cera 0.5.5

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

@binding(3) @group(0) var<storage, read> params_0 : array<vec4<u32>>;

@binding(1) @group(0) var<storage, read> x_0 : array<f32>;

@binding(0) @group(0) var<storage, read> w_0 : array<u32>;

@binding(2) @group(0) var<storage, read_write> y_0 : array<f32>;

fn get_wid_0( wid_0 : vec3<u32>) -> u32
{
    return wid_0.x + wid_0.y * u32(65535);
}

fn block_scale_0( blk_byte_0 : u32) -> f32
{
    var word_0 : u32 = w_0[(blk_byte_0 >> (u32(2)))];
    var scale_bits_0 : u32;
    if(((blk_byte_0 & (u32(2)))) != u32(0))
    {
        scale_bits_0 = (word_0 >> (u32(16)));
    }
    else
    {
        scale_bits_0 = (word_0 & (u32(65535)));
    }
    return (unpack2x16float((scale_bits_0)).x);
}

fn q_pair_0( qs_byte_0 : u32) -> u32
{
    var word_1 : u32 = w_0[(qs_byte_0 >> (u32(2)))];
    var _S1 : u32;
    if(((qs_byte_0 & (u32(2)))) != u32(0))
    {
        _S1 = (word_1 >> (u32(16)));
    }
    else
    {
        _S1 = (word_1 & (u32(65535)));
    }
    return _S1;
}

var<workgroup> partials_0 : array<f32, i32(256)>;

@compute
@workgroup_size(32, 1, 1)
fn gemv_q4_0_fast(@builtin(local_invocation_id) lid_0 : vec3<u32>, @builtin(workgroup_id) wid_1 : vec3<u32>)
{
    var _S2 : u32 = params_0[i32(0)].x;
    var nb_0 : u32 = params_0[i32(0)].y / u32(32);
    var _S3 : u32 = nb_0 * u32(18);
    var _S4 : u32 = get_wid_0(wid_1) * u32(8);
    var tid_0 : u32 = lid_0.x;
    var ix_0 : u32 = tid_0 / u32(2);
    var il_0 : u32 = ((tid_0 & (u32(1)))) * u32(8);
    var sumf_0 : array<f32, i32(8)>;
    var r_0 : u32 = u32(0);
    for(;;)
    {
        if(r_0 < u32(8))
        {
        }
        else
        {
            break;
        }
        sumf_0[r_0] = 0.0f;
        r_0 = r_0 + u32(1);
    }
    var _S5 : u32 = ix_0 * u32(32) + il_0;
    var ib_0 : u32 = ix_0;
    var yb_off_0 : u32 = _S5;
    for(;;)
    {
        if(ib_0 < nb_0)
        {
        }
        else
        {
            break;
        }
        var a0_0 : f32 = x_0[yb_off_0];
        var a1_0 : f32 = x_0[yb_off_0 + u32(1)];
        var a2_0 : f32 = x_0[yb_off_0 + u32(2)];
        var a3_0 : f32 = x_0[yb_off_0 + u32(3)];
        var a4_0 : f32 = x_0[yb_off_0 + u32(4)];
        var a5_0 : f32 = x_0[yb_off_0 + u32(5)];
        var a6_0 : f32 = x_0[yb_off_0 + u32(6)];
        var a7_0 : f32 = x_0[yb_off_0 + u32(7)];
        var a8_0 : f32 = x_0[yb_off_0 + u32(16)];
        var a9_0 : f32 = x_0[yb_off_0 + u32(17)];
        var a10_0 : f32 = x_0[yb_off_0 + u32(18)];
        var a11_0 : f32 = x_0[yb_off_0 + u32(19)];
        var a12_0 : f32 = x_0[yb_off_0 + u32(20)];
        var a13_0 : f32 = x_0[yb_off_0 + u32(21)];
        var a14_0 : f32 = x_0[yb_off_0 + u32(22)];
        var a15_0 : f32 = x_0[yb_off_0 + u32(23)];
        var _S6 : f32 = a1_0 / 256.0f;
        var _S7 : f32 = a3_0 / 256.0f;
        var _S8 : f32 = a5_0 / 256.0f;
        var _S9 : f32 = a7_0 / 256.0f;
        var _S10 : f32 = a8_0 / 16.0f;
        var _S11 : f32 = a9_0 / 4096.0f;
        var _S12 : f32 = a10_0 / 16.0f;
        var _S13 : f32 = a11_0 / 4096.0f;
        var _S14 : f32 = a12_0 / 16.0f;
        var _S15 : f32 = a13_0 / 4096.0f;
        var _S16 : f32 = a14_0 / 16.0f;
        var _S17 : f32 = a15_0 / 4096.0f;
        var _S18 : f32 = a0_0 + a1_0 + (a2_0 + a3_0) + (a4_0 + a5_0) + (a6_0 + a7_0) + (a8_0 + a9_0 + (a10_0 + a11_0) + (a12_0 + a13_0) + (a14_0 + a15_0));
        r_0 = u32(0);
        for(;;)
        {
            if(r_0 < u32(8))
            {
            }
            else
            {
                break;
            }
            var _S19 : u32 = _S4 + r_0;
            if(_S19 >= _S2)
            {
                r_0 = r_0 + u32(1);
                continue;
            }
            var blk_byte_1 : u32 = _S19 * _S3 + ib_0 * u32(18);
            var qs_byte_1 : u32 = blk_byte_1 + u32(2) + il_0;
            var q0_0 : u32 = q_pair_0(qs_byte_1);
            var q1_0 : u32 = q_pair_0(qs_byte_1 + u32(2));
            var q2_0 : u32 = q_pair_0(qs_byte_1 + u32(4));
            var q3_0 : u32 = q_pair_0(qs_byte_1 + u32(6));
            sumf_0[r_0] = sumf_0[r_0] + block_scale_0(blk_byte_1) * (_S18 * -8.0f + (a0_0 * f32((q0_0 & (u32(15)))) + a2_0 * f32((q1_0 & (u32(15)))) + a4_0 * f32((q2_0 & (u32(15)))) + a6_0 * f32((q3_0 & (u32(15))))) + (_S6 * f32((q0_0 & (u32(3840)))) + _S7 * f32((q1_0 & (u32(3840)))) + _S8 * f32((q2_0 & (u32(3840)))) + _S9 * f32((q3_0 & (u32(3840))))) + (_S10 * f32((q0_0 & (u32(240)))) + _S12 * f32((q1_0 & (u32(240)))) + _S14 * f32((q2_0 & (u32(240)))) + _S16 * f32((q3_0 & (u32(240))))) + (_S11 * f32((q0_0 & (u32(61440)))) + _S13 * f32((q1_0 & (u32(61440)))) + _S15 * f32((q2_0 & (u32(61440)))) + _S17 * f32((q3_0 & (u32(61440))))));
            r_0 = r_0 + u32(1);
        }
        var yb_off_1 : u32 = yb_off_0 + u32(512);
        ib_0 = ib_0 + u32(16);
        yb_off_0 = yb_off_1;
    }
    r_0 = u32(0);
    for(;;)
    {
        if(r_0 < u32(8))
        {
        }
        else
        {
            break;
        }
        partials_0[r_0 * u32(32) + tid_0] = sumf_0[r_0];
        r_0 = r_0 + u32(1);
    }
    workgroupBarrier();
    var stride_0 : u32 = u32(16);
    for(;;)
    {
        if(stride_0 > u32(0))
        {
        }
        else
        {
            break;
        }
        if(tid_0 < stride_0)
        {
            r_0 = u32(0);
            for(;;)
            {
                if(r_0 < u32(8))
                {
                }
                else
                {
                    break;
                }
                var idx_0 : u32 = r_0 * u32(32) + tid_0;
                partials_0[idx_0] = partials_0[idx_0] + partials_0[idx_0 + stride_0];
                r_0 = r_0 + u32(1);
            }
        }
        workgroupBarrier();
        stride_0 = (stride_0 >> (u32(1)));
    }
    if(tid_0 == u32(0))
    {
        r_0 = u32(0);
        for(;;)
        {
            if(r_0 < u32(8))
            {
            }
            else
            {
                break;
            }
            var _S20 : u32 = _S4 + r_0;
            if(_S20 < _S2)
            {
                y_0[_S20] = partials_0[r_0 * u32(32)];
            }
            r_0 = r_0 + u32(1);
        }
    }
    return;
}