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

@binding(2) @group(0) var<storage, read> par_buf_0 : array<vec4<u32>>;

@binding(0) @group(0) var<storage, read_write> x_buf_0 : array<f32>;

@binding(1) @group(0) var<storage, read> weight_buf_0 : array<f32>;

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

fn block_sum_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] = 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;
}

@compute
@workgroup_size(256, 1, 1)
fn per_head_rmsnorm(@builtin(local_invocation_id) lid_0 : vec3<u32>, @builtin(workgroup_id) wid_0 : vec3<u32>)
{
    var tid_1 : u32 = lid_0.x;
    var head_dim_0 : u32 = par_buf_0[i32(0)].x;
    var eps_0 : f32 = (bitcast<f32>((par_buf_0[i32(0)].y)));
    var _S2 : u32 = wid_0.x * head_dim_0;
    var i_0 : u32 = tid_1;
    var partial_0 : f32 = 0.0f;
    for(;;)
    {
        if(i_0 < head_dim_0)
        {
        }
        else
        {
            break;
        }
        var partial_1 : f32 = partial_0 + x_buf_0[_S2 + i_0] * x_buf_0[_S2 + i_0];
        i_0 = i_0 + u32(256);
        partial_0 = partial_1;
    }
    var sum_sq_0 : f32 = block_sum_0(tid_1, partial_0);
    var _S3 : f32 = 1.0f / sqrt(sum_sq_0 / f32(head_dim_0) + eps_0);
    i_0 = tid_1;
    for(;;)
    {
        if(i_0 < head_dim_0)
        {
        }
        else
        {
            break;
        }
        var _S4 : u32 = _S2 + i_0;
        x_buf_0[_S4] = x_buf_0[_S4] * _S3 * weight_buf_0[i_0];
        i_0 = i_0 + u32(256);
    }
    return;
}