XNNPACK: Skip the bf16 GEMM runtime tests in fbcode

backends/xnnpack/test/BUCK

@@ -109,6 +109,7 @@ fbcode_target(_kind = runtime.python_test,
         "//executorch/exir:lib",
         "//executorch/exir/capture:config",
         "//executorch/exir/passes:constant_prop_pass",
+        "//executorch/export:export_utils",
         "//executorch/runtime:runtime",
         "//pytorch/ao:torchao",  # @manual
     ],
@@ -191,6 +192,7 @@ fbcode_target(_kind = runtime.python_test,
         "//caffe2:torch",
         "//executorch/backends/xnnpack/partition:xnnpack_partitioner",
         "//executorch/exir:lib",
+        "//executorch/export:export_utils",
         "//executorch/extension/pybindings:portable_lib",
     ],
 )

backends/xnnpack/test/ops/test_bmm.py

@@ -9,6 +9,7 @@
 import torch
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
 from executorch.backends.xnnpack.test.tester import Tester, ToEdgeTransformAndLower
+from executorch.export.utils import is_fbcode
 
 
 class TestBMM(unittest.TestCase):
@@ -49,6 +50,10 @@ def test_fp32_bmm(self):
         )
         self._test_bmm(inputs)
 
+    @unittest.skipIf(
+        is_fbcode(),
+        "skipping for fbcode",
+    )
     def test_bf16_bmm(self):
         inputs = (
             torch.randn(2, 3, 4).to(torch.bfloat16),

backends/xnnpack/test/test_xnnpack_partitioner.py

@@ -13,6 +13,7 @@
 
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
 from executorch.exir import to_edge, to_edge_transform_and_lower
+from executorch.export.utils import is_fbcode
 from executorch.extension.pybindings.portable_lib import (
     _load_for_executorch_from_buffer,
 )
@@ -219,6 +220,10 @@ def lower(module, x):
                 ).forward((bf16_x,))[0]
                 torch.testing.assert_close(output, module(bf16_x))
 
+    @unittest.skipIf(
+        is_fbcode(),
+        "skipping for fbcode",
+    )
     def test_enable_bf16_fuses_batch_norm_into_a_linear(self):
         """
         XNNPACK cannot define a bf16 BatchNorm on its own, but one that follows a
@@ -259,6 +264,10 @@ def forward(self, x):
                 ).forward((x,))[0]
                 torch.testing.assert_close(output, module(x), atol=1e-2, rtol=1e-2)
 
+    @unittest.skipIf(
+        is_fbcode(),
+        "skipping for fbcode",
+    )
     def test_enable_bf16_runs_an_attention_block(self):
         """
         The linear layers and both batch matrix multiplications of a bf16 attention