Pre-configured CUDA-enabled Docker Devcontainer for Transformers Development #1
Description
Feature request
A pre-configured Docker devcontainer with all necessary dependencies, CUDA support, and pre-commit hooks. This standardizes the development environment, simplifying onboarding for new contributors and ensuring consistency across projects.
Motivation
Contributing to Transformers can be slowed by complex local setup—CUDA dependencies, version mismatches, and pre-commit tooling often create a steep onboarding curve and inconsistent contributor environments.
A pre-configured Docker devcontainer would:
- Standardize the development environment across platforms.
- Eliminate common setup errors and “works on my machine” issues.
- Reduce onboarding time for new contributors.
- Ensure consistent pre-commit and linting enforcement.
This proposal builds on lessons learned from xformers-devcontainer, which has shown that packaging all dependencies (including CUDA) into a ready-to-use container can dramatically accelerate contributor productivity.
In addition, a reproducible devcontainer could enable AI agents to perform automated testing, prepare pull requests, or even contribute autonomously—something only possible with a deterministic, standardized environment.
The MVP will target Windows 11 Pro (WSL) + CUDA, with a design that can be extended to:
- Linux + CUDA
- Linux CPU-only
- macOS on Apple Silicon (CPU-only)
- WSL CPU-only
Your contribution
I am the creator of xformers-devcontainer, a pre-configured development container for Meta’s xformers library. That project gave me experience with building GPU-enabled, lint-enforced development environments that work across different contributor setups.
I would like to bring a similar approach to transformers, adapting the patterns and tooling from xformers-devcontainer to meet the specific needs of this repository. I’m happy to take ownership of the initial implementation and work with the community to ensure it’s easy to maintain and extend.
Potential alternatives
- Continue with contributors setting up their local environments manually, which works well for many but can present a learning curve for some.
- Offer a partial Dockerfile without a devcontainer configuration, which would simplify dependency management but not fully standardize the development experience.