From 0c1b2b773e6bb1014b5a570c9dff60847265c164 Mon Sep 17 00:00:00 2001 From: Christine Long Date: Sun, 17 May 2026 09:06:06 -0700 Subject: [PATCH 1/3] Add a16w8 per-op test for var (#19596) Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.var` on Ethos-U55 and Ethos-U85. ## Changes - Add `test_var_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True`, reusing existing `Var.test_parameters` (4 test configurations) - Add `test_var_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_var.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532362 --- backends/arm/test/ops/test_var.py | 30 ++++++++++++++++++++++++++++++ backends/arm/test/targets.bzl | 1 + 2 files changed, 31 insertions(+) diff --git a/backends/arm/test/ops/test_var.py b/backends/arm/test/ops/test_var.py index a7943bfc19b..c65f4991ca3 100644 --- a/backends/arm/test/ops/test_var.py +++ b/backends/arm/test/ops/test_var.py @@ -224,6 +224,36 @@ def test_var_dim_vgf_quant_no_dim(test_data: Tuple): pipeline.run() +@common.parametrize("test_data", Var.test_parameters) +@common.XfailIfNoCorstone300 +def test_var_a16w8_u55_INT(test_data: Tuple): + test_data, keepdim, correction = test_data() + pipeline = EthosU55PipelineINT[input_t1]( + Var(keepdim, correction), + (test_data,), + aten_ops=[], + exir_ops=[], + a16w8_quantization=True, + symmetric_io_quantization=True, + ) + pipeline.run() + + +@common.parametrize("test_data", Var.test_parameters) +@common.XfailIfNoCorstone320 +def test_var_a16w8_u85_INT(test_data: Tuple): + test_data, keepdim, correction = test_data() + pipeline = EthosU85PipelineINT[input_t1]( + Var(keepdim, correction), + (test_data,), + aten_ops=[], + exir_ops=[], + a16w8_quantization=True, + symmetric_io_quantization=True, + ) + pipeline.run() + + ############# ## VarDim ### ############# diff --git a/backends/arm/test/targets.bzl b/backends/arm/test/targets.bzl index b8030ae7ba8..30fa348414f 100644 --- a/backends/arm/test/targets.bzl +++ b/backends/arm/test/targets.bzl @@ -39,6 +39,7 @@ def define_arm_tests(): "ops/test_exp.py", "ops/test_reciprocal.py", "ops/test_mean_dim.py", + "ops/test_var.py", ] # Quantization From 64cd4d2986ac191e8fe7c81cc7826ab56ed56a91 Mon Sep 17 00:00:00 2001 From: Christine Long Date: Sun, 17 May 2026 09:06:06 -0700 Subject: [PATCH 2/3] Add a16w8 per-op test for conv1d Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.conv1d` on Ethos-U55 and Ethos-U85. ## Changes - Add `test_conv1d_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True`, reusing existing `test_data_INT` parameters - Add `test_conv1d_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_conv1d.py` in `fbcode/` and `xplat/` `targets.bzl` Reviewed By: Ninja91 Differential Revision: D104532360 --- backends/arm/test/ops/test_conv1d.py | 31 ++++++++++++++++++++++++++++ backends/arm/test/targets.bzl | 1 + 2 files changed, 32 insertions(+) diff --git a/backends/arm/test/ops/test_conv1d.py b/backends/arm/test/ops/test_conv1d.py index 486b6b3ce7c..0e75eab621f 100644 --- a/backends/arm/test/ops/test_conv1d.py +++ b/backends/arm/test/ops/test_conv1d.py @@ -399,3 +399,34 @@ def test_convolution_1d_vgf_quant_a8w4(test_data): get_symmetric_a8w4_quantization_config(is_per_channel=per_channel_quantization) ) pipeline.run() + + +@common.parametrize("test_data", test_data_INT) +@common.XfailIfNoCorstone300 +def test_conv1d_a16w8_u55_INT(test_data): + model, per_channel_quantization = test_data() + pipeline = EthosU55PipelineINT[input_t]( + model, + model.get_inputs(), + aten_op, + exir_op, + a16w8_quantization=True, + symmetric_io_quantization=True, + per_channel_quantization=per_channel_quantization, + ) + pipeline.run() + +@common.parametrize("test_data", test_data_INT) +@common.XfailIfNoCorstone320 +def test_conv1d_a16w8_u85_INT(test_data): + model, per_channel_quantization = test_data() + pipeline = EthosU85PipelineINT[input_t]( + model, + model.get_inputs(), + aten_op, + exir_op, + a16w8_quantization=True, + symmetric_io_quantization=True, + per_channel_quantization=per_channel_quantization, + ) + pipeline.run() diff --git a/backends/arm/test/targets.bzl b/backends/arm/test/targets.bzl index 30fa348414f..1b2dfa8e2c6 100644 --- a/backends/arm/test/targets.bzl +++ b/backends/arm/test/targets.bzl @@ -40,6 +40,7 @@ def define_arm_tests(): "ops/test_reciprocal.py", "ops/test_mean_dim.py", "ops/test_var.py", + "ops/test_conv1d.py", ] # Quantization From 9781827632a0f3d8c77c60ad6281bb9431ae05f5 Mon Sep 17 00:00:00 2001 From: Christine Long Date: Sun, 17 May 2026 09:06:06 -0700 Subject: [PATCH 3/3] Add a16w8 per-op test for gelu Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.gelu` on Ethos-U55 and Ethos-U85. ## Changes - Add `test_gelu_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16`, reusing existing `Gelu.test_data` parameters (12 test configurations covering both \`none\` and \`tanh\` approximation modes) - Add `test_gelu_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_gelu.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Reviewed By: Ninja91 Differential Revision: D104532359 --- backends/arm/test/ops/test_gelu.py | 35 ++++++++++++++++++++++++++++++ backends/arm/test/targets.bzl | 1 + 2 files changed, 36 insertions(+) diff --git a/backends/arm/test/ops/test_gelu.py b/backends/arm/test/ops/test_gelu.py index 82e3cb1e53e..84626c70b4d 100644 --- a/backends/arm/test/ops/test_gelu.py +++ b/backends/arm/test/ops/test_gelu.py @@ -6,6 +6,7 @@ from typing import Tuple import torch + from executorch.backends.arm.test import common from executorch.backends.arm.test.tester.test_pipeline import ( EthosU55PipelineINT, @@ -176,3 +177,37 @@ def test_gelu_vgf_quant(test_data: input_t1): quantize=True, ) pipeline.run() + + +@common.parametrize("test_data", Gelu.test_data) +@common.XfailIfNoCorstone300 +def test_gelu_a16w8_u55_INT(test_data: input_t1): + approximate, data = test_data() + pipeline = EthosU55PipelineINT[input_t1]( + Gelu(approximate), + (data,), + Gelu.aten_op, + Gelu.exir_op, + a16w8_quantization=True, + symmetric_io_quantization=True, + qtol=128, + epsilon=2**-16, + ) + pipeline.run() + + +@common.parametrize("test_data", Gelu.test_data) +@common.XfailIfNoCorstone320 +def test_gelu_a16w8_u85_INT(test_data: input_t1): + approximate, data = test_data() + pipeline = EthosU85PipelineINT[input_t1]( + Gelu(approximate), + (data,), + Gelu.aten_op, + Gelu.exir_op, + a16w8_quantization=True, + symmetric_io_quantization=True, + qtol=128, + epsilon=2**-16, + ) + pipeline.run() diff --git a/backends/arm/test/targets.bzl b/backends/arm/test/targets.bzl index 1b2dfa8e2c6..16a554da3b3 100644 --- a/backends/arm/test/targets.bzl +++ b/backends/arm/test/targets.bzl @@ -41,6 +41,7 @@ def define_arm_tests(): "ops/test_mean_dim.py", "ops/test_var.py", "ops/test_conv1d.py", + "ops/test_gelu.py", ] # Quantization