vLLM for AMD enables native Windows multi-GPU

The transport bridge is the real breakthrough, turning native Windows AMD inference from a single-GPU workaround into a modular multi-GPU experiment.

• –Externalizing Direct-RCCL and D3D12 AllReduce keeps the vLLM fork easier to update.
• –Hybrid routing can select communication paths based on collective type and payload size.
• –The reference setup uses two Radeon AI PRO R9700 GPUs, so broader hardware validation remains outstanding.
• –Upstream vLLM still officially requires Linux and does not natively support Windows.
• –Developers should treat this as a compile-and-benchmark project, not production-ready infrastructure.