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
|
/*
* qwen_kernels.h - Math kernels for Qwen3-architecture inference
*
* Low-level math operations. All operate on float32 tensors in row-major order.
*/
#ifndef QWEN_KERNELS_H
#define QWEN_KERNELS_H
/* Public-API export annotation. The libraries are built with
* -fvisibility=hidden, so only declarations carrying EMBED_API are exported
* from the shared library; everything else stays internal. */
#ifndef EMBED_API
#if defined(__GNUC__)
#define EMBED_API __attribute__((visibility("default")))
#else
#define EMBED_API
#endif
#endif
#include <stddef.h>
#include <stdint.h>
/* ========================================================================
* Basic Operations
* ======================================================================== */
void qwen_add_inplace(float *a, const float *b, int n);
/* ========================================================================
* Matrix Operations
* ======================================================================== */
/* C = A @ B^T: A[M,K], B[N,K], C[M,N] */
void qwen_matmul_t(float *C, const float *A, const float *B, int M, int K, int N);
/* y = x @ W^T + b: x[seq,in], W[out,in], b[out], y[seq,out] */
void qwen_linear(float *y, const float *x, const float *W, const float *b,
int seq_len, int in_dim, int out_dim);
void qwen_linear_nobias(float *y, const float *x, const float *W,
int seq_len, int in_dim, int out_dim);
/* bf16 weight variants */
void qwen_bf16_to_f32_buf(float *dst, const uint16_t *src, size_t n);
void qwen_linear_nobias_bf16(float *y, const float *x, const uint16_t *W_bf16,
int seq_len, int in_dim, int out_dim);
/* Small-sequence BF16 path: compute two projections with one threaded dispatch. */
void qwen_linear_nobias_bf16_pair(float *a, float *b, const float *x,
const uint16_t *Wa_bf16,
const uint16_t *Wb_bf16,
int seq_len, int in_dim, int a_dim,
int b_dim);
/* Small-sequence BF16 path: compute Q/K/V matvecs with one threaded dispatch. */
void qwen_linear_nobias_bf16_qkv(float *q, float *k, float *v, const float *x,
const uint16_t *Wq_bf16,
const uint16_t *Wk_bf16,
const uint16_t *Wv_bf16,
int seq_len, int in_dim, int q_dim,
int kv_dim);
/* Dot product using the best available local SIMD implementation. */
float qwen_dot_f32(const float *a, const float *b, int n);
/* ========================================================================
* Normalization
* ======================================================================== */
/* RMS Normalization: out = x / rms(x) * weight */
void qwen_rms_norm(float *out, const float *x, const float *weight,
int seq_len, int hidden, float eps);
/* Per-head RMS Normalization for Q/K norms in decoder
* x: [seq, n_heads, head_dim], weight: [head_dim]
* Normalizes each head independently */
void qwen_rms_norm_per_head(float *x, const float *weight,
int seq_len, int n_heads, int head_dim, float eps);
/* ========================================================================
* Activation Functions
* ======================================================================== */
void qwen_softmax(float *x, int rows, int cols);
/* gate = SiLU(gate) * up */
void qwen_silu_mul_inplace(float *gate, const float *up, int n);
/* ========================================================================
* Attention Operations
* ======================================================================== */
/*
* Bidirectional GQA attention over a packed/ragged batch.
*
* Q: [total_seq, n_heads * head_dim]
* K,V: [total_seq, n_kv_heads * head_dim]
* offsets: [batch + 1], token spans for each sequence.
*
* Attention is block-diagonal: tokens only attend within their own
* offsets[b]..offsets[b+1] span.
*/
void qwen_bidirectional_gqa_attention_packed(float *out, const float *Q,
const float *K, const float *V,
const int *offsets, int batch,
int n_heads, int n_kv_heads,
int head_dim, float scale);
/* Optional reusable scratch for the BLAS-tiled packed attention path. */
size_t qwen_bidirectional_gqa_attention_packed_scratch_bytes(
const int *offsets, int batch);
void qwen_bidirectional_gqa_attention_packed_with_scratch(
float *out, const float *Q, const float *K, const float *V,
const int *offsets, int batch, int n_heads, int n_kv_heads,
int head_dim, float scale, float *scratch, size_t scratch_bytes);
/*
* Causal GQA attention over a packed/ragged batch. Query position i attends
* only to positions at or before i within its sequence.
*/
void qwen_causal_gqa_attention_packed(float *out, const float *Q,
const float *K, const float *V,
const int *offsets, int batch,
int n_heads, int n_kv_heads,
int head_dim, float scale);
void qwen_causal_gqa_attention_packed_with_scratch(
float *out, const float *Q, const float *K, const float *V,
const int *offsets, int batch, int n_heads, int n_kv_heads,
int head_dim, float scale, float *scratch, size_t scratch_bytes);
/* ========================================================================
* Position Embeddings
* ======================================================================== */
/*
* NeoX-style RoPE: compute cos/sin for positions.
* cos_out, sin_out: [seq, head_dim]
* cos[d] and cos[half+d] are the same (duplicated for full head_dim).
*/
void qwen_compute_rope_neox(float *cos_out, float *sin_out, const int *positions,
int seq, int head_dim, float theta);
/*
* Apply NeoX-style RoPE to Q or K.
* x: [seq, n_heads * head_dim] (in-place)
* cos_vals, sin_vals: [seq, head_dim]
*/
void qwen_apply_rope_neox(float *x, const float *cos_vals, const float *sin_vals,
int seq, int n_heads, int head_dim);
/* ========================================================================
* Threading
* ======================================================================== */
/* Set number of threads for parallel operations (default: 1).
* Creates a persistent thread pool. Call before inference. */
EMBED_API void qwen_set_threads(int n);
/* Get number of available CPU cores */
EMBED_API int qwen_get_num_cpus(void);
/* Global verbose flag */
EMBED_API extern int qwen_verbose;
#endif /* QWEN_KERNELS_H */
|