-
Notifications
You must be signed in to change notification settings - Fork 169
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Composing autoquant with compile (#175)
* Composing autoquant with compile Summary: this PR rewrites how torchao.autoquant works so that it works with torch.compile. Previously you had to do: torchao.autoquant(model, input) mod=torch.compile(model) mod(input) now you can do torchao.autoquant(torch.compile(model)) model(input) The new method works with/without compile. Also this is BC so the old path also works. We use a forward_prehook to intercept the model call before torch.compile tracing occurs at which point we do the autoquantization and clean up all remaining hooks before passing things off to the normal torch.compile tracing functionality. note: in the case of multiple inputs, you can also do: model.forward_log_only(input) to run the model forward with autoquant shape logging and prevent the torch.compile tracing/autoquant quantization from occuring. Test Plan: python test/integration/test_integration.py -k "autoquant" Reviewers: Subscribers: Tasks: Tags: * Fused DoRA kernels (#216) * add dora kernels * allowing error_on_unseen in autoquant func Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * Unified AffineQuantizedTensor subclass (#214) Summary: Creatd a `AffineQuantizedTensor` subclass that works for both weight and input (for dynamic quantization), for all granularities (levering the recently added choose_qparams_affine, quantize_affine and dequantize_affine ops) only verified for 8da4w right now, we can make it work for other types of quantization (mostly the operator dispatching part) later Test Plan: python test/quantization/test_quant_api.py -k test_quantized_tensor_subclass_8da4w Reviewers: Subscribers: Tasks: Tags: Co-authored-by: Mark Saroufim <marksaroufim@meta.com> * add expecttest to requirements.txt (#225) * add expecttest to requirements.txt * update * Install dev-requirements.txt in doc build (#224) Install dev-requirements.txt --------- Co-authored-by: Mark Saroufim <marksaroufim@meta.com> * Fix an error in subclass impl (#226) Summary: Accidently changed the device check code for old subclass instead of the new one, forgot to fix before landing Test Plan: CI Reviewers: Subscribers: Tasks: Tags: * update readme.md Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * trying to fix the error in CI on cleanup hooks Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * correct docs Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * Some follow up fixes for quant primitives (#220) Summary: att Test Plan: python test/quantization/test_quant_primitives.py -k test_raises Reviewers: Subscribers: Tasks: Tags: * Composing autoquant with compile Summary: this PR rewrites how torchao.autoquant works so that it works with torch.compile. Previously you had to do: torchao.autoquant(model, input) mod=torch.compile(model) mod(input) now you can do torchao.autoquant(torch.compile(model)) model(input) The new method works with/without compile. Also this is BC so the old path also works. We use a forward_prehook to intercept the model call before torch.compile tracing occurs at which point we do the autoquantization and clean up all remaining hooks before passing things off to the normal torch.compile tracing functionality. note: in the case of multiple inputs, you can also do: model.forward_log_only(input) to run the model forward with autoquant shape logging and prevent the torch.compile tracing/autoquant quantization from occuring. Test Plan: python test/integration/test_integration.py -k "autoquant" Reviewers: Subscribers: Tasks: Tags: * allowing error_on_unseen in autoquant func Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * update readme.md Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * trying to fix the error in CI on cleanup hooks Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: * correct docs Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: --------- Co-authored-by: jeromeku <jerome.ku@gmail.com> Co-authored-by: Jerry Zhang <jerryzh168@gmail.com> Co-authored-by: Mark Saroufim <marksaroufim@meta.com> Co-authored-by: Svetlana Karslioglu <svekars@meta.com>
- Loading branch information
1 parent
63c5ac5
commit f6d56ca
Showing
5 changed files
with
133 additions
and
33 deletions.
There are no files selected for viewing
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
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
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
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
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