Hermetic Toolchains for ML

This project provides Bazel rules for ML project to achieve hermetic builds.

C++ and CUDA hermetic builds benefits:

• Reproducibility: Every build produces identical results regardless of the developer's machine environment.
• Consistency: Eliminates "works on my machine" issues, ensuring builds are consistent across different development environments.
• Isolation: Builds are isolated from the host system, minimizing unexpected dependencies and side effects.

Configure C++ toolchains in rules_ml_toolchain

Add the following code to WORKSPACE file:

http_archive(
    name = "rules_ml_toolchain",
    sha256 = "dd6035b2aa22ec22c7598c0c78d1f593a74606d787d1059d19ab0f9b581e513d",
    strip_prefix = "rules_ml_toolchain-4d9fa39eda9c769db86770a13ce2c2e2090bced8",
    urls = [
        "https://github.com/google-ml-infra/rules_ml_toolchain/archive/4d9fa39eda9c769db86770a13ce2c2e2090bced8.tar.gz",
    ],
)

load(
    "@rules_ml_toolchain//cc/deps:cc_toolchain_deps.bzl",
    "cc_toolchain_deps",
)

cc_toolchain_deps()

register_toolchains("@rules_ml_toolchain//cc:linux_x86_64_linux_x86_64")
register_toolchains("@rules_ml_toolchain//cc:linux_aarch64_linux_aarch64")

If CUDA or SYCL initialization is required, ensure this block is inserted before either initialization occurs.

It must be ensured that builds for Linux x86_64 / aarch64 are run without the --noincompatible_enable_cc_toolchain_resolution flag. Furthermore, reliance on environment variables like CLANG_COMPILER_PATH, BAZEL_COMPILER, CC, or CXX must be avoided.

For diagnosing the utility set being used during build or test execution, the --subcommands flag should be appended to the Bazel command. This will facilitate checking that the compiler or linker are not being used from your machine.

Configure hermetic CUDA, CUDNN, NCCL and NVSHMEM

For detailed instructions on how to configure hermetic CUDA, CUDNN, NCCL and NVSHMEM, click this link.

Configure the LLVM / Sysroot in rules_ml_toolchain

LLVM 18 and the linux_glibc_2_27 sysroot are used for compilation by default. To change these defaults, specify the required LLVM version and sysroot distribution in .bazelrc file.

For example, to configure LLVM 20 with linux_glibc_2_31, update your .bazelrc with below lines

common --enable_platform_specific_config

build:linux --repo_env=LLVM_VERSION=20
build:linux --repo_env=SYSROOT_DIST=linux_glibc_2_31

Supported versions of LLVM

Version	Execution OS
18	Linux x86_64 / aarch64, macOS aarch64
19	Linux x86_64
20	Linux x86_64 / aarch64
21	Linux x86_64

Available sysroots

Name	Architecture	GCC	GLIBC	C++ Standard	Used OS
linux_glibc_2_27	x86_64, aarch64	GCC 8	2.27	C++17	Ubuntu 18.04
linux_glibc_2_31	x86_64, aarch64	GCC 10	2.31	C++20	Ubuntu 20.04
linux_glibc_2_35	x86_64	GCC 12	2.35	C++23 partial support	Ubuntu 22.04

Run rules_ml_toolchain tests

CPU hermetic tests

Project supports CPU hermetic builds on:

• Linux x86_64 / aarch64
• macOS aarch64 - In Development

The command allows you to run hermetic build tests:

bazel test //cc/tests/cpu:all

Non-hermetic CPU builds

When executor and a target are the same, you still can run non-hermetic build. Command should look like:

bazel build //cc/tests/cpu:all --config=clang_local

For details, look at the .bazelrc file, specifically the clang_local configuration.

CUDA and hermetic toolchains tests

Project supports GPU hermetic builds on Linux x86_64 / aarch64. Running tests requires a machine with an NVIDIA GPU.

Hermetic tests could be run with the help of the command:

Build by Clang

bazel test //cc/tests/gpu:all --config=build_cuda_with_clang --config=cuda --config=cuda_libraries_from_stubs

Build by NVCC

bazel test //cc/tests/gpu:all --config=build_cuda_with_nvcc --config=cuda --config=cuda_libraries_from_stubs

CUDA and non-hermetic toolchains tests

When the executor and the target are the same, a non-hermetic GPU build can still be run.

Build by Clang

bazel test //cc/tests/gpu:all --config=build_cuda_with_clang --config=cuda_clang_local --config=cuda_libraries_from_stubs

Build by NVCC

bazel test //cc/tests/gpu:all --config=build_cuda_with_nvcc --config=cuda_clang_local --config=cuda_libraries_from_stubs

For details, look at the .bazelrc file, specifically the cuda_clang_local configuration.