Hardware-Agnostic Models in vLLM

TL;DR To achieve state-of-the-art performance at the frontier, vLLM is changing its internal implementation in ways that make it incompatible with fullgraph torch.compile. This may have consequences for users who…