[BMG] `Triton GEMM + PostOp (add matrix) kernel benchmark int8` failure

[whitneywhtsang]

Last pass: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19329583123
First failure: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19362789627

[dev-tomek]

Analyzing the history of runs, looks like the issue is related to this particular runner triton-bmg-499320.
@kwasd is it possible to add a temporary label to this runner so it can be targeted when testing potential fixes?
Also is it possible to get access to this machine?

[kwasd]

relabeled from b580 to b580-issue-5481

will ping you about access

[dev-tomek]

Reproduced the issue. Looks like this particular runner has more RAM which results in this test case '[4096, 8, 16384, 128, torch.int8]' not being skipped by runtime checks as in the case of other runners.

Next steps: Need to figure out why the additional test case results in error

[dev-tomek]

Resolved the issue with the reported GEMM benchmark. I see however that a similar error shows up on FlexAttention (batch_size=16) Causal Mask fwd kernel benchmark. I've applied the same fix as for GEMM but the issue persists. Will investigate further as part of #5603.