Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Utilize to_numpy=True in quantize if available #1725

Merged
merged 1 commit into from
Jul 31, 2024
Merged

Utilize to_numpy=True in quantize if available #1725

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 14 additions & 5 deletions keras_nlp/src/layers/modeling/reversible_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,6 +253,7 @@ def _int8_call(self, inputs, reverse=False):

def quantize(self, mode, type_check=True):
import gc
import inspect

assert_quantization_support()
if type_check and type(self) is not ReversibleEmbedding:
Expand All @@ -262,19 +263,27 @@ def quantize(self, mode, type_check=True):
)
self._check_quantize_args(mode, self.compute_dtype)

def abs_max_quantize(inputs, axis):
sig = inspect.signature(keras.quantizers.abs_max_quantize)
if "to_numpy" in sig.parameters:
return keras.quantizers.abs_max_quantize(
inputs, axis=axis, to_numpy=True
)
else:
# `keras<=3.4.1` doesn't support `to_numpy`
return keras.quantizers.abs_max_quantize(inputs, axis=axis)

self._tracker.unlock()
if mode == "int8":
embeddings, embeddings_scale = keras.quantizers.abs_max_quantize(
embeddings, embeddings_scale = abs_max_quantize(
self._embeddings, axis=-1
)
embeddings_scale = ops.squeeze(embeddings_scale, axis=-1)
self._untrack_variable(self._embeddings)
del self._embeddings
if not self.tie_weights:
reverse_embeddings, reverse_embeddings_scale = (
keras.quantizers.abs_max_quantize(
self.reverse_embeddings, axis=0
)
reverse_embeddings, reverse_embeddings_scale = abs_max_quantize(
self.reverse_embeddings, axis=0
)
reverse_embeddings_scale = ops.squeeze(
reverse_embeddings_scale, axis=0
Expand Down
Loading