[PyTorch Upstream][Regression] PyTorch benchmark torchbench sam got fail_accuracy

[anmyachev]

Command to run: python benchmarks/dynamo/torchbench.py --accuracy --bfloat16 -d xpu -n10 --inference --only sam --cold-start-latency --backend=inductor

c6efadaf99068e70da7634487ce7952e00e28b32 is the first bad commit
commit c6efadaf99068e70da7634487ce7952e00e28b32
Author: Whitney Tsang <whitney.tsang@intel.com>
Date:   Thu Jan 8 10:09:52 2026 -0500

    Enable `TRITON_INTEL_PREDICATED_STORE` by default (#5776)

[anmyachev]

@whitneywhtsang please take a look

[whitneywhtsang]

Fail with TRITON_INTEL_PREDICATED_STORE=0 too if disable llvm opt and/or igc opt.

[whitneywhtsang]

Pass when triton_poi_fused_add_addmm_constant_pad_nd_native_layer_norm_view_14 doesn't use predicated store.

[whitneywhtsang]

Pass with TRITON_INTEL_FAST_MATH=0 or TRITON_DEFAULT_FP_FUSION=0.

[whitneywhtsang]

When predicated store is enabled, the reduced control flow allows for increased mul+add fusion operations. While mul+add fusion is the expected behavior in Triton across all platforms (including NVIDIA), the increased usage of fused multiply-add (FMA) instructions is causing accuracy test failures.

Attached shader dumps for three configurations:
Current implementation - predicated store + FP fusion enabled → fails accuracy
FP fusion disabled - predicated store enabled, FP fusion disabled → passes accuracy
Predicated store disabled - FP fusion enabled, predicated store disabled → passes accuracy
shader dumps: wi5917.zip

@EikanWang @etaf would it be possible to either modify the reference implementation or increase the tolerance thresholds to accommodate the expected FMA behavior?

Related Issues:
#4626
#4576