vLLM issue #13281: '[Installation]: Compiling and installing from source code is too slow, taking at least 3 hours.'

Filed 2025-02-14, CLOSED. The user's own measurement: a Docker build `docker build --pull --build-arg torch_cuda_arch_list=9.0+PTX --build-arg max_jobs=32 --build-arg vllm_fa_cmake_gpu_arches=90-real -t vllm:0.7.2 --target vllm-base` takes 'at least 3 hours on A100', and the open question is specifically how to speed up the vllm-flash-attn compilation. Conditions that matter and are stated: single SM 9.0 target with +PTX, max_jobs=32, A100-class host, vLLM 0.7.2, Docker build of the vllm-base target. Community issue, not a vendor measurement.