The Torch upstream packages on PyPI and PyTorch's custom index have a dependency on a pinned Triton version. "pip install torch==2.7.1" installs "triton==3.3.1".

$ grep triton lib64/python3.12/site-packages/torch-2.7.1+cu128.dist-info/METADATA  
Requires-Dist: triton==3.3.1; platform_system == "Linux"

Our downstream builds lack this dependency. It took me a bit to figure out why we are lacking the dependency on Triton. The dependency is injected at build time through the magic env var "PYTORCH_EXTRA_INSTALL_REQUIREMENTS".

• https://github.com/pytorch/pytorch/blob/v2.7.1/setup.py#L999-L1006 
• https://github.com/pytorch/pytorch/blob/v2.7.1/.circleci/scripts/binary_populate_env.sh
  • I think "pytorch-triton-rocm" is just a build of regular Triton for ROCm.
• https://github.com/pytorch/pytorch/blob/v2.7.1/.ci/docker/triton_version.txt

We should implement the same logic in our build plugin for Torch, then remove the vLLM Triton patch and Triton version constraints.