From a63a96b5aea031f9dabf1e7c8d5ee28af170e9e7 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Tue, 10 Dec 2024 18:13:47 +0100 Subject: Q5_K_R4 (#132) * q5_k_r4: WIP * q5_k_r4: Zen4 and AVX2 We get PP-512(LLaMA-3.1-8B) = 248.3 t/s on Zen4. Q5_K_S has PP-512 = 190 t/s. * q5_k_r4: NEON We get PP-512(LLaMA-3.1-8B) = 96.1 t/s. --------- Co-authored-by: Iwan Kawrakow --- ggml/src/ggml-common.h | 9 +++++++++ 1 file changed, 9 insertions(+) (limited to 'ggml/src/ggml-common.h') diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 8ace3d6f..2d73d3f8 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -319,6 +319,15 @@ typedef struct { } block_q5_K; static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding"); +typedef struct { + ggml_half d[8]; + uint8_t scales_h[QK_K/16];// scales and mins, quantized with 6 bits + uint8_t scales_l[QK_K/8]; // scales and mins, quantized with 6 bits + uint8_t qh[QK_K/2]; // quants, high bit + uint8_t qs[QK_K*2]; // quants, low 4 bits +} block_q5_k_r4; +static_assert(sizeof(block_q5_k_r4) == 8*sizeof(ggml_half) + QK_K/16 + QK_K/8 + QK_K/2 + QK_K*2, "wrong q5_k_r4 block size/padding"); + // 6-bit quantization // weight is represented as x = a * q // 16 blocks of 16 elements each -- cgit v1.2.3