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

List decompositions for torch.export #26878

Open
wants to merge 1 commit into
base: master
Choose a base branch
from

Conversation

mvafin
Copy link
Contributor

@mvafin mvafin commented Oct 1, 2024

Details:

  • item1
  • ...

Tickets:

  • ticket-id

@mvafin mvafin requested review from a team as code owners October 1, 2024 16:32
@github-actions github-actions bot added category: Python API OpenVINO Python bindings category: PyTorch FE OpenVINO PyTorch Frontend category: OVC OVC tool labels Oct 1, 2024
def get_export_decomposition_list():
# List of decompositions from torch._decomp.core_aten_decompositions
# removed _backward ops and ops supported without decomposition
decomp = [
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How are we going to maintain this list from one version to another?
Should we really need it for models conversion or is it for NNCF tool only?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This list changes not very often, we do not really need to always have it up to date. We may even reduce it in future. The main reason this list is needed is that SDPA gets decomposed by default but we want it to exist as a single operation.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we check in the test that SDPA op is really created in the original graph?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The decompositions happen inside convert_model. Checking that operation exist would require patching ovc, also would need to decide how to split ts ops from fx ops when writing a test. This solution is outside of scope of this PR. I suggest solving that with a different PR

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
category: OVC OVC tool category: Python API OpenVINO Python bindings category: PyTorch FE OpenVINO PyTorch Frontend
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants