生成速度的评估

[Vincent131499]

感谢杰出的工作！
对于下列表格的生成速度有疑问，请问这个速度(ms/token)是怎么计算出来的呢？

[li-plus]

耗时都是用这个 benchmark 在一台 Linux 服务器上实测的：

chatglm.cpp/chatglm_test.cpp

Lines 575 to 617 in 12f6865

	static void run_benchmark(const fs::path &model_path) {
	if (!fs::exists(model_path)) {
	GTEST_SKIP() << "Skipping benchmark test (model " << model_path << " not found)";
	}
	ggml_time_init();
	int64_t start_ms = ggml_time_ms();
	Pipeline pipeline(model_path.string());
	int64_t load_model_ms = ggml_time_ms() - start_ms;
	start_ms = ggml_time_ms();
	std::vector<std::string> history{"你好", "你好👋！我是人工智能助手 ChatGLM-6B，很高兴见到你，欢迎问我任何问题。",
	"晚上睡不着应该怎么办"};
	GenerationConfig gen_config;
	gen_config.do_sample = false;
	char *num_threads_env = getenv("CHATGLM_NUM_THREADS");
	if (num_threads_env) {
	gen_config.num_threads = std::stoi(num_threads_env);
	}
	PerfStreamer streamer;
	start_ms = ggml_time_ms();
	pipeline.chat(history, gen_config, &streamer);
	int64_t gen_s = (ggml_time_ms() - start_ms) / 1000.f;
	std::cout << "======== benchmark results for " << model_path.filename() << " ========\n"
	<< "using #threads: " << gen_config.num_threads << "\n"
	<< "model loaded within: " << load_model_ms << " ms\n"
	<< "generation finished within: " << gen_s << " s\n"
	<< streamer.to_string() << "\n"
	<< "===========================================================\n";
	}
	TEST(Benchmark, ChatGLM) {
	fs::path model_path = fs::path(__FILE__).parent_path() / "chatglm-ggml.bin";
	run_benchmark(model_path);
	}
	TEST(Benchmark, ChatGLM2) {
	fs::path model_path = fs::path(__FILE__).parent_path() / "chatglm2-ggml.bin";
	run_benchmark(model_path);
	}