Commit 016c1d914 for llama.cpp

commit 016c1d914c6e110318d20d2a8dbc2037a77e95f7
Author: ManuelPDM <153872964+ManuelPDM@users.noreply.github.com>
Date:   Thu Nov 5 05:18:38 2026 -0700

    CUDA: use highest compiled arch in get_mmvq_mmid_max_batch (#31679)

    Co-authored-by: DukeNukem-dev <54203477+DukeNukem-dev@users.noreply.github.com>

diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 1a9cf55cb..8f79123af 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -285,6 +285,7 @@ static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type
 int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
     // NVIDIA: P100, Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
     if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        cc = ggml_cuda_highest_compiled_arch(cc);
         if (cc == GGML_CUDA_CC_PASCAL || cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
             return MMVQ_MAX_BATCH_SIZE;
         }