Do not skip validation between consecutive Elemwise inplace replacements

[ricardoV94]

Closes #1420

There was a performance-related hack in the ElemwiseInplaceOptimizer where it tried to avoid validating the graph after replacing each node.

This is a bad idea because it may revert valid rewrites that happen to be caught in the same "check" window as an invalid rewrite. More importantly since we started inplacing on multi-output Elemwise, it could trigger an exception when trying to call has_destroyers on subsequent nodes, due to a previous invalid replacement.

It was hard to track down this issue because the special behavior was only triggered once a graph had more than 500 nodes.

After the refactor I noticed that most of the logic of the rewrite can be shared with Blockwise, so I went ahead and refactored it, which also closes #1457

---

📚 Documentation preview 📚: https://pytensor--1494.org.readthedocs.build/en/1494/

[codecov]

Codecov Report

Attention: Patch coverage is 92.85714% with 12 lines in your changes missing coverage. Please review.

Project coverage is 82.02%. Comparing base (236e50d) to head (1438953).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/tensor/rewriting/elemwise.py	92.66%	5 Missing and 3 partials ⚠️
pytensor/tensor/rewriting/blockwise.py	86.20%	2 Missing and 2 partials ⚠️

❌ Your patch check has failed because the patch coverage (92.85%) 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    #1494      +/-   ##
==========================================
+ Coverage   81.98%   82.02%   +0.04%     
==========================================
  Files         231      231              
  Lines       52192    52216      +24     
  Branches     9185     9186       +1     
==========================================
+ Hits        42790    42831      +41     
+ Misses       7094     7080      -14     
+ Partials     2308     2305       -3     

Files with missing lines	Coverage Δ
pytensor/graph/destroyhandler.py	73.05% <100.00%> (+3.60%)	⬆️
pytensor/tensor/rewriting/blockwise.py	96.07% <86.20%> (-1.28%)	⬇️
pytensor/tensor/rewriting/elemwise.py	92.70% <92.66%> (+0.93%)	⬆️

... and 1 file with indirect coverage changes

🚀 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 ensures validation after each in-place rewrite in the Elemwise optimizer, refactors shared logic for Blockwise in-place rewrites, and adds regression tests for both cases.

• Enforce full graph validation between consecutive Elemwise in-place replacements
• Refactor Blockwise in-place optimizer to extend the shared InplaceGraphOptimizer
• Add tests for regression in Elemwise and partial in-place behavior in Blockwise

Reviewed Changes

Copilot reviewed 4 out of 5 changed files in this pull request and generated no comments.

File	Description
tests/tensor/test_blockwise.py	Add test_partial_inplace to cover Blockwise in-place behavior
tests/tensor/rewriting/test_elemwise.py	Add test_InplaceElemwiseOptimizer_bug regression test for #1420
pytensor/tensor/rewriting/blockwise.py	Refactor Blockwise in-place rewrite using InplaceGraphOptimizer
pytensor/graph/destroyhandler.py	Introduce inplace_candidates helper for unified candidate logic

Comments suppressed due to low confidence (4)

tests/tensor/rewriting/test_elemwise.py:1535

• [nitpick] Test function names should use snake_case. Consider renaming this to test_inplace_elemwise_optimizer_bug for consistency with pytest conventions.

def test_InplaceElemwiseOptimizer_bug():

tests/tensor/rewriting/test_elemwise.py:1544

• The test references Elemwise but there is no corresponding import. Please add from pytensor.tensor.rewriting.elemwise import Elemwise at the top of the file.

    out1, out2 = Elemwise(ps.Composite([z1, z2], [z1 + z2, z2 - z1]))(z[1:], z[:-1])

pytensor/tensor/rewriting/blockwise.py:4

• The imported vectorize_node is no longer used in this file. Consider removing this import to clean up unused code.

from pytensor.graph.replace import vectorize_node

pytensor/tensor/rewriting/blockwise.py:5

• Neither copy_stack_trace nor out2in are used after the refactor. Removing these unused imports will improve code clarity.

from pytensor.graph.rewriting.basic import copy_stack_trace, out2in

[ricardoV94]

Actually I suspect with my second commit the inplace optimizer can't accidentally introduce invalid graphs anymore. If that's the case we could do a single validate at the end of the rewrite. I didn't actually try to see what's so slow about validate.

That's less priority than just the fix here