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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
|
/* tests/test_workspace.c - model/workspace API lifecycle tests.
* Runs via `make test`. Uses a tiny synthetic model. */
#include "embed.h"
#include "qwen_tokenizer.h"
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
static float max_abs_diff(const float *a, const float *b, int n)
{
float m = 0.0f;
for (int i = 0; i < n; i++) {
float d = fabsf(a[i] - b[i]);
if (d > m) m = d;
}
return m;
}
int main(int argc, char **argv)
{
if (argc != 2) {
fprintf(stderr, "usage: %s MODEL_DIR\n", argv[0]);
return 2;
}
const char *model_dir = argv[1];
char vocab_path[4096];
snprintf(vocab_path, sizeof(vocab_path), "%s/vocab.json", model_dir);
pplx_model_t *model = pplx_model_load(model_dir);
if (!model) {
fprintf(stderr, "failed to load model\n");
return 1;
}
const pplx_config_t *cfg = pplx_model_config(model);
if (!cfg || cfg->hidden_size <= 0) {
fprintf(stderr, "invalid model config\n");
pplx_model_free(model);
return 1;
}
pplx_workspace_t *batch_ws = pplx_workspace_new(model);
pplx_workspace_t *single_ws = pplx_workspace_new(model);
qwen_tokenizer_t *tok = qwen_tokenizer_load(vocab_path);
if (!batch_ws || !single_ws || !tok) {
fprintf(stderr, "failed to allocate workspace or tokenizer\n");
qwen_tokenizer_free(tok);
pplx_workspace_free(single_ws);
pplx_workspace_free(batch_ws);
pplx_model_free(model);
return 1;
}
size_t initial_ws_bytes = pplx_workspace_nbytes(batch_ws);
if (initial_ws_bytes == 0) {
fprintf(stderr, "workspace byte accounting returned zero\n");
goto fail;
}
const char *texts[] = {
"query: what is the capital of France?",
"document: Paris is the capital of France.",
"document: Berlin is the capital of Germany.",
"document: Istanbul is a major city in Turkey."
};
const int batch = (int)(sizeof(texts) / sizeof(texts[0]));
int *ids[4] = {0};
int ntok[4] = {0};
pplx_input_t inputs[4];
for (int i = 0; i < batch; i++) {
ids[i] = qwen_tokenizer_encode(tok, texts[i], &ntok[i]);
if (!ids[i] || ntok[i] <= 0) {
fprintf(stderr, "failed to tokenize input %d\n", i);
goto fail;
}
inputs[i].ids = ids[i];
inputs[i].n_tokens = ntok[i];
}
int dim = cfg->hidden_size;
float *batched = (float *)calloc((size_t)batch * dim, sizeof(float));
float *single = (float *)calloc((size_t)dim, sizeof(float));
float *normalized = (float *)calloc((size_t)dim, sizeof(float));
float *states = (float *)calloc((size_t)ntok[0] * dim, sizeof(float));
if (!batched || !single || !normalized || !states) {
fprintf(stderr, "allocation failure\n");
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
if (pplx_model_embed_batch(model, batch_ws, inputs, batch, batched) != 0) {
fprintf(stderr, "pplx_model_embed_batch failed\n");
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
size_t batch_ws_bytes = pplx_workspace_nbytes(batch_ws);
if (batch_ws_bytes <= initial_ws_bytes) {
fprintf(stderr, "workspace byte accounting did not grow after embedding\n");
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
float worst_diff = 0.0f;
float worst_cos = 1.0f;
for (int i = 0; i < batch; i++) {
if (pplx_model_embed_into(model, single_ws, ids[i], ntok[i], single) != 0) {
fprintf(stderr, "pplx_model_embed_into failed at %d\n", i);
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
const float *row = batched + (size_t)i * dim;
float diff = max_abs_diff(row, single, dim);
float cos = pplx_cosine_similarity(row, single, dim);
if (diff > worst_diff) worst_diff = diff;
if (cos < worst_cos) worst_cos = cos;
}
memcpy(normalized, batched, (size_t)dim * sizeof(float));
if (pplx_l2_normalize(normalized, dim) != 0) {
fprintf(stderr, "pplx_l2_normalize failed\n");
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
float normalized_norm_sq = 0.0f;
for (int i = 0; i < dim; i++)
normalized_norm_sq += normalized[i] * normalized[i];
if (fabsf(normalized_norm_sq - 1.0f) > 0.00005f) {
fprintf(stderr, "pplx_l2_normalize produced norm_sq=%g\n",
normalized_norm_sq);
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
if (pplx_model_forward_into(model, batch_ws, ids[0], ntok[0], states) != 0) {
fprintf(stderr, "pplx_model_forward_into failed\n");
free(states);
free(normalized);
free(single);
free(batched);
goto fail;
}
free(states);
free(normalized);
free(single);
free(batched);
for (int i = 0; i < batch; i++) free(ids[i]);
qwen_tokenizer_free(tok);
pplx_workspace_free(single_ws);
pplx_workspace_free(batch_ws);
pplx_model_free(model);
if (worst_diff > 0.00005f || worst_cos < 0.99999f) {
fprintf(stderr, "workspace API parity failed: max_abs_diff=%g cosine=%g\n",
worst_diff, worst_cos);
return 1;
}
printf("ok: workspace API dim=%d batch=%d workspace=%zu bytes max_abs_diff=%g cosine=%g\n",
dim, batch, batch_ws_bytes, worst_diff, worst_cos);
return 0;
fail:
for (int i = 0; i < batch; i++) free(ids[i]);
qwen_tokenizer_free(tok);
pplx_workspace_free(single_ws);
pplx_workspace_free(batch_ws);
pplx_model_free(model);
return 1;
}
|