|
| 1 | +#include "llama.h" |
| 2 | +#include <cstdio> |
| 3 | +#include <cstring> |
| 4 | +#include <iostream> |
| 5 | +#include <string> |
| 6 | +#include <vector> |
| 7 | + |
| 8 | +static void print_usage(int, char ** argv) { |
| 9 | + printf("\nexample usage:\n"); |
| 10 | + printf("\n %s -m model.gguf [-c context_size] [-ngl n_gpu_layers]\n", argv[0]); |
| 11 | + printf("\n"); |
| 12 | +} |
| 13 | + |
| 14 | +int main(int argc, char ** argv) { |
| 15 | + // path to the model gguf file |
| 16 | + std::string model_path; |
| 17 | + // number of layers to offload to the GPU |
| 18 | + int ngl = 99; |
| 19 | + int n_ctx = 2048; |
| 20 | + |
| 21 | + // parse command line arguments |
| 22 | + for (int i = 1; i < argc; i++) { |
| 23 | + try { |
| 24 | + if (strcmp(argv[i], "-m") == 0) { |
| 25 | + if (i + 1 < argc) { |
| 26 | + model_path = argv[++i]; |
| 27 | + } else { |
| 28 | + print_usage(argc, argv); |
| 29 | + return 1; |
| 30 | + } |
| 31 | + } else if (strcmp(argv[i], "-c") == 0) { |
| 32 | + if (i + 1 < argc) { |
| 33 | + n_ctx = std::stoi(argv[++i]); |
| 34 | + } else { |
| 35 | + print_usage(argc, argv); |
| 36 | + return 1; |
| 37 | + } |
| 38 | + } else if (strcmp(argv[i], "-ngl") == 0) { |
| 39 | + if (i + 1 < argc) { |
| 40 | + ngl = std::stoi(argv[++i]); |
| 41 | + } else { |
| 42 | + print_usage(argc, argv); |
| 43 | + return 1; |
| 44 | + } |
| 45 | + } else { |
| 46 | + print_usage(argc, argv); |
| 47 | + return 1; |
| 48 | + } |
| 49 | + } catch (std::exception & e) { |
| 50 | + fprintf(stderr, "error: %s\n", e.what()); |
| 51 | + print_usage(argc, argv); |
| 52 | + return 1; |
| 53 | + } |
| 54 | + } |
| 55 | + if (model_path.empty()) { |
| 56 | + print_usage(argc, argv); |
| 57 | + return 1; |
| 58 | + } |
| 59 | + |
| 60 | + // only print errors |
| 61 | + llama_log_set([](enum ggml_log_level level, const char * text, void * /* user_data */) { |
| 62 | + if (level >= GGML_LOG_LEVEL_ERROR) { |
| 63 | + fprintf(stderr, "%s", text); |
| 64 | + } |
| 65 | + }, nullptr); |
| 66 | + |
| 67 | + // initialize the model |
| 68 | + llama_model_params model_params = llama_model_default_params(); |
| 69 | + model_params.n_gpu_layers = ngl; |
| 70 | + |
| 71 | + llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params); |
| 72 | + if (!model) { |
| 73 | + fprintf(stderr , "%s: error: unable to load model\n" , __func__); |
| 74 | + return 1; |
| 75 | + } |
| 76 | + |
| 77 | + // initialize the context |
| 78 | + llama_context_params ctx_params = llama_context_default_params(); |
| 79 | + ctx_params.n_ctx = n_ctx; |
| 80 | + ctx_params.n_batch = n_ctx; |
| 81 | + |
| 82 | + llama_context * ctx = llama_new_context_with_model(model, ctx_params); |
| 83 | + if (!ctx) { |
| 84 | + fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__); |
| 85 | + return 1; |
| 86 | + } |
| 87 | + |
| 88 | + // initialize the sampler |
| 89 | + llama_sampler * smpl = llama_sampler_chain_init(llama_sampler_chain_default_params()); |
| 90 | + llama_sampler_chain_add(smpl, llama_sampler_init_min_p(0.05f, 1)); |
| 91 | + llama_sampler_chain_add(smpl, llama_sampler_init_temp(0.8f)); |
| 92 | + llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED)); |
| 93 | + |
| 94 | + // generation helper |
| 95 | + auto generate = [&](const std::string & prompt) { |
| 96 | + std::string response; |
| 97 | + |
| 98 | + // tokenize the prompt |
| 99 | + const int n_prompt = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true); |
| 100 | + std::vector<llama_token> prompt_tokens(n_prompt); |
| 101 | + if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) { |
| 102 | + GGML_ABORT("failed to tokenize the prompt\n"); |
| 103 | + } |
| 104 | + |
| 105 | + // prepare a batch for the prompt |
| 106 | + llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size()); |
| 107 | + llama_token new_token_id; |
| 108 | + while (true) { |
| 109 | + // check if we have enough context space to evaluate this batch |
| 110 | + int n_ctx = llama_n_ctx(ctx); |
| 111 | + int n_ctx_used = llama_get_kv_cache_used_cells(ctx); |
| 112 | + if (n_ctx_used + batch.n_tokens > n_ctx) { |
| 113 | + printf("\033[0m\n"); |
| 114 | + fprintf(stderr, "context size exceeded\n"); |
| 115 | + exit(0); |
| 116 | + } |
| 117 | + |
| 118 | + if (llama_decode(ctx, batch)) { |
| 119 | + GGML_ABORT("failed to eval\n"); |
| 120 | + } |
| 121 | + |
| 122 | + // sample the next token |
| 123 | + new_token_id = llama_sampler_sample(smpl, ctx, -1); |
| 124 | + |
| 125 | + // is it an end of generation? |
| 126 | + if (llama_token_is_eog(model, new_token_id)) { |
| 127 | + break; |
| 128 | + } |
| 129 | + |
| 130 | + // add the token to the response |
| 131 | + char buf[128]; |
| 132 | + int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true); |
| 133 | + if (n < 0) { |
| 134 | + GGML_ABORT("failed to convert token to piece\n"); |
| 135 | + } |
| 136 | + std::string piece(buf, n); |
| 137 | + response += piece; |
| 138 | + printf("%s", piece.c_str()); |
| 139 | + fflush(stdout); |
| 140 | + |
| 141 | + // prepare the next batch with the sampled token |
| 142 | + batch = llama_batch_get_one(&new_token_id, 1); |
| 143 | + } |
| 144 | + |
| 145 | + return response; |
| 146 | + }; |
| 147 | + |
| 148 | + std::vector<llama_chat_message> messages; |
| 149 | + std::vector<char> formatted(2048); |
| 150 | + int prev_len = 0; |
| 151 | + while (true) { |
| 152 | + std::string user; |
| 153 | + std::getline(std::cin, user); |
| 154 | + messages.push_back({"user", strdup(user.c_str())}); |
| 155 | + |
| 156 | + // format the messages |
| 157 | + int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size()); |
| 158 | + if (new_len > (int)formatted.size()) { |
| 159 | + formatted.resize(new_len); |
| 160 | + new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size()); |
| 161 | + } |
| 162 | + |
| 163 | + // remove previous messages and obtain a prompt |
| 164 | + std::string prompt(formatted.begin() + prev_len, formatted.begin() + new_len); |
| 165 | + |
| 166 | + // generate a response |
| 167 | + printf("\033[31m"); |
| 168 | + std::string response = generate(prompt); |
| 169 | + printf("\n\033[0m"); |
| 170 | + |
| 171 | + // add the response to the messages |
| 172 | + messages.push_back({"assistant", strdup(response.c_str())}); |
| 173 | + prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, formatted.data(), formatted.size()); |
| 174 | + } |
| 175 | + |
| 176 | + |
| 177 | + llama_sampler_free(smpl); |
| 178 | + llama_free(ctx); |
| 179 | + llama_free_model(model); |
| 180 | + |
| 181 | + return 0; |
| 182 | +} |
0 commit comments