Clear initializers in constant folding pass

[justinchuby]

Clear unused initializers on the fly to prevent memory usage jump due to intermediate folded tensors.

[Copilot]

Pull Request Overview

This PR adds automatic cleanup of unused initializers during constant folding and bumps the package version to 0.5.6.

• Adds functionality to detect and remove unused initializers when nodes are removed during constant folding
• Updates package version from 0.5.5 to 0.5.6

Reviewed Changes

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

File	Description
onnxscript/optimizer/_constant_folding.py	Adds _clear_unused_initializers helper function and integrates it into the replace_node method to clean up unused initializers after node replacement
VERSION	Bumps package version from 0.5.5 to 0.5.6

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 70.45%. Comparing base (ee9a6e8) to head (289a8ba).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2668      +/-   ##
==========================================
+ Coverage   70.42%   70.45%   +0.02%     
==========================================
  Files         224      224              
  Lines       26729    26751      +22     
  Branches     2673     2678       +5     
==========================================
+ Hits        18825    18847      +22     
  Misses       6973     6973              
  Partials      931      931              

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

[justinchuby]

@titaiwangms fixed tests. PTAL. Also cc @gramalingam

[gramalingam]

Do we need this assertion? This might be true in the current implementation ... but just wondering if we might ever want to replace [oldnode] by [newnode1, oldnode, newnode2], for example ... not sure of replace_nodes_and_values supports such a scenario ... but the logic here seems to be robust even in that case. But the assert would fail in that case. Anyway ... just a minor suggestion.

Also: in principle, this logic could go into replace_nodes_and_values, can't it?

[justinchuby]

this logic could go into replace_nodes_and_values, can't it?

Good point - I agree. But also replace_nodes_and_values is more general. So checking value users may be more expensive there.

replace [oldnode] by [newnode1, oldnode, newnode2]

Not sure? replace_nodes_and_values will disconnect the old nodes (which breaks all the value connections) so I don't think it will work?