vLLM in CUDA does not use it, it uses vllm's fork of flash-attn:

https://github.com/vllm-project/vllm/blob/ae87ddd040b793fd9f4f05cb660a4728c81d7670/cmake/external_projects/vllm_flash_attn.cmake#L13-L26

We don't need to build it anymore for RHAIIS, so we need to update collections/rhaiis/cuda-ubi9/requirements.txt