rlx-runtime 0.2.1

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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184

// 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/>.

//! One BERT-style transformer block end-to-end via MPSGraph vs thunks.
//!
//! Mirrors the in-tree BERT layer construction (`models` / `rlx::models`):
//!   QKV proj (matmul + add) → narrow ×3 → attention → out proj → +residual → LN
//!   FFN1 (matmul + add + GELU) → FFN2 (matmul + add) → +residual → LN
//!
//! cargo run --example bert_block_isolation --release \
//!   --features "cpu,metal,blas-accelerate" -p rlx-runtime

#[cfg(all(feature = "cpu", feature = "metal", target_os = "macos"))]
fn main() {
    use rlx_ir::*;
    use rlx_runtime::{Device, Session};

    let b = 4;
    let s = 8;
    let nh = 12;
    let dh = 32;
    let h = nh * dh;
    let int_dim = 4 * h;

    let build = || {
        let mut g = Graph::new("block");
        let x = g.input("x", Shape::new(&[b, s, h], DType::F32));
        let mask = g.input("mask", Shape::new(&[b, s], DType::F32));

        let qkv_w = g.param("qkv_w", Shape::new(&[h, 3 * h], DType::F32));
        let qkv_b = g.param("qkv_b", Shape::new(&[3 * h], DType::F32));
        let out_w = g.param("out_w", Shape::new(&[h, h], DType::F32));
        let out_b = g.param("out_b", Shape::new(&[h], DType::F32));
        let ln1_g = g.param("ln1_g", Shape::new(&[h], DType::F32));
        let ln1_b = g.param("ln1_b", Shape::new(&[h], DType::F32));
        let fc1_w = g.param("fc1_w", Shape::new(&[h, int_dim], DType::F32));
        let fc1_b = g.param("fc1_b", Shape::new(&[int_dim], DType::F32));
        let fc2_w = g.param("fc2_w", Shape::new(&[int_dim, h], DType::F32));
        let fc2_b = g.param("fc2_b", Shape::new(&[h], DType::F32));
        let ln2_g = g.param("ln2_g", Shape::new(&[h], DType::F32));
        let ln2_b = g.param("ln2_b", Shape::new(&[h], DType::F32));

        let qkv_mm = g.mm(x, qkv_w);
        let qkv = g.add(qkv_mm, qkv_b);
        let q = g.narrow_(qkv, 2, 0, h);
        let k = g.narrow_(qkv, 2, h, h);
        let v = g.narrow_(qkv, 2, 2 * h, h);

        let attn = g.attention_(q, k, v, mask, nh, dh);
        let attn_o = g.mm(attn, out_w);
        let attn_o = g.add(attn_o, out_b);
        let res1 = g.add(x, attn_o);
        let h1 = g.layer_norm(
            res1,
            ln1_g,
            ln1_b,
            -1,
            1e-12,
            Shape::new(&[b, s, h], DType::F32),
        );

        let fc1 = g.mm(h1, fc1_w);
        let fc1 = g.add(fc1, fc1_b);
        let fc1 = g.activation(
            op::Activation::Gelu,
            fc1,
            Shape::new(&[b, s, int_dim], DType::F32),
        );
        let fc2 = g.mm(fc1, fc2_w);
        let fc2 = g.add(fc2, fc2_b);
        let res2 = g.add(h1, fc2);
        let out = g.layer_norm(
            res2,
            ln2_g,
            ln2_b,
            -1,
            1e-12,
            Shape::new(&[b, s, h], DType::F32),
        );
        g.set_outputs(vec![out]);
        g
    };

    // Reproducible param data.
    let fill = |n: usize, seed: f32| -> Vec<f32> {
        (0..n)
            .map(|i| ((i as f32 + seed) * 0.01).sin() * 0.05)
            .collect()
    };
    let x_data = fill(b * s * h, 0.0);
    let mask_data: Vec<f32> = (0..b * s)
        .map(|i| if (i % s) < 6 { 1.0 } else { 0.0 })
        .collect();
    let qkv_w_data = fill(h * 3 * h, 1.0);
    let qkv_b_data = vec![0.0f32; 3 * h];
    let out_w_data = fill(h * h, 2.0);
    let out_b_data = vec![0.0f32; h];
    let ln1_g_data = vec![1.0f32; h];
    let ln1_b_data = vec![0.0f32; h];
    let fc1_w_data = fill(h * int_dim, 3.0);
    let fc1_b_data = vec![0.0f32; int_dim];
    let fc2_w_data = fill(int_dim * h, 4.0);
    let fc2_b_data = vec![0.0f32; h];
    let ln2_g_data = vec![1.0f32; h];
    let ln2_b_data = vec![0.0f32; h];

    let run_with = |use_mpsg: bool, dev: Device| -> Vec<f32> {
        if use_mpsg {
            rlx_ir::env::set("RLX_USE_MPSGRAPH", "1");
            rlx_ir::env::set("RLX_MPSGRAPH_ATTENTION", "1");
        } else {
            unsafe {
                rlx_ir::env::unset("RLX_USE_MPSGRAPH");
            }
            unsafe {
                rlx_ir::env::unset("RLX_MPSGRAPH_ATTENTION");
            }
        }
        let session = Session::new(dev);
        let mut compiled = session.compile(build());
        for (n, d) in [
            ("qkv_w", &qkv_w_data),
            ("qkv_b", &qkv_b_data),
            ("out_w", &out_w_data),
            ("out_b", &out_b_data),
            ("ln1_g", &ln1_g_data),
            ("ln1_b", &ln1_b_data),
            ("fc1_w", &fc1_w_data),
            ("fc1_b", &fc1_b_data),
            ("fc2_w", &fc2_w_data),
            ("fc2_b", &fc2_b_data),
            ("ln2_g", &ln2_g_data),
            ("ln2_b", &ln2_b_data),
        ] {
            compiled.set_param(n, d);
        }
        let outs = compiled.run(&[("x", &x_data), ("mask", &mask_data)]);
        outs.into_iter().next().unwrap_or_default()
    };

    let cpu = run_with(false, Device::Cpu);
    let metal_thunk = run_with(false, Device::Metal);
    let metal_mpsg = run_with(true, Device::Metal);

    println!("CPU thunk[..6]:    {:?}", &cpu[..6]);
    println!("Metal thunk[..6]:  {:?}", &metal_thunk[..6]);
    println!("Metal MPSG[..6]:   {:?}", &metal_mpsg[..6]);

    let diff = |a: &[f32], b: &[f32]| -> f32 {
        a.iter()
            .zip(b)
            .map(|(x, y)| (x - y).abs())
            .fold(0f32, f32::max)
    };
    println!(
        "\nmax_err CPU vs Metal-thunk: {:.3e}",
        diff(&cpu, &metal_thunk)
    );
    println!(
        "max_err CPU vs Metal-MPSG:  {:.3e}",
        diff(&cpu, &metal_mpsg)
    );
    println!(
        "max_err Metal-thunk vs MPSG:{:.3e}",
        diff(&metal_thunk, &metal_mpsg)
    );
}

#[cfg(not(all(feature = "cpu", feature = "metal", target_os = "macos")))]
fn main() {
    eprintln!("requires cpu + metal on macOS");
}