llama.cpp — Shutdown & Cleanup

4.1 Signal Handling

The server catches the following signals:

Signal	Platform	Handler Location	Behavior
SIGINT	Unix/macOS	server.cpp:302-306 sigaction()	Graceful shutdown
SIGTERM	Unix/macOS	server.cpp:307 sigaction()	Graceful shutdown
CTRL_C_EVENT	Windows	server.cpp:309-312 SetConsoleCtrlHandler()	Graceful shutdown

Signal Handler Implementation (server.cpp:27-36)

static void signal_handler(int signal) {
    if (is_terminating.test_and_set()) {
        // Second Ctrl+C: force exit
        fprintf(stderr, "Received second interrupt, terminating immediately.\n");
        exit(1);
    }
    shutdown_handler(signal);
}

Key design: an atomic_flag (is_terminating) ensures that:

• First SIGINT/SIGTERM: Triggers graceful shutdown via shutdown_handler
• Second SIGINT/SIGTERM: Calls exit(1) immediately — escape hatch if graceful shutdown hangs

4.2 Shutdown Sequence

Single-Model Server (server.cpp:294-344)

1. Signal received → signal_handler() called
2. Set terminating flag — is_terminating.test_and_set() prevents double-handling
3. Invoke shutdown_handler — calls ctx_server.terminate() (server.cpp:296)
4. ctx_server.terminate() — calls queue_tasks.terminate() which unblocks start_loop()
5. Main loop exits — start_loop() returns (server.cpp:336)
6. Execute cleanup (server.cpp:257-262):
   • ctx_http.stop() — stops HTTP server, closes listening socket
   • ctx_server.terminate() — waits for in-flight inference tasks
   • llama_backend_free() — releases GGML backend resources
7. Join HTTP thread (server.cpp:339-340) — ctx_http.thread.join()
8. Join monitor thread (server.cpp:342-343) — if child server mode
9. Print timing stats (server.cpp:346-353)
10. Exit with code 0

Router Server (server.cpp:233-248)

1. Signal received → signal_handler() → ctx_http.stop()
2. HTTP server stops — main thread unblocks from thread.join() (server.cpp:319-321)
3. Cleanup (server.cpp:236-242):
   • models_routes->models.unload_all() — terminates all child server processes
   • llama_backend_free()
4. Exit

4.3 Resource Cleanup Inventory

Resource	Cleanup Method	Location
HTTP server	ctx_http.stop()	server-http.cpp
Inference queue	queue_tasks.terminate()	server-queue.cpp
llama_context	llama_free()	llama.cpp (via server_context destructor)
llama_model	llama_model_free()	llama.cpp (via server_context destructor)
KV cache	Implicit in context free	llama-context.cpp:~llama_context
GGML backends	llama_backend_free()	ggml-backend.cpp
GPU VRAM	ggml_backend_buffer_free()	per-backend (cuda, metal, etc.)
Memory-mapped model	munmap() / UnmapViewOfFile()	llama-mmap.cpp:~llama_mmap
Child processes	models.unload_all()	server-models.cpp
Thread pool	ggml_threadpool_free()	ggml-threading.cpp
Sampler chain	llama_sampler_free()	llama-sampler.cpp
Grammar	llama_grammar_free()	llama-grammar.cpp
LoRA adapters	llama_adapter_lora_free()	llama-adapter.cpp