[BUG] Unable to set max-local-prefill-length on VllmDecodeWorker ("unknown field")

[PoWeiShen]

Describe the Bug

I’m trying to run a disaggregated service with the vLLM backend.

According to the docs Disaggregation and Performance Tuning (section “Set the local prefill length”), max-local-prefill-length should be configurable on the vLLM decode worker so that short prompts stay local and longer prompts are sent to prefill engines.

However, when I add max-local-prefill-length under spec.services.VllmDecodeWorker in my DynamoGraphDeployment YAML, the CRD validation fails with:

Error from server (BadRequest): error when creating "disagg.yaml": DynamoGraphDeployment in version "v1alpha1" cannot be handled as a DynamoGraphDeployment: strict decoding error: unknown field "spec.services.VllmDecodeWorker.max-local-prefill-length"

So it looks like the CRD schema does not accept this field, even though the docs say it should be supported.

I’m not sure if:

• this is a bug in the CRD / operator,
• or if max-local-prefill-length is now supposed to be set in a different place (for example in a nested config / env var / JSON),
• or if the option was renamed in newer versions.

Steps to Reproduce

Add max-local-prefill-length: 1500 in disagg.yaml file

spec:
  services:
    VllmDecodeWorker:
      max-local-prefill-length: 1500

kubectl apply -f disagg.yaml

Expected Behavior

Based on the docs, I expected one of the following:

1. max-local-prefill-length to be a valid field under spec.services.VllmDecodeWorker in the DynamoGraphDeployment CRD, or

2. Clear documentation showing the correct place / key to configure max-local-prefill-length in the DGD spec (for example under some config/params section that is passed to dynamo.vllm).

Actual Behavior

The CRD validation rejects max-local-prefill-length as an unknown field on VllmDecodeWorker, so there is currently no obvious way (from the docs) to configure this option via the DGD YAML.

Environment

• Dynamo version: 0.6.1
• Backend: vLLM (nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.6.1)
• Deployment: Kubernetes, using DynamoGraphDeployment CRD
• Mode: disaggregated (VllmDecodeWorker + VllmPrefillWorker)

Additional Context

1. Is max-local-prefill-length still a supported option in the current Dynamo + vLLM integration?
2. If yes, what is the correct way to configure it for VllmDecodeWorker in a DynamoGraphDeployment?
3. If it has been renamed / moved, could the docs be updated to reflect the new parameter name and location?

Thanks!

Screenshots

No response