Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Revist data type work arounds for onnxruntime >1.19.0 #42

Open
cbourjau opened this issue Aug 1, 2024 · 1 comment
Open

Revist data type work arounds for onnxruntime >1.19.0 #42

cbourjau opened this issue Aug 1, 2024 · 1 comment

Comments

@cbourjau
Copy link
Collaborator

cbourjau commented Aug 1, 2024

The next release of onnxruntime will sport some more kernels that we need for reduction operations (related PR). Let's revisit some of our data type work around after the next release.

@adityagoel4512
Copy link
Member

adityagoel4512 commented Aug 1, 2024

One broader note/thought on this. The type gymnastics required to ensure things work smoothly on the onnxruntime backend is really best suited for a separate package entirely (which we may or may not import here) that consolidates the rewrite rules being performed at the graph level.

A few reasons:

  • It means you can consolidate the techniques required and unit test this more systematically. You might even be able to automate much of the kernel diffing.
  • It is not exclusively a challenge for ndonnx.
  • It simplifies this codebase dramatically.
  • There is no clear timeline when ndonnx should remove any of these "hacks" once upstream changes are made but a rewrite library with a versioned API would make it possible to specify this explicitly.
  • onnxruntime isn't the only backend. Many of these workarounds simply are not relevant for other backends. Ultimately this is a library that should just focus on standard compliant ONNX export.

@cbourjau cbourjau changed the title Revist data type work arounds for onnxruntime >=1.18.1 Revist data type work arounds for onnxruntime >1.19.0 Aug 2, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants