Bazel + Python rules for cross-platform external dependencies
rules_pycross — Python + cross platform
rules_pycross lets you use Python lock files with Bazel, enabling cross-platform, hermetic builds of Python dependencies — including native extensions.
[!NOTE]
#243 merged a major "v2" overhaul. The prior release can be found at
the v1 branch.
Features
- Import lock files from uv, PDM, Poetry, or PEP 751 pylock.toml
- Build source distributions inside Bazel build actions, not during workspace initialization
- Pluggable build backends: setuptools, meson, cmake, maturin, and generic PEP 517
- Cross-platform sdist builds — build wheels for Linux and macOS from either host with an appropriate cross-compilation toolchain (e.g., toolchains_llvm)
- Multi-workspace support for monorepos with shared dependency deduplication
- Conflict/variant resolution for mutually exclusive dependencies (e.g., torch CPU vs. CUDA)
- Compatible with
rules_pythonand Gazelle
See the CI results for cross-platform build and test evidence.
Getting Started
Add your lock file import to MODULE.bazel:
uv = useextension("@rulespycross//pycross/extensions:uv.bzl", "uv")
uv.workspace( name = "pypi", lock_file = "//:uv.lock", ) use_repo(uv, "pypi")
For a single-project lock file, this is all you need — rules_pycross auto-discovers the project from the lock file and creates a repo named "pypi" with default dependencies.
To customize which projects or dependency groups are included, add a uv.repo() tag:
uv.workspace(
name = "pypi",
lock_file = "//:uv.lock",
)
uv.repo(
dependency_groups = ["default", "optional:grpc"],
workspace = "pypi",
)
use_repo(uv, "pypi")
After this, packages are available as @pypi//package_name. A requirement() macro is generated in @pypi//:requirements.bzl.
Other lock formats work the same way via their respective extensions: pdm.bzl, poetry.bzl, or pylock.bzl.
Repository Defaults and Auto-Generation
The Default Workspace Repository
For simple, single-project lock files, you can omit the uv.repo() tag entirely. rules_pycross will automatically synthesize a repository for you with the following defaults:
- Name: Matches the workspace name (e.g.,
@pypi). - Content: Includes only the
"default"dependency group of the single discovered project.
uv.repo() tags.
Project File Discovery
rules_pycross automatically discovers your pyproject.toml files by inspecting the workspace members defined in the lock file. If it finds none (e.g. for a standalone lock file), it falls back to looking for a pyproject.toml next to the lock file.
If you have additional pyproject.toml files that aren't part of the lock file's defined workspace members, but contain build settings or dependency definitions you need rulespycross to see, you can explicitly add them using extraproject_files on the workspace() tag:
uv.workspace(
name = "pypi",
lock_file = "//:uv.lock",
extraprojectfiles = ["//:pyproject.toml", "//tools:pyproject.toml"],
)
These explicitly specified files are appended to the auto-discovered files.
Transitive Aliases
By default, rulespycross only generates top-level aliases for packages that are explicitly defined as dependencies in your project. If you want to be able to depend on transitive dependencies directly using requirement("transitive-package"), you can enable createtransitive_aliases on your uv.repo() tag:
uv.repo(
workspace = "pypi",
createtransitivealiases = True,
)
If a transitive package has multiple versions in the lock file, rules_pycross will print a warning and alias to the highest version.
The Internal Build Tools Repository (__build)
For every workspace, rules_pycross also auto-generates an internal companion repository named <workspace>build (e.g., @pypibuild).
- Purpose: Provides build-time tools (like
setuptools,hatchling, etc.) required to build source distributions (sdists) hermetically.
) from the workspace.
This repository is managed automatically. However, if you need to customize its settings (such as restricting its dependency groups), you can override it by explicitly declaring a repo with the <workspace>__build name:
uv.repo(
name = "pypi__build",
dependency_groups = ["default", "group:build"],
workspace = "pypi",
)
Migrating from the legacy two-extension pattern
The previous approach used lockimport / lockrepos (or lock) extensions. These have been removed. Migrate by replacing them with the per-format extension:
# Before (removed):
lockimport = useextension("@rulespycross//pycross/extensions:lockimport.bzl", "lock_import")
lockimport.importuv(
lock_file = "//:uv.lock",
project_file = "//:pyproject.toml",
repo = "pypi",
)
lock_import.package(
name = "numpy",
always_build = True,
repo = "pypi",
)
lockrepos = useextension("@rulespycross//pycross/extensions:lockrepos.bzl", "lock_repos")
userepo(lockrepos, "pypi")
After:
uv = useextension("@rulespycross//pycross/extensions:uv.bzl", "uv")
uv.workspace(
name = "pypi",
lock_file = "//:uv.lock",
)
uv.repo(
workspace = "pypi",
)
uv.package(
name = "numpy",
always_build = True,
workspace = "pypi", # was: repo = "pypi"
)
use_repo(uv, "pypi")
[!TIP]
If you are migrating from a 1.x target layout where packages were referenced as@pypi//:packagename(with a colon), you can enablelegacycreaterootaliases = Trueon youruv.repo()tag to generate these aliases in the 2.x repo.
Toolchain Configuration
Python versions are auto-discovered from registered rulespython toolchains, and all supported platforms are included by default. You can restrict or customize this behavior using pycross.configuretoolchains() in your MODULE.bazel:
pycross = useextension("@rulespycross//pycross/extensions:pycross.bzl", "pycross")
pycross.configure_toolchains(
# Restrict supported platforms
platforms = [
"x86_64-unknown-linux-gnu",
"aarch64-apple-darwin",
],
# Restrict supported Python versions
python_versions = [
"3.11",
"3.12",
],
# Set platform version constraints
glibc_version = "2.28",
macos_version = "15.0",
musl_version = "1.2",
)
By default, rulespycross will automatically register toolchains for all configured platforms and versions. You can disable this by setting registertoolchains = False if you prefer to register them manually.
How It Works
A pip install operation can be broken down into:
- Determine the target environment (OS, CPU, Python version)
- Resolve dependencies from a lock file
- Select pre-built wheels or source distributions
- Download and build
rules_pycross maps each step to Bazel primitives:
- Native Bazel Platforms — target environments are determined by standard Bazel
@platformsconstraints andrules_pythontoolchain flags, mapped directly to PEP 508 markers at analysis time. - Lock extensions (
uv,pdm, etc.) — translates a lock file and resolves dependencies into Bazel repository rules:http_filefor downloads, build rules for source distributions. - Build backends (
setuptoolsbuild,mesonbuild, etc.) — build sdists into wheels inside sandboxed Bazel actions with remote execution support. pycrosswheellibrary— extracts a wheel (downloaded or built) and provides it as apy_library.
Dependency Groups
The dependency_groups attribute on uv.repo() controls which dependency groups are included. It accepts a list of group specifiers:
"default"— the project's default dependencies"optional:<name>"— a specific optional dependency group ([project.optional-dependencies])"group:<name>"— a specific dependency group ([dependency-groups])
"optional:" / "group:*" — all optional or all dependency groups
"" — all groups (default + all optional + all development)
The default is ["default"].
uv.repo(
dependency_groups = ["default", "optional:grpc", "group:test"],
workspace = "pypi",
)
Extras
When a dependency is used with extras (e.g., google-api-core[asyncrest,grpc]), rulespycross generates separate targets for the base package and each extra:
@pypi//googleapicore # Full package with all requested extras
@pypi//googleapicore:[] # Base package only (no extra dependencies)
@pypi//googleapicore:[asyncrest] # Just the asyncrest extra and its dependencies
@pypi//googleapicore:[grpc] # Just the grpc extra and its dependencies
The requirement() macro supports this syntax directly:
load("@pypi//:requirements.bzl", "requirement")
py_library( name = "my_lib", deps = [ requirement("google-api-core[grpc]"), ], )
Multi-Workspace Lock Import
rules_pycross supports importing multiple members of a single workspace lock file into a shared backing repository. This is useful for monorepos where different subprojects need different dependency subsets.
UV Workspace (Single Lock, Multiple Members)
uv = useextension("@rulespycross//pycross/extensions:uv.bzl", "uv")
1. Declare the workspace (shared lock file and settings)
uv.workspace(
name = "shared",
lock_file = "//:uv.lock",
)
2. Import all projects into a single repo
uv.repo(
name = "lock_all",
projects = ["*"],
workspace = "shared",
)
userepo(uv, "lockall")
All members share a single backing package_repo — overlapping packages are downloaded and built only once.
Per-Member Repos
To create separate repos per workspace member with different dependency selections:
uv.repo(
name = "lock_a",
projects = ["project-a"],
workspace = "shared",
)
uv.repo(
name = "lock_b",
projects = ["project-b"],
dependency_groups = ["default", "optional:grpc", "group:testing"],
workspace = "shared",
)
userepo(uv, "locka", "lock_b")
Package Annotations in a Workspace
uv.package() annotations target a workspace. The workspace attribute can be omitted if the module declares only one workspace:
# Apply to all members of the "shared" workspace
uv.package(
name = "regex",
installexcludeglobs = ["test_regex.py"],
workspace = "shared",
)
Use name = "*" to set defaults for all packages. A specific annotation for a package fully replaces the wildcard:
# Force all packages to build from source by default
uv.package(
name = "*",
always_build = True,
workspace = "shared",
)
Override the wildcard for a specific package
uv.package(
name = "requests",
always_build = False,
workspace = "shared",
)
Multiple Independent Lock Files
If your projects use separate lock files (not a shared workspace lock), declare separate workspaces:
uv.workspace(
name = "frontend_deps",
lock_file = "//frontend:uv.lock",
)
uv.repo(
workspace = "frontend_deps",
)
uv.workspace(
name = "ml_deps",
lock_file = "//ml:uv.lock",
)
uv.repo(
workspace = "ml_deps",
)
userepo(uv, "frontenddeps", "ml_deps")
Sdist Builds and Build Overrides
rules_pycross uses a pluggable build backend architecture. Build backends are automatically detected from the build-system.build-backend value in each package's pyproject.toml.
Supported Backends
| Backend | Detected from build-backend | Use case | |---|---|---| | pep517build | hatchling, flitcore, pdm.backend, poetry.core.masonry.api | Pure-Python packages (default fallback) | | setuptoolsbuild | setuptools.buildmeta | C extension packages using setuptools | | setuptoolsrustbuild | setuptools.build_meta (when setuptools-rust is in build-system.requires) | Rust+Python packages using setuptools-rust | | meson_build | mesonpy | Scientific packages (numpy, pandas, etc.) | | cmakebuild | scikitbuild_core.build, skbuild | Packages using CMake/scikit-build | | maturin_build | maturin | Rust+Python packages via maturin |
Forcing a Package to Build from Source
By default, rulespycross uses pre-built wheels when available. To force building from source, set alwaysbuild = True:
uv.package(
name = "numpy",
always_build = True,
workspace = "pypi",
)
Extra Build Tools
When building a package from source, rulespycross automatically includes the build-system.requires packages from the sdist's pyproject.toml. If a package needs additional Python packages at build time (e.g., cython, numpy, setuptools-scm), declare them with extrabuild_tools:
uv.package(
name = "pandas",
extrabuildtools = ["cython@0.29.36", "numpy@1.26.4"],
workspace = "pypi",
)
These package keys must match entries in the lock file. Only packages that aren't already runtime dependencies are added as build-only deps.
Custom Build Tools Repository
By default, build tools are resolved from the internal <workspace>_build repository. If a specific package needs to resolve its build dependencies from a different repository, you can specify buildtools_repo in its package() annotation:
uv.package(
name = "my-complex-package",
buildtoolsrepo = "mycustombuild_deps",
workspace = "pypi",
)
Default Extra Build Tools
Use name = "*" to set default extra build tools for all packages in a workspace. A specific extrabuildtools on an individual package fully replaces the wildcard:
# Default: every sdist build gets cython available
uv.package(
name = "*",
extrabuildtools = ["cython@0.29.36"],
workspace = "pypi",
)
numpy gets its own specific set instead
uv.package(
name = "numpy",
extrabuildtools = ["cython@0.29.36", "oldest-supported-numpy@0.9"],
workspace = "pypi",
)
Build Overrides
When packages need native dependencies, compiler flags, environment variables, or other build customizations, use the backend-specific override extensions. Use name = "*" to set defaults for all packages built with that backend.
Setuptools
setuptools = useextension("@rulespycross//pycross/backends:setuptools.bzl", "setuptools")
setuptools.override( name = "psycopg2", repo = "pypi", copts = ["-O2"], tooldeps = {"pgconfig": "@@//deps/psycopg2:pg_config"}, build_env = {"LDFLAGS": "-L/usr/lib"}, )
Meson
Building numpy with OpenBLAS, using pycrossccpkg_config to bridge Bazel CC deps into meson:
load("@rulespycross//pycross:defs.bzl", "pycrossccpkgconfig")
load("@pypi//backend:mesonbuild.bzl", "meson_build")
Generate a pkg-config .pc file so meson can find OpenBLAS
pycrossccpkg_config(
name = "genopenblaspc_file",
dep = "//third_party/openblas",
lib_name = "scipy-openblas",
version = "0.3.20",
)
meson_build( name = "wheel", build_deps = [ "@pypi//meson:pkg", "@pypi//ninja:pkg", "@pypi//cython:pkg", ], config_settings = { "setup-args": [], "compile-args": ["-v"], }, copts = ["-Wl,-s"], nativedeps = ["//thirdparty/openblas"], path_tools = [":cython"], pkgconfigfiles = [":genopenblaspc_file"], sdist = "@pypi//numpy:sdist", )
Maturin
For Rust+Python packages, use the rulespycrossbackend_maturin module:
bazeldep(name = "rulespycrossbackendmaturin", version = "0.0.0")
bazeldep(name = "rulesrust", version = "0.68.0")
Register Rust toolchains
rust = useextension("@rulesrust//rust:extensions.bzl", "rust")
rust.toolchain(
edition = "2021",
extratargettriples = [
"aarch64-apple-darwin",
"aarch64-unknown-linux-gnu",
"x86_64-unknown-linux-gnu",
],
)
Mark packages for source build
uv.package(
name = "rpds-py",
always_build = True,
workspace = "pypi",
)
uv.package(
name = "jiter",
always_build = True,
workspace = "pypi",
)
Provide a Cargo.lock for jiter (when the sdist doesn't include one)
maturin = useextension("@rulespycrossbackendmaturin//extensions:maturin.bzl", "maturin")
maturin.override(
name = "jiter",
cargo_lock = "//:jiter.lock",
repo = "pypi",
)
userepo(maturin, "pypicargo")
Using a Custom Build Target
For full control, provide your own build target:
uv.package(
name = "psycopg2",
build_target = "@//deps/psycopg2:wheel",
workspace = "pypi",
)
Conflict and Variant Resolution
When a project needs mutually exclusive dependency versions—for example, torch for CPU vs. CUDA—rules_pycross supports uv's conflict declarations.
Declaring Conflicts
In your pyproject.toml:
[project.optional-dependencies]
cpu = ["torch==2.6.0"]
cu124 = ["torch==2.7.0"]
[tool.uv] conflicts = [ [ { extra = "cpu" }, { extra = "cu124" }, ], ]
How Variants Work in Bazel
When rules_pycross processes a lock file with conflicts, it generates:
boolflagtargets under@<repo>//variants:— one per conflict member (e.g.,extracpu,extracu124).configsettingtargets —@<repo>//variants:isextracpu,@<repo>//variants:isextra_cu124.select()expressions on the package aliases — so@<repo>//:torchresolves to the correct version based on which flag is set.
Selecting a Variant
Set the variant flag on the command line:
# Build with CPU torch
bazel build //my:target --@pypi//variants:extracpu=True
Build with CUDA torch
bazel build //my:target --@pypi//variants:extracu124=True
Or embed the flag in a platform():
platform(
name = "linux_cuda",
constraint_values = [
"@platforms//os:linux",
"@platforms//cpu:x86_64",
],
flags = [
"--@pypi//variants:extracu124=True",
],
)
Then build with --platforms=//:linux_cuda.
Default Groups
If uv's default-groups is set, the corresponding variant is used as the select() default—building without flags resolves to that variant. Extras never have a default; building without an explicit flag produces a build error, preventing accidental misresolution.
Dependency Group Conflicts
Conflicts also work with [dependency-groups]:
[dependency-groups]
test-fast = ["pytest==7.0.0"]
test-slow = ["pytest==8.0.0"]
[tool.uv] conflicts = [ [ { group = "test-fast" }, { group = "test-slow" }, ], ]
The generated flags follow the pattern group<name> (e.g., --@pypi//variants:group_test-fast=True).
Platform Transitions
When a workspace member needs to be built under a specific platform configuration—for example, to pin a variant flag or target a particular architecture—you can declare a platform transition on the member import. This causes all proxy targets in the thin repo to apply a Bazel --platforms transition, ensuring the backing workspace targets are analyzed under the specified platform.
There are three ways to specify the transition:
1. Using flags — embed --flag=value settings into a generated platform:
uv.repo(
workspace = "shared",
name = "ml-pipeline",
flags = [
"--@pypi//variants:extracu124=True",
],
)
2. Using constraint_values — generate a platform with specific constraints:
uv.repo(
workspace = "shared",
name = "ml-pipeline",
constraint_values = [
"@platforms//os:linux",
"@platforms//cpu:x86_64",
],
)
3. Using platform — reference an existing platform target directly:
uv.repo(
workspace = "shared",
name = "ml-pipeline",
platform = "@//platforms:linux_cuda",
)
[!NOTE]
flagsandconstraint_valuescan be combined (they are merged into a single generated platform), butplatformis mutually exclusive with both.
These attributes are available on uv.repo() and its PDM/Poetry/Pylock equivalents.
When constraintvalues alone are specified, rulespycross generates an internal platform() target and uses pycrosstransitioninglibraryproxy / pycrosstransitioningfileproxy at each package level to apply the --platforms transition.
When flags are specified (with or without constraintvalues), rulespycross additionally generates a custom _transition.bzl in the thin repo. This is necessary because Bazel's platform(flags=[...]) mechanism only applies during top-level platform mapping — it does not take effect when --platforms is set via a Starlark transition. The generated transition directly sets both --platforms and the individual flag values. Root-level targets become transitioning proxies so that select() expressions in per-package BUILD files resolve in the transitioned configuration where the flags are set.
This is particularly useful for locking variant selections to a member without requiring --flag arguments on every bazel build invocation.
rules_python Compatibility
rulespycross integrates with rulespython. The generated target layout (@<repo>//<package>) is compatible with rules_python conventions.
- Venv support — when
rulespythonvenvs are enabled,pycrosswheellibrarypopulates the symlinks needed for a correctsite-packageslayout. Auto-detected paths can be overridden viauv.package(sitepaths = [...]), and additional path categories (binpaths,datapaths,include_paths) are also supported. pyconsolescriptbinary— eachpycrosswheellibraryproduces a:distinfooutput group for entry point discovery. Usepyconsolescript_binary(pkg = "@pypi//cython", script = "cython")directly.
Gazelle Integration
rulespycross is compatible with rulespythongazelleplugin. The target layout (@<repo>//<package>) matches the plugin's default label conventions, so no gazelle:pythonlabelconvention directives are needed.
The pycrossmodulesmapping rule generates modules_mapping.json from package metadata at build time — wheels do not need to be downloaded or extracted during analysis.
load("@gazelle//:def.bzl", "gazelle")
load("@rulespycross//pycross:defs.bzl", "pycrossmodules_mapping")
load("@rulespythongazelleplugin//manifest:defs.bzl", "gazellepython_manifest")
load("@pypi//:requirements.bzl", "all_requirements")
pycrossmodulesmapping( name = "modules_map", deps = all_requirements, )
gazellepythonmanifest( name = "gazellepythonmanifest", modulesmapping = ":modulesmap", piprepositoryname = "pypi", )
gazelle:python_extension enabled
gazelle:python_root //
gazelle(
name = "gazelle",
gazelle = "@rulespythongazelleplugin//python:gazellebinary",
)
> bazel run //:gazellepythonmanifest.update # Update gazelle_python.yaml
> bazel run //:gazelle # Apply BUILD file changes
See the API reference docs and e2e tests for more examples.