From b757f0d5cf7648a916ff5f44c869a5ddd26e1023 Mon Sep 17 00:00:00 2001
From: Oliver Simons <osimons@nvidia.com>
Date: Fri, 2 Oct 2026 18:48:50 +0200
Subject: [PATCH] Avoid races in fp8_fallback

---
 ggml/src/ggml-cuda/fp8.cu | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ggml/src/ggml-cuda/fp8.cu b/ggml/src/ggml-cuda/fp8.cu
index c7fa563501..82313caf4b 100644
--- a/ggml/src/ggml-cuda/fp8.cu
+++ b/ggml/src/ggml-cuda/fp8.cu
@@ -24,6 +24,7 @@ static __global__ void mul_mat_fp8_fallback(
         if (threadIdx.x == 0) {
             *(float *) (dst + i0*sizeof(float) + i1*nb1 + i2*nb2 + i3*nb3) = sum;
         }
+        __syncthreads();
     }
 }