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
|
/* safetensors parser limit and rejection tests.
* Hermetic: synthesizes every fixture in a temp directory.
*/
#include "safetensors.h"
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/stat.h>
#include <unistd.h>
static int failures = 0;
static char tmproot[1024];
/* Write u64-LE header length + JSON header + payload (safetensors layout). */
static void write_st(const char *path, const char *header, const void *payload, size_t payload_len) {
FILE *f = fopen(path, "wb");
if (!f) {
perror(path);
exit(2);
}
/* Pad the header to an 8-byte boundary with trailing spaces so the data
* section is 8-aligned, matching the real safetensors serializer; otherwise
* tensor data lands unaligned and typed reads of it are UBSan-flagged UB. */
uint64_t n = strlen(header);
uint64_t pad = (8 - (n % 8)) % 8;
uint64_t total = n + pad;
unsigned char len8[8];
for (int i = 0; i < 8; i++)
len8[i] = (unsigned char)(total >> (8 * i));
fwrite(len8, 1, 8, f);
fwrite(header, 1, n, f);
for (uint64_t i = 0; i < pad; i++)
fputc(' ', f);
if (payload_len)
fwrite(payload, 1, payload_len, f);
fclose(f);
}
static const char *path_in(const char *rel) {
static char buf[2048];
snprintf(buf, sizeof(buf), "%s/%s", tmproot, rel);
return buf;
}
static void check(const char *mode, const char *path, int want_ok) {
int got_ok;
if (strcmp(mode, "file") == 0) {
safetensors_file_t *sf = safetensors_open(path);
got_ok = sf != NULL;
safetensors_close(sf);
} else {
multi_safetensors_t *ms = multi_safetensors_open(path);
got_ok = ms != NULL;
multi_safetensors_close(ms);
}
if (got_ok != want_ok) {
fprintf(stderr, "%s %s: got %s want %s\n", mode, path, got_ok ? "ok" : "fail",
want_ok ? "ok" : "fail");
failures++;
}
}
static const char ONE_TENSOR_A[] = "{\"a\":{\"dtype\":\"F32\",\"shape\":[1],\"data_offsets\":[0,4]}}";
static const char ONE_TENSOR_B[] = "{\"b\":{\"dtype\":\"F32\",\"shape\":[1],\"data_offsets\":[0,4]}}";
static const char ONE_TENSOR_X[] = "{\"x\":{\"dtype\":\"F32\",\"shape\":[1],\"data_offsets\":[0,4]}}";
int main(void) {
snprintf(tmproot, sizeof(tmproot), "%s/ffwd-st-test-XXXXXX",
getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp");
if (!mkdtemp(tmproot)) {
perror("mkdtemp");
return 2;
}
static const char zeros[4 * 1025] = {0};
write_st(path_in("valid.safetensors"), ONE_TENSOR_X, zeros, 4);
check("file", path_in("valid.safetensors"), 1);
/* 1025 tensors exceeds the parser's tensor-count limit. */
{
size_t cap = 1025 * 64 + 16;
char *hdr = malloc(cap);
if (!hdr)
return 2;
size_t off = 0;
off += (size_t)snprintf(hdr + off, cap - off, "{");
for (int i = 0; i < 1025; i++) {
off += (size_t)snprintf(hdr + off, cap - off,
"%s\"t%d\":{\"dtype\":\"F32\",\"shape\":[1],"
"\"data_offsets\":[%d,%d]}",
i ? "," : "", i, i * 4, i * 4 + 4);
}
snprintf(hdr + off, cap - off, "}");
write_st(path_in("too_many.safetensors"), hdr, zeros, sizeof(zeros));
free(hdr);
check("file", path_in("too_many.safetensors"), 0);
}
write_st(path_in("too_many_dims.safetensors"),
"{\"x\":{\"dtype\":\"F32\",\"shape\":[1,1,1,1,1,1,1,1,1],"
"\"data_offsets\":[0,4]}}",
zeros, 4);
check("file", path_in("too_many_dims.safetensors"), 0);
write_st(path_in("bad_offsets.safetensors"),
"{\"x\":{\"dtype\":\"F32\",\"shape\":[1],"
"\"data_offsets\":[4,0]}}",
zeros, 4);
check("file", path_in("bad_offsets.safetensors"), 0);
mkdir(path_in("valid_multi"), 0755);
write_st(path_in("valid_multi/model-00001-of-00002.safetensors"), ONE_TENSOR_A, zeros, 4);
write_st(path_in("valid_multi/model-00002-of-00002.safetensors"), ONE_TENSOR_B, zeros, 4);
check("multi", path_in("valid_multi"), 1);
/* Payload accounting across shards: two 4-byte F32 tensors. */
{
multi_safetensors_t *ms = multi_safetensors_open(path_in("valid_multi"));
size_t nbytes = 0;
if (!ms || multi_safetensors_data_nbytes(ms, &nbytes) != 0 || nbytes != 8) {
fprintf(stderr, "multi data_nbytes: got %zu want 8\n", nbytes);
failures++;
}
multi_safetensors_close(ms);
}
mkdir(path_in("missing"), 0755);
write_st(path_in("missing/model-00001-of-00002.safetensors"), ONE_TENSOR_X, zeros, 4);
check("multi", path_in("missing"), 0);
mkdir(path_in("too_many_shards"), 0755);
write_st(path_in("too_many_shards/model-00001-of-00009.safetensors"), ONE_TENSOR_X, zeros, 4);
check("multi", path_in("too_many_shards"), 0);
mkdir(path_in("bad_name"), 0755);
write_st(path_in("bad_name/model-foo.safetensors"), ONE_TENSOR_X, zeros, 4);
check("multi", path_in("bad_name"), 0);
if (failures) {
fprintf(stderr, "safetensors tests: %d failure(s)\n", failures);
return 1;
}
puts("ok: safetensors parser limit checks passed");
return 0;
}
|