[Bug]: cached_tokens = prompt_tokens in prefill-decode disagg with stream=true

[romain-primalabs]

Your current environment

Environment Information

vLLM Router version:
Rust version:
Python version:
Operating System:
Deployment environment:
vLLM backend version:

Component

Prefill/Decode Disaggregation

Router Configuration

vllm-router --policy round_robin \
  --vllm-pd-disaggregation \
  --prefill http://$P:8000 \
  --decode http://$D:8000 \
  --host 0.0.0.0 \
  --port 30000

🐛 Describe the bug

Hello,

I am using vllm-router for prefill-decode disagg where I am using vllm servers with multi connector (lmcache + nixl). It works, but the cached_tokens reported is always the one from decode (getting 100% from prefill server producer).

Additional context

No response