Adopt uv for Python environment management

[tongyx361]

Feature request

Adopt uv as the standard Python environment and dependency manager for verl. This includes resolving existing dependency conflicts and migrating from pip to uv across the project.

Motivation

uv offers faster installs and reproducible environments, but uv sync requires the full dependency tree to resolve cleanly. Currently it fails due to dependency conflicts such as:

• pyext==0.7 is incompatible with Python >= 3.12 (uses removed inspect.getargspec)
• sglang and vllm optional dependencies have conflicting torchaudio version requirements
• There may be other undiscovered conflicts

These conflicts also affect pip users but are silently ignored — uv surfaces them explicitly.

Your contribution

Attempted in #5846 but reverted — needs dedicated work to:

1. Audit and resolve all dependency conflicts in setup.py / pyproject.toml
2. Fix or replace incompatible dependencies (e.g., pyext)
3. Add PEP 735 [dependency-groups] (e.g., dev group for pre-commit, py-spy, hydra-core)
4. Update CI workflows and docs to use uv

[Kirrito-k423]

Thank you for this well-structured proposal.

Adopting uv is indeed a sensible direction for modern Python projects, offering significant performance improvements in dependency resolution and installation.

Technical considerations:

1. Dependency conflicts: As noted, the current uv sync failure suggests dependency tree issues. A phased migration approach would help:

   • First: audit and resolve existing conflicts (e.g., version pinning mismatches)
   • Then: add pyproject.toml with uv-compatible specifications
   • Finally: update CI/CD pipelines

2. Backward compatibility: During transition, maintaining both pip and uv paths ensures existing workflows continue working. Consider adding uv.lock alongside requirements.txt temporarily.

3. Developer experience: uv's faster installs benefit CI pipelines and contributor setup. The uv sync workflow also improves reproducibility across environments.

4. CI integration: GitHub Actions supports uv via astral-sh/setup-uv. This would reduce CI job setup time significantly.

This would be a valuable infrastructure improvement for the project. Would you be interested in preparing an initial migration PR for the core dependencies?

[w568w]

+1 for this, given that many AI infrastructure libraries have adopted uv as their only supported dependency management / packaging tool.

	Dependency Management	Building / Packaging Tools	Environment Management
Former Gen	easy_install	setuptools	/
Current Gen	pip, poetry, etc.	setuptools, etc.	conda
Next Gen	uv	uv_build, hatchling, setuptools	pixi, mamba

Let's prepare for the future…