skip some tests

jeffdaily · jeffdaily · commit c5d30d61f069 · 2026-03-11T18:34:40.000Z
diff --git a/test/distributed/test_cupy_as_tensor.py b/test/distributed/test_cupy_as_tensor.py
@@ -8,7 +8,10 @@
 import torch
 from torch.multiprocessing.reductions import reduce_tensor
 from torch.testing._internal.common_cuda import SM100OrLater
-from torch.testing._internal.common_distributed import MultiProcContinuousTest
+from torch.testing._internal.common_distributed import (
+    MultiProcContinuousTest,
+    skip_if_rocm_multiprocess,
+)
 from torch.testing._internal.common_utils import (
     requires_cuda_p2p_access,
     run_tests,
@@ -67,6 +70,7 @@ def _init_device(self) -> None:
     def device(self) -> torch.device:
         return torch.device(device_type, self.rank)
 
+    @skip_if_rocm_multiprocess  # RuntimeError: pidfd_getfd Operation not permitted"
     @skip_but_pass_in_sandcastle_if(
         SM100OrLater,
         "Fails if ran in docker environment without privileged access (https://github.com/pytorch/pytorch/issues/165170)",
diff --git a/test/test_cuda.py b/test/test_cuda.py
@@ -5060,6 +5060,8 @@ def cb(device, alloc, device_alloc, device_free):
 
     def test_allocator_fuzz(self):
         # fuzz
+        if torch.version.hip and "expandable_segments:True" in torch._C._accelerator_getAllocatorSettings():
+            raise unittest.SkipTest("ROCm needs https://github.com/ROCm/rocm-systems/pull/3023")
         state = random.getstate()
         random.seed(123)
         N = 10000
diff --git a/torch/_C/__init__.pyi.in b/torch/_C/__init__.pyi.in
@@ -2608,6 +2608,7 @@ def _accelerator_getDeviceStats(device_index: _int) -> dict[str, Any]: ...
 def _accelerator_resetAccumulatedStats(device_index: _int) -> None: ...
 def _accelerator_resetPeakStats(device_index: _int) -> None: ...
 def _accelerator_getMemoryInfo(device_index: _int) -> tuple[_int, _int]: ...
+def _accelerator_getAllocatorSettings() -> str: ...
 def _accelerator_setAllocatorSettings(env: str) -> None: ...
 
 # Defined in torch/csrc/jit/python/python_tracer.cpp
diff --git a/torch/_dynamo/trace_rules.py b/torch/_dynamo/trace_rules.py
@@ -468,6 +468,7 @@
         "torch._C._accelerator_getAccelerator",
         "torch._C._accelerator_getDeviceIndex",
         "torch._C._accelerator_getStream",
+        "torch._C._accelerator_getAllocatorSettings",
         "torch._C._accelerator_setAllocatorSettings",
         "torch._C._accelerator_setStream",
         "torch._C._accelerator_synchronizeDevice",
diff --git a/torch/csrc/DeviceAccelerator.cpp b/torch/csrc/DeviceAccelerator.cpp
@@ -164,6 +164,10 @@ void initModule(PyObject* module) {
     return at::accelerator::getMemoryInfo(device_index);
   });
 
+  m.def("_accelerator_getAllocatorSettings", []() {
+    return c10::CachingAllocator::getAllocatorSettings();
+  });
+
   m.def("_accelerator_setAllocatorSettings", [](std::string env) {
     c10::CachingAllocator::setAllocatorSettings(env);
   });
