[Mayuri (mayuriphad)]

Fixes #3154.

For a pattern with several output nodes, SimplePatternMatcher.match fixed the first output to the candidate node and tried every graph node with the same op type for each other output. It also rebuilt an index of all graph nodes for every candidate. On a graph with many Slice nodes, SlicesSplit (the only multi-output rule in the default set) ends up far slower than every other rule together.

Changes in _matcher.py:

• The first output node is matched once, up front. Its result does not depend on the other candidates, so if it fails no combination can succeed and the loop is skipped.
• For each other output node, the candidates are the consumers of a value the first match already bound to one of its inputs, at the same input position. A node that does not read that value cannot match, so this does not drop any match. Named pattern variables are looked up in bindings, the others in value_bindings.
• The op-type index is only built when no input of the other output node is bound, and then once per call.

Timing SlicesSplit alone on n independent Slices of one input (the graph from the issue, axis 0 so nothing is rewritten):

n	main	this PR
250	0.26 s	0.03 s
500	1.00 s	0.10 s
1000	4.55 s	0.34 s

Here every Slice reads the same x, so each one still has all the others as candidates and the time still grows faster than linearly; avoiding that would need rule-specific knowledge.

Tests: two cases in pattern_test.py. One checks that only the pair sharing an input is rewritten. The other counts _multi_match calls on a graph where each Relu shares its input with one Neg: 20 calls with this change, 799 on main. pytest onnxscript/rewriter passes (595 passed, 2 skipped; ort_fusions/gqa_test.py skipped locally because torch is not installed), and ruff check / ruff format --check are clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[Copilot]

🟡 Changes recommended

A critical early return in _match_node leaves the match result truthy when a candidate has too few outputs, allowing an incomplete match to be treated as valid for rewriting.

1 open issue

• Mark output-count mismatches as failed matches · New

_{Review effort: Balanced}