rlx-wgpu 0.2.1

Cross-platform GPU backend for RLX via wgpu (Metal/Vulkan/DX12/WebGPU)
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

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.
//! Host-side training backward ops for wgpu arenas (readback → CPU → writeback).

use crate::buffer::Arena;

#[inline]
fn run_on_arena(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    f: impl FnOnce(*mut u8),
) {
    let mut host = arena.read_bytes_range(device, queue, 0, arena.size);
    f(host.as_mut_ptr());
    arena.write_bytes_range(queue, 0, &host);
}

pub fn run_rms_norm_backward_input(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    x: usize,
    gamma: usize,
    beta: usize,
    dy: usize,
    dx: usize,
    rows: u32,
    h: u32,
    eps: f32,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_rms_norm_backward_input_f32(
            x, gamma, beta, dy, dx, rows, h, eps, base,
        );
    });
}

pub fn run_rms_norm_backward_gamma(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    x: usize,
    gamma: usize,
    beta: usize,
    dy: usize,
    dgamma: usize,
    rows: u32,
    h: u32,
    eps: f32,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_rms_norm_backward_gamma_f32(
            x, gamma, beta, dy, dgamma, rows, h, eps, base,
        );
    });
}

pub fn run_rms_norm_backward_beta(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    x: usize,
    gamma: usize,
    beta: usize,
    dy: usize,
    dbeta: usize,
    rows: u32,
    h: u32,
    eps: f32,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_rms_norm_backward_beta_f32(
            x, gamma, beta, dy, dbeta, rows, h, eps, base,
        );
    });
}

pub fn run_rope_backward(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    dy: usize,
    cos: usize,
    sin: usize,
    dx: usize,
    batch: u32,
    seq: u32,
    hidden: u32,
    head_dim: u32,
    n_rot: u32,
    cos_len: u32,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_rope_backward_f32(
            dy, cos, sin, dx, batch, seq, hidden, head_dim, n_rot, cos_len, base,
        );
    });
}

pub fn run_cumsum_backward(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    dy: usize,
    dx: usize,
    rows: u32,
    cols: u32,
    exclusive: bool,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_cumsum_backward_f32(dy, dx, rows, cols, exclusive, base);
    });
}

pub fn run_gather_backward(
    arena: &Arena,
    device: &wgpu::Device,
    queue: &wgpu::Queue,
    dy: usize,
    indices: usize,
    dst: usize,
    outer: u32,
    axis_dim: u32,
    num_idx: u32,
    trailing: u32,
) {
    run_on_arena(arena, device, queue, |base| unsafe {
        rlx_cpu::thunk::execute_gather_backward_f32(
            dy, indices, dst, outer, axis_dim, num_idx, trailing, base,
        );
    });
}