add cuda backend to backend test infra

.ci/scripts/test_backend.sh

@@ -56,6 +56,15 @@ if [[ "$FLOW" == *vulkan* ]]; then
     EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_VULKAN=ON"
 fi
 
+if [[ "$FLOW" == *cuda* ]]; then
+    # Fix libstdc++ GLIBCXX version for CUDA backend.
+    # The embedded .so files in the CUDA blob require GLIBCXX_3.4.30
+    # which the default conda libstdc++ doesn't have.
+    echo "Installing newer libstdc++ for CUDA backend..."
+    conda install -y -c conda-forge 'libstdcxx-ng>=12'
+    export LD_LIBRARY_PATH="${CONDA_PREFIX}/lib:${LD_LIBRARY_PATH:-}"
+fi
+
 if [[ "$FLOW" == *arm* ]]; then
 
     # Setup ARM deps.

.github/workflows/test-backend-cuda.yml

@@ -0,0 +1,30 @@
+name: Test CUDA Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    branches:
+      - release/*
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-cuda.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-cuda:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: cuda
+      flows: '["cuda"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true
+      runner-linux: linux.g5.4xlarge.nvidia.gpu

backends/cuda/test/tester.py

@@ -0,0 +1,71 @@
+# 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.
+
+from typing import Any, List, Optional, Tuple
+
+import executorch
+import executorch.backends.test.harness.stages as BaseStages
+import torch
+from executorch.backends.cuda.cuda_backend import CudaBackend
+from executorch.backends.cuda.cuda_partitioner import CudaPartitioner
+from executorch.backends.test.harness import Tester as TesterBase
+from executorch.backends.test.harness.stages import StageType
+from executorch.exir import EdgeCompileConfig
+from executorch.exir.backend.partitioner import Partitioner
+
+
+def _create_default_partitioner() -> CudaPartitioner:
+    """Create a CudaPartitioner with default compile specs."""
+    compile_specs = [CudaBackend.generate_method_name_compile_spec("forward")]
+    return CudaPartitioner(compile_specs)
+
+
+class ToEdgeTransformAndLower(BaseStages.ToEdgeTransformAndLower):
+    """CUDA-specific ToEdgeTransformAndLower stage."""
+
+    def __init__(
+        self,
+        partitioners: Optional[List[Partitioner]] = None,
+        edge_compile_config: Optional[EdgeCompileConfig] = None,
+    ):
+        if partitioners is None:
+            partitioners = [_create_default_partitioner()]
+
+        super().__init__(
+            default_partitioner_cls=_create_default_partitioner,
+            partitioners=partitioners,
+            edge_compile_config=edge_compile_config
+            or EdgeCompileConfig(_check_ir_validity=False),
+        )
+
+
+class CudaTester(TesterBase):
+    """
+    Tester subclass for CUDA backend.
+
+    This tester defines the recipe for lowering models to the CUDA backend
+    using AOTInductor compilation.
+    """
+
+    def __init__(
+        self,
+        module: torch.nn.Module,
+        example_inputs: Tuple[torch.Tensor],
+        dynamic_shapes: Optional[Tuple[Any]] = None,
+    ):
+        stage_classes = (
+            executorch.backends.test.harness.Tester.default_stage_classes()
+            | {
+                StageType.TO_EDGE_TRANSFORM_AND_LOWER: ToEdgeTransformAndLower,
+            }
+        )
+
+        super().__init__(
+            module=module,
+            stage_classes=stage_classes,
+            example_inputs=example_inputs,
+            dynamic_shapes=dynamic_shapes,
+        )

backends/test/suite/flow.py

@@ -147,4 +147,13 @@ def all_flows() -> dict[str, TestFlow]:
     except Exception as e:
         logger.info(f"Skipping ARM flow registration: {e}")
 
+    try:
+        from executorch.backends.test.suite.flows.cuda import CUDA_TEST_FLOW
+
+        flows += [
+            CUDA_TEST_FLOW,
+        ]
+    except Exception as e:
+        logger.info(f"Skipping CUDA flow registration: {e}")
+
     return {f.name: f for f in flows if f is not None}

backends/test/suite/flows/cuda.py

@@ -0,0 +1,21 @@
+# 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.
+
+from executorch.backends.cuda.test.tester import CudaTester
+from executorch.backends.test.suite.flow import TestFlow
+
+
+def _create_cuda_flow(name: str = "cuda") -> TestFlow:
+    """Create a test flow for the CUDA backend."""
+    return TestFlow(
+        name,
+        backend="cuda",
+        tester_factory=CudaTester,
+        quantize=False,
+    )
+
+
+CUDA_TEST_FLOW = _create_cuda_flow("cuda")