onnxruntime
onnxruntime copied to clipboard
microsoft
ORT 1.17.0 Release Candidates available for testing
ORT 1.17 will be released in late January. Release candidate builds are available now for testing. If you encounter issues, please report them by responding in this issue.
Release branch: rel-1.17.0 Release manager: @YUNQIUGUO
| Pypi | Nuget | npm | Maven (Java) |
|---|---|---|---|
| CPU: 1.17.0.dev20240118001 GPU: 1.17.0.dev20240118001 |
CPU: 1.17.0-dev-20240119-0139-a63b71eadb GPU (CUDA/TRT): 1.17.0-dev-20240118-2301-a63b71eadb DirectML: 1.17.0-dev-20240119-0131-a63b71eadb WindowsAI: 1.17.0-dev-20240119-0131-a63b71eadb |
onnxruntime-node: 1.17.0-dev.20240118-a63b71eadb onnxruntime-react-native 1.17.0-dev.20240118-a63b71eadb onnxruntime-web 1.17.0-dev.20240118-a63b71eadb |
CPU: 1.17.0-rc1 GPU: 1.17.0-rc1 |
Describe scenario use case
not applicable.
Agree, prebuilt CUDA 12.1 packages would be really appreciated. Like you started doing with this nightly: https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/ort-cuda-12-nightly/PyPI/ort-nightly-gpu/versions/1.17.0.dev20231205004
Here's the CUDA 12 version RC packages for Python and Nuget: (based on the same commit id)
ort-nightly-gpu : 1.17.0.dev20240118002
Microsoft.ML.OnnxRuntime.Managed:1.17.0-dev-20240118-2235-a63b71eadb
Microsoft.ML.OnnxRuntime.Gpu: 1.17.0-dev-20240118-2235-a63b71eadb
Microsoft.ML.OnnxRuntime.Gpu.Linux:1.17.0-dev-20240118-2235-a63b71eadb
Microsoft.ML.OnnxRuntime.Gpu.Windows:1.17.0-dev-20240118-2235-a63b71eadb
@HectorSVC why was this issue unpinned? We pin our release candidates in GitHub issues so our partners + community members have easier access and can test them.
Thanks for managing the release @YUNQIUGUO. Given a PyPI release has been made for 1.17, is there a plan to tag and release it on GitHub as well? For context, the onnxruntime feedstock on conda-forge typically uses the tagged github release to build ORT.
Thanks for managing the release @YUNQIUGUO. Given a PyPI release has been made for 1.17, is there a plan to tag and release it on GitHub as well?
yep, we are still waiting for the last couple packages to be uploaded to the package management repo. after everything completed, will do a release announcement with the 1.17.0 package assets on Github as well.
Thanks for the new release! However, I see that a CUDA 12 version of the python onnxruntime-gpu package wasn't included in the pypy release. I tested the CUDA 12 RC packages that @YUNQIUGUO published in this thread last week and everything was working well, so I believe releasing a 1.17 version would be very useful.
Do you have plans to upload them at a later date or is there a particular reason why you are choosing not to do it?
Thank you very much!
The cuda 12 instructions are here: https://onnxruntime.ai/docs/install/#install-onnx-runtime-gpu-cuda-12x It seems the onnxruntime-gpu package doesn't exist on AzDo, I only found ort-nightly-gpu==1.17.0.dev20240130002 so far. Could we release a properly named artifact for 1.17, as opposed to the nightly build?
For pypi, i'm not sure pypi can release onnxruntime-gpu for more than one major version of cuda. E.g. pytorch hosts a separate repo for its cuda11 py packages.
@YUNQIUGUO, please upload onnxruntime-gpu 1.17.0 CUDA 12 python package to https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/onnxruntime-cuda-12/
@fvdnabee
The cuda 12 instructions are here: https://onnxruntime.ai/docs/install/#install-onnx-runtime-gpu-cuda-12x It seems the onnxruntime-gpu package doesn't exist on AzDo, I only found ort-nightly-gpu==1.17.0.dev20240130002 so far. Could we release a properly named artifact for 1.17, as opposed to the nightly build?
Thanks for reporting this. I am guessing our Python Cuda 12 packaging pipeline probably lacks a release version configuration/nightly build option. I actually uploaded onnxruntime-gpu to the Cuda 12 official feed but looks like it still contains a nightly in the naming though it should be an official one instead. I will contact the pipeline owner and we'll look into address this issue and re-upload an official named 1.17.0 package.
Since it's the first time release cuda 12 version wheels to the feed, sorry about that - the issue hasn't be identified before.
@dbuades I am not aware of a plan to upload it to an official repo like pypi.org yet. will ask around.
No. DML is ongoing (https://github.com/microsoft/onnxruntime/pull/19274) and CUDA support is the next.
I'm curious if onnxruntime-node now supports dml and cuda?
@dbuades @fvdnabee Hey, the issue has been resolved now. and here's the official onnxruntime-gpu package for cuda 12 version: https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/onnxruntime-cuda-12/PyPI/onnxruntime-gpu/overview/1.17.0
No. DML is ongoing (#19274) and CUDA support is the next.
I'm curious if onnxruntime-node now supports dml and cuda?
#19274 is merged in main and marked for patch release 1.17.1, pending approval.
I'm curious if there is any work towards DML for onnxruntime-java? Where might I be pointed to make that a thing? (I did search issues... seems not discussed yet?)
I'm curious if there is any work towards DML for onnxruntime-java? Where might I be pointed to make that a thing? (I did search issues... seems not discussed yet?)
@CaelumF 🤔 I haven't heard desire for that combination of language + backend before (more often DML is accessed via C++, Python, and C#), but is it already supported? I see this enum ai.onnxruntime.OrtProvider.DIRECT_ML. (alas I don't know who on ORT owns the Java language layer to verify)
@fdwr hey thanks for responding. The pre built jars do have functions for enabling DirectML, but the binary wasn't compiled with DirectML enabled and errors when attempting to use it, building with it enabled has some trouble. Issue posted here https://github.com/microsoft/onnxruntime/issues/19656 which also includes my use case in case you're curious
I'm the maintainer of the Java layer, I'll have a look at what's going on.
