search

microsoft / onnxruntime

ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
https://onnxruntime.ai
MIT License
14.1k stars 2.84k forks source link

[DO NOT UNPIN] ORT 1.19 release candidates available for testing #21678

Open MaanavD opened 1 month ago

MaanavD commented 1 month ago

ORT 1.19 will be released soon, and release candidate builds are now available for testing. If you encounter issues, please report them ASAP by responding to this issue and tagging myself and @prathikr.

Release branch: rel-1.19.0 Release manager: @prathikr

Here's the updated table with the entries shortened to just the build number and relevant descriptors added:

Python Whls Nuget NPM CocoaPods Maven (Java)
CPU: 1.19.0.dev20240805002 Managed: 1.19.0-dev-20240805-1630 Node: 1.19.0-dev.20240804 pod 'onnxruntime-c', '1.19.0-dev+20240805002.ee2fe87' CPU: 1.19.0-rc1
GPU: 1.19.0.dev20240731002 CPU: 1.19.0-dev-20240805-1630 React Native: 1.19.0-dev.20240804 pod 'onnxruntime-objc', '1.19.0-dev+20240805002.ee2fe87' GPU: 1.19.0-rc1
DirectML: 1.19.0.dev20240805002 GPU: 1.19.0-dev-20240805-0337 Web: 1.19.0-dev.20240804 pod 'onnxruntime-training-c', '1.19.0-dev+20240805002.ee2fe87'
WindowsAI: 1.19.0-dev-20240805-0131 pod 'onnxruntime-training-objc', '1.19.0-dev+20240805002.ee2fe87'
DirectML: 1.19.0-dev-20240805-1630
Training: 1.19.0-dev-20240805-0942
Djdefrag commented 1 month ago

20713

@MaanavD @prathikr Hi, this issue has been fixed? It was introduced with version 1.18 and seems to affect both C++ and Python

MaanavD commented 1 month ago

@Djdefrag currently looking into it :) Will let you know soon! If it hasn't been, we can discuss a patch release as I see this has been around for a while and I'd hate to see it not being addressed for you

henryruhs commented 3 weeks ago

DirectML (and probably ROCm) multi threading still being broken in that release is a stopper (in my opinion).

https://github.com/microsoft/onnxruntime/issues/20713

Djdefrag commented 3 weeks ago

@henryruhs

For me too. I am staying on 1.17.3 for now.

Have you also noticed a slowdown in inference speed with 1.19 compared to 1.17.3? Using DirectML (and only 1 thread) I noticed a regression in speed of 10%.