[WS1][CI] The GPU architecture unmatch in GPU CI.

[frank-2077]

Problem

GPU CI currently falls back from an unavailable requested GPU topology to a different GPU architecture and GPU count.

For example, the A100 matrix job requests:

2 × NVIDIA A100 80GB PCIe
TARGET_SM=8.0

When capacity is unavailable, ci/run_gpu_ci.sh falls back to:

1 × NVIDIA A40
actual SM=8.6

The script then correctly rejects the mismatch:

FATAL: requested TARGET_SM=8.0 (sm_8.0) but provisioned GPU is sm_8.6. Refusing to build mismatched kernels.

This exits before the extension build or pytest suite runs.

[hongleng]

Hi, I’d like to take this issue. I propose making fallback selection architecture-specific in the workflow matrix: every architecture-targeted job will explicitly provide a fallback GPU with the same compute capability as TARGET_SM. If no same-SM fallback is configured or available, the job will fail clearly and can be retried instead of provisioning a cross-architecture pod. I’ll keep the existing runtime SM assertion as the final safety check and limit the change to the workflow/script plus a small runnable selection check. Does this scope match the intended CI policy?