I probably find out the issue why CUDA build was not triggered. We have a new option cxx11_abi intorduced, but the field is not added to the CUDA build. This results in the error when call b.cxx11_abi == "1", false.

Apart from that, those .tf configs are quite out dated and not align with the currect GCP config. I have redo the terraform apply to keep everything up to date.

Hi @lsy323 , I have confirmed the syntax correctness with terraform plan and already applied the update to the GCP. Let's see if those wheels can be successfully built or not before merge in. Thanks!

I come to notice that our build will fail if we set BUNDLE_LIBTPU=1. The fail is due to the builder script will look for libtpu from the following address:

https://storage.googleapis.com/libtpu-nightly-releases/wheels/libtpu/libtpu-0.0.11.dev20250303+nightly-py3-none-linux_x86_64.whl

However, the address is not valid anymore. Will fix this issue later.