PyTorch 2.11 resolves a long-standing packaging issue on aarch64 Linux: CUDA-enabled wheels are now published to the default PyPI index, so `pip install torch` on NVIDIA GH200, GB200, and GB300 systems no longer silently installs a CPU-only build. The post traces the problem from a 2024 hackathon frustration through vLLM's interim workarounds (use_existing_torch.py and uv build isolation config), to a formal fix coordinated via the PyTorch Foundation's Technical Advisory Committee. With PyTorch 2.11, standard install instructions now work out of the box on Grace Blackwell hardware, eliminating the need for custom index URLs and post-install sanity checks for most users.
Table of contents
An issue I first hit at a hackathonThe workarounds vLLM carried in the meantimeFrom a hackathon headache to a TAC agenda itemThe fix has landedWhy this is worth writing aboutSort: