My previous commit can only build with TRT 8. To build with TRT 9 there are some necessary changes in https://github.com/HolyWu/TensorRT/commit/f644baa4f073afcb7e27ee655d62771607c4a7c2. Besides, if you prefer building from archives rather than using local sources, it's also doable in https://github.com/HolyWu/TensorRT/commit/b54a3bf9962b91a22f445596a53e24e7be740bfe.

I don't think you can install/upgrade the display driver via conda because the nvidia packages in conda seems to not include the display driver. Only the official CUDA Toolkit installer includes the display driver component. I'm not sure if you can successfully execute the network installer in silent mode to only install the display driver on the windows runner, because it may require admin privilege or rebooting.

On the other hand, maybe poke the pytorch guys to upgrade the driver on their windows runner is a better solution.

Thanks for this feedback - I am looking into getting the driver version for Windows updated here.

The default shell on Windows runner is powershell. You need to change the shell in Pack script to bash like https://github.com/gs-olive/TensorRT/blob/2b25c988970a8e4282cf5f5e58a2a116e74436b0/.github/workflows/windows-test.yml#L137 since most commands in the workflow are not supported by powershell.

I have managed to fix the CI errors in my fork, including successful C++ building. Because I cannot use PyTorch-hosted runner in my fork, I have to run on GitHub-hosted runner. But GitHub-hosted runner doesn't have GPU support, so I cannot run pytest.

Some things you need to do besides reverting back to use pytorch's repo in https://github.com/HolyWu/TensorRT/commit/fa7ecc973d58c960fd4d0f8ecb05467f772c583a:

• build-test-windows.yml#L3: change to pull_request only.
• build-test-windows.yml#L184: uncomment
• windows-test.yml#L56: change to use PyTorch's runner.
• windows-test.yml#L92: change to false.
• pre_build_script_windows.sh#L13: change to only install driver since PyTorch's runner already have CUDA Toolkit installed.
• smoke_test_windows.py#L33: uncomment.

Don't forget to set setup-miniconda to false in windows-test.yml when using pytorch's Windows runner.

Hmm...maybe you need to set https://github.com/gs-olive/test-infra/blob/0c5cc5b410cfaa6953a15fdd564c2399a95b185e/.github/workflows/build_wheels_windows.yml#L80 back to windows-2019? It's the culprit of the Called workflows cannot be queued onto self-hosted runners across organizations/enterprises. error.

That worked - thank you! Now looking into why this error is appearing during install:

ERROR: torch_tensorrt-2.3.0.dev0+6031b50-cp39-cp39-win_amd64.whl is not a supported wheel on this platform.

It seems that even though the builds are named py38, they are building for py39, which might be contributing