The versions of triton are defined in the variant constraints files:

collections/accelerated/cpu-ubi9/constraints.txt:6:triton==3.2.0
collections/accelerated/cuda-ubi9/constraints.txt:7:triton==3.3.1
collections/accelerated/gaudi-ubi9/constraints.txt:21:triton==3.1.0
collections/accelerated/gaudi-ubi9/requirements.txt:25:triton==3.1.0
collections/accelerated/rocm-ubi9/constraints.txt:6:triton==3.3.1
collections/accelerated/spyre-ubi9/constraints.txt:7:triton==3.2.0

Christian has pointed out that we should be using version of triton that is defined in pytorch in, for example,

https://github.com/pytorch/pytorch/blob/v2.7.1/.ci/docker/triton_version.txt

instead of defining it in the collections. To do this, we have to define

PYTORCH_EXTRA_INSTALL_REQUIREMENTS=torch

for each variant that requires torch. This would result in us being able to remove the vllm '0003-Add-Triton-Dependency.patch' patch from the repo which ends up building triton for all variants even when the variant doesn't require triton (spyre and cpu).