Skip to content

Commit 3a561bb

Browse files
thomasantonythomasantony
committed
Update main.cpp to use new llama library
1 parent 5319183 commit 3a561bb

File tree

1 file changed

+36
-154
lines changed

1 file changed

+36
-154
lines changed

‎main.cpp

+36-154
Original file line numberDiff line numberDiff line change
@@ -50,25 +50,6 @@ void sigint_handler(int signo) {
5050
}
5151
#endif
5252

53-
const char * llama_print_system_info(void) {
54-
static std::string s;
55-
56-
s = "";
57-
s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | ";
58-
s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | ";
59-
s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | ";
60-
s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | ";
61-
s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | ";
62-
s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | ";
63-
s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | ";
64-
s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | ";
65-
s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
66-
s += "BLAS = " + std::to_string(ggml_cpu_has_blas()) + " | ";
67-
s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | ";
68-
s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | ";
69-
70-
return s.c_str();
71-
}
7253

7354
int main(int argc, char ** argv) {
7455
ggml_time_init();
@@ -97,49 +78,20 @@ int main(int argc, char ** argv) {
9778

9879
int64_t t_load_us = 0;
9980

100-
gpt_vocab vocab;
101-
llama_model model;
102-
10381
// load the model
104-
{
105-
const int64_t t_start_us = ggml_time_us();
106-
if (!llama_model_load(params.model, model, vocab, params.n_ctx)) {
107-
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
108-
return 1;
109-
}
110-
111-
t_load_us = ggml_time_us() - t_start_us;
112-
}
82+
llama_context* ctx_ptr = llama_init_from_params(params);
83+
llama_context & ctx = *ctx_ptr;
84+
gpt_vocab & vocab = llama_context_get_vocab(ctx);
11385

11486
// print system information
115-
{
116-
fprintf(stderr, "\n");
117-
fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
118-
params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
119-
}
120-
121-
int n_past = 0;
122-
123-
int64_t t_sample_us = 0;
124-
int64_t t_predict_us = 0;
125-
126-
std::vector<float> logits;
87+
llama_print_context_info(ctx);
12788

12889
// tokenize the prompt
129-
std::vector<gpt_vocab::id> embd_inp = ::llama_tokenize(vocab, params.prompt, true);
130-
131-
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
90+
std::vector<gpt_vocab::id> embd_inp = llama_tokenize_text(ctx, params.prompt);
13291

13392
// tokenize the reverse prompt
134-
std::vector<gpt_vocab::id> antiprompt_inp = ::llama_tokenize(vocab, params.antiprompt, false);
93+
std::vector<gpt_vocab::id> antiprompt_inp = llama_tokenize_text(ctx, params.prompt);
13594

136-
fprintf(stderr, "\n");
137-
fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
138-
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
139-
for (int i = 0; i < (int) embd_inp.size(); i++) {
140-
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
141-
}
142-
fprintf(stderr, "\n");
14395
if (params.interactive) {
14496
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
14597
struct sigaction sigint_action;
@@ -165,17 +117,6 @@ int main(int argc, char ** argv) {
165117
fprintf(stderr, "sampling parameters: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
166118
fprintf(stderr, "\n\n");
167119

168-
std::vector<gpt_vocab::id> embd;
169-
170-
// determine the required inference memory per token:
171-
size_t mem_per_token = 0;
172-
llama_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
173-
174-
int last_n_size = params.repeat_last_n;
175-
std::vector<gpt_vocab::id> last_n_tokens(last_n_size);
176-
std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
177-
178-
179120
if (params.interactive) {
180121
fprintf(stderr, "== Running in interactive mode. ==\n"
181122
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
@@ -185,8 +126,6 @@ int main(int argc, char ** argv) {
185126
" - If you want to submit another line, end your input in '\\'.\n");
186127
}
187128

188-
int remaining_tokens = params.n_predict;
189-
int input_consumed = 0;
190129
bool input_noecho = false;
191130

192131
// prompt user immediately after the starting prompt has been loaded
@@ -199,81 +138,39 @@ int main(int argc, char ** argv) {
199138
printf(ANSI_COLOR_YELLOW);
200139
}
201140

202-
while (remaining_tokens > 0) {
203-
// predict
204-
if (embd.size() > 0) {
205-
const int64_t t_start_us = ggml_time_us();
206-
207-
if (!llama_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
208-
fprintf(stderr, "Failed to predict\n");
209-
return 1;
210-
}
211-
212-
t_predict_us += ggml_time_us() - t_start_us;
213-
}
214-
215-
n_past += embd.size();
216-
embd.clear();
217-
218-
if (embd_inp.size() <= input_consumed) {
219-
// out of user input, sample next token
220-
const float top_k = params.top_k;
221-
const float top_p = params.top_p;
222-
const float temp = params.temp;
223-
const float repeat_penalty = params.repeat_penalty;
224-
225-
const int n_vocab = model.hparams.n_vocab;
226-
227-
gpt_vocab::id id = 0;
228-
229-
{
230-
const int64_t t_start_sample_us = ggml_time_us();
231-
232-
id = llama_sample_top_p_top_k(vocab, logits.data() + (logits.size() - n_vocab), last_n_tokens, repeat_penalty, top_k, top_p, temp, rng);
233-
234-
last_n_tokens.erase(last_n_tokens.begin());
235-
last_n_tokens.push_back(id);
141+
if(!llama_injest_input(ctx, params.prompt))
142+
{
143+
fprintf(stderr, "Failed to injest prompt\n");
144+
return 1;
145+
};
236146

237-
t_sample_us += ggml_time_us() - t_start_sample_us;
238-
}
147+
// display text
148+
input_noecho = false;
149+
const std::vector<gpt_vocab::id>& embd = llama_context_get_embd(ctx);
150+
if (!input_noecho) {
151+
for (auto id : embd) {
152+
printf("%s", vocab.id_to_token[id].c_str());
153+
}
154+
fflush(stdout);
155+
}
239156

240-
// add it to the context
241-
embd.push_back(id);
157+
if (!input_noecho && params.use_color) {
158+
printf(ANSI_COLOR_RESET);
159+
}
242160

243-
// echo this to console
244-
input_noecho = false;
161+
const std::vector<gpt_vocab::id>& last_n_tokens = llama_context_get_last_n_tokens(ctx);
245162

246-
// decrement remaining sampling budget
247-
--remaining_tokens;
248-
} else {
249-
// some user input remains from prompt or interaction, forward it to processing
250-
// Copy at most n_batch elements from embd_inp to embd
251-
size_t num_copied = std::min((size_t) params.n_batch, embd_inp.size() - input_consumed);
252-
std::copy(embd_inp.begin() + input_consumed, embd_inp.begin() + input_consumed + num_copied, std::back_inserter(embd));
253-
input_consumed += num_copied;
254-
255-
// Copy the last `last_n_size` elements copied into embd to last_n_tokens
256-
size_t num_copied_last_n = std::min(num_copied, (size_t) last_n_size);
257-
last_n_tokens.erase(last_n_tokens.begin(), last_n_tokens.begin()+num_copied_last_n);
258-
last_n_tokens.insert(last_n_tokens.end(), embd.end() - num_copied_last_n, embd.end());
259-
260-
// reset color to default if we there is no pending user input
261-
if (!input_noecho && params.use_color && embd_inp.size() == input_consumed) {
262-
printf(ANSI_COLOR_RESET);
263-
}
264-
}
265-
266-
// display text
267-
if (!input_noecho) {
268-
for (auto id : embd) {
269-
printf("%s", vocab.id_to_token[id].c_str());
270-
}
163+
while (llama_context_not_finished(ctx) > 0) {
164+
std::optional<gpt_vocab::id> model_output = llama_inference(ctx);
165+
if (model_output.has_value()) {
166+
printf("%s", vocab.id_to_token[model_output.value()].c_str());
271167
fflush(stdout);
272168
}
273169

170+
274171
// in interactive mode, and not currently processing queued inputs;
275172
// check if we should prompt the user for more
276-
if (params.interactive && embd_inp.size() <= input_consumed) {
173+
if (params.interactive) {
277174
// check for reverse prompt
278175
if (antiprompt_inp.size() && std::equal(antiprompt_inp.rbegin(), antiprompt_inp.rend(), last_n_tokens.rbegin())) {
279176
// reverse prompt found
@@ -303,13 +200,8 @@ int main(int argc, char ** argv) {
303200
buf[n_read] = '\n';
304201
buf[n_read+1] = 0;
305202
}
306-
307-
std::vector<gpt_vocab::id> line_inp = ::llama_tokenize(vocab, buf, false);
308-
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
309-
310-
remaining_tokens -= line_inp.size();
311-
312-
input_noecho = true; // do not echo this again
203+
// Do not clear existing context in interactive mode
204+
llama_init_context_with_prompt(ctx, buf, false);
313205
}
314206

315207
is_interacting = false;
@@ -322,24 +214,14 @@ int main(int argc, char ** argv) {
322214
break;
323215
}
324216
}
325-
326-
#if defined (_WIN32)
327-
signal(SIGINT, SIG_DFL);
328-
#endif
329-
330-
// report timing
217+
218+
// report timing from context
331219
{
332220
const int64_t t_main_end_us = ggml_time_us();
333-
334-
fprintf(stderr, "\n\n");
335-
fprintf(stderr, "%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
336-
fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
337-
fprintf(stderr, "%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
338-
fprintf(stderr, "%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
221+
llama_print_end_stats(ctx);
339222
fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
340223
}
341-
342-
ggml_free(model.ctx);
224+
llama_free_context(ctx_ptr);
343225

344226
if (params.use_color) {
345227
printf(ANSI_COLOR_RESET);

0 commit comments

Comments
 (0)