forge-ml 0.1.0

A WebGPU-native machine learning runtime in Rust: train and run GPT-2 on any GPU wgpu reaches, with no CUDA and no Python
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

// Fused token + positional embedding gather:
// out[t, c] = wte[ids[t], c] + wpe[t + pos, c]   (wpe skipped if use_wpe == 0)

struct Params {
    t: u32,        // number of tokens
    c: u32,        // embedding dim
    pos: u32,      // position offset of ids[0]
    use_wpe: u32,
    // The bound wte buffer holds rows [row_start, row_end) of the full table
    // (row-chunked to respect max_storage_buffer_binding_size). Each token is
    // written by exactly the dispatch whose chunk owns its id.
    row_start: u32,
    row_end: u32,
    _pad0: u32,
    _pad1: u32,
}

@group(0) @binding(0) var<uniform> p: Params;
@group(0) @binding(1) var<storage, read> ids: array<u32>;
@group(0) @binding(2) var<storage, read> wte: array<f32>;
@group(0) @binding(3) var<storage, read> wpe: array<f32>;
@group(0) @binding(4) var<storage, read_write> out: array<f32>;

@compute @workgroup_size(256)
fn main(
    @builtin(workgroup_id) wid: vec3<u32>,
    @builtin(num_workgroups) nwg: vec3<u32>,
    @builtin(local_invocation_index) li: u32,
) {
    let i = (wid.y * nwg.x + wid.x) * 256u + li;
    let n = p.t * p.c;
    if (i >= n) { return; }
    let t = i / p.c;
    let ch = i % p.c;
    let id = ids[t];
    if (id < p.row_start || id >= p.row_end) { return; }
    var v = wte[(id - p.row_start) * p.c + ch];
    if (p.use_wpe == 1u) {
        v = v + wpe[(t + p.pos) * p.c + ch];
    }
    out[i] = v;
}