[ExecuTorch][WebGPU] gelu op tests

backends/webgpu/test/op_tests/cases.py

@@ -1,3 +1,3 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
@@ -286,3 +286,38 @@
         ],
         golden_dtype="float32",  # concatenation copies values; fp64 bit-identical
     )
+from executorch.backends.webgpu.test.ops.test_gelu import (
+    _det_input as _gelu_det_input,
+    GeluModule,
+    N as _GELU_N,
+)
+
+
+def _gelu_full_range(_shape) -> torch.Tensor:
+    # Reuse the deterministic linspace(-6, 6) spanning negatives/zero/positives.
+    return _gelu_det_input()
+
+
+@register_op_test("gelu")
+def _gelu_suite() -> WebGPUTestSuite:
+    # erf ("none") is the Florence-2/BART + PyTorch default; tanh is the approx.
+    return WebGPUTestSuite(
+        module_factory=lambda approximate: GeluModule(approximate),
+        cases=[
+            Case(name="erf_vec", construct={"approximate": "none"}, inputs=((M1,),)),
+            Case(name="erf_mat", construct={"approximate": "none"}, inputs=((M1, M2),)),
+            Case(
+                name="erf_rank3",
+                construct={"approximate": "none"},
+                inputs=((S1, M1, M2),),
+            ),
+            Case(name="tanh_mat", construct={"approximate": "tanh"}, inputs=((M1, M2),)),
+            Case(
+                name="erf_range",
+                construct={"approximate": "none"},
+                inputs=(InputSpec(shape=(_GELU_N,), gen=_gelu_full_range),),
+            ),
+        ],
+        atol=1e-4,
+        rtol=1e-3,
+    )

backends/webgpu/test/ops/test_gelu.py

@@ -0,0 +1,59 @@
+# 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.
+
+"""`aten.gelu.default` module + input for the WebGPU op-test framework.
+
+`GeluModule`, `N`, and `_det_input` are imported by `cases.py` to drive the
+declarative op-test suite. `GeluTest` is the export-delegation smoke test. The
+`approximate` kwarg selects the exact (erf) path ("none", PyTorch's default and
+the Florence-2/BART path) or the tanh approximation; the deterministic input
+spans negatives, zero, and the saturation region.
+"""
+
+import unittest
+
+import torch
+
+from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
+from executorch.exir import to_edge_transform_and_lower
+
+# Input length; the deterministic input spans negatives, zero, and positives.
+N = 64
+
+
+class GeluModule(torch.nn.Module):
+    def __init__(self, approximate: str = "none") -> None:
+        super().__init__()
+        self.approximate = approximate
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.nn.functional.gelu(x, approximate=self.approximate)
+
+
+def _det_input() -> torch.Tensor:
+    """Deterministic fp32 input spanning negatives, zero, and positives."""
+    return torch.linspace(-6.0, 6.0, N, dtype=torch.float32)
+
+
+def _export(m: torch.nn.Module, x: torch.Tensor):
+    ep = torch.export.export(m, (x,))
+    return to_edge_transform_and_lower(
+        ep, partitioner=[VulkanPartitioner()]
+    ).to_executorch()
+
+
+class GeluTest(unittest.TestCase):
+    def test_export_delegates(self) -> None:
+        for approximate in ("none", "tanh"):
+            et = _export(GeluModule(approximate).eval(), _det_input())
+            found = any(
+                d.id == "VulkanBackend"
+                for plan in et.executorch_program.execution_plan
+                for d in plan.delegates
+            )
+            self.assertTrue(
+                found, f"Expected a VulkanBackend delegate (gelu {approximate})"
+            )