vLLM hits Qwen3.6 MTP pipeline wall

The sharp edge here is not the model alone, but the interaction between vLLM’s speculative decoding path and PP support for this serving stack.

• –This looks like a genuine compatibility gap, not a misconfigured container, because the failure matches vLLM’s own PP support checks and the linked upstream bug.
• –The user-facing pain is real: PP is often needed to fit a 27B model across GPUs, while MTP is the lever that would make generation faster.
• –If you need the model to run today, the choice is basically “fit it with PP” or “speed it up with MTP,” but not both together in this configuration.
• –The post is most useful to operators evaluating vLLM for multi-GPU Qwen deployments, especially when comparing it with llama.cpp setups that may already fit their workflow.