TheRock/rocm-libraries/dnn-providers/hip-kernel-provider/config/HIP_MLOPS_ENGINE.toml

# Integration test configuration for HIP_MLOPS_ENGINE
# Filters use GTest-style globs: * matches any sequence of characters
# Later entries take precedence over earlier ones when multiple filters match

[meta]
version = 1

# -- Test Skips --
# Skip tests on specific architectures (or globally if 'archs' is omitted).
# Each entry requires:
#   archs:   optional array of substrings matched against the device's raw
#            gcnArchName (e.g. "gfx942" matches "gfx942:sramecc+:xnack-").
#            Omit or leave empty to skip on every arch.
#   filters: array of GTest globs matched against the full GTest name
#   reason:  human-readable string surfaced in the GTEST_SKIP message
# First matching entry wins.
#
# -- Tolerance Overrides --
# Override default tolerances for specific tests or groups of tests.
# Each entry requires:
#   filters: array of glob patterns matched against the full GTest name
#   atol: absolute tolerance
#   rtol: relative tolerance
#
# RMSNormBackward Pure-Bfp16 (dy/x/dx all bf16) fails CPU-reference tolerance
# at the suite's large-reduction shapes (x:[4096,128,1,1], x:[5,256,14,14],
# x:[16,3,8,14,14]) — smaller shapes in the same suites pass reliably. The
# exact failing case (which training/bias/layout combo) varies between runs
# at a fixed shape, pointing at run-to-run nondeterminism in the GPU kernel's
# reduction order rather than a fixed tolerance shortfall, so per-case index
# filters aren't reliable here — skipping both fixtures outright. Root cause
# not yet identified; revisit once the CPU reference / GPU kernel numerics
# have been compared directly.
[[test_skips]]
filters = [
    "Smoke/IntegrationGpuRMSNormBackwardPureBfp16.Correctness/*",
    "Smoke/IntegrationGpuRMSNormBackward3dPureBfp16.Correctness/*",
]
reason = "RMSNormBackward Pure-Bfp16 is flaky at large reduction shapes — root cause under investigation"

# Layernorm forward batch-256/512 volumetric shapes (Full5dLargeBatch prefix). Each case runs
# ~17-29 s and the set roughly doubles the 5D layernorm full-tier runtime without exercising a
# code path the batch-32 Full shapes don't already cover. Skipped for now to keep the tier fast;
# drop this entry once per-test tier filtering is fully wired.
[[test_skips]]
filters = ["Full5dLargeBatch/*Layernorm*"]
reason = "Layernorm large-batch shapes are too slow for the current full tier; re-enable once tier filtering is fully set up"