vLLM 官方博客T1·Open Source
62 pts
Starting with v0.33.0, vLLM ships CUDA wheels built against the LibTorch Stable ABI, so upgrading LibTorch no longer requires rebuilding or relinking vLLM; users can swap the torch wheel and import it directly. The change reduces coupling between vLLM's C++ backend and PyTorch's binary interface, though compatibility still depends on Python APIs, minimum CUDA versions and third-party dependencies.