fix: Bugfix in convNd_to_convolution lowering pass by gs-olive · Pull Request #1693 · pytorch/TensorRT (original) (raw)

Description

Lowering pass did not respect prim::If block boundaries
Refactor convNd implementation to use more precise guard-insert paradigm instead of subgraph rewriting
Write general function to apply for all convolution replacements
When replacing a subgraph that occurs within an "If" block, the rewriter places the actual logic of the code outside of the block, so the rewrite makes the code execute both the "if" and the "else" path regardless of what the condition is
Add regression test case to validate refactoring on aten::conv1d

Addresses first bug in #1597

[ x ] My code follows the style guidelines of this project (You can use the linters)
[ x ] I have performed a self-review of my own code
[ x ] I have commented my code, particularly in hard-to-understand areas and hacks
[ x ] I have made corresponding changes to the documentation
[ x ] I have added tests to verify my fix or my feature
[ x ] New and existing unit tests pass locally with my changes
[ x ] I have added the relevant labels to my PR in so that relevant reviewers are notified

Lowering pass did not respect prim::If block boundaries
Refactor convNd implementation to use more precise guard-insert paradigm instead of subgraph rewriting
Write general function to apply for all convolution replacements
When replacing a subgraph that occurs within an "If" block, the rewriter places the actual logic of the code outside of the block, so the rewrite makes the code execute both the "if" and the "else" path regardless of what the condition is
Add a test case to validate refactoring on conv1d

Partitioning test incorrectly expected 1 conditional engine, but got 2 since log_sigmoid operator is not currently supported