aboutsummaryrefslogtreecommitdiff
path: root/tests/tiny_model.h
blob: 0a4e67b659dbd6a5018c7d2c5105fd5160e47a16 (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
/* tests/tiny_model.h - synthesize a tiny 1-layer model directory (config.json
 * + model.safetensors) for hermetic tests. Values are BF16-representable so an
 * F32 and a BF16 copy of the same model produce identical embeddings. */

#ifndef TINY_MODEL_H
#define TINY_MODEL_H

#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

static uint16_t tm_f32_to_bf16(float x)
{
    uint32_t u;
    memcpy(&u, &x, sizeof(u));
    return (uint16_t)(u >> 16);
}

static float tm_bf16_to_f32(uint16_t b)
{
    uint32_t u = ((uint32_t)b) << 16;
    float f;
    memcpy(&f, &u, sizeof(f));
    return f;
}

/* Deterministic, BF16-exact value for element i of tensor `name`. */
static float tm_value(const char *name, size_t i)
{
    float v;
    size_t len = strlen(name);
    if ((len >= 11 && strcmp(name + len - 11, "norm.weight") == 0) ||
        strstr(name, "layernorm.weight"))
        v = 0.75f + 0.03f * (float)((i % 5) + 1);
    else if (strcmp(name, "embed_tokens.weight") == 0)
        v = sinf((float)i * 0.17f) * 0.2f;
    else
        v = sinf((float)i * 0.11f + (float)len) * 0.08f;
    return tm_bf16_to_f32(tm_f32_to_bf16(v));
}

typedef struct {
    const char *name;
    int rows, cols;     /* cols == 0 means a 1-D tensor of `rows` values */
} tm_spec_t;

/* Always one layer; everything else is free so a test can match a real
 * model family's dimensions (e.g. hidden 1024 for the 0.6B server slot). */
typedef struct {
    int hidden, heads, kv_heads, head_dim, intermediate, vocab;
} tm_dims_t;

enum { TM_N_SPECS = 13 };

/* Write config.json + model.safetensors with the given dimensions into dir.
 * dtype: "F32" or "BF16". Returns 0 on success. */
static int tm_write_model_dims(const char *dir, const char *dtype,
                               const tm_dims_t *d)
{
    int q = d->heads * d->head_dim, kv = d->kv_heads * d->head_dim;
    const tm_spec_t specs[TM_N_SPECS] = {
        {"embed_tokens.weight", d->vocab, d->hidden},
        {"layers.0.self_attn.q_proj.weight", q, d->hidden},
        {"layers.0.self_attn.k_proj.weight", kv, d->hidden},
        {"layers.0.self_attn.v_proj.weight", kv, d->hidden},
        {"layers.0.self_attn.o_proj.weight", d->hidden, q},
        {"layers.0.self_attn.q_norm.weight", d->head_dim, 0},
        {"layers.0.self_attn.k_norm.weight", d->head_dim, 0},
        {"layers.0.input_layernorm.weight", d->hidden, 0},
        {"layers.0.post_attention_layernorm.weight", d->hidden, 0},
        {"layers.0.mlp.gate_proj.weight", d->intermediate, d->hidden},
        {"layers.0.mlp.up_proj.weight", d->intermediate, d->hidden},
        {"layers.0.mlp.down_proj.weight", d->hidden, d->intermediate},
        {"norm.weight", d->hidden, 0},
    };

    char path[2048];
    snprintf(path, sizeof(path), "%s/config.json", dir);
    FILE *f = fopen(path, "w");
    if (!f) return -1;
    fprintf(f, "{\"hidden_size\":%d,\"num_hidden_layers\":1,"
            "\"num_attention_heads\":%d,\"num_key_value_heads\":%d,"
            "\"head_dim\":%d,\"intermediate_size\":%d,\"vocab_size\":%d,"
            "\"rms_norm_eps\":1e-6,\"rope_theta\":10000.0}",
            d->hidden, d->heads, d->kv_heads, d->head_dim,
            d->intermediate, d->vocab);
    fclose(f);

    int bf16 = strcmp(dtype, "BF16") == 0;
    size_t esize = bf16 ? 2 : 4;

    char header[4096];
    size_t hoff = 0, doff = 0;
    hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, "{");
    for (int t = 0; t < TM_N_SPECS; t++) {
        const tm_spec_t *s = &specs[t];
        size_t n = (size_t)s->rows * (s->cols ? (size_t)s->cols : 1);
        size_t end = doff + n * esize;
        if (s->cols)
            hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff,
                "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d,%d],"
                "\"data_offsets\":[%zu,%zu]}",
                t ? "," : "", s->name, dtype, s->rows, s->cols, doff, end);
        else
            hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff,
                "%s\"%s\":{\"dtype\":\"%s\",\"shape\":[%d],"
                "\"data_offsets\":[%zu,%zu]}",
                t ? "," : "", s->name, dtype, s->rows, doff, end);
        doff = end;
    }
    hoff += (size_t)snprintf(header + hoff, sizeof(header) - hoff, "}");
    if (hoff >= sizeof(header) - 1) return -1;

    snprintf(path, sizeof(path), "%s/model.safetensors", dir);
    f = fopen(path, "wb");
    if (!f) return -1;
    unsigned char len8[8];
    for (int i = 0; i < 8; i++) len8[i] = (unsigned char)(hoff >> (8 * i));
    fwrite(len8, 1, 8, f);
    fwrite(header, 1, hoff, f);
    for (int t = 0; t < TM_N_SPECS; t++) {
        const tm_spec_t *s = &specs[t];
        size_t n = (size_t)s->rows * (s->cols ? (size_t)s->cols : 1);
        for (size_t i = 0; i < n; i++) {
            float v = tm_value(s->name, i);
            if (bf16) {
                uint16_t h = tm_f32_to_bf16(v);
                fwrite(&h, sizeof(h), 1, f);
            } else {
                fwrite(&v, sizeof(v), 1, f);
            }
        }
    }
    fclose(f);
    return 0;
}

/* The historical default fixture: 4-dim hidden, 16-token vocab. */
static inline int tm_write_model(const char *dir, const char *dtype)
{
    static const tm_dims_t d = {4, 2, 1, 2, 8, 16};
    return tm_write_model_dims(dir, dtype, &d);
}

#endif /* TINY_MODEL_H */