[Bug]: Gemma model is giving empty responses with new version of docker image vllm-openai:v.8.5

[FernandoDorado]

Current environment

Kubernetes Cluster on Azure with A100 GPUs

Bug

Hello team,

After upgrading the Docker image from vllm-openai:v0.8.4 to v0.8.5, I observed one issue when running the google/gemma-3-27b-it model (Hugging Face Model Link).

The model successfully returns metadata (e.g., finish reason, token usage), but the content field in the response is consistently an empty string. No changes were made to the Kubernetes deployment manifest apart from the image version bump.

When reverting to v0.8.4, the model responds correctly with expected text completions, confirming that the issue is specific to the new image version.

Steps to Reproduce:

1. Deploy vllm-openai:v0.8.5 with the gemma-3-27b-it model.

2. Send a chat completion request.

3. Observe that the content field is empty in the response.

[DarkLight1337]

It should be fixed by #17629 , can you use the latest code?

[FernandoDorado]

Hello @DarkLight1337, thanks for reporting the solution. It's working but the latest version of the docker image is still giving the error. I recommend to update the tag to avoid further issues from other users.