decompose con1d to conv2d for cuda backend

.github/workflows/cuda.yml

@@ -71,7 +71,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        model: [linear, add, add_mul, resnet18]
+        model: [linear, add, add_mul, resnet18, conv1d]
     with:
       timeout: 90
       runner: linux.g5.4xlarge.nvidia.gpu

backends/cuda/cuda_backend.py

@@ -24,9 +24,14 @@
 )
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from torch._inductor.codegen.cpp_wrapper_cpu import CppWrapperCpu
+from torch._inductor.decomposition import conv1d_to_conv2d
 from torch.export.passes import move_to_device_pass
 from torch.nn.attention import SDPBackend
 
+cuda_decomposition_table = {
+    torch.ops.aten.conv1d.default: conv1d_to_conv2d,
+}
+
 # exist fallback operators in et namespace;
 supported_fallback_kernels: Dict[str, Any] = {}
 
@@ -119,6 +124,10 @@ def preprocess(
         # replace slice_copy with slice
         ReplaceSliceCopyWithSlicePass()(cuda_edge_program.graph_module)
 
+        cuda_edge_program = cuda_edge_program.run_decompositions(
+            cuda_decomposition_table
+        )
+
         edge_program_module = cuda_edge_program.module()
 
         # Grab all input placeholders from the graph

backends/cuda/tests/test_cuda_export.py

@@ -251,3 +251,22 @@ def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
         self.assertIsNotNone(
             edge_program_manager, "Mathematical operations export failed"
         )
+
+    def test_conv1d(self):
+        """Test CUDA export for 1D convolution."""
+
+        class Conv1dModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.conv = torch.nn.Conv1d(3, 16, kernel_size=3, padding=1)
+
+            def forward(self, x: torch.Tensor) -> torch.Tensor:
+                return self.conv(x)
+
+        module = Conv1dModule()
+        module.eval()
+        inputs = (torch.randn(1, 3, 10),)
+
+        # Test export
+        edge_program_manager = self._export_to_cuda_with_lower(module, inputs)
+        self.assertIsNotNone(edge_program_manager, "Conv1d operation export failed")

examples/models/__init__.py

@@ -14,6 +14,7 @@ class Model(str, Enum):
     Add = "add"
     AddMul = "add_mul"
     Softmax = "softmax"
+    Conv1d = "conv1d"
     Dl3 = "dl3"
     Edsr = "edsr"
     EmformerTranscribe = "emformer_transcribe"
@@ -59,6 +60,7 @@ def __str__(self) -> str:
     str(Model.Add): ("toy_model", "AddModule"),
     str(Model.AddMul): ("toy_model", "AddMulModule"),
     str(Model.Softmax): ("toy_model", "SoftmaxModule"),
+    str(Model.Conv1d): ("toy_model", "Conv1dModule"),
     str(Model.Dl3): ("deeplab_v3", "DeepLabV3ResNet50Model"),
     str(Model.Edsr): ("edsr", "EdsrModel"),
     str(Model.EmformerTranscribe): ("emformer_rnnt", "EmformerRnntTranscriberModel"),

examples/models/toy_model/__init__.py

@@ -4,11 +4,19 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from .model import AddModule, AddMulModule, LinearModule, MulModule, SoftmaxModule
+from .model import (
+    AddModule,
+    AddMulModule,
+    Conv1dModule,
+    LinearModule,
+    MulModule,
+    SoftmaxModule,
+)
 
 __all__ = [
     AddModule,
     AddMulModule,
+    Conv1dModule,
     LinearModule,
     MulModule,
     SoftmaxModule,

examples/models/toy_model/model.py

@@ -88,3 +88,20 @@ def get_eager_model(self) -> torch.nn.Module:
 
     def get_example_inputs(self):
         return (torch.ones(2, 2),)
+
+
+class Conv1dModule(torch.nn.Module, EagerModelBase):
+    def __init__(self):
+        super().__init__()
+        self.conv1d = torch.nn.Conv1d(
+            in_channels=3, out_channels=16, kernel_size=3, padding=1
+        )
+
+    def forward(self, x):
+        return self.conv1d(x)
+
+    def get_eager_model(self) -> torch.nn.Module:
+        return self
+
+    def get_example_inputs(self):
+        return (torch.randn(1, 3, 10),)