As part of AIPCC-8444, a new builder collection was created for torch v2.9.1. This was necessary because the CPU build for vLLM v0.13 has torch v2.9.1 as a requirement. Some additional effort is required to build vLLLM for aarch64 as AIPCC-8444 builds only for x86_64.

Changes are most likely needed to vLLM's cpu_extensions.cmake:

• It customizes a path for libgomp which ships with upstream torch. It should use libgomp already installed with gcc.
  • https://gitlab.com/redhat/rhel-ai/core/mirrors/github/neuralmagic/nm-vllm-ent/-/blob/main/cmake/cpu_extension.cmake?ref_type=heads#L223
• It will either fetch oneDNN from GitHub or use a source directory if variable FETCHCONTENT_SOURCE_DIR_ONEDNN is set. A mirror for oneDNN is needed on GitLab
  • https://github.com/vllm-project/vllm/blob/28ae32a5d3717d4d5fc73c2338cca909775910c0/cmake/cpu_extension.cmake#L291-L306