KBaba7/llama.cpp

1#include "llama.h"2#include "common.h"3#include "unicode.h"4#include "console.h"5 6#include <cassert>7#include <codecvt>8#include <cstdio>9#include <cstring>10#include <locale>11#include <string>12#include <thread>13#include <vector>14#include <atomic>15 16int main(int argc, char ** argv) {17    if (argc < 2) {18        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);19        return 1;20    }21 22    const std::string fname = argv[1];23 24    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());25 26    llama_model * model;27    llama_context * ctx;28 29    llama_backend_init();30 31    // load the vocab32    {33        auto mparams = llama_model_default_params();34 35        mparams.vocab_only = true;36 37        model = llama_model_load_from_file(fname.c_str(), mparams);38 39        if (model == NULL) {40            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());41            return 1;42        }43 44        auto cparams = llama_context_default_params();45 46        ctx = llama_init_from_model(model, cparams);47 48        if (ctx == NULL) {49            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());50            llama_model_free(model);51            return 1;52        }53    }54 55    const llama_vocab * vocab = llama_model_get_vocab(model);56 57    //GGML_ASSERT(llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);58    if (llama_vocab_type(vocab) != LLAMA_VOCAB_TYPE_SPM) {59        return 99;60    }61 62#ifdef _WIN3263    // We need this for unicode console support64    console::init(false, false);65    atexit([]() { console::cleanup(); });66#endif67 68    const int n_vocab = llama_vocab_n_tokens(vocab);69 70    for (int i = 0; i < n_vocab; ++i) {71        std::string str = common_detokenize(ctx, std::vector<int>(1, i), true);72        std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);73        std::string check = common_detokenize(ctx, tokens);74        if (check != str) {75            fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",76                __func__, i, str.c_str(), str.length(), check.c_str(), check.length());77            return 2;78        }79    }80 81    // unicode82    {83        const int nthread = std::thread::hardware_concurrency();84 85        std::vector<std::thread> threads(nthread);86 87        std::atomic_int errcode = {};88 89        for (int i = 0; i < nthread; ++i) {90            threads[i] = std::thread([i, nthread, ctx, &errcode]() {91                for (uint32_t cp = i; !errcode && cp < 0x00110000; cp += nthread) {92                    if ((0x0000D800 <= cp && cp <= 0x0000DFFF) ||  // surrogates \p{Cs}93                        (0x00040000 <= cp && cp <= 0x000E0000)) {  // undefined \p{Cn}94                        continue;95                    }96 97                    std::string str = unicode_cpt_to_utf8(cp);98                    std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);99                    std::string check = common_detokenize(ctx, tokens);100                    if (cp != 9601 && str != check) {101                        fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",102                                cp, check.c_str(), check.length(), str.c_str(), str.length());103                        errcode = 3;104                    }105                }106            });107        }108 109        for (auto & t : threads) {110            t.join();111        }112 113        if(errcode) {114            return errcode;115        }116    }117 118    llama_model_free(model);119    llama_free(ctx);120 121    llama_backend_free();122 123    return 0;124}125