vLLM-Omni crashes on Hindi ref_audio with VoxCPM2 (/v1/audio/speech)

Describe the bug

When running vllm-omni with the VoxCPM2 model, the server crashes while processing requests to /v1/audio/speech if the ref_audio contains Hindi speech.

This results in an AssertionError inside the engine (StageEngineCoreProc), causing the entire service to terminate and requiring a manual restart.

The same workflow works correctly when the ref_audio contains English speech, suggesting a language-specific issue.

The same workflow also works correctly when tested using vLLM-Nano.

Root cause (from logs)

AssertionError: voxcpm2 prefill length mismatch: scaffold_len=182 tts_len=22;
caller must pad prompt_token_ids to the full prefill length

Version

vLLM Version: 0.19.0
vllm-omni version: main branch (post PR#2911)

Questions

1.Multilingual Support
Is VoxCPM2 expected to support non-English ref_audio (e.g., Hindi)?
2.Is this a known limitation of VoxCPM2 or an issue in vLLM-Omni?

Start by reproducing the crash through /v1/audio/speech with VoxCPM2 and Hindi ref_audio, then inspect the StageEngineCoreProc assertion and the reported prefill-length mismatch. Compare the same workflow with English ref_audio and vLLM-Nano. Done means determining whether non-English ref_audio is supported and preventing the service from terminating, with the expected behavior documented.