Broadcast operations in aten::masked_fill and element_wise operations #1654
Conversation
This file contains 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
|
@apbose we need tests to verify these changes. Also the TS tests are failing |
|
Yes @narendasan , I am looking at debugging the failing ts test |
gs-olive
reviewed
Sep 11, 2023
Comment on lines
157
to
163
| int diff = selfDim.size() - otherDim.size(); | ||
| bool canSqueeze = true; | ||
| for (int i = 0; i < diff; i++) { | ||
| if (selfDim[i] != 1) { | ||
| canSqueeze = false; | ||
| } | ||
| } |
There was a problem hiding this comment.
Does this work if diff < 0?
There was a problem hiding this comment.
Yes it probably should since the above condition ensures that the self has greater dimension than the other. But this is leading to dimensions getting mishandled in bert model, so I need to remove this logic
|
The above model fails in torchscript. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Description
This PR addresses the broadcasting issue encountered in the NeMo Citrinet model. (Bug 1573).
Fixes # (issue)
This PR makes two main changes-
The TensorRT broadcast rules are more limited as compared to Torch broadcast rules - #1629. Hence the aten::masked_fill is modified to make mask and self Tensors broadcastable according to TensorRT rules.At present only the mask tensor is padded if its no of dimensions is lesser than the self Tensor, but not vice versa.
Modification : The mask tensor is unpadded if its dimensions is greater than the self Tensor and it has leading 1 dimensions
The element_wise operations broadcasts the two tensors to get to the same dimensions.
Modification : The tensor having extra dimensions is checked if it requires the leading 1 dimensions and accordingly unpadded. Else the other tensor is padded.
The above model fails with the runtime error-RuntimeError: expand(CUDABoolType{[1, 1, 1, 1488]}, size=[1, 256, 1488]): the number of sizes provided (3) must be greater or equal to the number of dimensions in the tensor (4)
out dimension = 1, 256, 1488
mask dimension = 1,1,1,1488
Cases with output = out.masked_fill_(mask,2 ) fail while output = torch.masked_fill(mask,2) passes