deps: bump deep_ep pin to V2 merge commit b306af0

[dmvevents]

Summary

Bumps deep_ep git pin in pyproject.toml from bfded348 (2025-10-29, pre-V2) to b306af0 (2026-04-29), the merge commit of deepseek-ai/DeepEP#605 "Introducing EPv2: faster EP, and Engram/PP/CP supports".

Why

The current pin predates the DeepEP V2 API reshape (Buffer -> ElasticBuffer, PP/CP/Engram support). Without this bump, consumers who follow the Megatron-side V2 adoption (NVIDIA/Megatron-LM#4632, "Shape Y") cannot import deep_ep.ElasticBuffer because the NeMo-RL virtualenv still installs the pre-V2 tree.

This PR is intentionally scoped to the pin only - it does not change any NeMo-RL code path. End-to-end V2 enablement requires pairing this bump with Megatron-LM#4632.

Scope

Single-file change to pyproject.toml:

• Three deep_ep @ git+.../DeepEP.git@bfded348... pins become ...@b306af06afd412c88e51e71802951606e40b7358.

Validation / reproduction

End-to-end reproduction (Dockerfile, Kubernetes manifests, 2-node p5en.48xlarge dispatch+combine smoke bench, vLLM MoE end-to-end chat completion) is public at:

https://github.com/antonai-work/nemo-rl-deepep-v2-efa

Verified inside the reproduction tree:

• deep_ep installs at commit b306af0 with ElasticBuffer import succeeding.
• DeepEP/tests/elastic/test_ep.py runs on 2x H200 EFA at p50 ~740us dispatch+combine.

Related PRs

• deepseek-ai/DeepEP#605: V2 merge (upstream source).
• NVIDIA/Megatron-LM#4632: Megatron-side V2 adoption.
• NVIDIA-NeMo/RL#2410: separate Dockerfile change to re-export LD_LIBRARY_PATH so AWS EFA's OFI plugin is discoverable. Same fleet, different concern - filed independently so each can be reviewed on its own merits.

Risk / compatibility

• The pre-V2 deep_ep.Buffer symbol is gone in V2. If a downstream NeMo-RL code path still imports it, that path breaks with or without Megatron#4632. Searches of NVIDIA-NeMo/RL show DeepEP is consumed through Megatron-Core, not directly, so NeMo-RL itself has no direct deep_ep.Buffer call sites.
• vLLM's V2 integration (vllm-project/vllm#41183) is still open; NeMo-RL installs vLLM independently, so this pin does not force the vLLM venv.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[dmvevents]

Closing as superseded. This PR bumped the deep_ep pin to the EPv2 public-release commit
(b306af0, 2026-04-29) as a single pin across all extras. Since then main has reorganized the
deep_ep dependency into a platform-split pin (separate x86_64 and aarch64 commits, last
updated 2026-06-20 in the SgLang-rollout refactor #2267).

Rebasing this PR as-is would regress that — it would drop the aarch64 handling and move the pin
backward relative to the current structure — so it's not the right thing to force-merge. The EPv2
intent here is better served by a fresh change against main's platform-split layout if/when an EPv2
bump is desired on both arches. Closing to keep the queue clean; happy to re-open a properly-rebased
version if the maintainers want EPv2 evaluated against the current pins.