Move some experimental tests

.github/workflows/torchao_experimental_test.yml → .github/workflows/regression_test_aarch64.yml

@@ -1,4 +1,4 @@
-name: Run TorchAO Experimental Tests
+name: Run Regression Tests (aarch64)
 
 on:
   push:
@@ -44,17 +44,19 @@ jobs:
         if: runner.os == 'Linux'
         run: |
           conda activate venv
+          pip install coremltools
           pip install torch==2.7.0 --index-url https://download.pytorch.org/whl/cpu --force-reinstall
           pip install -r dev-requirements.txt
           BUILD_TORCHAO_EXPERIMENTAL=1 TORCHAO_BUILD_CPU_AARCH64=1 TORCHAO_BUILD_KLEIDIAI=1 TORCHAO_ENABLE_ARM_NEON_DOT=1 TORCHAO_PARALLEL_BACKEND=OPENMP pip install .
       - name: Run python tests
         run: |
           conda activate venv
-          pytest -s test/quantization/test_int8_dynamic_activation_intx_weight_config_v1.py
-          pytest -s torchao/experimental/tests/test_embedding_xbit_quantizer.py
           pytest -s torchao/experimental/tests/test_quant_passes.py
-          pytest -s test/prototype/test_dynamic_activation_lut.py
+          pytest -s test/quantization/test_int8_dynamic_activation_intx_weight_config_v1.py
+          pytest -s test/quantization/test_embedding_xbit_quantizer.py
           pytest -s test/quantization/quantize_/workflows/intx/test_intx_opaque_tensor.py
+          pytest -s test/prototype/test_dynamic_activation_lut.py
+          pytest -s test/prototype/test_groupwise_lowbit_weight_lut_quantizer.py
       - name: torchao/csrc/cpu - build and run C++ tests
         if: runner.os == 'macOS'
         run: |

torchao/experimental/tests/test_groupwise_lowbit_weight_lut_quantizer.py → test/prototype/test_groupwise_lowbit_weight_lut_quantizer.py

@@ -20,8 +20,12 @@
     group_size_to_block_shapes,
 )
 from torchao.quantization.quant_api import quantize_
+from torchao.quantization.quantize_.workflows.intx.intx_opaque_tensor import (
+    _is_kernel_library_loaded,
+)
 
 
+@unittest.skipIf(not _is_kernel_library_loaded(), "Need torchao lowbit kernels")
 class TestGroupwiseLowbitWeightLut(unittest.TestCase):
     """
     Test suite for the GroupwiseLutWeight quantization scheme, updated for the

torchao/experimental/tests/test_embedding_xbit_quantizer.py → test/quantization/test_embedding_xbit_quantizer.py

@@ -32,9 +32,13 @@
     MappingType,
     quantize_,
 )
+from torchao.quantization.quantize_.workflows.intx.intx_opaque_tensor import (
+    _is_kernel_library_loaded,
+)
 from torchao.quantization.utils import compute_error
 
 
+@unittest.skipIf(not _is_kernel_library_loaded(), "Need torchao lowbit kernels")
 class TestEmbeddingQuantizer(unittest.TestCase):
     def test_accuracy(self):
         granularity = PerGroup(128)