NVIDIA-NeMo / NVIDIA-NeMo/RL

Unify duplicated deep_ep dependency commits in pyproject.toml

Open
#2,522 2 comments 0 reactions 1 assignee Claimed by @seonjinn View on GitHub
Documentation
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

## Context

PR #1942 added platform-conditional `deep_ep` pins (x86_64 vs aarch64) to support HybridEP. The same two commit hashes are now copy-pasted across **4 separate locations** in `pyproject.toml`:

1. `[project.optional-dependencies] automodel`
2. `[project.optional-dependencies] vllm`
3. `[project.optional-dependencies] mcore`
4. `[tool.uv] override-dependencies`

Each location carries the identical pair:
```
deep_ep @ git+...@bfded348... ; platform_machine == 'x86_64'
deep_ep @ git+...@a4849360... ; platform_machine == 'aarch64'
```

## Problem

- Updating the deep_ep commit requires changing 4 places (8 lines) — easy to miss one and create a silent divergence.
- As more platform/arch conditions are added, the duplication grows.

## Proposed cleanup

Consolidate the `deep_ep` pin so the commit hash(es) are defined once and referenced by all dependency groups. Options include:
- A dedicated `deep_ep` optional-dependency group that other groups depend on
- A `uv` workspace/override that covers all groups from a single source of truth
- Any other approach that eliminates the duplication

## Origin

Tracked from review of #1942.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.