TheRock/rocm-libraries/dnn-providers/hipblaslt-provider/config/HIPBLASLT_ENGINE.toml

# Integration test configuration for HIPBLASLT_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

# -- 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

# MatmulBiasActiv Bf16 (bias-add then SWISH_FWD) is flaky against the default
# bf16 tolerance (atol=0.01, rtol=0.01): a local 40-run sweep of the whole
# suite showed ~20% of runs mismatching by 1-2 bf16 ULP on a single output
# element, and the failing parameterization/index/tensor-position varied
# from run to run (e.g. Correctness/7, /19, /23, /31 all observed failing
# across different sweeps; never more than one element per run). Observed
# worst case: ref=4.5, GPU=4.5625, absolute diff 0.0625 vs default
# threshold 0.055. This is nondeterministic GPU reduction/epilogue rounding
# landing on the bf16 boundary just past bias-add before the nonlinear
# SWISH slope amplifies it, not a fixed per-shape bug — a single-index
# override would leave other indices intermittently red. Not a regression
# from this port: same synthesis range and default tolerance as pre-port.
# atol/rtol bumped to 0.02 covers the worst observed diff (0.0625 at
# ref=4.5, threshold becomes 0.11) with ~75% margin, and covers the
# smaller 0.03125-at-ref=2.0625 case with ~95% margin.
[[tolerance_overrides]]
filters = ["Smoke/IntegrationGpuMatmulBiasActivBf16.Correctness/*"]
atol = 0.02
rtol = 0.02

# -- 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.

# ROCm/rocm-libraries#8033 — hipblasLtMatmulAlgoGetHeuristic returns 0
# algorithms on gfx12 (gfx1200, gfx1201, gfx1250, ...) for FP16 with
# transA=T, transB=T, and a bias-fused epilogue (BIAS, RELU_BIAS,
# GELU_BIAS, SWISH_BIAS_EXT, CLAMP_BIAS_EXT). Every neighboring cell (no
# bias, BF16/FP32 with bias, FP16 with any other transpose pair) returns
# at least one algorithm.
#
# Replaces the same-arch carve-out previously enforced at runtime by
# IS_HIPBLASLT_GFX12_FP16_TT_BIAS / shouldSkipOnEngineConfigResult in
# dnn-providers/hipblaslt-provider/integration_tests/TestWorkarounds.hpp.
[[test_skips]]
archs   = ["gfx12"]
filters = ["*MatmulBias*Fp16*"]
reason  = "ROCm/rocm-libraries#8033 — hipblasLtMatmulAlgoGetHeuristic returns no algorithms for FP16 with transA=T, transB=T and any bias-fused epilogue on gfx12"