feat(vllm): default DP launch to per-node hybrid LB

[alec-flowers]

Summary

• Make one hybrid-LB vLLM parent process per node the default for DP endpoints.
• Derive --data-parallel-size-local, --data-parallel-start-rank, DP address, and DP RPC port from the allocated topology.
• Add backend.legacy_dp_per_gpu: true as a compatibility escape hatch for the previous one-process-per-DP-rank layout.
• Derive Dynamo health expectations from the generated process topology.
• Reserve non-overlapping KV-event, NIXL, DP-RPC, HTTP, and system-port ranges across the generated layout.

This builds on and supersedes #90, but makes the per-node layout opinionated rather than adding per-role launch-mode tuning.

Registration and routing model

The Dynamo routing boundary is a registered generate engine, not each internal vLLM DP rank.

• Default per-node hybrid: one Dynamo generate registration per node/parent process. vLLM owns the node-local DP ranks below that registration and performs hybrid load balancing within the parent.
• Legacy per-GPU: one independent Dynamo generate registration per vLLM DP process. vLLM is still launched with its global --data-parallel-rank, but Dynamo routes to each process as an independent engine.

For DEP8 on two 4-GPU nodes, the default therefore creates two Dynamo engines. The two vLLM parents own global DP ranks 0-3 and 4-7 respectively. The Dynamo router does not see eight independently routable engines in this layout.

Configuration

Per-node hybrid is automatic when a vLLM role sets data-parallel-size:

backend:
  type: vllm
  vllm_config:
    aggregated:
      data-parallel-size: 8

The previous layout remains available as a deployment-wide compatibility switch:

backend:
  type: vllm
  legacy_dp_per_gpu: true
  vllm_config:
    decode:
      data-parallel-size: 4

Validation

Local

• 131 focused topology, config, and health tests passed.
• Ruff check passed.
• Ruff format check passed.

Real GB200 smoke tests

All tests used local DeepSeek-V4-Flash weights (not the Hugging Face cache), vLLM 0.25.0, and matching aarch64 nightly wheels:

• ai-dynamo==1.3.0.dev20260718
• ai-dynamo-runtime==1.3.0.dev20260718

Job	Layout	Observed topology and registration	Result
2424943	Aggregated DEP8, 2 nodes x 4 GB200	Two parent processes with local size 4 and start ranks 0/4; two Dynamo decode registrations; vLLM logs global DP ranks 0-7	COMPLETED 0:0; server healthy; smoke requests passed
2424942	PD, DEP4 prefill + DEP4 decode	One parent and one Dynamo registration per role; four internal DP/NIXL identities per parent (_dp0-_dp3); NIXL compatibility passed on all decode ranks	COMPLETED 0:0; server healthy at 1P + 1D; smoke requests passed
2424944	PD with legacy_dp_per_gpu: true	Four independent processes and Dynamo registrations per role; vLLM global DP ranks 0-3	COMPLETED 0:0; server healthy at 4P + 4D; smoke requests passed

The DEP8 logs also confirm NCCL_SYMM_MEM was selected for the DP and EP groups with the per-node parent layout.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 93.33333% 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	92.85%	4 Missing ⚠️
src/srtctl/core/topology.py	87.50%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #272   +/-   ##
=======================================
  Coverage        ?   67.40%           
=======================================
  Files           ?       69           
  Lines           ?     9108           
  Branches        ?        0           
=======================================
  Hits            ?     6139           
  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.