[release/2.13] [Bug Skip] test_philox_rand on XPU: CUDARngStateHelper has no XPU equivalent

Bug Description

test_prims.test_philox_rand is collected for the XPU device via test_prims_xpu.py but fails because the test body unconditionally calls CUDARngStateHelper.get_torch_state_as_tuple(), which raises RuntimeError: CUDA not available on XPU-only machines. The upstream test is decorated @onlyCUDA, but the XPU port file (or the XPUPatchForImport pre-import shim) re-routes it to XPU and the body itself remains CUDA-hardcoded.

Affected Tests

Cases:
op_ut,third_party.torch-xpu-ops.test.xpu.test_prims_xpu.TestPrimsXPU,test_philox_rand_xpu_float32

Error Message

RuntimeError: CUDA not available

Test Code Snippet

# pytorch/test/test_prims.py:281-310 (test_philox_rand)
@onlyCUDA
@dtypes(torch.float32)
def test_philox_rand(self, device, dtype):
    sizes = (1000, 1000000)
    repeats = 2
    for size in sizes:
        torch.cuda.manual_seed(123)
        references = []
        results = []
        rng_states = []
        for _ in range(repeats):
            rng_states.append(CUDARngStateHelper.get_torch_state_as_tuple())  # FAILS here on XPU
            ...

# pytorch/torch/_prims_common/__init__.py: CUDARngStateHelper.get_torch_state_as_tuple
@staticmethod
def get_torch_state_as_tuple(fake_mode=nullcontext()):
    if not torch.cuda.is_available():
        raise RuntimeError("CUDA not available")
    ...

Traceback

pytest_command:
cd ~/daisy_pytorch/third_party/torch-xpu-ops/test/xpu
pytest -v test_prims_xpu.py -k test_philox_rand_xpu_float32

Traceback (most recent call last):
  File ".../test/test_prims.py", line 286, in test_philox_rand
    rng_states.append(CUDARngStateHelper.get_torch_state_as_tuple())
  File ".../torch/_prims_common/__init__.py", line 2198, in get_torch_state_as_tuple
    raise RuntimeError("CUDA not available")
RuntimeError: CUDA not available

Root Cause Analysis

Two factors combine:

1. The XPU port re-routes test_philox_rand to the XPU device despite the upstream @onlyCUDA decorator — likely because XPUPatchForImport neutralises @onlyCUDA so the test body is dispatched on XPU.
2. The test body is CUDA-specific by construction: it uses CUDARngStateHelper, which gates on torch.cuda.is_available(). There is no XPU equivalent of CUDARngStateHelper exposed by torch._prims_common.

This is therefore a coverage gap, not a numeric bug:

• Either the XPU port should keep the @onlyCUDA decorator effective and skip the test on XPU, or
• An XPU equivalent of CUDARngStateHelper.get_torch_state_as_tuple() (e.g. XPURngStateHelper) should be added so the test can be ported faithfully and exercise XPU's philox RNG path.

Related Intel/torch-xpu-ops Issues

• (none found for philox on XPU as of 2026-04-25)

Versions

• pytorch: source build (HEAD)
• torch-xpu-ops: HEAD
• env: pytorch_opencode_env on PVC