Merge pull request #2143 from dhiltgen/llm_verbosity

Refine debug logging for llm

llm/dyn_ext_server.go

@@ -136,6 +136,12 @@ func newDynExtServer(library, model string, adapters, projectors []string, opts
 
 	sparams.n_threads = C.uint(opts.NumThread)
 
+	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+		sparams.verbose_logging = C.bool(true)
+	} else {
+		sparams.verbose_logging = C.bool(false)
+	}
+
 	slog.Info("Initializing llama server")
 	initResp := newExtServerResp(128)
 	defer freeExtServerResp(initResp)

llm/ext_server/ext_server.cpp

@@ -30,16 +30,17 @@ std::atomic<bool> ext_server_running(false);
 std::thread ext_server_thread;
 
 void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
-#if SERVER_VERBOSE != 1
-  log_disable();
-#endif
-  LOG_TEE("system info: %s\n", llama_print_system_info());
   assert(err != NULL && sparams != NULL);
+  log_set_target(stderr);
+  if (!sparams->verbose_logging) {
+    log_disable();
+  }
+
+  LOG_TEE("system info: %s\n", llama_print_system_info());
   err->id = 0;
   err->msg[0] = '\0';
   try {
     llama = new llama_server_context;
-    log_set_target(stdout);
     gpt_params params;
     params.n_ctx = sparams->n_ctx;
     params.n_batch = sparams->n_batch;

llm/ext_server/ext_server.h

@@ -45,6 +45,7 @@ typedef struct ext_server_params {
   bool embedding;        // get only sentence embedding
   ext_server_lora_adapter_t *lora_adapters;
   char *mmproj;
+  bool verbose_logging;  // Enable verbose logging of the server
 } ext_server_params_t;
 
 typedef struct ext_server_task_result {