*Multiple identical GPU's.

For context, vLLM is mostly used for small to midscale commercial setups. It's heavily geared towards the GPU rich and squeezing performance out of large inference clusters. Nearly always deployed on a production Kubernetes GPU cluster.

Llama.cpp/Ollama/lmstudio are geared towards small scale hobbyist use on consumer hardware. It's way better at running on CPU/RAM or mixed card situations. Those aren't very common in production clusters, because production clusters generally use GPU-rich uniform hardware.

Generally speaking, vLLM is designed to shine at larger scales. You can run itsy bitsy single user vLLM servers, and if you know how to set it up correctly you might be able to squeeze a little more performance out if you have Blackwell cards. But it is definitely going to be significantly harder to set up, and may not fare well long term because single-user setups are really not a big priority for vLLM. Same with funny quants like Q6_K_M, mismatched cards, CPU offload, etc. They're more focused on things like autoscaling, KV prefix routing, dynamic batching, RDMA networking, and all those sorts of things that matter a lot for large industrial-scale deployments, but don't matter at all for personal use.

I would honestly recommend sticking with the consumer stuff unless you have professional experience working with production GPU Kubernetes clusters. I actually do, and I mostly use Llama.cpp myself. I do use vLLM for some parts of my home automation setup, where the management benefits of Kubernetes outweigh the pain of setting up DCGM and nvidia container toolkit and all that.