Conversation
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2477 +/- ##
==========================================
- Coverage 70.00% 69.90% -0.11%
==========================================
Files 215 215
Lines 25992 26035 +43
Branches 2606 2614 +8
==========================================
+ Hits 18196 18199 +3
- Misses 6896 6936 +40
Partials 900 900 ☔ View full report in Codecov by Sentry. |
titaiwangms
requested review from
Copilot,
gramalingam,
justinchuby and
titaiwangms
There was a problem hiding this comment.
Pull Request Overview
This PR implements the repeat_interleave operation for the torch library, adding support for both scalar and tensor variants. The implementation handles different cases where the repeat count is either an integer or a tensor, with optional dimension specification.
- Adds two new functions:
aten_repeat_interleave_intfor scalar repeats andaten_repeat_interleave_Tensorfor tensor repeats - Includes comprehensive end-to-end tests covering integer repeats, tensor repeats, and tensor repeats with no dimension specified
- Comments out the TorchLibOpInfo entries temporarily with a note about splitting into two cases
Reviewed Changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
| onnxscript/function_libs/torch_lib/ops/core.py | Implements the core repeat_interleave functionality with two variants for scalar and tensor inputs |
| tests/function_libs/torch_lib/e2e_ops_tests.py | Adds three comprehensive test cases covering different repeat_interleave scenarios |
| tests/function_libs/torch_lib/ops_test_data.py | Comments out TorchLibOpInfo entries with explanation about splitting into separate cases |
Signed-off-by: xadupre <xadupre@microsoft.com>
Signed-off-by: xadupre <xadupre@microsoft.com>
| torch.arange(4, dtype=torch.float32).reshape((2, 2)), | ||
| torch.tensor([1, 2, 3, 2], dtype=torch.int64), | ||
| ) | ||
| onnx_program = torch.onnx.export( |
Check warning
Code scanning / CodeQL
Variable defined multiple times Warning test
|
I think we can simplify the index computation logic like previous comments suggested. |
|
And then Expand can be leveraged to simplify the graph |
Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
| final_shape = op.Concat( | ||
| op.Shape(self, start=0, end=dim), | ||
| op.Constant(value_ints=[-1]), | ||
| op.Shape(self, start=dim + 1), |
There was a problem hiding this comment.
I suggest using pos_dim instead of dim ... otherwise, dim+1 can cause problems when dim == -1
There was a problem hiding this comment.
Would be good to test-cases for negative dim, including -1.
| if dim is None: | ||
| # flatten | ||
| self = op.Reshape(self, [-1]) | ||
| rk = 1 |
There was a problem hiding this comment.
| rk = 1 | |
| rank = 1 |
Similar to #2464. Does not support all the cases but we can add them in other PRs.