Add support asymmetric fake-quantization to AQTv2. #675
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
copybara-service
bot
force-pushed
the
test_651580879
branch
4 times, most recently
from
0071d1a to
33de2e9
Compare
copybara-service
bot
force-pushed
the
test_651580879
branch
from
33de2e9 to
b47fe35
Compare
copybara-service
bot
force-pushed
the
test_651580879
branch
2 times, most recently
from
ef818dc to
cef287a
Compare
copybara-service
bot
force-pushed
the
test_651580879
branch
from
cef287a to
097249a
Compare
copybara-service
bot
force-pushed
the
test_651580879
branch
9 times, most recently
from
b066596 to
4694998
Compare
copybara-service
bot
force-pushed
the
test_651580879
branch
from
4694998 to
e4fa804
Compare
Integration of native quantization with biases will require computing the cross terms. See [#725](#725) Itemized changes: - Add `IntAsymmetric` to handle asymmetric integer numerics. - this class forgoes some of the more research-y parameters present on `IntSymmetric`. - Add `MinMaxCalibration` to calculate the scale and bias for asymmetric quantization. I additionally tested this change by training MNIST models using `flax_e2e_model`. With symmetric quantization the model fails to converge for `config.config_v4(fwd_bits=2, dlhs_bits=None, drhs_bits=None)` (due to `NaN` losses). With asymmetric quantization the model converges even with `config.config_v4(fwd_bits=2, dlhs_bits=2, drhs_bits=4)`. PiperOrigin-RevId: 651580879
copybara-service
bot
force-pushed
the
test_651580879
branch
from
e4fa804 to
ba94cf8
Compare
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Outdated suggestions cannot be applied.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.
Add support asymmetric fake-quantization to AQTv2.
Integration of native quantization with biases will require computing the cross terms. See #725
Itemized changes:
IntAsymmetricto handle asymmetric integer numerics.IntSymmetric.MinMaxCalibrationto calculate the scale and bias for asymmetric quantization.I additionally tested this change by training MNIST models using
flax_e2e_model. With symmetric quantization the model fails to converge forconfig.config_v4(fwd_bits=2, dlhs_bits=None, drhs_bits=None)(due toNaNlosses). With asymmetric quantization the model converges even withconfig.config_v4(fwd_bits=2, dlhs_bits=2, drhs_bits=4).