From 5a5f1038337b1b38db76872ecc0c19545bce45ed Mon Sep 17 00:00:00 2001
From: Zonghang Li <lizhuestc@gmail.com>
Date: Wed, 16 Apr 2025 08:55:07 +0400
Subject: [PATCH] fix q6k and q80

---
 common/profiler.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/common/profiler.cpp b/common/profiler.cpp
index 48af0950..69a20af0 100644
--- a/common/profiler.cpp
+++ b/common/profiler.cpp
@@ -370,7 +370,7 @@ float device_inp_embd_delay(struct llama_model * model, enum ggml_type src0t, in
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_Q8_0:
-            QK_K = 256;
+            QK_K = ggml_blck_size(src0t);
             matrix_B = malloc((embd_size / QK_K) * ggml_type_size(src0t));
             break;
         default: