[AOTI] add C shim for QConvPointWise #138540
Closed
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
chunyuan-w
requested review from
digantdesai,
jerryzh168,
jianyuh,
kimishpatel and
salilsdesai
as code owners
pytorch-bot
bot
added
ciflow/inductor
module: cpu
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138540
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit 8177bed with merge base 07b0d63 ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
chunyuan-w
removed request for
digantdesai,
jerryzh168,
jianyuh,
kimishpatel and
salilsdesai
jgong5
approved these changes
Oct 24, 2024
desertfire
approved these changes
Oct 25, 2024
|
|
||
| // Conv2D with binary postop | ||
| m.def(TORCH_SELECTIVE_SCHEMA("onednn::qconv2d_pointwise.binary(Tensor qx, float x_scale, int x_zero_point, Tensor qaccum, float accum_scale, int accum_zero_point, Tensor qw, Tensor w_scale, Tensor w_zero_point, Tensor? bias, int[] stride, int[] padding, int[] dilation, int groups, float output_scale, int output_zero_point, ScalarType? output_dtype, str binary_attr, Scalar? alpha, str? unary_attr, Scalar?[] unary_scalars, str? unary_algorithm) -> Tensor")); | ||
| m.def(TORCH_SELECTIVE_SCHEMA("onednn::qconv2d_pointwise.binary(Tensor qx, float x_scale, int x_zero_point, Tensor qw, Tensor w_scale, Tensor w_zero_point, Tensor qaccum, Tensor? bias, int[] stride, int[] padding, int[] dilation, int groups, float output_scale, int output_zero_point, ScalarType? output_dtype, float accum_scale, int accum_zero_point, str binary_attr, Scalar? alpha, str? unary_attr, Scalar?[] unary_scalars, str? unary_algorithm) -> Tensor")); |
There was a problem hiding this comment.
Is this safe to directly change this one?
There was a problem hiding this comment.
I guess my question is, what is the BC policy for onednn ops?
There was a problem hiding this comment.
Another option is to add something like a v2 version of this OP to avoid the BC-breaking.
Regarding the qconv2d_pointwise, since the cpu quantization support in inductor is still a prototype feature, I guess it should be fine to make API change since it's not yet stable. May I know if the current way looks fine to you or it's more preferred to add a v2 version?
There was a problem hiding this comment.
Since this is an ondnn specific op, I am ok with your decision.
chunyuan-w
added a commit
to chunyuan-w/pytorch
that referenced
this pull request
Oct 30, 2024
ghstack-source-id: ae23a41 Pull Request resolved: pytorch#138540
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Nov 5, 2024
This PR adds C shim for `QConvPointWisePT2E` and `QConvPointWiseBinaryPT2E` similar to pytorch#138439. Besides that, we aligned the implementation of `qconv_pointwise` with `qlinear_pointwise` in the following aspects: 1. The parameter order of `qconv_pointwise` and `qlinear_pointwise` are quite different, we aligned the schema of `qconv_pointwise` to have similar parameter order as `qlinear_pointwise` to make it more consistent. 2. We always converted `x_scale` and `x_zero_point` to Tensors, just like in the lowering of `qlinear_pointwise`. This avoids the need to create two separate C APIs (one for `double x_scale` and `int64_t x_zero_point`, and another for `Tensor` versions). Instead, we only need one API for `Tensor`-based `x_scale` and `x_zero_point`. If we later add dynamic quantization for qconv (which will use `Tensor` for `x_scale` and `x_zero_point`), we can reuse the code from this PR and don't need to change the C shim layer API. Pull Request resolved: pytorch#138540 Approved by: https://github.com/jgong5, https://github.com/desertfire ghstack dependencies: pytorch#138691, pytorch#138806
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
Stack from ghstack (oldest at bottom):
len(serialized_weights)when calculatingconsts_size#139054This PR adds C shim for
QConvPointWisePT2EandQConvPointWiseBinaryPT2Esimilar to #138439. Besides that, we aligned the implementation ofqconv_pointwisewithqlinear_pointwisein the following aspects:qconv_pointwiseandqlinear_pointwiseare quite different, we aligned the schema ofqconv_pointwiseto have similar parameter order asqlinear_pointwiseto make it more consistent.x_scaleandx_zero_pointto Tensors, just like in the lowering ofqlinear_pointwise. This avoids the need to create two separate C APIs (one fordouble x_scaleandint64_t x_zero_point, and another forTensorversions). Instead, we only need one API forTensor-basedx_scaleandx_zero_point. If we later add dynamic quantization for qconv (which will useTensorforx_scaleandx_zero_point), we can reuse the code from this PR and don't need to change the C shim layer API.cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov