rlx-runtime 0.2.13

RLX runtime — feature-gated backends, session API, compile+execute pipeline
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

// 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/>.

//! Tiny DiT-style block train step: adaLN → linear → gated residual → sum loss.

#![cfg(feature = "cpu")]

use rlx_ir::infer::GraphExt;
use rlx_ir::op::AdaNormKind;
use rlx_ir::{DType, Graph, Op, Shape};
use rlx_opt::autodiff::grad_with_loss;
use rlx_runtime::{Device, Session};

const B: usize = 2;
const S: usize = 3;
const D: usize = 4;
const EPS: f32 = 1e-5;

fn fill(n: usize, seed: f32) -> Vec<f32> {
    (0..n)
        .map(|i| seed * (0.11 * (i as f32) - 0.07 * ((i % 3) as f32)))
        .collect()
}

/// Build `out = gated(x, linear(ada(x)), gate)` then scalar sum loss.
fn dit_block_loss() -> Graph {
    let f = DType::F32;
    let mut g = Graph::new("dit_block");
    let x = g.param("x", Shape::new(&[B, S, D], f));
    let scale = g.input("scale", Shape::new(&[B, 1, D], f));
    let shift = g.input("shift", Shape::new(&[B, 1, D], f));
    let gate = g.input("gate", Shape::new(&[B, 1, D], f));
    let w = g.param("w", Shape::new(&[D, D], f));
    let n = g.ada_layer_norm(x, scale, shift, AdaNormKind::LayerNorm, EPS);
    // [B,S,D] @ [D,D] — reshape to 2-D matmul then back.
    let flat = g.reshape_(n, vec![(B * S) as i64, D as i64]);
    let y2 = g.mm(flat, w);
    let y = g.reshape_(y2, vec![B as i64, S as i64, D as i64]);
    let out = g.gated_residual(x, y, gate);
    let loss = g.sum(out, vec![0, 1, 2], false);
    g.set_outputs(vec![loss]);
    g
}

#[test]
fn dit_block_one_train_step_cpu() {
    let g = dit_block_loss();
    let x_id = g
        .nodes()
        .iter()
        .find(|n| matches!(&n.op, Op::Param { name } if name == "x"))
        .map(|n| n.id)
        .unwrap();
    let w_id = g
        .nodes()
        .iter()
        .find(|n| matches!(&n.op, Op::Param { name } if name == "w"))
        .map(|n| n.id)
        .unwrap();
    let bwd = grad_with_loss(&g, &[x_id, w_id]);
    assert!(
        bwd.nodes()
            .iter()
            .any(|n| matches!(n.op, Op::AdaLayerNormBackward { .. }))
    );
    assert!(
        bwd.nodes()
            .iter()
            .any(|n| matches!(n.op, Op::GatedResidualBackward))
    );

    let x0 = fill(B * S * D, 1.0);
    let w0 = fill(D * D, 0.05);
    let scale = fill(B * D, 0.2);
    let shift = fill(B * D, -0.1);
    let gate = fill(B * D, 0.3);
    let d_output = [1.0f32];

    let mut c = Session::new(Device::Cpu).compile(bwd);
    c.set_param("x", &x0);
    c.set_param("w", &w0);
    let outs = c.run(&[
        ("scale", &scale),
        ("shift", &shift),
        ("gate", &gate),
        ("d_output", &d_output),
    ]);
    let loss = outs[0][0];
    let dx = &outs[1];
    let dw = &outs[2];
    assert!(loss.is_finite(), "loss={loss}");
    assert!(dx.iter().all(|v| v.is_finite()));
    assert!(dw.iter().all(|v| v.is_finite()));
    // Non-trivial gradient: at least one entry moves.
    assert!(
        dx.iter().any(|v| v.abs() > 1e-8) || dw.iter().any(|v| v.abs() > 1e-8),
        "expected non-zero grads"
    );
}

#[test]
fn hir_ada_layer_norm_lowers() {
    use rlx_ir::HirGraphExt;
    use rlx_ir::hir::{HirModule, HirMut};

    let f = DType::F32;
    let mut hir = HirModule::new("hir_ada");
    let mut gb = HirMut::new(&mut hir);
    let x = gb.input("x", Shape::new(&[B, S, D], f));
    let scale = gb.input("scale", Shape::new(&[B, 1, D], f));
    let shift = gb.input("shift", Shape::new(&[B, 1, D], f));
    let out = gb.ada_layer_norm(x, scale, shift, AdaNormKind::RmsNorm, EPS);
    gb.set_outputs(vec![out]);
    let mir = hir.lower_to_mir().expect("HIR lower");
    let g = mir.into_graph();
    assert!(
        g.nodes()
            .iter()
            .any(|n| matches!(n.op, Op::AdaLayerNorm { .. }))
    );
}