aboutsummaryrefslogtreecommitdiff
path: root/bench/bench_tokenizer.c
blob: 1033f2ea5cb3d219a03a8f33faaf33a8bd67e00a (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
/* bench/bench_tokenizer.c - tokenizer encode-path benchmark.
 * Build via `make bench-tokenizer`; needs a model dir (vocab.json). */

#include "qwen_tokenizer.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

int qwen_verbose = 0;

static double now_ms(void)
{
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC, &ts);
    return (double)ts.tv_sec * 1000.0 + (double)ts.tv_nsec / 1000000.0;
}

static int same_ids(const int *a, int na, const int *b, int nb)
{
    if (na != nb) return 0;
    for (int i = 0; i < na; i++) {
        if (a[i] != b[i]) return 0;
    }
    return 1;
}

int main(int argc, char **argv)
{
    int status = 1;
    int **gold = NULL;
    int *gold_n = NULL;
    int **bufs = NULL;
    int *caps = NULL;

    if (argc < 4) {
        fprintf(stderr, "usage: %s VOCAB_JSON RUNS TEXT...\n", argv[0]);
        return 2;
    }

    const char *vocab = argv[1];
    int runs = atoi(argv[2]);
    if (runs <= 0) return 2;
    int n_texts = argc - 3;
    char **texts = argv + 3;

    qwen_tokenizer_t *tok = qwen_tokenizer_load(vocab);
    qwen_tokenizer_workspace_t *ws = qwen_tokenizer_workspace_new();
    if (!tok || !ws) {
        fprintf(stderr, "failed to initialize tokenizer\n");
        goto cleanup;
    }

    gold = (int **)calloc((size_t)n_texts, sizeof(*gold));
    gold_n = (int *)calloc((size_t)n_texts, sizeof(*gold_n));
    bufs = (int **)calloc((size_t)n_texts, sizeof(*bufs));
    caps = (int *)calloc((size_t)n_texts, sizeof(*caps));
    if (!gold || !gold_n || !bufs || !caps) goto cleanup;

    int tokens_per_pass = 0;
    for (int i = 0; i < n_texts; i++) {
        gold[i] = qwen_tokenizer_encode(tok, texts[i], &gold_n[i]);
        if (!gold[i] || gold_n[i] <= 0) goto cleanup;
        tokens_per_pass += gold_n[i];

        int n_ws = 0;
        int *ids_ws = qwen_tokenizer_encode_with_workspace(tok, ws, texts[i],
                                                           &n_ws);
        if (!ids_ws || !same_ids(gold[i], gold_n[i], ids_ws, n_ws)) {
            free(ids_ws);
            goto cleanup;
        }
        free(ids_ws);

        caps[i] = (int)(strlen(texts[i]) * 2 + 8);
        if (caps[i] < gold_n[i]) caps[i] = gold_n[i];
        bufs[i] = (int *)malloc((size_t)caps[i] * sizeof(int));
        if (!bufs[i]) goto cleanup;
        int n_into = 0;
        if (qwen_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i],
                                       &n_into) != 0 ||
            !same_ids(gold[i], gold_n[i], bufs[i], n_into))
            goto cleanup;
    }

    double t0 = now_ms();
    long long old_tokens = 0;
    for (int r = 0; r < runs; r++) {
        for (int i = 0; i < n_texts; i++) {
            int n = 0;
            int *ids = qwen_tokenizer_encode(tok, texts[i], &n);
            if (!ids || n != gold_n[i]) {
                free(ids);
                goto cleanup;
            }
            old_tokens += n;
            free(ids);
        }
    }
    double old_ms = now_ms() - t0;

    t0 = now_ms();
    long long ws_tokens = 0;
    for (int r = 0; r < runs; r++) {
        for (int i = 0; i < n_texts; i++) {
            int n = 0;
            int *ids = qwen_tokenizer_encode_with_workspace(tok, ws, texts[i],
                                                            &n);
            if (!ids || n != gold_n[i]) {
                free(ids);
                goto cleanup;
            }
            ws_tokens += n;
            free(ids);
        }
    }
    double ws_ms = now_ms() - t0;

    t0 = now_ms();
    long long into_tokens = 0;
    for (int r = 0; r < runs; r++) {
        for (int i = 0; i < n_texts; i++) {
            int n = 0;
            int rc = qwen_tokenizer_encode_into(tok, ws, texts[i], bufs[i],
                                                caps[i], &n);
            if (rc != 0 || n != gold_n[i]) goto cleanup;
            into_tokens += n;
        }
    }
    double into_ms = now_ms() - t0;

    printf("texts=%d runs=%d tokens/pass=%d\n", n_texts, runs,
           tokens_per_pass);
    printf("old_malloc:       %.3f ms  %.0f tokens/s\n", old_ms,
           (double)old_tokens / (old_ms / 1000.0));
    printf("workspace_malloc: %.3f ms  %.0f tokens/s  speedup=%.2fx\n",
           ws_ms, (double)ws_tokens / (ws_ms / 1000.0), old_ms / ws_ms);
    printf("workspace_into:   %.3f ms  %.0f tokens/s  speedup=%.2fx\n",
           into_ms, (double)into_tokens / (into_ms / 1000.0),
           old_ms / into_ms);

    status = 0;

cleanup:
    for (int i = 0; i < n_texts; i++) {
        if (gold) free(gold[i]);
        if (bufs) free(bufs[i]);
    }
    free(gold);
    free(gold_n);
    free(bufs);
    free(caps);
    qwen_tokenizer_workspace_free(ws);
    qwen_tokenizer_free(tok);
    return status;
}