fix(vllm): OPS-8696 let the EPD encode worker report healthy

[dmitry-tokarev-nv]

Linear: OPS-8696

Fixes #15432.

Summary

The vLLM E/P/D launcher's encode worker never reports healthy on its system port. Three
post-merge serve tests poll every worker's /health, so they wait out their timeout:
mm_epd_qwen3-vl-2b, mm_epd_qwen3-vl-2b_b64_frontend_decoding and
mm_epd_video_qwen3-vl-2b_mixed_frontend_decoding.

The encode worker serves generate without a health-check payload, so the runtime
registers no canary target for it. With canaries off, which is the default,
SystemHealth::get_health_status then falls back to the process-wide status. That status
starts NotReady, and the vLLM worker never sets it, so /health answers 503 forever. The
runtime test ready_endpoint_without_a_registered_target_still_reports_unhealthy
documents this fallback.

Deployments through the operator do not hit it. WorkerDefaults.GetBaseContainer in
deploy/operator/internal/dynamo/component_worker.go sets
DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS=["generate"] for the worker, prefill and decode
component types. Encode workers are declared as componentType: worker, as in
recipes/qwen3-vl-32b-fp8/vllm/hetero_hardware_disagg/deploy.yaml. Readiness then follows
the generate endpoint, which turns ready on registration. This PR sets the same variable
for the encode worker in examples/backends/vllm/launch/disagg_multimodal_epd.sh.

The tests started polling the encode worker in #14751, which added
health_check_worker_count=3 to the epd and epd_video topologies. Main at
5e21f9c618 (2026-09-16), before #14751, passed the video test.

Validation

All runs used the CI vllm-runtime-test image for main 3f46e3d3c5 on one RTX 6000 Ada.
The patched runs mount only the changed launcher into that image.

Test	Launcher	Result
mm_epd_qwen3-vl-2b	unpatched	failed after 309 s, encode worker /health 503 throughout
mm_epd_video_qwen3-vl-2b_mixed_frontend_decoding	unpatched	failed at its 600 s timeout, encode worker /health 503 throughout
all three affected tests	patched	3 passed, 184 s together

mm_epd_qwen3-vl-2b_b64_frontend_decoding uses the same epd topology as
mm_epd_qwen3-vl-2b. It was not run unpatched. The llava E/P/D profile does not poll
worker health, so it does not hit this. disagg_multimodal_e_pd.sh also starts an encode
worker, but the e_pd topology does not poll worker health either. Every other topology
that polls worker health polls only workers that register a health-check payload.

PR CI does not exercise this change: all three affected tests are marked post_merge. A
green PR run is therefore no evidence for the fix. The GPU runs above are.

Not changed: xpu/disagg_multimodal_epd_xpu.sh starts its encode worker the same way, but
no XPU test polls its health, and no XPU hardware was available to run it.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • The encode worker now uses the generate endpoint’s health status when determining readiness.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: ai-dynamo/dynamo/.coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 731ee80b-a583-4726-93e5-1c69da25158f

📥 Commits

Reviewing files that changed from the base of the PR and between 8ad295a and e1ddf2d.

📒 Files selected for processing (1)

• examples/backends/vllm/launch/disagg_multimodal_epd.sh

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

Walkthrough

The encode-worker launch sets DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS to ["generate"] so readiness uses the generate endpoint health status.

Changes

Encode worker readiness

Layer / File(s)	Summary
Configure endpoint health status examples/backends/vllm/launch/disagg_multimodal_epd.sh	The encode-worker launch sets DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS to ["generate"].

Priority: ⬆️ High

Estimated code review effort: 2 (Simple) | ~5 minutes

Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to e1ddf

The encode worker’s /health should report healthy after its generate endpoint registers, addressing the reported persistent 503. No concrete merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 7

✅ Passed checks (7 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The description links the pull request to OPS-8696 and issue #15432. The linked issue matches the reported encode-worker health failure and the proposed fix.
Out of Scope Changes check	✅ Passed	The change is limited to the vLLM EPD launcher and directly supports the stated objective. No unrelated changes are reported.
Linked Issues check	✅ Passed	The directly linked issue [#15432] requires the vLLM E/P/D encode worker to report healthy so the video serve test can complete. The PR sets DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS='["generate"]' for …
Out of Scope Changes check	✅ Passed	The PR changes one vLLM E/P/D launcher and adds comments that explain the health-status fallback. These changes directly support [#15432]. No unrelated code or unrelated behavior is included. The unch…
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1…
Title check	✅ Passed	The title clearly identifies the vLLM fix and its purpose: allowing the EPD encode worker to report healthy. It is concise and specific.
Description check	✅ Passed	The description explains the problem, implementation, affected tests, validation results, and linked issue. It omits the template headings and does not explicitly identify where reviewers should start…

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

_{Comment @coderabbitai help to get the list of available commands.}

[JulienDarve]

:ship-it-parrot: