1  #include "common.h"
  2  #include "llama.h"
  3  
  4  #include <cmath>
  5  #include <cstdio>
  6  #include <string>
  7  #include <vector>
  8  
  9  static void print_usage(int argc, char ** argv, const gpt_params & params) {
 10      gpt_params_print_usage(argc, argv, params);
 11  
 12      LOG_TEE("\nexample usage:\n");
 13      LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
 14      LOG_TEE("\n");
 15  }
 16  
 17  int main(int argc, char ** argv) {
 18      gpt_params params;
 19  
 20      params.prompt = "Hello my name is";
 21      params.n_predict = 32;
 22  
 23      if (!gpt_params_parse(argc, argv, params)) {
 24          print_usage(argc, argv, params);
 25          return 1;
 26      }
 27  
 28      // total length of the sequence including the prompt
 29      const int n_predict = params.n_predict;
 30  
 31      // init LLM
 32  
 33      llama_backend_init();
 34      llama_numa_init(params.numa);
 35  
 36      // initialize the model
 37  
 38      llama_model_params model_params = llama_model_params_from_gpt_params(params);
 39  
 40      llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 41  
 42      if (model == NULL) {
 43          fprintf(stderr , "%s: error: unable to load model\n" , __func__);
 44          return 1;
 45      }
 46  
 47      // initialize the context
 48  
 49      llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 50  
 51      llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 52  
 53      if (ctx == NULL) {
 54          fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
 55          return 1;
 56      }
 57  
 58      // tokenize the prompt
 59  
 60      std::vector<llama_token> tokens_list;
 61      tokens_list = ::llama_tokenize(ctx, params.prompt, true);
 62  
 63      const int n_ctx    = llama_n_ctx(ctx);
 64      const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size());
 65  
 66      LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
 67  
 68      // make sure the KV cache is big enough to hold all the prompt and generated tokens
 69      if (n_kv_req > n_ctx) {
 70          LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
 71          LOG_TEE("%s:        either reduce n_predict or increase n_ctx\n", __func__);
 72          return 1;
 73      }
 74  
 75      // print the prompt token-by-token
 76  
 77      fprintf(stderr, "\n");
 78  
 79      for (auto id : tokens_list) {
 80          fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
 81      }
 82  
 83      fflush(stderr);
 84  
 85      // create a llama_batch with size 512
 86      // we use this object to submit token data for decoding
 87  
 88      llama_batch batch = llama_batch_init(512, 0, 1);
 89  
 90      // evaluate the initial prompt
 91      for (size_t i = 0; i < tokens_list.size(); i++) {
 92          llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
 93      }
 94  
 95      // llama_decode will output logits only for the last token of the prompt
 96      batch.logits[batch.n_tokens - 1] = true;
 97  
 98      if (llama_decode(ctx, batch) != 0) {
 99          LOG_TEE("%s: llama_decode() failed\n", __func__);
100          return 1;
101      }
102  
103      // main loop
104  
105      int n_cur    = batch.n_tokens;
106      int n_decode = 0;
107  
108      const auto t_main_start = ggml_time_us();
109  
110      while (n_cur <= n_predict) {
111          // sample the next token
112          {
113              auto   n_vocab = llama_n_vocab(model);
114              auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
115  
116              std::vector<llama_token_data> candidates;
117              candidates.reserve(n_vocab);
118  
119              for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
120                  candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
121              }
122  
123              llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
124  
125              // sample the most likely token
126              const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
127  
128              // is it an end of generation?
129              if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
130                  LOG_TEE("\n");
131  
132                  break;
133              }
134  
135              LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
136              fflush(stdout);
137  
138              // prepare the next batch
139              llama_batch_clear(batch);
140  
141              // push this new token for next evaluation
142              llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
143  
144              n_decode += 1;
145          }
146  
147          n_cur += 1;
148  
149          // evaluate the current batch with the transformer model
150          if (llama_decode(ctx, batch)) {
151              fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
152              return 1;
153          }
154      }
155  
156      LOG_TEE("\n");
157  
158      const auto t_main_end = ggml_time_us();
159  
160      LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
161              __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
162  
163      llama_print_timings(ctx);
164  
165      fprintf(stderr, "\n");
166  
167      llama_batch_free(batch);
168  
169      llama_free(ctx);
170      llama_free_model(model);
171  
172      llama_backend_free();
173  
174      return 0;
175  }