Build on OpenStack server for TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0+PTX

[weiji14]

Originally posted by @carterbox in #1 (comment)

In parallel, if you feel there are significant performance to be gained by building for '8.0,9.0+PTX' or even more archs, then please start doing the steps in this checklist in order to get this feedstock and yourself added to the allowlist for the gpu-long-running runners which have a time limit longer than 6 hours.

CUDA Compute Capability table (https://developer.nvidia.com/cuda-gpus) for server-grade GPUs:

GPU	Compute Capability
NVIDIA RTX PRO 6000 Blackwell Server Edition	12.0
NVIDIA B200	10.0
NVIDIA H100	9.0
NVIDIA L4	8.9
NVIDIA L40	8.9
NVIDIA A100	8.0
NVIDIA A40	8.6
NVIDIA A30	8.0
NVIDIA A10	8.6
NVIDIA A16	8.6
NVIDIA A2	8.6
NVIDIA T4	7.5
NVIDIA V100	7.0

Notes:

• flash-attn is only supported on Ampere generation and above (see https://github.com/Dao-AILab/flash-attention/tree/v2.6.3?tab=readme-ov-file#installation-and-features)
• you can find your GPU's compute capability by running nvidia-smi --query-gpu=gpu_name,compute_cap --format=csv (ref).