vLLM tuning tames Qwen 3.5 lag

Community tuning posts usually are not major news, but this one is useful because it turns fresh vLLM engine features into a practical fix for a real serving pain point.

• –The key knobs are `--performance-mode interactivity`, `--mamba-cache-mode align`, and `--mamba-block-size 8`, which the poster says stop the model from effectively reprocessing the full prompt every turn
• –This matters most for long-context chat, coding agents, and tool-use loops, where latency compounds fast and makes otherwise capable models feel broken
• –vLLM 0.17.0’s release notes explicitly mention Qwen3.5 support, Mamba cache align mode, and chunk alignment for prefix caching, so this is grounded in recent engine work rather than random folklore
• –It is still a field report, not a benchmark-backed release claim, so developers should treat it as a high-signal tuning recipe and validate against their own hardware and workloads