bzlmod: Incompatible with pure MODULE.bazel setup due to "generate-then-load" pattern

[hugomano]

The Goal

I want to configure rules_ml_toolchain in a pure bzlmod project so I started to write an overlay to declare it as publishable module.

The Problem: "Generate-then-Load" Pattern

The current setup process is incompatible with a pure bzlmod workflow due to its reliance on a "generate-then-load" pattern. The setup requires a sequence of operations that cannot be expressed declaratively in MODULE.bazel.

For example, the CUDA setup follows these steps:

1. Generate: The cuda_json_init_repository rule is called. It runs during the Analysis Phase and generates a new repository (@cuda_redist_json) containing a distributions.bzl file.
2. Load: The script must then execute load("@cuda_redist_json//:distributions.bzl", "CUDA_REDISTRIBUTIONS") to get a variable from the file that was just created.
3. Consume: This variable is then passed to another repository rule, cuda_redist_init_repositories.

This creates a fundamental conflict with Bazel's architecture:

• All load() statements are processed during the Loading Phase.
• Repository rules are executed later, during the Analysis Phase.

A pure MODULE.bazel setup cannot load() a file that will only be created in a future phase. This forces users to adopt the WORKSPACE.bzlmod file, which prevents a fully modern and clean Bazel setup.

Thank you for considering this enhancement.

[rickeylev]

cc @rickeylev

I'll be wanting this via bzlmod, too, so big +1.

writing an overlay

Lets avoid this :). Lets just add bzlmod support directly here. It'll be necessary in a year or so anyways, when Bazel drops workspace support.

Maybe start simple? First make bazel_dep(name="rules_ml_toolchain") valid.

For that all we need is:

1. A basic MODULE.bazel file here.
2. Create the PR in the BCR repo

I don't think we even need (1) to be actually functional. Just working enough so that Bazel doesn't immediately throw an error and single_version_override() can be used in the interim until it's working in a more proper manner.

It sounds like you are familiar with BCR? Would you be able to do those two items to kick start the process?

[hugomano]

Hey @rickeylev, sorry for the late reply — and thanks a lot for your guidance, I really appreciate it.
I agree with you: it's going to be necessary anyway, and it's definitely better if we upgrade the Bazel workspace/module here.
I'll check with my team (@zml) what bandwidth we have for contributing upstream, since we rely heavily on OSS projects like XLA and PJRT GPU runtimes — and we do vendor some of them, like here: https://github.com/zml/zml/tree/master/third_party/modules