From: Johannes Gäßler <redacted>
Date: Sat, 23 Dec 2023 08:16:33 +0000 (+0100)
Subject: CUDA: fixed row rounding for 0 tensor splits (llama/4594)
X-Git-Tag: upstream/0.0.1642~1163
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=a7db4da893414f79052fe11e93d71b69f8b19297;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: fixed row rounding for 0 tensor splits (llama/4594)
---

diff --git a/src/ggml-cuda.cu b/src/ggml-cuda.cu
index 7c2a834e..490081ca 100644
--- a/src/ggml-cuda.cu
+++ b/src/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;
+                }
             }
         }
     }