cera 0.1.1

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

// Element-wise operations on f32 buffers.
//
// Bind group 0:
//   @binding(0) a: array<f32>  (read-write, modified in-place)
//   @binding(1) b: array<f32>  (read-only)
//   @binding(2) params: vec2<u32>  (n, unused)
//
// Dispatch: (ceil(n / 256), 1, 1) workgroups

@group(0) @binding(0) var<storage, read_write> a: array<f32>;
@group(0) @binding(1) var<storage, read> b: array<f32>;
@group(0) @binding(2) var<storage, read> params: vec2<u32>;

@compute @workgroup_size(256, 1, 1)
fn add_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i = gid.x;
    let n = params.x;
    if i >= n { return; }
    a[i] = a[i] + b[i];
}

// Residual add with a scalar on the addend: a[i] += s * b[i], with `s`
// bit-cast in params.y. Granite 3.x sets `s` to its residual multiplier; every
// other arch passes s = 1.0, making this identical to `add_inplace`.
@compute @workgroup_size(256, 1, 1)
fn scaled_add_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i = gid.x;
    let n = params.x;
    if i >= n { return; }
    a[i] = a[i] + bitcast<f32>(params.y) * b[i];
}

@compute @workgroup_size(256, 1, 1)
fn mul_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i = gid.x;
    let n = params.x;
    if i >= n { return; }
    a[i] = a[i] * b[i];
}

@compute @workgroup_size(256, 1, 1)
fn silu_mul_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i = gid.x;
    let n = params.x;
    if i >= n { return; }
    let g = a[i];
    a[i] = (g / (1.0 + exp(-g))) * b[i];
}