Guard large-head nonpad Attention MEA dispatch

[Kevin-Li-2025]

Description

Fixes #28388.

The ONNX Attention CUDA path currently allows Memory Efficient Attention for the nonpad_kv_seqlen external-cache path with large head sizes. That path uses the CUTLASS custom right-padding variant, which can exceed the dynamic shared-memory opt-in limit on smaller architectures for head_size > 256 and crash instead of falling back.

This keeps MEA available for the normal path, but makes nonpad_kv_seqlen != nullptr && head_size > 256 fall through to the unified unfused path, which already supports large head sizes.

Tests

• python3 -m py_compile onnxruntime/test/python/transformers/test_onnx_attention/test_gqa.py
• git diff --check

I also attempted the targeted pytest locally:

python3 -m pytest -q onnxruntime/test/python/transformers/test_onnx_attention/test_gqa.py -k large_head_nonpad_seqlen_falls_back_from_mea_fp16

but local collection is blocked by a missing parameterized package before reaching ORT/CUDA execution.

[Kevin-Li-2025]

@microsoft-github-policy-service agree