default transpose false, cuda and cpu ep tests

olive/passes/onnx/mnb_to_qdq.py

@@ -31,11 +31,12 @@ def _default_config(cls, accelerator_spec: AcceleratorSpec) -> Dict[str, PassCon
  "use_transpose_op": PassConfigParam(
  type_=bool,
  # TODO(jambayk): decide whether to enable this by default or not
- # False gives same output on arm Mac/Windows, but not on x64 Linux/Windows
- default_value=True,
+ # CPU-EP: False gives same output on arm Mac/Windows, but not on x64 Linux/Windows
+ default_value=False,
  description=(
  "Whether to use a Transpose operator after the DequantizeLinear operator. If False, the weight"
- " initializer will be transposed instead."
+ " initializer will be transposed instead. Default is False. True might be more efficient on some"
+ " EPs such as DirectML."
  ),
  ),
  **get_external_data_config(),
@@ -76,8 +77,9 @@ def _run_for_config(
  block_size = node_attributes["block_size"]
  num_k_blocks = math.ceil(K / block_size)
 
- # only deal with 4 bits for now
+ # only deal with 4 bits (int4) for now
  if node_attributes["bits"] != 4:
+ logger.debug("%s uses %d bits, only 4 bits is supported", node_name, node_attributes["bits"])
  continue
 
  # we can only deal with trivial g_idx, dequantize linear does not support g_idx

test/unit_test/passes/onnx/test_mnb_to_qdq.py

@@ -61,7 +61,12 @@ def forward(self, x):
  reason="Int4 DQ is only supported in ORT >= 1.20",
 )
 @pytest.mark.parametrize("use_transpose_op", [True, False])
-def test_mnb_to_qdq(create_mnb_model, use_transpose_op, tmp_path):
+@pytest.mark.parametrize("execution_provider", ["CPUExecutionProvider", "CUDAExecutionProvider"])
+def test_mnb_to_qdq(create_mnb_model, execution_provider, use_transpose_op, tmp_path):
+ available_providers = onnxruntime.get_available_providers()
+ if execution_provider not in available_providers:
+ pytest.skip(f"{execution_provider} is not available on this system {available_providers}")
+
  mnb_path, in_dim = create_mnb_model
  input_model = ONNXModelHandler(mnb_path)
 
@@ -73,14 +78,20 @@ def test_mnb_to_qdq(create_mnb_model, use_transpose_op, tmp_path):
  qdq_model: ONNXModelHandler = p.run(input_model, output_folder)
 
  # validate
- original_session = onnxruntime.InferenceSession(str(mnb_path))
- qdq_session = onnxruntime.InferenceSession(str(qdq_model.model_path))
+ original_session = onnxruntime.InferenceSession(str(mnb_path), providers=[execution_provider])
+ original_session.disable_fallback()
+ qdq_session = onnxruntime.InferenceSession(str(qdq_model.model_path), providers=[execution_provider])
+ qdq_session.disable_fallback()
 
  input_data = {"input": np.random.randn(1, 1, in_dim).astype(np.float32)}
  original_output = original_session.run(None, input_data)[0]
  qdq_output = qdq_session.run(None, input_data)[0]
  assert original_output.shape == qdq_output.shape
  assert original_output.dtype == qdq_output.dtype
- if use_transpose_op:
- # Pre transposed DQ model does not match the expected output on x64
+ if execution_provider == "CPUExecutionProvider" and not use_transpose_op:
+ # Pre transposed DQ model does not match the expected output on x64 CPU
+ # check for assertion failure so we know when the test is fixed
+ with pytest.raises(AssertionError):
+ np.testing.assert_allclose(original_output, qdq_output, rtol=1e-3, atol=1e-3)
+ else:
  np.testing.assert_allclose(original_output, qdq_output, rtol=1e-3, atol=1e-3)