[jthomson04]

Fixes DYN-3545.

Summary

• Configure KVBM prefill and decode workers with --disable-hybrid-kv-cache-manager so both sides use the same vLLM KV-cache layout.
• Mark decode workers explicitly with --disaggregation-mode decode and apply the matching decode-side configuration to the LMCache and FlexKV PdConnector launchers.
• Give the disaggregated KVBM integration-test prefiller a distinct NIXL side-channel port.
• Apply the KVBM worker configuration consistently across local launchers, current DynamoGraphDeployment examples, and the existing disaggregated integration test.

The unpatched 1.3.0-rc.9 launcher reproduced a NIXL compatibility-hash mismatch followed by handshake and KV-load failures. PdConnector disables the vLLM hybrid KV-cache manager when a child connector does not support it, while the paired plain NixlConnector decoder previously kept it enabled.

Validation

• Ran the final patched examples/backends/vllm/launch/disagg_kvbm.sh with nvcr.io/nvstaging/ai-dynamo/vllm-runtime:1.3.0-rc.9, Qwen/Qwen3-0.6B, and a 1-prefill/1-decode two-GPU topology.
  • /v1/models listed Qwen/Qwen3-0.6B.
  • /v1/chat/completions returned HTTP 200 through distinct prefill and decode workers.
  • NIXL reported that the compatibility check passed, and transfer metrics recorded one successful KV transfer.
  • Complete logs contained no compatibility-hash mismatch, handshake failure, KV-load failure, or frontend finish-reason deserialization error.
  • The existing VRAM_SEG registration warnings remained but did not prevent successful inference; transport registration changes are outside this PR.
• Ran the patched LMCache launcher with the same image, model, and two-GPU topology; model readiness and chat completion passed with no compatibility-hash, handshake, or KV-load errors.
• Verified that the FlexKV launcher accepts the new worker arguments and that its decode worker initializes NIXL with the HND layout. Full FlexKV inference could not run because the 1.3.0-rc.9 image does not include the optional flexkv Python package.
• bash -n passed for all five modified PdConnector launchers.
• Parsed all three modified deployment manifests with yaml.safe_load and compiled the modified Python test source.
• Collected tests/kvbm_integration/test_determinism_disagg.py::TestDeterminismDisagg::test_determinism_disagg_with_cache_reset successfully in the vLLM runtime image.
• git diff --check

[devin-ai-integration]

Devin Review found 2 potential issues.

[coderabbitai]

Walkthrough

The changes update vLLM decode worker deployments, launch scripts, and the determinism integration test. Decode workers now disable the hybrid KV cache manager. The relevant commands also specify disaggregated decode mode.

Changes

Disaggregated decode configuration

Layer / File(s)	Summary
Decode worker runtime arguments examples/backends/vllm/deploy/disagg_kvbm*.yaml, examples/backends/vllm/launch/disagg_kvbm*.sh	Deployment manifests and launch scripts pass --disable-hybrid-kv-cache-manager. The primary launch script also passes --disaggregation-mode decode.
Determinism integration command tests/kvbm_integration/test_determinism_disagg.py	The decoder command passes disaggregated decode mode and disables the hybrid KV cache manager.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main change: aligning the KVBM decode KV-cache layout.
Description check	✅ Passed	The description explains the change, affected areas, related issue, validation, and known limitation in sufficient detail.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@examples/backends/vllm/launch/disagg_kvbm.sh`:
- Line 21: Ensure every prefill worker includes
--disable-hybrid-kv-cache-manager, matching the existing decode-worker
workaround. Update the prefill commands in
examples/backends/vllm/launch/disagg_kvbm.sh:21, disagg_kvbm_2p2d.sh:21-23, and
both prefill commands in disagg_kvbm_router.sh:30 and :38; update the prefill
components in examples/backends/vllm/deploy/disagg_kvbm.yaml:27,
disagg_kvbm_2p2d.yaml:27, and disagg_kvbm_tp2.yaml:27; and update
self.prefiller_cmd in tests/kvbm_integration/test_determinism_disagg.py:152-154.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 8a587876-ebc4-4eeb-9418-fa84ddb89926

📥 Commits

Reviewing files that changed from the base of the PR and between cd6b63b and fde8962.

📒 Files selected for processing (7)

• examples/backends/vllm/deploy/disagg_kvbm.yaml
• examples/backends/vllm/deploy/disagg_kvbm_2p2d.yaml
• examples/backends/vllm/deploy/disagg_kvbm_tp2.yaml
• examples/backends/vllm/launch/disagg_kvbm.sh
• examples/backends/vllm/launch/disagg_kvbm_2p2d.sh
• examples/backends/vllm/launch/disagg_kvbm_router.sh
• tests/kvbm_integration/test_determinism_disagg.py