Fix local_fill_sink rewrite for multiple output Elemwise Ops
#773
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ricardoV94
requested review from
tomicapretto,
jessegrabowski and
Dhruvanshu-Joshi
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #773 +/- ##
==========================================
- Coverage 80.85% 80.85% -0.01%
==========================================
Files 162 162
Lines 47019 47016 -3
Branches 11504 11501 -3
==========================================
- Hits 38018 38014 -4
- Misses 6750 6751 +1
Partials 2251 2251
|
ricardoV94
force-pushed
the
fix_local_fill_sink_multiple_outputs
branch
2 times, most recently
from
ba17e8f to
381b88f
Compare
The changes get rid of the eager sink at the local node rewriter level. This was actually not working because the nested replacements referenced variables that were never part of the original fgraph and those replacements were being ignored altogether. Instead we wrap the rewrite in an in2out that will safely achieve the intended behavior.
ricardoV94
commented
May 16, 2024
Comment on lines
+337
to
+342
| a, b = inp.owner.inputs | ||
| if b.type.dtype != inp.dtype: | ||
| # The input was implicitly casted by the fill operation | ||
| b = b.cast(inp.dtype) | ||
| models.append(a) | ||
| inputs.append(b) |
There was a problem hiding this comment.
This was also a potential source of bugs in the old rewrite. Ops may behave fundamentally different if the input types change so we shouldn't let that happen
ricardoV94
force-pushed
the
fix_local_fill_sink_multiple_outputs
branch
from
381b88f to
7019ddf
Compare
10 tasks
ricardoV94
commented
May 16, 2024
| @node_rewriter([Elemwise]) | ||
| def local_fill_sink(fgraph, node): | ||
| """ | ||
| f(fill(a, b), fill(c, d), e) -> fill(c, fill(a, f(b, d, e))) | ||
| f need to be an elemwise that isn't a fill. | ||
| """ | ||
| if not hasattr(node, "op") or not isinstance(node.op, Elemwise) or node.op == fill: | ||
| if isinstance(node.op.scalar_op, Second): |
There was a problem hiding this comment.
The extra checks were only needed for the recursive call of this rewrite. A default call will never call it on an node without Op that is not already Elemwise
|
For reference fill is identical to |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Reported by @tomicapretto in pymc-devs/pymc#7315 (review)
The rewrite implicitly assumed Elemwise nodes have a single output which is not true. The reported issue involved the gradient of BetaIncGrad which includes a ScalarLoop Op with multiple outputs.
The changes get rid of the eager sink at the local node rewriter level. This was actually not working because the nested replacements referenced variables that were never part of the original fgraph and those replacements were being ignored altogether. Instead we wrap the rewrite in an in2out that will safely achieve the intended behavior.
The new test works as a regression for the bug, in that if we were to call the old rewrite (or just try to compile the graph) it would lead to the reported issue when finding the MultiOutput Op.
Checklist
Type of change