Refactor QModuleMixin and Calibration and fix stream-lining bug #249
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
This reveals an overflow in qfloat8_e5m2 activations test, that is removed for now.
Only QLinear might request its inputs to be always quantized, as it is the only layer for which an optimized kernel exists.
By putting the input/output quantization code inside module forward hooks, it allows them to be called only after the calibration hooks. This simplifies a lot the calibration code, in particular avoiding several calls to forward during output calibration.
dacorvo
force-pushed
the
refactor_qmodule
branch
from
6ff9769
to
59b9983
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Outdated suggestions cannot be applied.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.
What does this PR do?
This heavily refactors
QModuleMixinandCalibrationto simplify the quantization of inputs and outputs.This also fixes a bug in
streamlinemode: instead of disabling only the quantization of the outputs of a quantized module if they are immediately dequantized, it also disabled the quantization of inputs.With that fix, the inference of int8/int8 models can be up to 50% faster because of the accelerated
torch._int_mm.