rlx-cuda 0.2.2

NVIDIA CUDA backend — cuBLAS for matmul + NVRTC-compiled kernels for everything else, via the pure-Rust `cudarc` crate.
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

// 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/>.

//! rlx-cuda build script.
//!
//! By default this is a no-op — the crate is pure Rust + cudarc and
//! NVRTC-compiles its `.cu` sources at runtime.
//!
//! With `--features hip-cpu-validate`, we compile a single C++ TU
//! (`cpp/cpu_dispatch.cpp`) against HIP-CPU's header-only runtime,
//! producing a static lib that the Rust crate links against. The CPU
//! path is *strictly* a development convenience — it lets us run the
//! same kernel sources on CPU threads on Mac/Docker for parity-check
//! purposes, without renting a CUDA box. Production paths still use
//! cudarc + real NVIDIA hardware.

fn main() {
    println!("cargo:rerun-if-changed=build.rs");

    #[cfg(feature = "hip-cpu-validate")]
    build_hip_cpu();
}

#[cfg(feature = "hip-cpu-validate")]
fn build_hip_cpu() {
    use std::path::Path;
    let hip_cpu_include = Path::new("vendor/HIP-CPU/include");
    if !hip_cpu_include.exists() {
        panic!(
            "rlx-cuda hip-cpu-validate: missing HIP-CPU headers at {}\n\
             Initialize the submodule with:\n\
             \n\
             \tgit submodule add https://github.com/ROCm-Developer-Tools/HIP-CPU.git \\\n\
             \t    rlx-cuda/vendor/HIP-CPU\n\
             \n\
             (or whatever upstream HIP-CPU mirror is current)",
            hip_cpu_include.display()
        );
    }

    let kernels_dir = Path::new(env!("CARGO_MANIFEST_DIR")).join("../rlx-gpu-kernels/kernels");

    println!("cargo:rerun-if-changed=cpp/cpu_dispatch.cpp");
    println!("cargo:rerun-if-changed={}", kernels_dir.display());

    cc::Build::new()
        .cpp(true)
        .std("c++17")
        .file("cpp/cpu_dispatch.cpp")
        .include(hip_cpu_include)
        .include(&kernels_dir)
        // HIP-CPU runtime mode — selects the CPU thread-pool backend
        // instead of any GPU runtime.
        .define("__HIP_CPU_RT__", None)
        // Allow `__global__`, `__device__`, etc. attributes used in
        // our `.cu` files to be treated as no-ops on the CPU side.
        .flag_if_supported("-Wno-unknown-attributes")
        .flag_if_supported("-Wno-deprecated-declarations")
        .compile("rlx_cuda_cpu_dispatch");

    // pthread for HIP-CPU's std::thread fallback.
    println!("cargo:rustc-link-lib=pthread");
}