[DO NOT UNPIN] ORT 1.20 release candidates available for testing

[sophies927]

ORT 1.20 will be available soon (more info, including timeline + planned features can be found here), and the release candidate builds are now available for testing. If you encounter any issues, please report them ASAP by responding to this issue and tagging me and @apsonawane.

Release branch: rel-1.20.0
Release manager: @apsonawane

PyPI (Python Wheels)	NuGet	NPM	CocoaPods	Maven (Java)
CPU: 1.20.0.dev20241022005	Managed: 1.20.0-dev-20241022-1606-2d00351d7b	Node: 1.20.0-dev.20241022-2d00351d7b	pod 'onnxruntime-c', '1.20.0-dev+20241022005.2d00351'	CPU: 1.20.0-rc1
GPU (CUDA 12): 1.20.0.dev20241022004	CPU: 1.20.0-dev-20241023-1635-2d00351d7b	React Native: 1.20.0-dev.20241022-2d00351d7b	pod 'onnxruntime-objc', '1.20.0-dev+20241022005.2d00351'	GPU: 1.20.0-rc1
GPU (CUDA 11): 1.20.0.dev20241022005	GPU: 1.20.0-dev-20241022-1606-2d00351d7b	Web: 1.20.0-dev.20241022-2d00351d7b		Android QNN: 1.20.0-beta-01
DirectML: 1.20.0.dev20241022005	GPU Linux: 1.20.0-dev-20241022-1606-2d00351d7b
QNN: 1.20.0.dev20241022005	GPU Windows: 1.20.0-dev-20241022-1606-2d00351d7b
	Windows AI: 1.20.0-dev-20241022-1050-2d00351
	DirectML: 1.20.0-dev-20241023-1635-2d00351d7b
	QNN: 1.20.0-dev-20241025-0238-2d00351
	ROCm: 1.20.0-dev-20241022-2122-2d00351d

Note: Our team completes independent testing of the Android Maven package, so we do not include it in the release candidates.

NuGet GPU packages with CUDA 11 can be found here:

• Microsoft.ML.OnnxRuntime.Managed: 1.20.0-dev-20241023-1207-2d00351d7b
• Microsoft.ML.OnnxRuntime.Gpu: 1.20.0-dev-20241023-1207-2d00351d7b
• Microsoft.ML.OnnxRuntime.Gpu.Linux: 1.20.0-dev-20241023-1207-2d00351d7b
• Microsoft.ML.OnnxRuntime.Gpu.Windows: 1.20.0-dev-20241023-1207-2d00351d7b

[Djdefrag]

#20713

Tested the RC diretcml 1.20.0.dev20241022005

the problem has not been solved, in fact, it have gotten worse and now with 1.20 the gpu driver crash

[skyline75489]

Is the NuGet package supposed to be CUDA11? It seems that both 1.20.0-dev-20241023-1635-2d00351d7b and 1.20.0-dev-20241023-1207-2d00351d7b are CUDA11

[sophies927]

Is the NuGet package supposed to be CUDA11? It seems that both 1.20.0-dev-20241023-1635-2d00351d7b and 1.20.0-dev-20241023-1207-2d00351d7b are CUDA11

@skyline75489 your first link is the CPU (Microsoft.ML.OnnxRuntime) NuGet pkg, so that one shouldn't include any version of CUDA. The following links should go to all 3 of the GPU CUDA 12 NuGet pkgs:

1. Microsoft.ML.OnnxRuntime.Gpu (GPU CUDA 12 NuGet pkg): 1.20.0-dev-20241022-1606-2d00351d7b
2. Microsoft.ML.OnnxRuntime.Gpu.Linux (GPU CUDA 12 NuGet pkg for Linux): 1.20.0-dev-20241022-1606-2d00351d7b
3. Microsoft.ML.OnnxRuntime.Gpu.Windows (GPU CUDA 12 NuGet pkg for Windows): 1.20.0-dev-20241022-1606-2d00351d7b
4. Microsoft.ML.OnnxRuntime.Managed (CUDA 12 Managed NuGet pkg): 1.20.0-dev-20241022-1606-2d00351d7b

[skyline75489]

@sophies927 Thanks. I'll try these packages out. Maybe the link in the chart should be updated.

[sophies927]

@sophies927 Thanks. I'll try these packages out. Maybe the link in the chart should be updated.

Done.

[jkbeavers]

With the deprecation of the training packages, will any of the 1.20 releases have training-supported builds (e.g. python package has onnxruntime.training)? I do confirm that the rc for the python cpu wheel does not.

[sophies927]

@jkbeavers we actually ended up publishing both of our training CocoaPods packages (onnxruntime-training-objc and onnxruntime-training-c) as part of ORT 1.20, but I am not aware of any other training-supported builds. We are currently in the process of deprecating our training pipelines, so we likely will not have any training-supported builds for future releases.

[jkbeavers]

@sophies927 Thanks for the update. Is there a target date for when the the training pipelines will be fully deprecated? To clarify, does this include on device training?

Is there any additional communication besides the notice of deprecation of training packages from this release? It wasn't quite clear to me if the deprecation of the packages also included the training pipelines in their entirety.