Feat (proxy): scale computation delegated to bias proxy #938
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
costigt-dev
reviewed
Apr 19, 2024
| input_scale: Optional[Tensor] = None) -> Union[Tensor, QuantTensor]: | ||
| def quant_output_scale_impl( | ||
| self, input: QuantTensor, weight: QuantTensor, module: torch.nn.Module) -> Tensor: | ||
| channel_dim = -1 if isinstance(module, torch.nn.Linear) else 1 |
There was a problem hiding this comment.
how come its -1 if linear and 1 otherwise?
There was a problem hiding this comment.
https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
For linear layer, the channel dimension is always the last one, otherwise for Conv, ConvTranpose etc. is always at dim 1, at least for the input tensor.
costigt-dev
approved these changes
Apr 19, 2024
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Outdated suggestions cannot be applied.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.
As part of decoupling QuantLayers and QuantTensor, the computation of output scale for bias quantization is now delegated to the underlying proxy, who takes as input the possibly quantized input and weights, and internally defines the output scale.
Compared to our current setup, the main change would happen in the case where:
Our current setup would have no issue with this case, while after this PR, the user would be required to use internally scaled bias and not externally defined.
I believe this to be a edge case that can be safely ignored.