Optimize aten::min/max.dim with TopK op by danielhumanmod · Pull Request #2780 · microsoft/onnxscript

danielhumanmod · 2026-01-25T18:31:12Z

Context

As mentioned in the issue, torch.max(dim=...) can be optimized with TopK to replace the current ReduceMax and ArgMax implementation. This optimization reduces redundant input scans and avoids potential performance overhead in certain execution providers (e.g., ONNX Runtime CUDA EP microsoft/onnxruntime#11348).

In additional, given the torch.min(dim=...) has the similar pattern with max, I also apply this optimization to it.

Verification

Successfully passed existing OpInfo consistency tests:

pytest tests/function_libs/torch_lib/ops_test.py
pytest tests/function_libs/torch_lib/e2e_ops_tests.py

danielhumanmod · 2026-01-25T18:34:41Z

@danielhumanmod please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.
@microsoft-github-policy-service agree [company="{your company}"]
Options:

(default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.
@microsoft-github-policy-service agree
(when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.
@microsoft-github-policy-service agree company="Microsoft"
Contributor License Agreement

@microsoft-github-policy-service agree

codecov · 2026-01-25T19:55:19Z

Codecov Report

❌ Patch coverage is 96.24060% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 70.72%. Comparing base (e06dd92) to head (301635e).
⚠️ Report is 7 commits behind head on main.

Files with missing lines	Patch %	Lines
.../rewriter/rules/common/_fuse_reduce_arg_to_topk.py	91.30%	6 Missing and 2 partials ⚠️
...iter/rules/common/_fuse_reduce_arg_to_topk_test.py	98.85%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2780      +/-   ##
==========================================
+ Coverage   70.46%   70.72%   +0.25%     
==========================================
  Files         228      230       +2     
  Lines       27258    27443     +185     
  Branches     2761     2757       -4     
==========================================
+ Hits        19208    19409     +201     
+ Misses       7100     7096       -4     
+ Partials      950      938      -12

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

justinchuby

Thanks for creating the PR. Reading it again it seems like topk is more general than ReduceMax and ArgMax. From a node count perspective this may be fewer nodes, but I wonder if the original is easier to optimize with.

danielhumanmod · 2026-01-26T00:14:28Z

Thanks for creating the PR. Reading it again it seems like topk is more general than ReduceMax and ArgMax. From a node count perspective this may be fewer nodes, but I wonder if the original is easier to optimize with.

Thanks so much for the review! That is a great point, I took some time to dig into the ONNX Runtime implementations to see how they handle this.

From ONNX runtime perspective,
1. CPU EP provide a fastline when k = 1, which performs a simple linear scan. So on CPU, it seems to behave identically to a fused max+argmax.
2. CUDA EP will walk through the whole Bitonic/Radix sort process, which can involve more complex instructions. But the upside is that these operations happen primarily in shared memory.
PyTorch Inductor (as an reference): it adopts a similar approach—splitting into reduce_max/arg_max in IR—but leaves it to the runtime (Scheduler) to fuse them. However, when I checked ONNX Runtime, it didn't seem to have an optimization rule to automatically fuse ReduceMax and ArgMax, which implies the split approach effectively incurs one more IO pass compared to TopK

So to the best of my knowledge, TopK might brings more instruction overhead but with less IO. I would appreciate your thoughts here—which approach aligns more with the community's needs? I am flexible to pivot to other tasks if we want to keep the original implementation.

justinchuby · 2026-01-26T06:28:08Z

I am not exactly sure what the actual usage of this operator looks like. Are the two outputs always used? One can imagine that if the second output is unused at all, computing it would be a waste of effort. I wonder if it would make sense for you to contribute a rewrite rule to https://github.com/microsoft/onnxscript/tree/main/onnxscript/rewriter/rules ? This way we can do fusion only when the two outputs are used (if not the second output will be removed by the dead code elimination pass)

danielhumanmod · 2026-01-27T02:57:32Z

I am not exactly sure what the actual usage of this operator looks like. Are the two outputs always used? One can imagine that if the second output is unused at all, computing it would be a waste of effort. I wonder if it would make sense for you to contribute a rewrite rule to https://github.com/microsoft/onnxscript/tree/main/onnxscript/rewriter/rules ? This way we can do fusion only when the two outputs are used (if not the second output will be removed by the dead code elimination pass)

Yeah, that's a good point. It makes more sense to handle this in the rewriter/optimizer. I will take a look at the rules and follow up. Thanks for the feedback!

danielhumanmod · 2026-02-03T02:41:17Z

Hey @justinchuby ,I’ve added a new rewrite rule to optimize this case based on our previous discussion. Whenever you have a moment, I’d appreciate your thoughts on it. Thanks!

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk_test.py

Copilot

Pull request overview

Adds a new ONNXScript rewriter rule to fuse Reduce{Max,Min} + Arg{Max,Min} patterns into a single TopK (plus optional Squeeze), aiming to improve performance for torch.min/max(dim=...)-style graphs.

Changes:

Introduces FuseReduce{Max,Min}Arg{Max,Min}ToTopK rewrite rules and a RewriteRuleSet.
Adds extensive unit tests covering success and failure conditions across opset 13 and 18.
Validates numerical equivalence and serialized-model correctness for rewritten graphs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
`onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py`	Implements the Reduce+Arg → TopK fusion rules for both max and min cases.
`onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk_test.py`	Adds unit tests for the new fusion rules, including opset and attribute/input variants.

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk_test.py

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

titaiwangms · 2026-02-06T22:51:56Z

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

+from onnxscript.rewriter._rewrite_rule import RewriteRuleClassBase, RewriteRuleSet
+
+
+class FuseReduceArgToTopKBase(RewriteRuleClassBase):


Suggested change

class FuseReduceArgToTopKBase(RewriteRuleClassBase):

class _FuseReduceArgToTopKBase(RewriteRuleClassBase):

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

titaiwangms · 2026-02-06T23:22:51Z

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

+        reduce_keepdims = (
+            reduce_keepdims_attr.as_int() if reduce_keepdims_attr is not None else 1
+        )
+        arg_keepdims = arg_keepdims_attr.as_int() if arg_keepdims_attr is not None else 1


titaiwangms · 2026-02-06T23:23:25Z

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

+
+        # ONNX default: keepdims = 1 for both Reduce and Arg operations
+        reduce_keepdims = (
+            reduce_keepdims_attr.as_int() if reduce_keepdims_attr is not None else 1


@justinchuby is there a more clean way to get the default? (just curious)

I think there should be a get_int method

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

titaiwangms · 2026-02-06T23:27:49Z

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

+            )
+
+        # Step 3: Get axes from Reduce operation
+        # In opset 18+, axes is an input; in opset 13-17, it's an attribute


I wonder if we would be interested in only supporting opset 18+ here to reduce the complexity? (we have version converter) It's just the matter whether we see the rule will be applied standalone or not I guess?

That makes sense to remove, I see this rule should be mostly used in pipeline, thanks for the suggestion!

@justinchuby What do you think?

Only opset 18+ is fine

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

titaiwangms · 2026-02-06T23:32:57Z

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

+
+        # Step 7: Normalize axes if rank is known (handle negative indices)
+        input_x = reduce_node.inputs[0]
+        rank = len(input_x.shape) if input_x.shape is not None else None


I wonder if symbolic shape could work on this case? @justinchuby

Can you elaborate?

Skipping none of shape means this does not support dynamic at the moment. But symbolic inference should be able to handle the eq

titaiwangms · 2026-02-06T23:34:08Z

You will have to enable it here:

onnxscript/onnxscript/rewriter/__init__.py

Line 52 in 0206a98

_DEFAULT_REWRITE_RULES: tuple[pattern.RewriteRule, ...] = (

justinchuby · 2026-02-07T22:56:23Z

You will have to enable it here:

onnxscript/onnxscript/rewriter/__init__.py

Line 52 in 0206a98

_DEFAULT_REWRITE_RULES: tuple[pattern.RewriteRule, ...] = (

I don’t think we want to enable this by default. It is unclear if this is generally more performant. @danielhumanmod you may simply expose the rule in https://github.com/microsoft/onnxscript/blob/main/onnxscript/rewriter/rules/common/__init__.py

justinchuby · 2026-02-09T17:33:21Z

onnxscript/rewriter/__init__.py

    *_fuse_pad_into_conv.rules,
    *_fuse_batchnorm.rules,
    *_remove_optional_bias.rules,
+    *_fuse_reduce_arg_to_topk.rules,


Suggested change

*_fuse_reduce_arg_to_topk.rules,

Do not expose as default unless it is measured to be performant

danielhumanmod added 2 commits January 24, 2026 22:22

optimize min/max dim with topk

d15e6ac

tests

39baa38

github-project-automation bot added this to ONNX Script Review Board Jan 25, 2026

github-project-automation bot moved this to Todo in ONNX Script Review Board Jan 25, 2026

justinchuby requested changes Jan 25, 2026

View reviewed changes

github-project-automation bot moved this from Todo to In Progress in ONNX Script Review Board Jan 25, 2026

rewrite rule to fuse ruduce + arg min/max with topk

312ef9e

danielhumanmod requested a review from justinchuby February 3, 2026 02:39

github-advanced-security bot found potential problems Feb 3, 2026

View reviewed changes

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py Fixed Show fixed Hide fixed

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py Fixed Show fixed Hide fixed

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py Fixed Show fixed Hide fixed

github-advanced-security bot found potential problems Feb 3, 2026

View reviewed changes

format

dde8d14

justinchuby assigned titaiwangms Feb 6, 2026

justinchuby added the module: rewriter label Feb 6, 2026

titaiwangms requested review from Copilot and titaiwangms February 6, 2026 20:22

Copilot started reviewing on behalf of titaiwangms February 6, 2026 20:22 View session

Copilot AI reviewed Feb 6, 2026

View reviewed changes

titaiwangms reviewed Feb 6, 2026

View reviewed changes

danielhumanmod added 2 commits February 6, 2026 22:55

solve comments - part 1

72fe1db

resolve comments from copilot

301635e

justinchuby reviewed Feb 9, 2026

View reviewed changes

		from onnxscript.rewriter._rewrite_rule import RewriteRuleClassBase, RewriteRuleSet


		class FuseReduceArgToTopKBase(RewriteRuleClassBase):

	class FuseReduceArgToTopKBase(RewriteRuleClassBase):
	class _FuseReduceArgToTopKBase(RewriteRuleClassBase):

Conversation

danielhumanmod commented Jan 25, 2026

Context

Verification

Uh oh!

danielhumanmod commented Jan 25, 2026

Uh oh!

codecov bot commented Jan 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

justinchuby left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

danielhumanmod commented Jan 26, 2026

Uh oh!

justinchuby commented Jan 26, 2026

Uh oh!

danielhumanmod commented Jan 27, 2026

Uh oh!

danielhumanmod commented Feb 3, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

titaiwangms commented Feb 6, 2026

Uh oh!

justinchuby commented Feb 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

codecov bot commented Jan 25, 2026 •

edited

Loading

justinchuby left a comment •

edited

Loading

justinchuby commented Feb 7, 2026 •

edited

Loading