vllm - 💡(How to fix) Fix [Bug]: Prefix caching fails for incremental multimodal requests on Mamba-Attention hybrid models (Qwen3.5) [2 pull requests]

Your current environment

vLLM version: 0.19.0 (V1 engine) Model: Qwen3.5-4B (Qwen3_5ForConditionalGeneration, Mamba-Attention hybrid / GDN + Attention) GPU: NVIDIA RTX A6000 OS: Linux (CentOS 8)

🐛 Describe the bug

In a multi-turn agentic evaluation setting, each turn sends a prompt with one more image than the previous turn (incremental multimodal). Despite the shared prefix blocks having identical block hashes, num_cached_tokens is always 0.

This issue is specific to the Mamba-Attention hybrid architecture. The same test with a pure-Attention model (Qwen3-VL-4B) works correctly.

Details in document.

vllm_prefix_cache_issue.md

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.