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
|
/*
* embed_distributed.h - Stateless two-node layer-sharded inference.
*/
#ifndef PPLX_DISTRIBUTED_H
#define PPLX_DISTRIBUTED_H
#include "embed.h"
typedef struct pplx_dist_client pplx_dist_client_t;
typedef struct {
int mlx_quantize_bits;
int mlx_quantize_group_size;
int activation_bits; /* 0 defaults to fp32; supported: 32, 16 */
} pplx_dist_options_t;
/*
* Open a coordinator-side connection. The coordinator owns layers
* [0, local_layer_end); the remote worker must own [local_layer_end, N).
*/
pplx_dist_client_t *pplx_dist_client_open(const char *model_dir,
const char *host, int port,
int local_layer_end, int use_mlx);
pplx_dist_client_t *pplx_dist_client_open_with_options(
const char *model_dir, const char *host, int port,
int local_layer_end, int use_mlx, const pplx_dist_options_t *opts);
void pplx_dist_client_free(pplx_dist_client_t *client);
const pplx_config_t *pplx_dist_client_config(const pplx_dist_client_t *client);
int pplx_dist_client_embed_batch(pplx_dist_client_t *client,
const pplx_input_t *inputs, int batch,
float *out_embeddings);
/*
* Run a final-stage worker until the process is terminated. The worker owns
* [layer_start, layer_end), and layer_end must be the model's final layer.
*/
int pplx_dist_run_worker(const char *model_dir, const char *host, int port,
int layer_start, int layer_end, int use_mlx);
int pplx_dist_run_worker_with_options(
const char *model_dir, const char *host, int port,
int layer_start, int layer_end, int use_mlx,
const pplx_dist_options_t *opts);
#endif /* PPLX_DISTRIBUTED_H */
|