I added TORCH_CUDA_ARCH_LIST and PYTORCH_ROCM_ARCH args and env variables in the hope that the information would be useful. I wanted to have one central place to configure and record supported GPU architectures.

The idea turned out to cause more problems than benefits:

• Wheel builder is changing CUDA arch list more often than base images are released.
• The presence of TORCH_CUDA_ARCH_LIST can slow down vLLM startup, see AIPCC-4016. Some operations and dependencies might compile kernels just-in-time. Without TORCH_CUDA_ARCH_LIST, only cubins for the current GPU arch are compiled. With TORCH_CUDA_ARCH_LIST present, code is compiled for additional archs.

Let's remove these env ars.