1#include "llama.h"
2#include "common.h"
3#include "console.h"
4
5#include "../src/unicode.h"
6
7#include <cassert>
8#include <codecvt>
9#include <cstdio>
10#include <cstring>
11#include <locale>
12#include <string>
13#include <thread>
14#include <vector>
15#include <atomic>
16
17int main(int argc, char ** argv) {
18 if (argc < 2) {
19 fprintf(stderr, format: "Usage: %s <vocab-file>\n", argv[0]);
20 return 1;
21 }
22
23 const std::string fname = argv[1];
24
25 fprintf(stderr, format: "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
26
27 llama_model * model;
28 llama_context * ctx;
29
30 llama_backend_init();
31
32 // load the vocab
33 {
34 auto mparams = llama_model_default_params();
35
36 mparams.vocab_only = true;
37
38 model = llama_model_load_from_file(path_model: fname.c_str(), params: mparams);
39
40 if (model == NULL) {
41 fprintf(stderr, format: "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
42 return 1;
43 }
44
45 auto cparams = llama_context_default_params();
46
47 ctx = llama_init_from_model(model, params: cparams);
48
49 if (ctx == NULL) {
50 fprintf(stderr, format: "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
51 llama_model_free(model);
52 return 1;
53 }
54 }
55
56 const llama_vocab * vocab = llama_model_get_vocab(model);
57
58 //GGML_ASSERT(llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
59 if (llama_vocab_type(vocab) != LLAMA_VOCAB_TYPE_SPM) {
60 return 99;
61 }
62
63#ifdef _WIN32
64 // We need this for unicode console support
65 console::init(false, false);
66 atexit([]() { console::cleanup(); });
67#endif
68
69 const int n_vocab = llama_vocab_n_tokens(vocab);
70
71 for (int i = 0; i < n_vocab; ++i) {
72 std::string str = common_detokenize(ctx, tokens: std::vector<int>(1, i), special: true);
73 std::vector<llama_token> tokens = common_tokenize(ctx, text: str, add_special: false, parse_special: true);
74 std::string check = common_detokenize(ctx, tokens);
75 if (check != str) {
76 fprintf(stderr, format: "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
77 __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
78 return 2;
79 }
80 }
81
82 // unicode
83 {
84 const int nthread = std::thread::hardware_concurrency();
85
86 std::vector<std::thread> threads(nthread);
87
88 std::atomic_int errcode = {};
89
90 for (int i = 0; i < nthread; ++i) {
91 threads[i] = std::thread([i, nthread, ctx, &errcode]() {
92 for (uint32_t cp = i; !errcode && cp < 0x00110000; cp += nthread) {
93 if ((0x0000D800 <= cp && cp <= 0x0000DFFF) || // surrogates \p{Cs}
94 (0x00040000 <= cp && cp <= 0x000E0000)) { // undefined \p{Cn}
95 continue;
96 }
97
98 std::string str = unicode_cpt_to_utf8(cpt: cp);
99 std::vector<llama_token> tokens = common_tokenize(ctx, text: str, add_special: false, parse_special: true);
100 std::string check = common_detokenize(ctx, tokens);
101 if (cp != 9601 && str != check) {
102 fprintf(stderr, format: "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
103 cp, check.c_str(), check.length(), str.c_str(), str.length());
104 errcode = 3;
105 }
106 }
107 });
108 }
109
110 for (auto & t : threads) {
111 t.join();
112 }
113
114 if(errcode) {
115 return errcode;
116 }
117 }
118
119 llama_model_free(model);
120 llama_free(ctx);
121
122 llama_backend_free();
123
124 return 0;
125}
126