML_TYPE_Q8_K32] = { + .type_name = "q8_K32", + .blck_size = QK_K, + .type_size = sizeof(block_q8_K), + .is_quantized = true, + .from_float = quantize_row_q8_K32, + .row_meta_size = 0, + }, [GGML_TYPE_BF16] = { .type_name = "bf16", .blck_size = 1, @@ -1292,7 +1300,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_iq4_xs_r4, .from_float_ref = (ggml_from_float_t)quantize_row_iq4_xs_r4_ref, .vec_dot = vec_dot_iq4_xs_r4_q8_k, - .vec_dot_type = GGML_TYPE_Q8_K, + .vec_dot_type = GGML_TYPE_Q8_K32, .nrows = 1, .row_meta_size = 0, }, @@ -15633,6 +15641,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q8_K: case GGML_TYPE_Q8_K64: case GGML_TYPE_Q8_K16: + case GGML_TYPE_Q8_K32: case GGML_TYPE_Q4_0_4_4: case GGML_TYPE_Q4_0_4_8: case GGML_TYPE_Q4_0_8_8: -- cgit v1.2.3