From 8ab016e2d54ce3d95aa04ef6e63fdb2b75fc0d99 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Thu, 26 Mar 2026 17:22:27 +0100 Subject: [PATCH] Fix bug in CPU floash attention for bf16 KV cache (#1521) --- ggml/src/iqk/fa/iqk_fa_templates.h | 1 + 1 file changed, 1 insertion(+) diff --git a/ggml/src/iqk/fa/iqk_fa_templates.h b/ggml/src/iqk/fa/iqk_fa_templates.h index 648e7308..2cd0c18f 100644 --- a/ggml/src/iqk/fa/iqk_fa_templates.h +++ b/ggml/src/iqk/fa/iqk_fa_templates.h @@ -2008,6 +2008,7 @@ struct FlashAttnBF16 { q += q_step*stride_q; mask += q_step*stride_m; qkv += q_step*stride_qkv; + if (M && S) { M += q_step; S += q_step; } } int n_left = nq1 - q_step*(nq1/q_step); if (n_left > 0) {