Skip the batch sampler runtime test on ROCm

Summary:
The ROCm AOTI unittest job on main fails during pytest collection: the newly added batch sampler test imports the Python runtime at module scope, and the ROCm CI build only ships executor_runner, not the pybindings extension. One collection error aborts the whole backends/cuda/tests run, so every ROCm test stops running.

Import the runtime inside the test and skip the class on ROCm, following the existing convention for tests that load through the Python runtime.

Differential Revision: D126231386

Pull Request resolved: #24013

Author: Gasoonjia

backends/cuda/tests/test_batch_sampler.py

@@ -20,7 +20,6 @@
 from executorch.exir.passes import MemoryPlanningPass
 from executorch.exir.passes.propagate_device_pass import PropagateDeviceConfig
 from executorch.extension.llm.batching.sampler import BatchSampler
-from executorch.runtime import Runtime
 
 
 def export_sampler(directory: Path, vocab: int):
@@ -56,8 +55,14 @@ def export_sampler(directory: Path, vocab: int):
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
+@unittest.skipIf(
+    torch.version.hip is not None,
+    "loads through the Python runtime, which the ROCm build does not include",
+)
 class CudaBatchSamplerTest(unittest.TestCase):
     def test_dynamic_rows_with_masked_reduction_lanes(self):
+        from executorch.runtime import Runtime
+
         # A vocabulary spanning several reduction tiles exposes padding lanes.
         vocab = 202048
         with tempfile.TemporaryDirectory() as temporary: