Commit 15759e0ca for llama.cpp

commit 15759e0ca9d5caa7f9d31cd999c2434f17e7a1f5
Author: Aman Gupta <amangupta052@gmail.com>
Date:   Tue Nov 10 18:53:47 2026 +0800

    CUDA: skip graph compat check when graph is being reused (#32004)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index becd33483..f2c484731 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -5262,7 +5262,9 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,

     ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
     if (graph->is_enabled()) {
-        const bool graph_compatible = ggml_cuda_graph_check_compability(cgraph);
+        // graph->uid is set only for a compatible graph (see ggml_cuda_graph_update_required), so a reused graph needs no new check
+        const bool graph_reused     = cgraph->uid != 0 && cgraph->uid == graph->uid;
+        const bool graph_compatible = graph_reused || ggml_cuda_graph_check_compability(cgraph);
         if (graph_compatible) {
             const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph);