Row: ENG-EXPERT-STREAM
Observed
The thread-sanitizer lane for PR #2856 failed the focused guard added by #2623 at merge ref 65c96d71ec4ef467119736e79560541cc01bd403:
test_gemma4_moe_fp8_native_arm_guard.cpp:301:
FATAL ERROR: REQUIRE( AnyNonZero(want) ) is NOT correct!
values: REQUIRE( false )
The failing case is gemma4 moe: the FP8-native expert upload refuses where its three ops have no arm. It produced an all-zero baseline, making the later equality check vacuous. The lane emitted no ThreadSanitizer race report.
The ordinary CPU lane ran the same test tree without this failure. The address/undefined-sanitizer lane did not list this test among its failures. The result is therefore specific to the thread-sanitized build or its execution conditions, not a generally failing assertion.
Attribution
PR #2856 changes no Gemma4, MoE, FP8, expert-streaming, backend-probe, or sanitizer path. Its complete diff is one ROCm evidence spec, two oracle scripts, one script test, four NumPy arrays, one manifest, and comments in the Qwen3.5 sacred test.
The case and AnyNonZero precondition came from 19ed51cf7, the #2623 repair. Current main 41b6aa2560a576c28fa15f2a3ff4505305f9b850 still carries them.
Owed
The ENG-EXPERT-STREAM row owns this guard and its non-vacuity guarantee. Reproduce the all-zero baseline in a focused thread-sanitizer build, determine whether the cause is the test fixture, CPU dispatch, or sanitizer-sensitive production behavior, and add a mutation that proves the repaired test still detects deletion of the three-op availability guard.
This is not fixed in #2856 because the cause is not yet localized and the repair belongs to the independently specified and reviewed expert-streaming row.
Row:
ENG-EXPERT-STREAMObserved
The thread-sanitizer lane for PR #2856 failed the focused guard added by #2623 at merge ref
65c96d71ec4ef467119736e79560541cc01bd403:The failing case is
gemma4 moe: the FP8-native expert upload refuses where its three ops have no arm. It produced an all-zero baseline, making the later equality check vacuous. The lane emitted no ThreadSanitizer race report.The ordinary CPU lane ran the same test tree without this failure. The address/undefined-sanitizer lane did not list this test among its failures. The result is therefore specific to the thread-sanitized build or its execution conditions, not a generally failing assertion.
Attribution
PR #2856 changes no Gemma4, MoE, FP8, expert-streaming, backend-probe, or sanitizer path. Its complete diff is one ROCm evidence spec, two oracle scripts, one script test, four NumPy arrays, one manifest, and comments in the Qwen3.5 sacred test.
The case and
AnyNonZeroprecondition came from19ed51cf7, the #2623 repair. Current main41b6aa2560a576c28fa15f2a3ff4505305f9b850still carries them.Owed
The
ENG-EXPERT-STREAMrow owns this guard and its non-vacuity guarantee. Reproduce the all-zero baseline in a focused thread-sanitizer build, determine whether the cause is the test fixture, CPU dispatch, or sanitizer-sensitive production behavior, and add a mutation that proves the repaired test still detects deletion of the three-op availability guard.This is not fixed in #2856 because the cause is not yet localized and the repair belongs to the independently specified and reviewed expert-streaming row.