[CI] oss:cutlass-4.8:sm107 gated-attention FROST GEMM plans missing after #1177

[liujane-dev]

Summary

oss:cutlass-4.8:sm107 regressed between FE 35fa1148 and 5757ef24.

• Passing anchor: pipeline 69042084, job 449584482, FE 35fa1148
• Failing job: pipeline 69242469, job 451155191, FE 5757ef24
• Component: FROST GEMM heuristics / SM107 dense FP8 plan selection
• Confirmed culprit: 322d30b59 (frost gemm heur: take the 64-byte-K MMA and the 256-tall CTA pair on parts that issue it #1177) frost gemm heur: take the 64-byte-K MMA and the 256-tall CTA pair on parts that issue it

Failure

SM107 gated-attention / projection GEMM tests fail because frost_gemm is not in the backend ranked plan list and the tests refuse fallback to non-FROST backend plans.

Representative failed test:

test/python/fe_api/gated_attention_block/test_proj_gemm.py::test_fp8_matmul_with_alpha_matches_fp32_reference[256-256-256]
_NoFrostPlan: fp8: the 'frost_gemm' plan is not in the ranked list [...]

A/B Confirmation

Targeted SM107 A/B repro using same wheel/runtime and overlaid FE Python files:

d518022b:  test_proj_gemm.py::test_fp8_matmul_with_alpha_matches_fp32_reference[256-256-256] PASSED
322d30b5: test_proj_gemm.py::test_fp8_matmul_with_alpha_matches_fp32_reference[256-256-256] FAILED with _NoFrostPlan

Notes

Later commit 885201ac appears intended to fix this class of issue; pipeline 69342374 job 451959339 was canceled before providing a green signal. Will verify in next pipeline.