[Feature Request] prebuilt package with cudnn 9 support

[LSC527]

Describe the feature request

NVIDIA released tensorrt 10 and cudnn 9, but onnxruntime's prebuilt package only support cudnn 8 by now.

Describe scenario use case

more convenient for users to install if using cudnn 9

[jywu-msft]

we plan to publish 1.18.1 gpu packages built against CUDA 12.x and CuDNN 9.x later this month.
(in addition to the default gpu packages built against CUDA 11.8 + CuDNN 8.x)

[jywu-msft]

Can you help test out these Release Candidate Packages which support CUDA 12.x , TensorRT 10 and CuDNN 9.x?
https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/onnxruntime-cuda-12/NuGet/Microsoft.ML.OnnxRuntime.Gpu.Linux/overview/1.18.1-rc
https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/onnxruntime-cuda-12/NuGet/Microsoft.ML.OnnxRuntime.Gpu.Windows/overview/1.18.1-rc

[LSC527]

@jywu-msft Not so sure how to use this package. Is there a pypi whl package？

[jywu-msft]

no, those were nuget packages.
here are whl's
https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/onnxruntime-cuda-12/PyPI/ort-nightly-gpu/overview/1.18.1.dev20240624003

[LSC527]

@jywu-msft Thx.
Everything went great when I test this whl in a docker env nvcr.io/nvidia/pytorch:24.05-py3 (cuda12.4, tensorrt 10.0.1, cudnn 9.1), using some simple models from torchvision.
Will report isssues if encountered.

[jywu-msft]

@jywu-msft Thx. Everything went great when I test this whl in a docker env nvcr.io/nvidia/pytorch:24.05-py3 (cuda12.4, tensorrt 10.0.1, cudnn 9.1), using some simple models from torchvision. Will report isssues if encountered.

Thank you!