cera 0.5.2

Rust-native LLM inference engine
Documentation 
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

@binding(7) @group(0) var<storage, read> par_buf_0 : array<vec4<u32>>;

@binding(0) @group(0) var<storage, read> q_buf_0 : array<f32>;

@binding(4) @group(0) var<storage, read> bu_buf_0 : array<f32>;

@binding(5) @group(0) var<storage, read> bv_buf_0 : array<f32>;

@binding(1) @group(0) var<storage, read> k_buf_0 : array<f32>;

@binding(3) @group(0) var<storage, read> p_buf_0 : array<f32>;

@binding(2) @group(0) var<storage, read> v_buf_0 : array<f32>;

@binding(6) @group(0) var<storage, read_write> out_buf_0 : array<f32>;

var<workgroup> qu_0 : array<f32, i32(128)>;

var<workgroup> qv_0 : array<f32, i32(128)>;

var<workgroup> scores_0 : array<f32, i32(1024)>;

var<workgroup> scratch_0 : array<f32, i32(256)>;

fn block_max_0( tid_0 : u32,  v_0 : f32) -> f32
{
    scratch_0[tid_0] = v_0;
    workgroupBarrier();
    var s_0 : u32 = u32(128);
    for(;;)
    {
        if(s_0 > u32(0))
        {
        }
        else
        {
            break;
        }
        if(tid_0 < s_0)
        {
            scratch_0[tid_0] = max(scratch_0[tid_0], scratch_0[tid_0 + s_0]);
        }
        workgroupBarrier();
        s_0 = (s_0 >> (u32(1)));
    }
    var _S1 : f32 = scratch_0[i32(0)];
    return _S1;
}

fn block_sum_0( tid_1 : u32,  v_1 : f32) -> f32
{
    scratch_0[tid_1] = v_1;
    workgroupBarrier();
    var s_1 : u32 = u32(128);
    for(;;)
    {
        if(s_1 > u32(0))
        {
        }
        else
        {
            break;
        }
        if(tid_1 < s_1)
        {
            scratch_0[tid_1] = scratch_0[tid_1] + scratch_0[tid_1 + s_1];
        }
        workgroupBarrier();
        s_1 = (s_1 >> (u32(1)));
    }
    var _S2 : f32 = scratch_0[i32(0)];
    return _S2;
}

@compute
@workgroup_size(256, 1, 1)
fn audio_xl_attention(@builtin(local_invocation_id) lid_0 : vec3<u32>, @builtin(workgroup_id) wid_0 : vec3<u32>)
{
    var ac_0 : f32;
    var bd_0 : f32;
    var tid_2 : u32 = lid_0.x;
    var tokens_0 : u32 = par_buf_0[i32(0)].x;
    var head_dim_0 : u32 = par_buf_0[i32(0)].z;
    var _S3 : f32 = (bitcast<f32>((par_buf_0[i32(0)].w)));
    var q_idx_0 : u32 = wid_0.x;
    var dim_0 : u32 = par_buf_0[i32(0)].y * head_dim_0;
    var head_base_0 : u32 = wid_0.y * head_dim_0;
    var _S4 : u32 = q_idx_0 * dim_0 + head_base_0;
    var d_0 : u32 = tid_2;
    for(;;)
    {
        if(d_0 < head_dim_0)
        {
        }
        else
        {
            break;
        }
        var qd_0 : f32 = q_buf_0[_S4 + d_0];
        var _S5 : u32 = head_base_0 + d_0;
        qu_0[d_0] = qd_0 + bu_buf_0[_S5];
        qv_0[d_0] = qd_0 + bv_buf_0[_S5];
        d_0 = d_0 + u32(256);
    }
    workgroupBarrier();
    var _S6 : u32 = tokens_0 - u32(1);
    var key_0 : u32 = tid_2;
    for(;;)
    {
        if(key_0 < tokens_0)
        {
        }
        else
        {
            break;
        }
        var _S7 : u32 = key_0 * dim_0 + head_base_0;
        var _S8 : u32 = (_S6 + key_0 - q_idx_0) * dim_0 + head_base_0;
        d_0 = u32(0);
        ac_0 = 0.0f;
        bd_0 = 0.0f;
        for(;;)
        {
            if(d_0 < head_dim_0)
            {
            }
            else
            {
                break;
            }
            var ac_1 : f32 = ac_0 + qu_0[d_0] * k_buf_0[_S7 + d_0];
            var bd_1 : f32 = bd_0 + qv_0[d_0] * p_buf_0[_S8 + d_0];
            d_0 = d_0 + u32(1);
            ac_0 = ac_1;
            bd_0 = bd_1;
        }
        scores_0[key_0] = (ac_0 + bd_0) * _S3;
        key_0 = key_0 + u32(256);
    }
    workgroupBarrier();
    ac_0 = -3.4028234663852886e+38f;
    key_0 = tid_2;
    for(;;)
    {
        if(key_0 < tokens_0)
        {
        }
        else
        {
            break;
        }
        var _S9 : f32 = max(ac_0, scores_0[key_0]);
        var key_1 : u32 = key_0 + u32(256);
        ac_0 = _S9;
        key_0 = key_1;
    }
    var _S10 : f32 = block_max_0(tid_2, ac_0);
    key_0 = tid_2;
    bd_0 = 0.0f;
    for(;;)
    {
        if(key_0 < tokens_0)
        {
        }
        else
        {
            break;
        }
        var e_0 : f32 = exp(scores_0[key_0] - _S10);
        scores_0[key_0] = e_0;
        var partial_0 : f32 = bd_0 + e_0;
        key_0 = key_0 + u32(256);
        bd_0 = partial_0;
    }
    workgroupBarrier();
    var _S11 : f32 = block_sum_0(tid_2, bd_0);
    var _S12 : f32 = 1.0f / _S11;
    d_0 = tid_2;
    for(;;)
    {
        if(d_0 < head_dim_0)
        {
        }
        else
        {
            break;
        }
        key_0 = u32(0);
        var acc_0 : f32 = 0.0f;
        for(;;)
        {
            if(key_0 < tokens_0)
            {
            }
            else
            {
                break;
            }
            var acc_1 : f32 = acc_0 + scores_0[key_0] * v_buf_0[key_0 * dim_0 + head_base_0 + d_0];
            key_0 = key_0 + u32(1);
            acc_0 = acc_1;
        }
        out_buf_0[_S4 + d_0] = acc_0 * _S12;
        d_0 = d_0 + u32(256);
    }
    return;
}