[Bug] TestDropoutAlignRandomEager.test_dropout_mask_parity_and_rng_offset_cuda: eager-vs-compiled dropout mask mismatch ~50% on XPU (recurrence of #4851)

Bug Description

TestDropoutAlignRandomEager.test_dropout_mask_parity_and_rng_offset_cuda
fails on XPU. The test asserts that torch.nn.Dropout run in eager mode and
the same module wrapped with torch.compile produce matching dropout masks
(and matching RNG-offset deltas) when seeded identically. On XPU, only
~49.95% of mask elements agree between eager and compiled output, i.e. the
masks are effectively uncorrelated (random ~50% agreement) rather than
identical, for all three tested dtypes (float32, float16, bfloat16).

Affected Test

op_ut,test/inductor/test_dropout_align_random_eager.py,TestDropoutAlignRandomEager.test_dropout_mask_parity_and_rng_offset_cuda

Note: the test method is only defined with a _cuda suffix in current
upstream pytorch/pytorch source (no _xpu-suffixed or generic variant
exists); it runs on XPU because GPU_TYPE/HAS_GPU_AND_TRITON resolve to
XPU in this environment and the class is gated only by
requires_gpu()/IS_LINUX and HAS_GPU_AND_TRITON, not by an explicit
CUDA-only check.

Reproduce

conda activate pytorch_xpu
cd /tmp   # avoid running from inside the pytorch source checkout (conftest.py double-load issue)
python -m pytest -sv "/home/daisyden/upstream/inductor_agnostic/test/inductor/test_dropout_align_random_eager.py::TestDropoutAlignRandomEager::test_dropout_mask_parity_and_rng_offset_cuda"

Observed Failure (this run)

FAILED test/inductor/test_dropout_align_random_eager.py::TestDropoutAlignRandomEager::test_dropout_mask_parity_and_rng_offset_cuda

Traceback (most recent call last):
  File "test/inductor/test_dropout_align_random_eager.py", line 257, in test_dropout_mask_parity_and_rng_offset_cuda
    self.assertLessEqual(
  File ".../unittest/case.py", line 1238, in assertLessEqual
    self.fail(self._formatMessage(msg, standardMsg))
  File ".../unittest/case.py", line 675, in fail
    raise self.failureException(msg)
AssertionError: 0.5005362033843994 not less than or equal to 0.0001
Dropout mask mismatch ratio too high: 0.50053620

1 failed in 15.03s

Measured mismatch ratio: 0.5005362 (threshold is 1e-4), i.e. eager and
compiled masks agree only at chance level (~50%), for the first dtype
(float32) tested in the loop; the assertion fails before later dtypes in
the loop are reached.

Root Cause Analysis

The test seeds eager and compiled runs identically via _set_seed(BASE_SEED)
and expects Inductor's Philox-based RNG to reproduce the same
stream/offset as the eager XPU RNG so that dropout masks match bit-for-bit.
Instead, the compiled-vs-eager mask agreement (~50%) indicates the two paths
are drawing effectively independent/uncorrelated random streams rather than
sharing the same seed/offset contract on XPU. This points to a divergence
between the XPU eager RNG stream and the Inductor-generated (Philox) RNG
offset/seed handling used for compiled dropout on XPU — the same failure
mode already described in #4851 for a set of related tests in this same
file (including this exact test method and error message).

Context

• intel/torch-xpu-ops#4851 — "[Bug] TestDropoutAlignRandomEager: compiled
  dropout mask ~50% mismatch vs eager on XPU" was filed for the same failure
  signature (same test method, same AssertionError: ... not less than or equal to 0.0001 / Dropout mask mismatch ratio too high: 0.5005...) across
  4 test methods in this class. That issue was closed on 2026-08-09 without
  a linked commit or merged PR (gh api .../issues/4851/timeline shows the
  closed event has commit_id: null), so the closure does not represent a
  verified fix. This new issue is filed because the bug still reproduces
  locally against the current checkout, per the reproduce steps above.
• pytorch/pytorch#190237 — "TestDropoutAlignRandomEager: 4 tests fail on
  main" (OPEN). Related context on this test class's general flakiness/known
  issues, though it covers different test methods
  (test_linear_block_compile_parity_forward/backward,
  test_dropout_parity_dynamic_shapes) which are already @unittest.skip'd
  in-file citing that issue. Low relevance to the specific RNG-offset root
  cause here but included for completeness since it is the same test class.

Environment

torch: 2.14.0.dev20260726+xpu
Platform: Linux
Python: 3.10.20
GPU: Intel(R) Data Center GPU Max 1550 (device_id=0xBD5)
Driver: 1.6.33578+57 (Level-Zero, oneAPI Unified Runtime)
Checkout: pytorch/pytorch @ deefeb3bd593d7a810baad416abbd9ce180dd98a (branch daisyden/inductor)