feat: [AMD Support via SemiAnalysis][2/4] vLLM MoRI-IO integration

[cquil11]

Summary

Integrate native vLLM Router with MoRI-IO prefill/decode workers. Static router registration and NIXL port handling do not describe MoRI's ZMQ discovery path; this PR supplies the connector-specific launch and readiness behavior.

Stacked on #1, which combines AMD hardware enablement and required cluster runtime. #7 follows this PR; ATOM uses Mooncake and has no engine dependency on MoRI.

Review map

• backends/vllm.py: generate role-aware MoRIIOConnector JSON from allocated addresses and ports; preserve upstream ephemeral port allocation for TP children instead of injecting a shared VLLM_PORT scan base. Other connectors retain their existing behavior.
• frontends/vllm_router.py, frontends/static_router.py, ports.py: select ZMQ discovery, own its registration port, and omit static P/D URL registration for the managed MoRI path.
• core/health.py, cli/mixins/benchmark_stage.py: use a one-token completion probe for dynamic discovery, whose registry is not represented by Router's static /workers endpoint.
• core/schema.py: restrict the managed integration to a P/D layout with one router on the head node.
• Tests and documentation move with the integration. No engine patches, image changes, workload changes, or new GPU tests are introduced by this split.

---

Note

Medium Risk
Changes P/D launch, router CLI, and readiness gating for moriio deployments; misconfiguration is caught at schema validation, but end-to-end probe behavior affects when benchmarks start.

Overview
Adds managed MoRI-IO for ROCm-style vLLM prefill/decode behind frontend.type: vllm-router when engine.connector: moriio.

Worker launch: srtctl builds role-aware MoRIIOConnector JSON (kv_producer / kv_consumer) from the head node IP, fixed discovery port 36367, and each worker’s HTTP port. MoRI workers skip NIXL side-channel env and the per-process VLLM_PORT scan base so TP children don’t collide on ephemeral listeners.

Router: ZMQ discovery mode passes --kv-connector moriio and --vllm-discovery-address and omits static --prefill / --decode URLs. NIXL bootstrap ports are not advertised for MoRI.

Readiness: Because Router’s /workers doesn’t reflect the MoRI registry, benchmarks wait on a one-token POST /v1/completions probe (requires model_name) instead of worker-count polling.

Config validation: MoRI + vllm-router requires P/D topology, enable_multiple_frontends: false, and orchestrator_placement: head. Docs and schema list moriio as a connector option; tests cover KV config, health probe, port env, and device binding for vllm-router.

^{Reviewed by Cursor Bugbot for commit 9935d90. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 9935d90. Configure here.}

[cursor]

MoRI workers share default control ports

High Severity

Managed MoRIIOConnector JSON omits per-worker handshake_port and notify_port, so every engine falls back to the same defaults. Co-located workers, including multiple endpoints on one node, then collide on those listeners and on engine_id.

 

^{Reviewed by Cursor Bugbot for commit 9935d90. Configure here.}