PR #184451: Un-xfail test_sdpa_rewriter_14_cpu on ARM64 SVE256

• Repository: pytorch/pytorch
• Author: izaitsevfb
• State: open | merged: False
• Link: https://github.com/pytorch/pytorch/pull/184451

Description (problem / solution / changelog)

Summary

Remove xfailIf(IS_ARM64 and IS_CPU_CAPABILITY_SVE256) from test_sdpa_rewriter_14_cpu in test/inductor/test_fused_attention.py.

The new OSDC aarch64 runner (mt-l-arm64g3-16-62, enabled trunk-wide via #183482) reports IS_CPU_CAPABILITY_SVE256=True but does not reproduce the numerical mismatch from #177244. The decorator now flips passing runs to FAILED CONSISTENTLY via "Unexpected success" whenever target determination pulls the file into the shard.

Failure first surfaced at 07734446eb on trunk / linux-jammy-aarch64-py3.10 / test-osdc (default, 4, 5, mt-l-arm64g3-16-62) and has been consistent across every subsequent main commit whose TD scope selected test_fused_attention.py.

Closes #177244 (if the issue still applies to the older m7g.4xlarge runner family, we can re-narrow the gate later).

Test plan

• pull / trunk aarch64 OSDC test-osdc shard 4/5 turns green on this branch.
• No regressions on x86 or other aarch64 shards.

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo

Changed files

• test/inductor/test_fused_attention.py (modified, +1/-2)