From: Johannes Gäßler <redacted>
Date: Mon, 9 Dec 2024 19:07:12 +0000 (+0100)
Subject: CUDA: fix shared memory access condition for mmv (llama/10740)
X-Git-Tag: upstream/0.0.1642~79
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=e6a76b1abc8f1b4b84aa54bc665703aeb45380ce;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: fix shared memory access condition for mmv (llama/10740)
---

diff --git a/src/ggml-cuda/mmv.cu b/src/ggml-cuda/mmv.cu
index cfe91f42..a4b4f6bc 100644
--- a/src/ggml-cuda/mmv.cu
+++ b/src/ggml-cuda/mmv.cu
@@ -57,7 +57,7 @@ static __global__ void mul_mat_vec(
     if (block_size > WARP_SIZE) {
         buf_iw[tid/WARP_SIZE] = sumf;
         __syncthreads();
-        if (tid > WARP_SIZE) {
+        if (tid >= WARP_SIZE) {
             return;
         }
         sumf = buf_iw[tid];