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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
|
#ifndef EMBED_SERVER_INTERNAL_H
#define EMBED_SERVER_INTERNAL_H
/* Declarations shared across the embed-server translation units (HTTP framing,
* JSON validation, output encoding, scheduling, model lifecycle, endpoint
* handlers). The server is split by concern but compiles as several objects
* linked into one binary; this header is what they share. Not a public API -
* the CLI and library do not include it. */
#include "embed.h"
#include "sbuf.h"
#include "tokenizer_bpe.h"
#include "tokenizer_sentencepiece.h"
#include "tokenizer_wordpiece.h"
#ifdef USE_MLX
#include "mlx.h"
#endif
#ifdef USE_CUDA
#include "cuda.h"
#endif
#include "deps/ae/ae.h"
#include <cjson/cJSON.h>
#include <pthread.h>
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#define EMBED_HTTP_IO_BUF 8192
#define EMBED_HTTP_MAX_HEADER (64u * 1024u)
#define EMBED_HTTP_MAX_BODY (64u * 1024u * 1024u)
#define EMBED_HTTP_MAX_PATH 255u
#define EMBED_HTTP_BACKLOG 128
#define EMBED_HTTP_SETSIZE 10240
#define EMBED_HTTP_CLIENT_TIMEOUT_MS 30000
#define EMBED_HTTP_SWEEP_MS 1000
#define EMBED_API_MAX_STANDARD_INPUTS 512
#define EMBED_API_MAX_CONTEXT_DOCS 512
#define EMBED_API_MAX_CONTEXT_CHUNKS 16000
#define EMBED_API_MAX_RERANK_DOCUMENTS 1000
#define EMBED_API_MAX_ITEM_TOKENS 32768
#define EMBED_API_MAX_TOTAL_TOKENS 120000
#define EMBED_LATE_QUERY_TOKENS 32
#define EMBED_LATE_MASK_TOKEN_ID 151642
#define EMBED_LATE_QUERY_PREFIX_ID 151669
#define EMBED_LATE_DOCUMENT_PREFIX_ID 151670
/* `text_type: query` prepends the model's published retrieval instruction.
* Documents and the default pass through unchanged. Callers needing a custom
* instruction prepend it themselves and leave text_type at "document". */
#define EMBED_QWEN3_QUERY_INSTRUCT \
"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
#define EMBED_GTE_QWEN2_QUERY_INSTRUCT \
"Instruct: Given a web search query, retrieve relevant passages that answer the " \
"query\nQuery: "
#define EMBED_E5_QUERY_INSTRUCT \
"Instruct: Given a web search query, retrieve relevant passages that answer the " \
"query\nQuery: "
#define EMBED_SNOWFLAKE_QUERY_INSTRUCT "query: "
/* Chosen from the L4 scheduler sweep: -b 32 gained ~24% concurrent
* short-request throughput over 8 with no long-document penalty, while 128
* was no faster and inflated long-document tail latency. */
#define EMBED_SERVER_DEFAULT_BATCH_SIZE 32
#define EMBED_SERVER_DEFAULT_MAX_BATCH_TOKENS 16384
/* Microseconds the worker waits for more requests before dispatching a batch.
* CUDA uses a 1 ms window so arrivals group into one launch and keep the GPU
* busy; MLX and CPU dispatch immediately, where waiting only adds latency.
* Override with --batch-wait-us. */
#define EMBED_SERVER_BATCH_WAIT_US 0
#define EMBED_SERVER_CUDA_BATCH_WAIT_US 1000
#define EMBED_SERVER_MICROBATCH_MAX_JOBS 128
#define EMBED_MLX_MEMORY_BUDGET_PERCENT 90
#define EMBED_MLX_RESIDENT_MULTIPLIER 2
typedef enum {
MODEL_STD_06,
MODEL_STD_4,
MODEL_QWEN3_06,
MODEL_QWEN3_4,
MODEL_QWEN3_8,
MODEL_GTE_QWEN2_15,
MODEL_MINILM_L6,
MODEL_BGE_SMALL,
MODEL_BGE_BASE,
MODEL_BGE_LARGE,
MODEL_E5_MULTI_LARGE_INSTRUCT,
MODEL_SNOWFLAKE_ARCTIC_L_V2,
MODEL_CTX_06,
MODEL_CTX_4,
MODEL_LATE_06,
MODEL_COUNT,
MODEL_UNKNOWN
} model_slot;
typedef enum { MODEL_KIND_STANDARD, MODEL_KIND_CONTEXTUAL, MODEL_KIND_LATE } model_kind;
/* Which embedding API a model speaks. pplx-embed models follow the Perplexity
* API: the canonical output is the tanh int8 quantization, so the default
* encoding is base64_int8 and "float" is the int8-decoded view (int8/128).
* Instruction embedding models follow the OpenAI-compatible API: "float" is
* the true float32 vector and "base64" is base64 of little-endian float32.
* int8 quantization underuses the range of L2-normalized vectors, so it is not
* used for these models. */
typedef enum {
EMBED_API_PERPLEXITY = 0,
EMBED_API_OPENAI = 1,
} embedding_api_t;
typedef struct {
const char *id;
model_kind kind;
int dim;
int min_dim;
int token_dim;
embed_attention_mode_t attention_mode;
embed_pooling_mode_t pooling_mode;
int normalize_embeddings;
embedding_api_t api;
const char *query_instruct;
} model_info;
typedef struct {
const model_info *info;
char *path;
embed_tokenizer_t *tok;
embed_tokenizer_workspace_t *tok_ws;
/* BERT-family WordPiece tokenizer, selected by file probe in load_one_model.
* When wp_tok is set, tokenize_one wraps the ids with [CLS]/[SEP] and the
* Qwen byte-level BPE fields above stay NULL. */
wordpiece_tokenizer_t *wp_tok;
wordpiece_workspace_t *wp_tok_ws;
sentencepiece_tokenizer_t *sp_tok;
sentencepiece_workspace_t *sp_tok_ws;
int cls_id;
int sep_id;
int context_separator_id;
int late_mask_id;
int late_query_prefix_id;
int late_document_prefix_id;
int late_skip_ids[64];
int n_late_skip_ids;
int append_terminal_token;
int terminal_token_id;
int renormalize_truncated;
embed_model_t *cpu_model;
embed_workspace_t *cpu_ws;
embed_late_model_t *cpu_late_model;
embed_late_workspace_t *cpu_late_ws;
#ifdef USE_MLX
embed_mlx_ctx_t *mlx_ctx;
embed_mlx_late_ctx_t *mlx_late_ctx;
#endif
#ifdef USE_CUDA
embed_cuda_ctx_t *cuda_ctx;
embed_cuda_late_ctx_t *cuda_late_ctx;
#endif
} loaded_model;
typedef struct client client;
typedef struct job job;
typedef struct {
aeEventLoop *loop;
int listen_fd;
int completion_pipe[2];
int signal_pipe[2];
const char *host;
int port;
int batch_size;
int max_batch_tokens;
int batch_wait_us;
char *api_key;
loaded_model models[MODEL_COUNT];
int use_mlx;
int use_cuda;
int mlx_quantize_bits;
int mlx_quantize_group_size;
pthread_t worker;
pthread_mutex_t mu;
pthread_cond_t cv;
int stopping;
int worker_ready;
int worker_init_rc;
job *job_head;
job *job_tail;
job *done_head;
job *done_tail;
client *clients;
int n_clients;
} http_server;
typedef struct {
char method[8];
char path[EMBED_HTTP_MAX_PATH + 1];
char version[16];
char *body;
size_t body_len;
char *auth;
} http_request;
struct client {
http_server *srv;
int fd;
int refcount;
int cancelled;
int linked;
uint64_t last_active_ms;
sbuf in;
sbuf out;
size_t sent;
int header_done;
size_t header_len;
size_t content_length;
int close_after_write;
http_request req;
client *prev;
client *next;
};
struct job {
http_server *srv;
client *c;
char method[8];
char path[EMBED_HTTP_MAX_PATH + 1];
char *body;
size_t body_len;
char *auth;
int status;
char *content_type;
char *extra_headers;
char *response;
size_t response_len;
uint64_t created_ns;
uint64_t started_ns;
uint64_t parse_ns;
uint64_t tokenize_ns;
uint64_t infer_ns;
uint64_t encode_ns;
job *next;
};
/* Parsed request state, shared by the prepare (handlers), execute (schedule),
* and render (encode) stages. Each carries its source cJSON root (freed when
* the request is freed), the resolved model, and the tokenized inputs. */
typedef struct {
int *ids;
int n_tokens;
} token_buf;
typedef struct {
job *j;
cJSON *root;
loaded_model *model;
int dims;
const char *encoding;
int n_inputs;
token_buf *tokens;
embed_input_t *inputs;
int total_tokens;
int ready;
} embedding_request;
typedef struct {
int *ids;
int n_tokens;
embed_span_t *spans;
int n_spans;
} contextual_doc;
typedef struct {
job *j;
cJSON *root;
loaded_model *model;
int dims;
const char *encoding;
contextual_doc *docs;
int n_docs;
int total_chunks;
int total_tokens;
int ready;
} contextual_request;
typedef struct {
int *ids;
int n_tokens;
int *keep;
int n_keep;
} late_tokens;
typedef struct {
job *j;
cJSON *root;
loaded_model *model;
late_tokens query;
late_tokens *documents;
int n_documents;
int top_n;
int return_documents;
int query_tokens;
int document_tokens;
int ready;
} rerank_request;
/* ---- server_json.c: request validation and error bodies ---- */
void append_json_string(sbuf *b, const char *s);
char *json_error_body(const char *message, const char *type, size_t *len);
void job_set_error(job *j, int status, const char *message, const char *type);
bool ve_add(cJSON *detail, const char *loc_json, const char *msg, const char *type);
void job_set_422(job *j, cJSON *detail);
cJSON *parse_json_body(job *j, cJSON *detail);
bool cjson_is_integer(cJSON *item);
const char *encoding_from_root(cJSON *root, cJSON *detail, embedding_api_t api);
int text_type_is_query(cJSON *root, cJSON *detail, const char *query_instruct);
int dimensions_from_root(
cJSON *root, cJSON *detail, int min_dim, int max_dim, const char *encoding);
/* ---- server_http.c: connection lifecycle, header parse, response framing ---- */
int set_nonblock(int fd);
void client_incref(client *c);
void client_decref_n(client *c, int n);
int close_client_unlink(client *c);
void close_client(client *c);
void append_http_response_ex(client *c,
int status,
const char *ctype,
const char *extra_headers,
const char *body,
size_t body_len);
void append_http_response(
client *c, int status, const char *ctype, const char *body, size_t body_len);
int queue_write(client *c);
bool route_is_inference(const char *method, const char *path);
void read_cb(aeEventLoop *loop, int fd, void *clientData, int mask);
/* Defined in server.c (request routing): framed requests are handed here. */
void dispatch_request(client *c);
/* ---- server_encode.c: embedding output encoding ---- */
signed char quantize_int8_tanh(float x);
char *encode_embedding(const float *emb, int dims, const char *encoding);
void append_embedding_value(
sbuf *b, int index, const float *emb, int dims, const char *encoding, embedding_api_t api);
void set_response_from_buf(job *j, sbuf *b);
int render_embedding_response(embedding_request *r, const float *embs);
void render_contextual_response(contextual_request *r, const float *embs);
/* ---- server_models.c: registry, load/lifecycle, tokenize/inference dispatch ---- */
extern const model_info k_models[];
model_slot model_slot_for_id(const char *id);
void free_token_bufs(token_buf *t, int n);
void embedding_request_free(embedding_request *r);
void contextual_request_free(contextual_request *r);
void late_tokens_free(late_tokens *t);
void rerank_request_free(rerank_request *r);
int tokenize_one(loaded_model *m, job *j, const char *text, token_buf *out);
int tokenize_input(
loaded_model *m, job *j, const char *text, const char *query_instruct, token_buf *out);
int tokenize_late_text(loaded_model *m, job *j, const char *text, int is_query, late_tokens *out);
int model_embed_batch(loaded_model *m, const embed_input_t *inputs, int batch, float *out);
int model_embed_spans_batch(loaded_model *m,
const embed_context_input_t *inputs,
int batch,
float *out);
int inference_batch_accepts_input(
const loaded_model *m, int batch, int packed_tokens, int next_tokens, int max_batch_tokens);
int configure_loaded_model(loaded_model *m, const embed_config_t *config, int token_dim);
int load_one_model(http_server *s, model_slot slot, const char *path);
void free_models(http_server *s);
/* ---- server_handlers.c: per-endpoint prepare / batch-group / execute ---- */
void prepare_embedding_request(job *j, cJSON *root, http_server *s, embedding_request *out);
int embedding_request_compatible(const embedding_request *a, const embedding_request *b);
int embedding_request_fits_group(const embedding_request *r,
int group_inputs,
int group_tokens,
int max_batch,
int max_batch_tokens);
void execute_embedding_request_list(embedding_request **reqs, int n_reqs);
void prepare_contextual_request(job *j, cJSON *root, http_server *s, contextual_request *out);
int contextual_request_compatible(const contextual_request *a, const contextual_request *b);
int contextual_request_fits_group(const contextual_request *r,
int group_docs,
int group_tokens,
int max_batch,
int max_batch_tokens);
void execute_contextual_request_list(contextual_request **reqs, int n_reqs);
void prepare_rerank_request(job *j, cJSON *root, http_server *s, rerank_request *out);
void execute_rerank_request(rerank_request *r);
/* ---- server_schedule.c: job queue, micro-batching, completion ---- */
void enqueue_job(job *j);
int collect_job_batch(http_server *s, job **jobs, int max_jobs);
void process_job_group(http_server *s, job **jobs, int n_jobs);
void completion_cb(aeEventLoop *loop, int fd, void *clientData, int mask);
#endif /* EMBED_SERVER_INTERNAL_H */
|