11
static void print_usage(int, char ** argv) {
12
LOG("\nexample usage:\n");
13
LOG("\n %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
17
int main(int argc, char ** argv) {
20
params.prompt = "Hello my name is";
21
params.n_predict = 32;
23
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
29
// number of parallel batches
30
int n_parallel = params.n_parallel;
32
// total length of the sequences including the prompt
33
int n_predict = params.n_predict;
38
llama_numa_init(params.numa);
40
// initialize the model
42
llama_model_params model_params = llama_model_params_from_gpt_params(params);
44
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
47
LOG_ERR("%s: error: unable to load model\n" , __func__);
51
// tokenize the prompt
53
std::vector<llama_token> tokens_list;
54
tokens_list = ::llama_tokenize(model, params.prompt, true);
56
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
58
// initialize the context
60
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
62
ctx_params.n_ctx = n_kv_req;
63
ctx_params.n_batch = std::max(n_predict, n_parallel);
65
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
67
auto sparams = llama_sampler_chain_default_params();
69
llama_sampler * smpl = llama_sampler_chain_init(sparams);
71
llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
72
llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
73
llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
74
llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
77
LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
81
const int n_ctx = llama_n_ctx(ctx);
83
LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
85
// make sure the KV cache is big enough to hold all the prompt and generated tokens
86
if (n_kv_req > n_ctx) {
87
LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__, n_kv_req);
88
LOG_ERR("%s: either reduce n_parallel or increase n_ctx\n", __func__);
92
// print the prompt token-by-token
96
for (auto id : tokens_list) {
97
LOG("%s", llama_token_to_piece(ctx, id).c_str());
100
// create a llama_batch
101
// we use this object to submit token data for decoding
102
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
104
std::vector<llama_seq_id> seq_ids(n_parallel, 0);
105
for (int32_t i = 0; i < n_parallel; ++i) {
109
// evaluate the initial prompt
110
for (size_t i = 0; i < tokens_list.size(); ++i) {
111
llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
113
GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
115
if (llama_model_has_encoder(model)) {
116
if (llama_encode(ctx, batch)) {
117
LOG_ERR("%s : failed to eval\n", __func__);
121
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
122
if (decoder_start_token_id == -1) {
123
decoder_start_token_id = llama_token_bos(model);
126
llama_batch_clear(batch);
127
llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
130
// llama_decode will output logits only for the last token of the prompt
131
batch.logits[batch.n_tokens - 1] = true;
133
if (llama_decode(ctx, batch) != 0) {
134
LOG_ERR("%s: llama_decode() failed\n", __func__);
138
//// assign the system KV cache to all parallel sequences
139
//// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
140
//for (int32_t i = 1; i < n_parallel; ++i) {
141
// llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
144
if (n_parallel > 1) {
145
LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
150
// we will store the parallel decoded sequences in this vector
151
std::vector<std::string> streams(n_parallel);
153
// remember the batch index of the last token for each parallel sequence
154
// we need this to determine which logits to sample from
155
std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
157
int n_cur = batch.n_tokens;
160
const auto t_main_start = ggml_time_us();
162
while (n_cur <= n_predict) {
163
// prepare the next batch
164
llama_batch_clear(batch);
166
// sample the next token for each parallel sequence / stream
167
for (int32_t i = 0; i < n_parallel; ++i) {
168
if (i_batch[i] < 0) {
169
// the stream has already finished
173
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
175
// is it an end of generation? -> mark the stream as finished
176
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
179
if (n_parallel > 1) {
180
LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
186
// if there is only one stream, we print immediately to stdout
187
if (n_parallel == 1) {
188
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
191
streams[i] += llama_token_to_piece(ctx, new_token_id);
193
i_batch[i] = batch.n_tokens;
195
// push this new token for next evaluation
196
llama_batch_add(batch, new_token_id, n_cur, { i }, true);
201
// all streams are finished
202
if (batch.n_tokens == 0) {
208
// evaluate the current batch with the transformer model
209
if (llama_decode(ctx, batch)) {
210
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
215
if (n_parallel > 1) {
218
for (int32_t i = 0; i < n_parallel; ++i) {
219
LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
223
const auto t_main_end = ggml_time_us();
225
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
226
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
229
llama_perf_sampler_print(smpl);
230
llama_perf_context_print(ctx);
232
fprintf(stderr, "\n");
234
llama_batch_free(batch);
236
llama_sampler_free(smpl);
238
llama_free_model(model);
240
llama_backend_free();