From e0a4002273907b2c414b6b5442d99e08bfe2df35 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sat, 23 Dec 2023 09:16:33 +0100
Subject: [PATCH] CUDA: fixed row rounding for 0 tensor splits (#4594)

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

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 7c2a834e3..490081cac 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -7937,12 +7937,16 @@ 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;
+                }
             }
         }
     }