From f9307d79071c2a1e8efe10ecb1e1304bf77c021a Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Sun, 23 Mar 2025 17:10:52 +0100 Subject: Improve DeepSeek batched processing speed (#282) * Improve DeepSeek batched processing speed * Revert the commented out section in iqk_mul_mat.cpp It does have some benefit at long contexts. --------- Co-authored-by: Iwan Kawrakow --- ggml/src/iqk/iqk_mul_mat.cpp | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-) (limited to 'ggml/src') diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index 8b6d6b1c..4d29e2f0 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -17265,13 +17265,25 @@ template inline void iqk_deepseek_helper(KHelper& kh, VHelper& vh, int nq1, int nk1, int stride_q, int stride_m, int stride_qkv, const float * q, const char * mask, float scale, float softcap, float * qkv, float * M, float * S) { - if (nq1 % 8 == 0) { + if (nq1 >= 8) { FlashAttn<576, 512, 8, step_k> fa(scale, softcap); fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S); - } else { + } + else if (nq1 >= 4) { + FlashAttn<576, 512, 4, step_k> fa(scale, softcap); + fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S); + } + else { FlashAttn<576, 512, 1, step_k> fa(scale, softcap); fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S); } + //if (nq1 % 8 == 0) { + // FlashAttn<576, 512, 8, step_k> fa(scale, softcap); + // fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S); + //} else { + // FlashAttn<576, 512, 1, step_k> fa(scale, softcap); + // fa.compute(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, qkv, M, S); + //} } template -- cgit v1.2.3