Test

.ci/docker/ci_commit_pins/nightly.txt

@@ -1 +1 @@
-dev20231002
+dev20231005

backends/xnnpack/operators/node_visitor.py

@@ -423,7 +423,7 @@ def define_nodes_tensor_inputs_outputs(
                     inp,
                     xnn_graph,
                     vals_to_ids,
-                    quant_params=QuantParams.from_inputs(inp),
+                    quant_params=QuantParams.from_inputs(inp, self._exported_program),
                     convert_to_nhwc=convert_to_nhwc,
                 )
         else:
@@ -434,7 +434,9 @@ def define_nodes_tensor_inputs_outputs(
             )
             # Define Input Node
             input_node = get_input_node(node, input_type_map.node_input)
-            input_quant_params = QuantParams.from_inputs(input_node)
+            input_quant_params = QuantParams.from_inputs(
+                input_node, self._exported_program
+            )
             self.define_tensor(
                 input_node,
                 xnn_graph,

backends/xnnpack/operators/op_add.py

@@ -43,7 +43,7 @@ def define_node(
             input1,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input1),
+            quant_params=QuantParams.from_inputs(input1, self._exported_program),
         )
         input1_id = vals_to_ids[input1]
 
@@ -53,7 +53,7 @@ def define_node(
             input2,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input2),
+            quant_params=QuantParams.from_inputs(input2, self._exported_program),
         )
         input2_id = vals_to_ids[input2]
 

backends/xnnpack/operators/op_cat.py

@@ -46,7 +46,9 @@ def define_node(
                 tensor_input,
                 xnn_graph,
                 vals_to_ids,
-                quant_params=QuantParams.from_inputs(tensor_input),
+                quant_params=QuantParams.from_inputs(
+                    tensor_input, self._exported_program
+                ),
             )
 
         self.define_tensor(

backends/xnnpack/operators/op_conv2d.py

@@ -46,7 +46,7 @@ def define_node(
         kwargs = {}
         # input
         input_node = get_input_node(node, 0)
-        input_quant_params = QuantParams.from_inputs(input_node)
+        input_quant_params = QuantParams.from_inputs(input_node, self._exported_program)
         self.define_tensor(
             input_node,
             xnn_graph,

backends/xnnpack/operators/op_multiply.py

@@ -43,7 +43,7 @@ def define_node(
             input1,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input1),
+            quant_params=QuantParams.from_inputs(input1, self._exported_program),
         )
         input1_id = vals_to_ids[input1]
 
@@ -53,7 +53,7 @@ def define_node(
             input2,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input2),
+            quant_params=QuantParams.from_inputs(input2, self._exported_program),
         )
         input2_id = vals_to_ids[input2]
 

backends/xnnpack/operators/op_sub.py

@@ -43,7 +43,7 @@ def define_node(
             input1,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input1),
+            quant_params=QuantParams.from_inputs(input1, self._exported_program),
         )
         input1_id = vals_to_ids[input1]
 
@@ -53,7 +53,7 @@ def define_node(
             input2,
             xnn_graph,
             vals_to_ids,
-            quant_params=QuantParams.from_inputs(input2),
+            quant_params=QuantParams.from_inputs(input2, self._exported_program),
         )
         input2_id = vals_to_ids[input2]
 

backends/xnnpack/operators/op_to_copy.py

@@ -51,7 +51,7 @@ def define_node(
         )
 
         input_node = get_input_node(node, 0)
-        input_quant_params = QuantParams.from_inputs(input_node)
+        input_quant_params = QuantParams.from_inputs(input_node, self._exported_program)
         output_quant_params = QuantParams.from_outputs(node)
 
         permute_order = PERM_NCHW_TO_NHWC if to_channels_last else PERM_NHWC_TO_NCHW

backends/xnnpack/operators/quant_params.py

@@ -11,8 +11,9 @@
 import torch
 from executorch.backends.xnnpack.passes.tag_implicit_q_dq_pass import TagImplicitQDqPass
 from executorch.backends.xnnpack.utils.quant_utils import is_dequant, is_quant
-from executorch.backends.xnnpack.utils.utils import check_or_raise
+from executorch.backends.xnnpack.utils.utils import check_or_raise, is_param_node
 from executorch.exir.dialects._ops import ops as exir_ops
+from torch.export import ExportedProgram
 
 
 class QuantParams:
@@ -178,11 +179,18 @@ def from_weights(cls, tensor_node: torch.fx.Node) -> Optional[QuantParams]:
         return cls.from_q_dq_node(q)
 
     @classmethod
-    def from_inputs(cls, tensor_node: torch.fx.Node) -> Optional[QuantParams]:
+    def from_inputs(
+        cls, tensor_node: torch.fx.Node, ep: ExportedProgram
+    ) -> Optional[QuantParams]:
         # tensor_node is quantized if it is produced by a dequant node
         if is_dequant(tensor_node) and TagImplicitQDqPass.is_tagged_as_implicit_q_dq(
             tensor_node
         ):
+            dq_input = cast(torch.fx.Node, tensor_node.args[0])
+            if is_quant(dq_input):
+                q_input = cast(torch.fx.Node, dq_input.args[0])
+                if is_param_node(ep, q_input):
+                    return cls.from_q_dq_node(dq_input)
             return cls.from_q_dq_node(tensor_node)
 
         return None

backends/xnnpack/test/TARGETS

@@ -127,6 +127,8 @@ python_unittest(
     ]),
     tags = ["long_running"],
     deps = [
+        "fbsource//third-party/pypi/timm:timm",
+        "fbsource//third-party/pypi/torchsr:torchsr",  # @manual
         "//caffe2:torch",
         "//executorch/backends/xnnpack/test/tester:tester",
         "//pytorch/vision:torchvision",

backends/xnnpack/test/models/edsr.py

@@ -0,0 +1,42 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+
+from executorch.backends.xnnpack.test.tester import Tester
+from torchsr.models import edsr_r16f64
+
+
+class TestEDSR(unittest.TestCase):
+    edsr = edsr_r16f64(2, False).eval()  # noqa
+    model_inputs = (torch.ones(1, 3, 224, 224),)
+
+    def test_fp32_edsr(self):
+        (
+            Tester(self.edsr, self.model_inputs)
+            .export()
+            .to_edge()
+            .partition()
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )
+
+    def test_qs8_edsr(self):
+        (
+            Tester(self.edsr, self.model_inputs)
+            .quantize()
+            .export()
+            .to_edge()
+            .partition()
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )

backends/xnnpack/test/models/inception_v3.py

@@ -0,0 +1,67 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+import torchvision.models as models
+from executorch.backends.xnnpack.test.tester import Tester
+
+
+class TestInceptionV3(unittest.TestCase):
+    # pyre-ignore
+    ic3 = models.inception_v3(weights="IMAGENET1K_V1").eval()  # noqa
+    model_inputs = (torch.ones(1, 3, 224, 224),)
+
+    all_operators = {
+        "executorch_exir_dialects_edge__ops_aten_addmm_default",
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor",
+        "executorch_exir_dialects_edge__ops_aten_cat_default",
+        "executorch_exir_dialects_edge__ops_aten_convolution_default",
+        "executorch_exir_dialects_edge__ops_aten_max_pool2d_with_indices_default",
+        # "executorch.exir.dialects.edge._ops.aten.avg_pool2d.default", Currently do not have avg_pool2d partitioned
+        "executorch_exir_dialects_edge__ops_aten_mean_dim",
+        "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        "executorch_exir_dialects_edge__ops_aten_permute_copy_default",
+        "executorch_exir_dialects_edge__ops_aten_relu_default",
+    }
+
+    def test_fp32_ic3(self):
+
+        (
+            Tester(self.ic3, self.model_inputs)
+            .export()
+            .to_edge()
+            .check(list(self.all_operators))
+            .partition()
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(self.all_operators))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )
+
+    def test_qs8_ic3(self):
+        # Quantization fuses away batchnorm, so it is no longer in the graph
+        ops_after_quantization = self.all_operators - {
+            "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        }
+
+        (
+            Tester(self.ic3, self.model_inputs)
+            .quantize()
+            .export()
+            .to_edge()
+            .check(list(ops_after_quantization))
+            .partition()
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(ops_after_quantization))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )

backends/xnnpack/test/models/inception_v4.py

@@ -0,0 +1,65 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+from executorch.backends.xnnpack.test.tester import Tester
+from timm.models import inception_v4
+
+
+class TestInceptionV4(unittest.TestCase):
+    ic4 = inception_v4(pretrained=False).eval()
+    model_inputs = (torch.ones(3, 299, 299).unsqueeze(0),)
+
+    all_operators = {
+        "executorch_exir_dialects_edge__ops_aten_addmm_default",
+        # "executorch.exir.dialects.edge._ops.aten.avg_pool2d.default", Currently do not have avg_pool2d partitioned
+        "executorch_exir_dialects_edge__ops_aten_cat_default",
+        "executorch_exir_dialects_edge__ops_aten_convolution_default",
+        "executorch_exir_dialects_edge__ops_aten_max_pool2d_with_indices_default",
+        "executorch_exir_dialects_edge__ops_aten_mean_dim",
+        "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        "executorch_exir_dialects_edge__ops_aten_permute_copy_default",
+        "executorch_exir_dialects_edge__ops_aten_relu_default",
+    }
+
+    def test_fp32_ic4(self):
+
+        (
+            Tester(self.ic4, self.model_inputs)
+            .export()
+            .to_edge()
+            .check(list(self.all_operators))
+            .partition()
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(self.all_operators))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )
+
+    def test_qs8_ic4(self):
+        # Quantization fuses away batchnorm, so it is no longer in the graph
+        ops_after_quantization = self.all_operators - {
+            "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        }
+
+        (
+            Tester(self.ic4, self.model_inputs)
+            .quantize()
+            .export()
+            .to_edge()
+            .check(list(ops_after_quantization))
+            .partition()
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(ops_after_quantization))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )

backends/xnnpack/test/tester/tester.py

@@ -36,7 +36,10 @@
     get_symmetric_quantization_config,
     XNNPACKQuantizer,
 )
-from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import QuantizationConfig
+from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import (
+    convert_scalars_to_attrs,
+    QuantizationConfig,
+)
 from torch.testing import FileCheck
 from torch.utils._pytree import tree_flatten
 
@@ -105,6 +108,7 @@ def run(
         self, artifact: torch.nn.Module, inputs: Optional[Tuple[torch.Tensor]]
     ) -> None:
         captured_graph = export.capture_pre_autograd_graph(artifact, inputs)
+        captured_graph = convert_scalars_to_attrs(captured_graph)
         prepared = prepare_pt2e(captured_graph, self.quantizer)
         converted = convert_pt2e(prepared)
         self.converted_graph = converted

install_requirements.sh

@@ -17,7 +17,7 @@ pip install .
 # models in executorch/examples/models.
 # The version in this file will be the correct version for the
 # corresponsing version of the repo.
-NIGHTLY_VERSION=dev20231002
+NIGHTLY_VERSION=dev20231005
 
 TORCH_VERSION=2.2.0.${NIGHTLY_VERSION}
 pip install --force-reinstall --pre torch=="${TORCH_VERSION}" -i https://download.pytorch.org/whl/nightly/cpu