Commit b019be0c8 for llama.cpp

commit b019be0c86e8adaf8b46627c091c24c1597b95f9
Author: Johannes Gäßler <johannesg@5d6.de>
Date:   Sat Nov 21 12:46:59 2026 +0100

    CUDA: abort on UB in fastdiv (#32698)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 053ed60f7..8cd2366ed 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -1046,6 +1046,12 @@ static const uint3 init_fastdiv_values(uint64_t d_64) {
     uint32_t L = 0;
     while (L < 32 && (uint32_t{ 1 } << L) < d) {
         L++;
+        if (L == 32) {
+            // Shifting uint32_t by 32 to the right is undefined behavior.
+            // On NVIDIA it will produce 0.
+            // On AMD the value by which is shifted is masked with & 31 so it is shifting by 0 bits.
+            GGML_ABORT("undefined behavior in fastdiv");
+        }
     }

     uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);