Fix optional-output guard in DecoderAttention/MultiHeadAttention shape inference#29243

Closed

titaiwangms wants to merge 36 commits into

microsoft:mainfrom

titaiwangms:fix/contrib-attn-shapeinf-oob

Commits on Jun 23, 2026

Require all present outputs before populating attention shape inference

titaiwangms
and
Copilot
committed

Commits on Jun 24, 2026

Add regression tests for attention contrib ops with optional present outputs omitted

titaiwangms
and
Copilot
committed
Add 3-output positive cases for attention contrib op shape inference

titaiwangms
and
Copilot
committed
Clarify optional-output guard comments in attention shape inference

titaiwangms
and
Copilot
committed
Strengthen MHA/DMMHA two-output shape-inference regression tests

titaiwangms
and
Copilot
committed
Apply clang-format to attention optional-output shape inference tests

titaiwangms
and
Copilot
committed
docs: add contrib-op shape-inference output-index safety skill + coding-convention note

titaiwangms
and
Copilot
committed
contrib: validate MaxpoolWithMask kernel rank matches input spatial rank

titaiwangms
and
Copilot
committed
contrib: reject unsupported MaxpoolWithMask pooling rank and DRY rank check

titaiwangms
and
Copilot
committed

Commits on Jun 25, 2026