B200 runners used CUDA 13.0 docker images with Python 3.10, which caused install_cutlass_api to skip installation and left NV Universal GEMM smoke tests untested. Add py3.12 cuda13.0 docker image variants and point B200 workflows at them so cutlass_api can install on the runner.
Authored with Cursor Composer 2.5 (composer-2.5-fast).
Fixes #186991
Pull Request resolved: #186997
Approved by: https://github.com/huydhn
SOCIAL SHARE CARD GENERATOR