8
static void print_usage(int, char ** argv) {
9
LOG("\nexample usage:\n");
10
LOG("\n %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
14
int main(int argc, char ** argv) {
17
params.prompt = "Hello my name is";
18
params.n_predict = 32;
20
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
26
// total length of the sequence including the prompt
27
const int n_predict = params.n_predict;
32
llama_numa_init(params.numa);
34
// initialize the model
36
llama_model_params model_params = llama_model_params_from_gpt_params(params);
38
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
41
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
45
// initialize the context
47
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
49
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
52
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
56
auto sparams = llama_sampler_chain_default_params();
58
sparams.no_perf = false;
60
llama_sampler * smpl = llama_sampler_chain_init(sparams);
62
llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
64
// tokenize the prompt
66
std::vector<llama_token> tokens_list;
67
tokens_list = ::llama_tokenize(ctx, params.prompt, true);
69
const int n_ctx = llama_n_ctx(ctx);
70
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size());
73
LOG_INF("%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
75
// make sure the KV cache is big enough to hold all the prompt and generated tokens
76
if (n_kv_req > n_ctx) {
77
LOG_ERR("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
78
LOG_ERR("%s: either reduce n_predict or increase n_ctx\n", __func__);
82
// print the prompt token-by-token
86
for (auto id : tokens_list) {
87
LOG("%s", llama_token_to_piece(ctx, id).c_str());
90
// create a llama_batch with size 512
91
// we use this object to submit token data for decoding
93
llama_batch batch = llama_batch_init(512, 0, 1);
95
// evaluate the initial prompt
96
for (size_t i = 0; i < tokens_list.size(); i++) {
97
llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
100
// llama_decode will output logits only for the last token of the prompt
101
batch.logits[batch.n_tokens - 1] = true;
103
if (llama_decode(ctx, batch) != 0) {
104
LOG("%s: llama_decode() failed\n", __func__);
110
int n_cur = batch.n_tokens;
113
const auto t_main_start = ggml_time_us();
115
while (n_cur <= n_predict) {
116
// sample the next token
118
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, -1);
120
// is it an end of generation?
121
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
127
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
130
// prepare the next batch
131
llama_batch_clear(batch);
133
// push this new token for next evaluation
134
llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
141
// evaluate the current batch with the transformer model
142
if (llama_decode(ctx, batch)) {
143
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
150
const auto t_main_end = ggml_time_us();
152
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
153
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
156
llama_perf_sampler_print(smpl);
157
llama_perf_context_print(ctx);
161
llama_batch_free(batch);
162
llama_sampler_free(smpl);
164
llama_free_model(model);
166
llama_backend_free();