• Sources: primary, discussion
  • Summary: Astral published prebuilt GPU wheel indexes for the PyTorch ecosystem, listing 23 packages across indexes from CUDA 11.8 to 13.2. Coverage varies by index, from 1 package at CUDA 11.8 to 21 at CUDA 12.8, with flash-attn, vLLM, DeepSpeed, and Transformer Engine among the packages listed.
  • Why it matters: Setup and CI can skip local CUDA compiles where an index carries the package, and still build from source where it does not.