| 1 | #include "arg.h" |
| 2 | #include "common.h" |
| 3 | #include "log.h" |
| 4 | #include "llama.h" |
| 5 | |
| 6 | #include <cmath> |
| 7 | #include <cstdio> |
| 8 | #include <string> |
| 9 | #include <vector> |
| 10 | #include <algorithm> |
| 11 | |
| 12 | static void print_usage(int, char ** argv) { |
| 13 | LOG("\nexample usage:\n" ); |
| 14 | LOG("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n" , argv[0]); |
| 15 | LOG("\n" ); |
| 16 | } |
| 17 | |
| 18 | int main(int argc, char ** argv) { |
| 19 | common_params params; |
| 20 | |
| 21 | params.n_junk = 250; |
| 22 | params.n_keep = 32; |
| 23 | params.i_pos = -1; |
| 24 | |
| 25 | if (!common_params_parse(argc, argv, params, ex: LLAMA_EXAMPLE_PASSKEY, print_usage)) { |
| 26 | return 1; |
| 27 | } |
| 28 | |
| 29 | common_init(); |
| 30 | |
| 31 | int n_junk = params.n_junk; |
| 32 | int n_keep = params.n_keep; |
| 33 | int n_grp = params.grp_attn_n; |
| 34 | int i_pos = params.i_pos; |
| 35 | |
| 36 | if (i_pos == -1) { |
| 37 | i_pos = rand() % n_junk; |
| 38 | } |
| 39 | |
| 40 | const std::string prompt_prefix = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there." ; |
| 41 | const std::string prompt_suffix = " What is the pass key? The pass key is" ; |
| 42 | |
| 43 | // generate junk text |
| 44 | params.prompt = prompt_prefix; |
| 45 | |
| 46 | const int passkey = rand() % 50000 + 1; |
| 47 | |
| 48 | for (int i = 0; i < n_junk; i++) { |
| 49 | if (i % n_junk == i_pos) { |
| 50 | params.prompt += " The pass key is " + std::to_string(val: passkey) + ". Remember it. " + std::to_string(val: passkey) + " is the pass key." ; |
| 51 | } |
| 52 | |
| 53 | params.prompt += " The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again." ; |
| 54 | } |
| 55 | |
| 56 | params.prompt += prompt_suffix; |
| 57 | |
| 58 | // init LLM |
| 59 | |
| 60 | llama_backend_init(); |
| 61 | llama_numa_init(numa: params.numa); |
| 62 | |
| 63 | // initialize the model |
| 64 | |
| 65 | llama_model_params model_params = common_model_params_to_llama(params); |
| 66 | |
| 67 | llama_model * model = llama_model_load_from_file(path_model: params.model.path.c_str(), params: model_params); |
| 68 | |
| 69 | if (model == NULL) { |
| 70 | LOG_ERR("%s: unable to load model\n" , __func__); |
| 71 | return 1; |
| 72 | } |
| 73 | |
| 74 | const llama_vocab * vocab = llama_model_get_vocab(model); |
| 75 | |
| 76 | // initialize the context |
| 77 | |
| 78 | llama_context_params ctx_params = common_context_params_to_llama(params); |
| 79 | |
| 80 | ctx_params.n_ctx = llama_model_n_ctx_train(model)*n_grp + n_keep; |
| 81 | |
| 82 | GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp" ); |
| 83 | |
| 84 | llama_context * ctx = llama_init_from_model(model, params: ctx_params); |
| 85 | if (ctx == NULL) { |
| 86 | LOG_ERR("%s: failed to create the llama_context\n" , __func__); |
| 87 | return 1; |
| 88 | } |
| 89 | |
| 90 | auto sparams = llama_sampler_chain_default_params(); |
| 91 | |
| 92 | llama_sampler * smpl = llama_sampler_chain_init(params: sparams); |
| 93 | |
| 94 | llama_sampler_chain_add(chain: smpl, smpl: llama_sampler_init_greedy()); |
| 95 | |
| 96 | // tokenize the prompt |
| 97 | std::vector<llama_token> tokens_list; |
| 98 | tokens_list = common_tokenize(ctx, text: params.prompt, add_special: true); |
| 99 | |
| 100 | // tokenize the prefix and use it as a sink |
| 101 | const int n_tokens_prefix = common_tokenize(ctx, text: prompt_prefix, add_special: true).size(); |
| 102 | |
| 103 | const int n_tokens_all = tokens_list.size(); |
| 104 | |
| 105 | // we leave a margin of 16 tokens for the generated text - it should contain just the passkey |
| 106 | const int n_predict = 16; |
| 107 | |
| 108 | // total length of the sequences including the prompt |
| 109 | const int n_len = n_tokens_all + n_predict; |
| 110 | |
| 111 | const int n_ctx = llama_n_ctx(ctx) - n_keep; |
| 112 | const int n_kv_req = llama_n_ctx(ctx); |
| 113 | const int n_batch = ctx_params.n_batch; |
| 114 | const int n_batch_grp = ctx_params.n_batch/n_grp; |
| 115 | |
| 116 | LOG_INF("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n" , __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos); |
| 117 | |
| 118 | // print the prompt token-by-token |
| 119 | |
| 120 | LOG_INF("\n" ); |
| 121 | LOG_INF("prefix tokens: %d\n" , n_tokens_prefix); |
| 122 | LOG_INF("prompt tokens: %d\n" , n_tokens_all); |
| 123 | //LOG_INF("prompt: %s\n", params.prompt.c_str()); |
| 124 | |
| 125 | llama_batch batch = llama_batch_init(n_tokens: params.n_batch, embd: 0, n_seq_max: 1); |
| 126 | |
| 127 | int n_past = 0; |
| 128 | |
| 129 | auto * mem = llama_get_memory(ctx); |
| 130 | |
| 131 | // fill the KV cache |
| 132 | for (int i = 0; i < n_ctx; i += n_batch) { |
| 133 | if (i > 0 && n_grp > 1) { |
| 134 | // if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp |
| 135 | const int ib = i/n_batch - 1; |
| 136 | const int bd = n_batch_grp*(n_grp - 1); |
| 137 | |
| 138 | llama_memory_seq_add(mem, seq_id: 0, p0: n_past - n_batch, p1: n_past, delta: ib*bd); |
| 139 | llama_memory_seq_div(mem, seq_id: 0, p0: n_past - n_batch + ib*bd, p1: n_past + ib*bd, d: n_grp); |
| 140 | |
| 141 | n_past = llama_memory_seq_pos_max(mem, seq_id: 0) + 1; |
| 142 | } |
| 143 | |
| 144 | common_batch_clear(batch); |
| 145 | |
| 146 | for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) { |
| 147 | common_batch_add(batch, id: tokens_list[i + j], pos: n_past++, seq_ids: { 0 }, logits: false); |
| 148 | } |
| 149 | |
| 150 | if (i + n_batch >= n_tokens_all) { |
| 151 | batch.logits[batch.n_tokens - 1] = true; |
| 152 | } |
| 153 | |
| 154 | if (llama_decode(ctx, batch) != 0) { |
| 155 | LOG_INF("%s: llama_decode() failed\n" , __func__); |
| 156 | return 1; |
| 157 | } |
| 158 | |
| 159 | LOG_INF("%s: processed: [%6d, %6d)\n" , __func__, i, std::min(i + n_batch, n_tokens_all)); |
| 160 | |
| 161 | if (i + n_batch >= n_tokens_all) { |
| 162 | break; |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | for (int i = n_ctx; i < n_tokens_all; i += n_batch) { |
| 167 | const int n_discard = n_batch; |
| 168 | |
| 169 | LOG_INF("%s: shifting KV cache with %d\n" , __func__, n_discard); |
| 170 | |
| 171 | llama_memory_seq_rm (mem, seq_id: 0, p0: n_keep , p1: n_keep + n_discard); |
| 172 | llama_memory_seq_add(mem, seq_id: 0, p0: n_keep + n_discard, p1: n_ctx, delta: -n_discard); |
| 173 | |
| 174 | n_past = llama_memory_seq_pos_max(mem, seq_id: 0) + 1; |
| 175 | |
| 176 | common_batch_clear(batch); |
| 177 | |
| 178 | for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) { |
| 179 | common_batch_add(batch, id: tokens_list[i + j], pos: n_past++, seq_ids: { 0 }, logits: false); |
| 180 | } |
| 181 | |
| 182 | if (i + n_batch >= n_tokens_all) { |
| 183 | batch.logits[batch.n_tokens - 1] = true; |
| 184 | } |
| 185 | |
| 186 | if (llama_decode(ctx, batch) != 0) { |
| 187 | LOG_ERR("%s: llama_decode() failed\n" , __func__); |
| 188 | return 1; |
| 189 | } |
| 190 | |
| 191 | LOG_INF("%s: processed: [%6d, %6d)\n" , __func__, i, std::min(i + n_batch, n_tokens_all)); |
| 192 | } |
| 193 | |
| 194 | { |
| 195 | const int n_discard = n_past - n_ctx + n_predict; |
| 196 | |
| 197 | if (n_discard > 0) { |
| 198 | LOG_INF("%s: shifting KV cache with %d to free space for the answer\n" , __func__, n_discard); |
| 199 | |
| 200 | llama_memory_seq_rm (mem, seq_id: 0, p0: n_keep , p1: n_keep + n_discard); |
| 201 | llama_memory_seq_add(mem, seq_id: 0, p0: n_keep + n_discard, p1: n_ctx, delta: -n_discard); |
| 202 | |
| 203 | n_past = llama_memory_seq_pos_max(mem, seq_id: 0) + 1; |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | LOG_INF("\n" ); |
| 208 | LOG_INF("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n" , __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk); |
| 209 | LOG_INF("\n" ); |
| 210 | |
| 211 | // main loop |
| 212 | |
| 213 | int n_cur = n_tokens_all; |
| 214 | int n_decode = 0; |
| 215 | |
| 216 | LOG_INF("%s" , prompt_suffix.c_str()); |
| 217 | |
| 218 | const auto t_main_start = ggml_time_us(); |
| 219 | |
| 220 | while (n_cur <= n_len) { |
| 221 | // sample the next token |
| 222 | { |
| 223 | const llama_token new_token_id = llama_sampler_sample(smpl, ctx, idx: batch.n_tokens - 1); |
| 224 | |
| 225 | // is it an end of generation? |
| 226 | if (llama_vocab_is_eog(vocab, token: new_token_id) || n_cur == n_len) { |
| 227 | LOG("\n" ); |
| 228 | |
| 229 | break; |
| 230 | } |
| 231 | |
| 232 | LOG("%s" , common_token_to_piece(ctx, new_token_id).c_str()); |
| 233 | |
| 234 | n_decode += 1; |
| 235 | |
| 236 | // prepare the next batch |
| 237 | common_batch_clear(batch); |
| 238 | |
| 239 | // push this new token for next evaluation |
| 240 | common_batch_add(batch, id: new_token_id, pos: n_past++, seq_ids: { 0 }, logits: true); |
| 241 | } |
| 242 | |
| 243 | n_cur += 1; |
| 244 | |
| 245 | // evaluate the current batch with the transformer model |
| 246 | if (llama_decode(ctx, batch)) { |
| 247 | LOG_ERR("%s : failed to eval, return code %d\n" , __func__, 1); |
| 248 | return 1; |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | LOG("\n" ); |
| 253 | |
| 254 | const auto t_main_end = ggml_time_us(); |
| 255 | |
| 256 | LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n" , |
| 257 | __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f)); |
| 258 | |
| 259 | LOG("\n" ); |
| 260 | llama_perf_context_print(ctx); |
| 261 | |
| 262 | LOG("\n" ); |
| 263 | |
| 264 | llama_sampler_free(smpl); |
| 265 | |
| 266 | llama_batch_free(batch); |
| 267 | |
| 268 | llama_free(ctx); |
| 269 | llama_model_free(model); |
| 270 | |
| 271 | llama_backend_free(); |
| 272 | |
| 273 | return 0; |
| 274 | } |
| 275 | |