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
|
/*
* pplx_embed.h - Pure C inference for Perplexity AI's pplx-embed models.
*
* Supports all pplx-embed-v1 and pplx-embed-context-v1 variants (0.6B, 4B).
*/
#ifndef PPLX_EMBED_H
#define PPLX_EMBED_H
#include <stddef.h>
#include <stdint.h>
/* ========================================================================
* Shared constants (identical across all models)
* ======================================================================== */
#define PPLX_VOCAB_SIZE 151936
#define PPLX_HEAD_DIM 128
#define PPLX_MAX_LAYERS 64 /* upper bound for stack arrays */
/* ========================================================================
* Model Configuration (populated from config.json at load time)
* ======================================================================== */
typedef struct {
int hidden_size; /* 0.6B: 1024 4B: 2560 */
int n_layers; /* 0.6B: 28 4B: 36 */
int n_heads; /* 0.6B: 16 4B: 32 */
int n_kv_heads; /* 0.6B: 8 4B: 8 */
int head_dim; /* always 128 */
int q_dim; /* n_heads * head_dim */
int kv_dim; /* n_kv_heads * head_dim */
int intermediate_size; /* 0.6B: 3072 4B: 9728 */
int vocab_size; /* 151936 */
float rms_norm_eps; /* 1e-6 */
float rope_theta; /* 1e6 */
} pplx_config_t;
/* ========================================================================
* Model + workspace ownership
* ======================================================================== */
typedef struct pplx_model pplx_model_t;
typedef struct pplx_workspace pplx_workspace_t;
typedef struct {
const int *ids;
int n_tokens;
} pplx_input_t;
/* ========================================================================
* API
* ======================================================================== */
/*
* Load immutable model data from a directory containing:
* config.json - model hyperparameters
* model.safetensors (or sharded) - F32 or BF16 weights
* vocab.json + merges.txt - BPE tokenizer
*
* Reads config.json to determine model size (0.6B vs 4B).
* Returns NULL on error.
*/
pplx_model_t *pplx_model_load(const char *model_dir);
void pplx_model_free(pplx_model_t *model);
/*
* Allocate/free mutable scratch buffers for a model. The model must outlive
* workspaces created from it.
*/
pplx_workspace_t *pplx_workspace_new(const pplx_model_t *model);
void pplx_workspace_free(pplx_workspace_t *ws);
/* Config accessor for opaque model ownership. */
const pplx_config_t *pplx_model_config(const pplx_model_t *model);
/*
* Compute embedding for a token sequence.
*
* 1. Token embedding lookup
* 2. N transformer layers (bidirectional GQA + RoPE + SwiGLU)
* 3. Final RMSNorm
* 4. Mean pooling over all positions
* 5. L2 normalization
*
* Returns malloc'd float[hidden_size] (caller frees). NULL on error.
*/
float *pplx_model_embed(const pplx_model_t *model, pplx_workspace_t *ws,
const int *token_ids, int n_tokens);
/*
* Compute one embedding into caller-provided out[hidden_size].
* Returns 0 on success, -1 on error.
*/
int pplx_model_embed_into(const pplx_model_t *model, pplx_workspace_t *ws,
const int *token_ids, int n_tokens,
float *out_embedding);
/*
* Compute embeddings for a true packed/ragged batch.
*
* out_embeddings is caller-provided [batch, hidden_size].
* Each sequence attends only to tokens from the same input, and RoPE
* positions restart at zero for every input.
*
* Returns 0 on success, -1 on error.
*/
int pplx_model_embed_batch(const pplx_model_t *model, pplx_workspace_t *ws,
const pplx_input_t *inputs, int batch,
float *out_embeddings);
/*
* Run the transformer forward pass WITHOUT pooling.
* Returns the full [n_tokens * hidden_size] output after final RMSNorm.
* Caller frees the returned buffer. NULL on error.
*
* Useful for contextual (late-chunking) models where you need
* per-token embeddings before splitting by separator positions.
*/
float *pplx_model_forward(const pplx_model_t *model, pplx_workspace_t *ws,
const int *token_ids, int n_tokens);
/*
* Run the transformer forward pass into caller-provided
* out_states[n_tokens * hidden_size]. Returns 0 on success, -1 on error.
*/
int pplx_model_forward_into(const pplx_model_t *model, pplx_workspace_t *ws,
const int *token_ids, int n_tokens,
float *out_states);
/*
* Cosine similarity between two L2-normalized vectors.
*/
float pplx_cosine_similarity(const float *a, const float *b, int dim);
/* Verbose level: 0=quiet, 1=info, 2=debug */
extern int pplx_verbose;
extern int qwen_verbose;
#endif /* PPLX_EMBED_H */
|