Use incsubtesor in gradient of repeat

[ricardoV94]

This should be better than a dense dot product, both memory and speed wise. It's also more elegant?

Also:

1. Remove unused scalar repeat logic. The helper function pt.repeat uses broadcast_to + alloc in this case, so it's never triggered unless users create their Op manually. Well no more
2. Remove special axis=None logic. This is a common motif as many operations in numpy have a special meaning for axis=None, which means, ravel the input, and then apply operation as if axis=0. I decided to do that out of the Op, and let the Op focus on just axis=int case. Same as in Handle axis=None symbolically instead of within CumOp #1574

This way we only have to support the code that is really specific to Repeat with vector repeats.

---

📚 Documentation preview 📚: https://pytensor--1621.org.readthedocs.build/en/1621/

[codecov]

Codecov Report

❌ Patch coverage is 78.00000% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.62%. Comparing base (96122d1) to head (2230e91).

Files with missing lines	Patch %	Lines
pytensor/tensor/extra_ops.py	73.80%	7 Missing and 4 partials ⚠️

❌ Your patch check has failed because the patch coverage (78.00%) is below the target coverage (100.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1621      +/-   ##
==========================================
- Coverage   81.64%   81.62%   -0.02%     
==========================================
  Files         231      231              
  Lines       52997    52968      -29     
  Branches     9395     9384      -11     
==========================================
- Hits        43267    43237      -30     
- Misses       7282     7285       +3     
+ Partials     2448     2446       -2     

Files with missing lines	Coverage Δ
pytensor/link/numba/dispatch/basic.py	77.86% <100.00%> (ø)
pytensor/link/numba/dispatch/extra_ops.py	95.21% <100.00%> (-0.29%)	⬇️
pytensor/tensor/extra_ops.py	88.15% <73.80%> (-0.41%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull Request Overview

This PR refactors the Repeat operation to use IncSubtensor for gradient computation instead of dense dot products, removing support for scalar repeats and axis=None at the Op level. These cases are now handled by the helper function pt.repeat using broadcast_to and reshape operations.

• Replace gradient computation with IncSubtensor approach for better memory/performance
• Remove scalar repeat and axis=None logic from Repeat Op, delegating to helper function
• Simplify Numba implementation by removing object mode for most cases

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
pytensor/tensor/extra_ops.py	Main refactoring of Repeat Op constructor, make_node, grad, and repeat helper function
tests/tensor/test_extra_ops.py	Updated tests to reflect new Op constraints and added static shape inference tests
tests/link/numba/test_extra_ops.py	Updated test parameters to match new Op requirements
pytensor/link/numba/dispatch/extra_ops.py	Simplified Numba implementation with fallback for unsupported cases
pytensor/link/numba/dispatch/basic.py	Made node parameter required in generate_fallback_impl
pytensor/utils.py	Removed outdated comment about Python < 2.6

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[jessegrabowski]

Why?

[ricardoV94]

Because it's not really optional, the code immediately assumes node is passed (and not None)

[jessegrabowski]

Check that the relevant error is raised?

[jessegrabowski]

Suggested change

	x = TensorType(config.floatX, shape=(None, 1, 3))()
	x = pt.tensor(shape=(None, 1, 3))

Why is TensorType being used directly here?

[ricardoV94]

someone wrote it like that at first

[jessegrabowski]

The static shape is this case is just the sum of the repeat values?

You want to do something like try to constant fold the sum of repeats? We could just check if its a constant and grab the data out?

[ricardoV94]

We could... we can always COULD