From d0b52076da0261f291b01f1ffa44884c8b2cdb1c Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Thu, 27 Mar 2025 05:49:16 +0100 Subject: Use bf16 instead of fp16 block scales for q8_1 (#292) * WIP - not working * q8_0 without bells and wistles works * It works for q8_0 * Use bf16 instead of f16,int16 * q4_0_r8 * q5_0_r4 * q6_0_r4 * Also q4_1 and q5_1 * q8_0_r8 on avx2 --------- Co-authored-by: Iwan Kawrakow --- ggml/include/ggml.h | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'ggml/include') diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 91219d4a..7cc9100d 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -396,8 +396,9 @@ extern "C" { // GGML_TYPE_I2_S = 36, // - GGML_TYPE_Q8_0_X4 = 98, - GGML_TYPE_Q8_1_X4 = 99, + GGML_TYPE_Q8_0_X4 = 97, + GGML_TYPE_Q8_1_X4 = 98, + GGML_TYPE_Q8_2_X4 = 99, GGML_TYPE_Q6_0 = 133, GGML_TYPE_IQ1_BN = 134, GGML_TYPE_IQ2_BN = 135, -- cgit v1.2.3