vllm + p2p driver hack + qwen 3.8 27B vs llamacpp + qwen flash next ?

Hi everyone

I'm running four rtx 4090, 64GB ram, on a threadripper pro motherboard so all PCIe x16 ports, as a homelab machine for coding.

I was migrating from vllm + qwen 3.8 27B (fp8+256k kv cache) to llamacpp + qwen flash next iq4xs + 8 bit cache 200k kv cache... until someone had to ruin my migration with this awesome patch to enable p2p between cards on the open source nvidia drivers. On my hardware configuration apparently this would matter a lot. But only VLLM would benefit from it, not llamacpp - and no 4 bit like quant available on vllm and not enough vram for qwen flash next fp8.

I hereby ask the grand experts, what should I do amd what are your opinions?

1. vllm + qwen 3.8 27B dense + p2p driver patch + 256k kv cache

2. llamacpp + qwen flash next MoE iq4xs + 200k kv cache

Is qwen flash next really worth the extra intelligence versus higher throughput with vllm and sticking to qwen 3.8 27B ?