Add pass for replacing dq-q patterns with rescale #8415
Merged
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/8415
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 1c39228 with merge base 8d96d74 ( 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
Erik-Lundell
added
partner: arm
facebook-github-bot
added
the
CLA Signed
When an int8 op meets an int32 op, the int8 is first dequantized to a float which is then quantized to the desired int8 dtype. This produces a (int8 dq -> q int32) pattern that we can replace with a TOSA.RESCALE since they are approximately mathematically equivalent, differing only in how the rounding is done. This requires a few changes: - Introduce custom rescale op - Create pass to replace the dq-q pattern with the rescale op - Implement node_visitor for rescale op The change makes it possible to mix int8 and int32 quantization, as showcased in the new test_add_i32_tosa_BI test. Signed-off-by: Erik Lundell <[email protected]> Change-Id: Ifd475ade488fbbeb8395ac883986b19f8edfae5a
Erik-Lundell
force-pushed
the
quantize_to_int32
branch
from
06d343f to
1c39228
Compare
zingo
approved these changes
Feb 13, 2025
dbort
reviewed
Feb 13, 2025
| @@ -9,13 +9,19 @@ | |||
| from typing import Tuple | |||
|
|
|||
| import torch | |||
| from executorch.backends.arm.arm_backend import get_tosa_version | |||
There was a problem hiding this comment.
This seems to have broken the unittest-arm job:
https://github.com/pytorch/executorch/actions/runs/13311971065/job/37176555490#step:15:11475
_______________________ ERROR collecting ops/test_add.py _______________________
ImportError while importing test module '/pytorch/executorch/backends/arm/test/ops/test_add.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
/opt/conda/envs/py_3.10/lib/python3.10/importlib/__init__.py:126: in import_module
return _bootstrap._gcd_import(name[level:], package, level)
backends/arm/test/ops/test_add.py:12: in <module>
from executorch.backends.arm.arm_backend import get_tosa_version
E ImportError: cannot import name 'get_tosa_version' from 'executorch.backends.arm.arm_backend' (/opt/conda/envs/py_3.10/lib/python3.10/site-packages/executorch/backends/arm/arm_backend.py)
There was a problem hiding this comment.
Yes i also think it did, sorry for the messup. I confirm with a revert to make sure it fixes it. If so lets merge the revert and fix/retry this PR again later.
dbort
pushed a commit
that referenced
this pull request
Feb 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.
When an int8 op meets an int32 op, the int8 is first dequantized to a float which is then quantized to the desired int8 dtype.
This produces a (int8 dq -> q int32) pattern that
we can replace with a TOSA.RESCALE since they are approximately mathematically equivalent, differing only in how the rounding is done.
This requires a few changes:
The change makes it possible to mix int8 and int32 quantization, as showcased in the new test_add_i32_tosa_BI test.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218