Skip to content

Enable 16-bit activations in Cadence Quantizer For fully_connected and linear #15010

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

Merged
merged 1 commit into from
Oct 14, 2025
Merged

Enable 16-bit activations in Cadence Quantizer For fully_connected and linear #15010

Show file tree
Hide file tree
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
13 changes: 13 additions & 0 deletions backends/cadence/aot/quantizer/quantizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -342,3 +342,16 @@ def __init__(self, quantizers: Optional[list[Quantizer]] = None) -> None:
quantizers = get_cadence_default_quantizers()
quantizers.append(CadenceAtenQuantizer(SoftmaxPattern(), qconfig_A16))
super().__init__(quantizers)


class CadenceWith16BitLinearActivationsQuantizer(CadenceQuantizer):
"""
Quantizer including A16 fully_connected
"""

def __init__(self, quantizers: Optional[list[Quantizer]] = None) -> None:
if quantizers is None:
quantizers = []
# Add 16-bit quantizers for LinearPattern
quantizers.append(CadenceAtenQuantizer(LinearPattern(), qconfig_A16))
super().__init__(quantizers)
2 changes: 1 addition & 1 deletion backends/cadence/aot/ref_implementations.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,7 +261,7 @@ def quantized_linear_common(
src = src.view(-1, K)

dtype = src.dtype
supported_dtypes = [torch.int8, torch.uint8, torch.int32]
supported_dtypes = [torch.int8, torch.uint8, torch.int16, torch.int32]
if dtype not in supported_dtypes:
raise ValueError(
f"Unsupported dtype to quantize to {dtype}. Supported dtypes must be one of {supported_dtypes}"
Expand Down
22 changes: 19 additions & 3 deletions backends/cadence/aot/tests/test_ref_implementations.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,6 +183,8 @@ def test_quantized_add(
(False, torch.int8),
(True, torch.int8),
(True, torch.uint8),
(True, torch.int16),
(False, torch.int16),
)
],
# Test case 2: 1x3 input, 2x3 weight (2 output features)
Expand All @@ -207,6 +209,8 @@ def test_quantized_add(
for (per_tensor, dtype) in (
(False, torch.int8),
(True, torch.int8),
(False, torch.int16),
(True, torch.int16),
)
],
*[
Expand Down Expand Up @@ -256,6 +260,8 @@ def test_quantized_add(
for (per_tensor, dtype) in (
(False, torch.int8),
(True, torch.int8),
(False, torch.int16),
(True, torch.int16),
)
],
# Test case 4: Non-zero zero points
Expand All @@ -280,6 +286,8 @@ def test_quantized_add(
for (per_tensor, dtype) in (
(False, torch.int8),
(True, torch.int8),
(False, torch.int16),
(True, torch.int16),
# (True, torch.uint8),
)
],
Expand All @@ -302,7 +310,10 @@ def test_quantized_add(
False,
False,
)
for dtype in (torch.int8,)
for dtype in (
torch.int8,
torch.int16,
)
],
# Test case 6: Non-zero out_shift (shift=1)
*[
Expand All @@ -325,7 +336,12 @@ def test_quantized_add(
False,
False,
)
for (per_tensor, dtype) in ((False, torch.int8), (True, torch.int8))
for (per_tensor, dtype) in (
(False, torch.int8),
(True, torch.int8),
(False, torch.int16),
(True, torch.int16),
)
],
*[
(
Expand All @@ -348,7 +364,7 @@ def test_quantized_add(
transposed_matmul,
)
for (matmul, transposed_matmul) in ((True, False), (True, True))
for (per_tensor, dtype) in ((True, torch.int8),)
for (per_tensor, dtype) in ((True, torch.int8), (True, torch.int16))
],
*[
(
Expand Down
Loading