[torchlib] Add missing dtype parameter to aten_mean_dim - #2885
Merged
Justin Chu (justinchuby) merged 2 commits intoApr 10, 2026
Conversation
The ATen schema for mean.dim documents dtype as an optional parameter, but aten_mean_dim and aten_mean_dim_complex did not accept it. This causes a TypeError when PyTorch lowers mean.dim with an explicit dtype (e.g. from GlobalAveragePooling2D in Keras). Add dtype: int = -1 to both functions, following the same pattern used by aten_sum_dim_IntList. Fixes microsoft#2884
Contributor
Author
|
Hey! This is my first contribution to onnxscript. We ran into this while exporting Keras models to ONNX at work - GlobalAveragePooling2D lowers through aten::mean.dim with an explicit dtype, which hits the missing parameter. Happy to adjust anything if needed😄 |
Justin Chu (justinchuby)
approved these changes
Apr 10, 2026
Justin Chu (justinchuby)
left a comment
Justin Chu (justinchuby)
left a comment
Collaborator
There was a problem hiding this comment.
Thanks!
Justin Chu (justinchuby)
added
the
module: torchlib
Justin Chu (justinchuby)
enabled auto-merge (squash)
April 10, 2026 15:16
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2885 +/- ##
==========================================
- Coverage 72.04% 72.03% -0.01%
==========================================
Files 239 239
Lines 29305 29309 +4
Branches 2880 2882 +2
==========================================
Hits 21112 21112
- Misses 7216 7218 +2
- Partials 977 979 +2 ☔ View full report in Codecov by Sentry. |
Justin Chu (justinchuby)
disabled auto-merge
April 10, 2026 15:39
Collaborator
|
Linus Juni (@linusjuni) could you fix lint? |
Contributor
Author
|
@microsoft-github-policy-service agree |
Justin Chu (justinchuby)
enabled auto-merge (squash)
April 10, 2026 15:55
Justin Chu (justinchuby)
approved these changes
Apr 10, 2026
Justin Chu (justinchuby)
merged commit Apr 10, 2026
12234f8
into
microsoft:main
28 of 32 checks passed
Justin Chu (justinchuby)
pushed a commit
that referenced
this pull request
Apr 17, 2026
Fixes #2884 `aten_mean_dim` and `aten_mean_dim_complex` are missing the `dtype` keyword argument from their signatures, even though the ATen schema documents it (`ScalarType? dtype=None`). This causes a `TypeError` when PyTorch lowers `aten::mean.dim` with an explicit `dtype` - which happens for any model using `GlobalAveragePooling2D` (Keras/PyTorch). - Add `dtype: int = -1` to `aten_mean_dim`, with `op.Cast` when dtype is specified - Add `dtype: int = -1` to `aten_mean_dim_complex`, raising `NotImplementedError` for complex tensors Follows the same pattern used by `aten_sum_dim_IntList` and `aten_sum_dim_IntList_complex`.
Justin Chu (justinchuby)
pushed a commit
that referenced
this pull request
Aug 24, 2026
) # Honor the dtype argument of aten::mean when exporting without dim Fixes #3008 ## Summary The no dim overload of `aten::mean` declares `dtype` in its schema (`mean(Tensor self, *, ScalarType? dtype=None)`), but `aten_mean` in `onnxscript/function_libs/torch_lib/ops/core.py` did not accept the argument. Exporting `torch.mean(x, dtype=torch.float64)` for a float32 input succeeded silently, emitted only `ReduceMean -> Squeeze`, declared a FLOAT output, and accumulated in float32. For the input `[[1e8, 1.0, -1e8]]` PyTorch returns `0.3333333333333333` as float64 while the exported model returned `0.0` as float32. ## Changes - `aten_mean` is now `trace_only=True`, takes `dtype: int = -1`, and when a dtype is given casts `self` before `ReduceMean`. The no dtype path is unchanged (`ReduceMean -> Squeeze`). - `aten_mean_complex` takes the same argument and raises `NotImplementedError` when it is supplied, matching `aten_mean_dim_complex` and `aten_sum_complex`. - New `ops.aten.mean.dtype` OpInfo in `tests/function_libs/torch_lib/extra_opinfo.py` (`sample_inputs_mean_dtype`) registered against `core_ops.aten_mean` in `ops_test_data.py`. It yields `make_tensor` samples of shapes `(5, 5)`, `(5,)` and `()` plus the precision sensitive tensor `[[1e8, 1.0, -1e8]]`, all with `dtype=torch.float64`. ## Why the cast happens before the reduction `aten_mean_dim` (#2885) and `aten_sum` cast the reduced result after the reduction. That is not sufficient here: PyTorch accumulates in the requested dtype, so for `[1e8, 1.0, -1e8]` the float32 mean is `0.0` (the `1.0` is lost when added to `1e8`) while the float64 mean is `1/3`. Casting after `ReduceMean` would produce a DOUBLE tensor holding `0.0`, which still mismatches PyTorch. Casting the input first reproduces PyTorch semantics. The new test includes this sample specifically so that a cast after reduction implementation cannot pass by accident. ## Verification Reporter's script (`torch.onnx.export(..., dynamo=True)` of `torch.mean(x, dtype=torch.float64)` with float32 input) on pristine main at a39c0a5: ``` torch eager result: 0.3333333333333333 dtype: torch.float64 ONNX output elem_type: 1 (FLOAT), expected 11 (DOUBLE) ORT result: 0.0 dtype: float32 ``` With this change: ``` torch eager result: 0.3333333333333333 dtype: torch.float64 ONNX output elem_type: 11 (DOUBLE) ORT result: 0.3333333333333333 dtype: float64 ``` `pytest tests/function_libs/torch_lib/ops_test.py -k mean`: | State | Result | | --- | --- | | Pristine main, without the new test | 8 passed, 48 skipped, 8 xfailed, 60 subtests passed | | New test with the source change stashed | 4 failed, 10 passed, 48 skipped, 8 xfailed, 60 subtests passed. All four `ops_aten_mean_dtype` samples fail with `TypeInferenceError: Inferred elem type differs from existing elem type: (1) vs (11)` | | New test with a cast placed after the reduction | 1 failed, 8 passed, 50 skipped, 8 xfailed, 63 subtests passed. Only the `[[1e8, 1.0, -1e8]]` sample fails: `Expected 0.3333333333333333 but got 0.0` | | New test with this change | 8 passed, 50 skipped, 8 xfailed, 64 subtests passed | The two additional skips in the fixed state are the function proto validity checks, which skip for traced functions. `ruff check` and `ruff format --check` (ruff 0.15.1, the lintrunner pinned version) pass on the three changed files. Environment: Python 3.10, torch 2.13.0 (CPU), onnx 1.22.0, onnxruntime 1.23.2, macOS arm64.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Fixes #2884
aten_mean_dimandaten_mean_dim_complexare missing thedtypekeyword argument from their signatures, even though the ATen schema documents it (ScalarType? dtype=None). This causes aTypeErrorwhen PyTorch lowersaten::mean.dimwith an explicitdtype- which happens for any model usingGlobalAveragePooling2D(Keras/PyTorch).dtype: int = -1toaten_mean_dim, withop.Castwhen dtype is specifieddtype: int = -1toaten_mean_dim_complex, raisingNotImplementedErrorfor complex tensorsFollows the same pattern used by
aten_sum_dim_IntListandaten_sum_dim_IntList_complex.