From 1b56724ae3bb3f05e0bc1602d99cca04600c9bbe Mon Sep 17 00:00:00 2001
From: JohannesGaessler <johannesg@5d6.de>
Date: Fri, 22 Dec 2023 10:32:39 +0100
Subject: [PATCH] CUDA: fixed row rounding for 0 tensor splits

---
 ggml-cuda.cu | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index ac91ee12e..51db2c045 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -7929,12 +7929,16 @@ static void ggml_cuda_op_mul_mat(
 
             if (id != 0) {
                 row_low[id]  = ne01*g_tensor_split[id];
-                row_low[id] -= row_low[id] % rounding;
+                if (row_low[id] < ne01) {
+                    row_low[id] -= row_low[id] % rounding;
+                }
             }
 
             if (id != g_device_count - 1) {
                 row_high[id]  = ne01*g_tensor_split[id + 1];
-                row_high[id] -= row_high[id] % rounding;
+                if (row_high[id] < ne01) {
+                    row_high[id] -= row_high[id] % rounding;
+                }
             }
         }
     }