[ET] enabling half dtype output for dequantization and making logic consistent #11669
Conversation
This file contains hidden or 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
…onsistent Pull Request resolved: #11552 # Context Currently the cpu implementation for the dequantization operator (which includes `dequantize_per_token`, `dequantize_per_tensor`, and `dequantize_per_channel`), does not inherently support half (fp16) input scalar types. In order to align with the PyTorch implementation that accepts fp16 and bfp16 inputs, this diff aims to enable half input dtype support for the quantization operators. We will be comparing this implementation against the vulkan operators. Furthermore, there is a casting bug when applying the zero_point, as only in the `dequantize_per_tensor` implementation does it cast the zero_point to int32, while comparitively for `dequantize_per_channel` and `dequantize_per_token` they do not cast the zero_point. In an environment that only supports 32bit integers, understandbly there will be some inconsistencies in dequantization logic as per_tensor will contain different overflow logic compared to its respective per_token and per_channel partner since the latter eliminates the overflow by utilizing a 64bit value. # Changes As defined in ExecuTorch [scalar_type_util.h](https://github.com/pytorch/executorch/blob/053686242c1687f0d51b3bb8befd14b047d7b025/runtime/core/exec_aten/util/scalar_type_util.h), the changes in this diff include adding a new macro `ET_FORALL_FLOATH_TYPES_WITH` to `util/scalar_type_util.h`, updating the `CALCULATE_INT_TYPE` macro to handle the new dtype. This enables support for Half (fp16), Float (fp32), and Double (fp64). I have also included more comprehensive testing against the input dtypes, including adding double testing since it didn't already exist before. Instead of just confirming that all the output dtypes are supported, we also have a check that all input dtypes are supported now as well. In order to align both dequantization implementations, we cast the zero_point to 32bit for both to maintain the overflow logic carried over from `dequantize_per_tensor`. ghstack-source-id: 290376483 @exported-using-ghexport Differential Revision: [D76289181](https://our.internmc.facebook.com/intern/diff/D76289181/)
pytorchbot
requested review from
JacobSzwejbka,
lucylq,
swolchok and
manuelcandales
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11669
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit a6ffb06 with merge base 56392aa ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
ahmtox
approved these changes
Jun 14, 2025
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #11552 by @ahmtox
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/17/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/17/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/ahmtox/12/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/ahmtox/17/orig
@diff-train-skip-merge