A few fixes #22
Merged
A few fixes #22
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
The torch.Tensor implementation seems to deal with non-contiguous inputs just fine, but if we don't overload it for QTensor and let instead the chain of aten operations flow, it starts with a sequence of expand/view calls that are not compatible with non-contiguous inputs. It is better to overload matmul anyway, as we might be able to call directly an optimized kernel here.
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Outdated suggestions cannot be applied.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.
While debugging the OPT model, I found out that a few things were not working as expected when batching inputs.
For a start,
is_contiguouswas never called (and dispatched) because it must be overloaded directly in subclasses (would need to check how many methods are actually needed inTensorsubclasses).Then, I realized that the chain of
atenoperations when torch Function is disabled in theTensorsubclass was not necessarily equivalent to their "functional" counterparts.A good example is
torch.matmul: if I disable torch Function forQTensorand invoketorch.matmul, I end up with a sequence of operations that does not check if inputs are contiguous, which leads to failures when the modeling code is not bullet-proof.Finally, the calibration code was not strictly correct, as during calibration we were always using the "optimal" scale for inputs and outputs where it should be the "current" scale evaluated using a momentum.