Skip to content

Arm backend: Make per-channel quantization default for VgfPipeline #12705

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 4 commits into from
Jul 23, 2025
Merged

Arm backend: Make per-channel quantization default for VgfPipeline #12705

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3868638
Arm backend: Make per-channel quantization default for VgfPipeline
YufengShi-dudu Jul 17, 2025
81f8d6f
Merge branch 'main' into enable-per-channel-quantization-for-VgfPipeline
YufengShi-dudu Jul 22, 2025
b674933
Arm backend: Skip multihead_attention VGF tests in upstream
YufengShi-dudu Jul 23, 2025
42e348e
Merge branch 'main' into enable-per-channel-quantization-for-VgfPipeline
YufengShi-dudu Jul 23, 2025
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
37 changes: 37 additions & 0 deletions backends/arm/test/ops/test_multihead_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@
EthosU85PipelineBI,
TosaPipelineBI,
TosaPipelineMI,
VgfPipeline,
)


Expand Down Expand Up @@ -105,3 +106,39 @@ def test_multihead_attention_u85_BI(test_data: input_t1):
per_channel_quantization=False,
)
pipeline.run()


@common.parametrize(
"test_data",
test_suite,
)
@common.SkipIfNoModelConverter
def test_multihead_attention_vgf_FP(test_data: input_t1):
test_data_vals, module = test_data()
pipeline = VgfPipeline[input_t1](
module,
(*test_data_vals, *test_data_vals, *test_data_vals),
[],
[],
tosa_version="TOSA-1.0+FP",
)
pipeline.run()


@common.parametrize(
"test_data",
test_suite,
)
@common.SkipIfNoModelConverter
def test_multihead_attention_vgf_INT(test_data: input_t1):
test_data_vals, module = test_data()
pipeline = VgfPipeline[input_t1](
module,
(*test_data_vals, *test_data_vals, *test_data_vals),
[],
[],
tosa_version="TOSA-1.0+INT",
# TODO: Per-channel quantization is broken (MLETORCH-1144)
per_channel_quantization=False,
)
pipeline.run()
12 changes: 2 additions & 10 deletions backends/arm/test/tester/test_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -854,7 +854,7 @@ def __init__(
vgf_compiler_flags: Optional[str] = "",
tosa_version: str = "TOSA-1.0+FP",
symmetric_io_quantization: bool = False,
per_channel_quantization: bool = False,
per_channel_quantization: bool = True,
use_to_edge_transform_and_lower: bool = True,
custom_path: str = None,
atol: float = 1e-03,
Expand All @@ -866,11 +866,6 @@ def __init__(
] = None,
):

if (
symmetric_io_quantization or per_channel_quantization
) and tosa_version == "TOSA-1.0+FP":
raise ValueError("Dont configure quantization with FP TOSA profile.")

tosa_profile = TosaSpecification.create_from_string(tosa_version)
compile_spec = common.get_vgf_compile_spec(
tosa_profile, compiler_flags=vgf_compiler_flags, custom_path=custom_path
Expand All @@ -887,18 +882,15 @@ def __init__(
transform_passes=transform_passes,
)

if symmetric_io_quantization or per_channel_quantization:
if "INT" in tosa_version:
quantizer = VgfQuantizer(compile_spec)
quantization_config = get_symmetric_quantization_config(
is_per_channel=per_channel_quantization
)
if symmetric_io_quantization:
quantizer.set_io(quantization_config)
quant_stage = Quantize(quantizer, quantization_config)
else:
quant_stage = None

if "INT" in tosa_version:
self.add_stage(self.tester.quantize, quant_stage, pos=0)

self.add_stage_after(
Expand Down
Loading