ggml-cuda: drop the CUDA graph cache count cap, as upstream does

ggml/src/ggml-cuda/common.cuh

@@ -1459,11 +1459,12 @@ struct ggml_backend_cuda_context {
     // Map from graph key to cuda_graph - allows multiple graphs per context when the
     // computation is split across CPU/GPU (e.g., with --n-cpu-moe), and when the same
     // split is called with different tensor shapes (e.g. a speculative verify batch)
-    std::unordered_map<uint64_t, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
-
     // a cuda graph instance is only valid for the shapes it captured, so a caller that
-    // alternates shapes needs one instance per shape to stay on the graph path
-    static const size_t max_cuda_graphs = 64;
+    // alternates shapes needs one instance per shape to stay on the graph path. the map is
+    // bounded only by the time based sweep below, as upstream does: tensor split mode alone
+    // needs about 2 * n_layers + 1 graphs per device and shape, so any count cap small enough
+    // to matter evicts graphs that are still in use and re-captures them on every token
+    std::unordered_map<uint64_t, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
 
     int64_t last_graph_eviction_sweep = 0;
 
@@ -1484,16 +1485,6 @@ struct ggml_backend_cuda_context {
 
         auto it = cuda_graphs.find(graph_key);
         if (it == cuda_graphs.end()) {
-            // a workload with many distinct shapes must not grow this without bound
-            while (cuda_graphs.size() >= max_cuda_graphs) {
-                auto lru = cuda_graphs.begin();
-                for (auto c = cuda_graphs.begin(); c != cuda_graphs.end(); ++c) {
-                    if (c->second->last_used_time < lru->second->last_used_time) {
-                        lru = c;
-                    }
-                }
-                cuda_graphs.erase(lru);
-            }
             it = cuda_graphs.emplace(graph_key, std::make_unique<ggml_cuda_graph>()).first;
         }
         it->second->last_used_time = time_now;