Implements aten_index_put if inputs are SymbolicTensor

[xadupre]

No description provided.

[codecov]

Codecov Report

❌ Patch coverage is 79.01235% with 17 lines in your changes missing coverage. Please review.
✅ Project coverage is 70.38%. Comparing base (75b3d42) to head (988e9f6).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
onnxscript/function_libs/torch_lib/ops/core.py	79.01%	9 Missing and 8 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2606      +/-   ##
==========================================
- Coverage   70.38%   70.38%   -0.01%     
==========================================
  Files         222      222              
  Lines       26288    26390     +102     
  Branches     2629     2647      +18     
==========================================
+ Hits        18503    18574      +71     
- Misses       6865     6886      +21     
- Partials      920      930      +10     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

check not isinstance(index, int) instead as we should not reference the private class.

[xadupre]

What are the possible types here? Only int and SymbolicTensor? I prefer to keep SymbolicTensor because I know exactly which type the function is supposed to handle.

[justinchuby]

Only int and SymbolicTensor for the exporter, but could be other ir.Value subclasses as well. As the type is an internal type it is not meant for public use, the current usage is not supported and is brittle.

If preferred you may check for ir.Value instead, but really we just assume a type that has shape and dtype fields

[xadupre]

Feel free to suggest. I still prefer to keep this one since this is what shows up in the error message.

[justinchuby]

Prefer clear variable names and avoid abbreviations.

[xadupre]

The variable are used just after. It is short to make the code shorter and esier to read.

[justinchuby]

Any abbreviation adds cognitive load and is not preferred. All names should be thoughtfully created to be precise. Creating variables used right after is a plus but I don’t feel warranting the use of ambiguous names.

https://google.github.io/styleguide/pyguide.html#316-naming

[xadupre]

I let you choose the name.

[justinchuby]

As commented:

1. avoid referencing internal members
2. follow existing coding style for choosing names for consistency and readability

[gramalingam]

Is there a pointer to the real use-cases that show up? I thought each index in indices is supposed to be a 1D tensor? What are the ways in which SymbolicTensors are created?

[gramalingam]

Is a SymbolicTensor a 1-element tensor version of a SymbolicInt?

[justinchuby]

SymbolicTensor is just ir.Value with magic methods defined to support python operators.

[justinchuby]

Here the symbolic tensors are created from the symints.