For the partitioning part, the docs say NVIDIA vGPU isn't supported on XCP-ng, since it depends on a proprietary XenServer component that can't be redistributed: https://docs.xcp-ng.org/compute#vgpu.
So as far as I can tell, the "20, 40 or 80 GB per VM" split won't work through vGPU. Passing a whole GPU to one VM is the documented path, on the same page.
I haven't seen anyone report on H100 or H200 cards here, though, so I can't tell you if that class of card brings its own surprises.
There's also a blog post doing LLM inference that way on a much smaller Tesla T4: https://xcp-ng.org/blog/2025/07/28/your-own-gpu-powered-llms-with-xcp-ng/.
There's an open design discussion upstream in XAPI about GPU partitioning, MIG included (https://github.com/xapi-project/xen-api/pull/7260), but that's the XenServer side, and I don't know if or when it would reach XCP-ng.
As GPU partitioning matters for your project, it might be worth opening a request on https://feedback.vates.tech/ with what you described here (H200s, LLM inference, several tenants per card, fixed VRAM per VM). From what I've seen, a concrete use case like yours carries more weight there than a plain "please support vGPU", and other people needing the same thing can vote on it.
