fix(eplb): preserve static maps and reject offline pipeline parallelism

[llseek]

Loading an offline expert map still allowed periodic EPLB to replace it. Recording with pipeline parallelism also let separate EP rank-zero workers overwrite the same model entry and temporary file.

This change suppresses scheduled rearrangement when expert_map_path is set, while retaining step statistics and explicit elastic-EP remapping. It rejects offline recording/loading with pipeline_parallel_size > 1 before model registration or file writes. Online EPLB is unchanged. Regression coverage checks static, recording, online, and pipeline-parallel paths.

Targets integrate/pr61-into-hy4-pcp-ep to keep the diff limited to these two fixes (22 production lines changed, plus tests).

Validation:

• .venv/bin/python -m pytest tests/runtime_patch/test_offline_eplb.py -q: 16 passed.
• An isolated CPU check executed the local vLLM v0.25.1 EplbState.step implementation at a rearrangement interval boundary: online mode rearranged once; static mode did not rearrange.
• VLLM_V0251_SOURCE_ROOT=/data/users/xiaojie.yuan/vllm .venv/bin/python -m pytest tests/runtime_patch/test_offline_eplb.py tests/patch/test_worker_dispatcher.py -q: 23 passed, 10 failed during vLLM imports because NumPy was missing from the CPU test environment.
• git diff --check: passed.
• Ruff check and format check report existing issues in these files. Comparing Ruff diagnostics against the unmodified base found no new lint diagnostics. No HCU/multi-rank runtime validation was performed.

Duplicate-work check: inspected #61 and the open PR searches for 61 in:body, eplb, static, and offline, including #34 and #62. Those PRs provide the underlying HY V4/EPLB integration; none addresses these static-map lifecycle fixes. This PR adds only the fixes on top of #62's branch.

AI assistance: Codex assisted with implementation, review, testing, and PR preparation. Human review and relevant hardware validation remain necessary before merge.

[Copilot]

🔵 Needs a closer look

It changes distributed/runtime EPLB behavior and enforces a new parallelism constraint, which warrants multi-rank/hardware validation beyond the included unit tests.

Pull request overview

This PR adjusts the offline/static EPLB patching layer to (1) prevent periodic EPLB rearrangement from overwriting a loaded static expert map, and (2) reject offline EPLB recording/loading when pipeline parallelism is enabled to avoid multi-rank clobbering of shared model/file state.

Changes:

• Suppress EplbState.rearrange() when expert_map_path is set (static/offline load), while still allowing explicit elastic-EP remapping via rank_mapping.
• Reject offline recording/loading early when pipeline_parallel_size > 1, before any model registration or file writes occur.
• Extend regression tests to cover static vs. online rearrangement behavior and the pipeline-parallel rejection path.

File summaries

File	Description
vllm_hcu/patch/worker/framework_opt/patch_offline_eplb.py	Adds rearrange wrapping to preserve loaded static maps; adds early PP rejection for offline paths.
tests/runtime_patch/test_offline_eplb.py	Adds coverage for static-map non-rearrangement, continued online rearrangement, and PP>1 offline rejection.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.