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

[FRONTEND][MXNET 2.0] Support three missing operators: _npi_subtract_scalar, _npi_stack, _npi_advanced_indexing_multiple #6721

Closed
sxjscience opened this issue Oct 20, 2020 · 8 comments

Comments

@sxjscience
Copy link
Member

Recently, TVM has supported the incoming MXNet 2.0 frontend via a series of PRs: #6054, #6699. This enables the new GluonNLP 1.0, which has upgraded from MXNet 1.x to MXNet 2.0 to convert to TVM graph (as added in dmlc/gluon-nlp#1390). However, I noticed that there are three missing operators when I profile against the BART model.

  • _npi_subtract_scalar
  • _npi_stack
  • _npi_advanced_indexing_multiple

The first two operators are pretty straightforward to support while the _npi_advanced_indexing_multiple is more complicated and is triggered when we call a[idx1, idx2]. The MXNet-side implementation of _npi_advanced_indexing_multiple is here: https://github.com/apache/incubator-mxnet/blob/6bbd53107aa16fc41e8d462cf5dc46fb70d592df/src/operator/numpy/np_indexing_op.cc#L479-L491

Since I do not have the bandwidth in supporting this operator. Help is really appreciated.

@sxjscience
Copy link
Member Author

Also cc @yzhliu

@sxjscience
Copy link
Member Author

To provide more context on the _npi_advanced_indexing_multiple usage in GluonNLP, it's used here: https://github.com/dmlc/gluon-nlp/blob/149270a7b357e2a34972e6a596fbeea2bbd6869e/src/gluonnlp/models/bart.py#L277

@junrushao
Copy link
Member

Hey Xingjian, thanks for bringing this up! It is certainly more than welcome for future contribution. Shall we put up a list of MXNet operators to be supported and consolidate them into a single thread instead? If there is no actionable item, shall we close this issue? Thanks!

@sxjscience
Copy link
Member Author

I think we can create a thread about that. Do you think that we should create a new one?

@junrushao
Copy link
Member

Yeah we can have a consolidated "[Frontend][MXNet] Importer Missing Operators" thread and mark it as "help wanted". What do you think?

@sxjscience
Copy link
Member Author

Sounds good to me. Should I change the title of this issue so we may reuse that? (Or create a new one?)

@junrushao
Copy link
Member

What about creating a new one, categorizing operators like #1799 does

@sxjscience
Copy link
Member Author

Move to #7186

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