feat(vllm): add per-node DP launch mode

[esmeetu]

Summary

• Add opt-in backend.dp_launch_mode: per_node for vLLM data-parallel endpoints.
• Launch one process per physical node with the full local GPU set while preserving the existing per_gpu default.
• Derive --data-parallel-size-local, --data-parallel-start-rank, --data-parallel-address, and DP-RPC settings from the allocated topology.
• Reserve the per-local-rank KV-event port range and validate that configured global DP size matches allocated GPUs.

Per-role mixed launch overrides were split into the separate stacked follow-up #271.

Motivation

The existing per-GPU launcher restricts every process to one GPU. DeepGEMM MegaMoE and CUDA symmetric-memory paths can require node-local DP ranks to share one process and CUDA namespace. The opt-in per-node layout provides that namespace without changing existing recipes.

backend:
  type: vllm
  dp_launch_mode: per_node
  vllm_config:
    prefill:
      data-parallel-size: 8
      data-parallel-hybrid-lb: true
    decode:
      data-parallel-size: 16
      data-parallel-hybrid-lb: true

Validation

• make check
• Ruff and formatting pass.
• Full test suite: 865 passed, 2 skipped, 6 deselected.
• DeepSeek V4 Pro AgentX job 9700 validated uniform per-node launch for 3xPDEP8 + 1xDDEP16 with decode AMXF4 MegaMoE: 26.85 req/s, 91,193 TPGS on 40 GPUs, p50 TTFT 8.24s, and p50 TPOT 58.73ms.

[alec-flowers]

Is there a reason to mix the launch modes? Or is 1 way better than the other? If we have a reason to choose one way then I would prefer to be opinionated.

[esmeetu]

Split as suggested: #90 now contains only the opt-in global per_node launch mode. The per-role mixed-DP overrides moved to the separate stacked draft #271. per_node is useful when node-local ranks need a shared CUDA namespace (for example MegaMoE/symmetric-memory paths); this does not assert a universal performance advantage, and the existing per_gpu behavior remains the default.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 90.19608% with 5 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@c1b6b5c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/srtctl/backends/vllm.py	90.69%	4 Missing ⚠️
src/srtctl/core/topology.py	87.50%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #90   +/-   ##
=======================================
  Coverage        ?   67.34%           
=======================================
  Files           ?       69           
  Lines           ?     9091           
  Branches        ?        0           
=======================================
  Hits            ?     6122           
  Misses          ?     2969           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.