Improves implementation of aten_index_put #2641
Conversation
… xadupre/input_put
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2641 +/- ##
==========================================
- Coverage 70.46% 70.42% -0.04%
==========================================
Files 224 224
Lines 26572 26678 +106
Branches 2637 2658 +21
==========================================
+ Hits 18723 18789 +66
- Misses 6928 6956 +28
- Partials 921 933 +12 ☔ View full report in Codecov by Sentry. |
| isinstance(index, torch.onnx._internal.exporter._tensors.SymbolicTensor) # pylint: disable=protected-access | ||
| for index in indices | ||
| ) | ||
| and len(values.shape) == 1 |
There was a problem hiding this comment.
What is this condition for? I am just trying to understand the assumptions/conditions for this special case.
There was a problem hiding this comment.
ONNX does not have similar operator than index_put. So the goal was to convert different cases in different ways. Finding a generic way to convert all the cases may end up with inefficient conversions. So I chose to filter out some cases where I know a simple conversion. At the end of the function, you still have an implementation working for everything left. About this condition, I don't know when SymbolicTensor appears, I just know it appears in some models (Qwen) and in that case, the conversion I implemented works. If the generic case is difficult to implement, let's go through the list of all cases we face. That's the logic I followed.
|
Adding some pointers/info for my own clarification: |
From #2606.