Fix CUDA build with contrib ops disabled

[Copilot]

Description

The CUDA Attention kernel (core/providers/cuda/llm/attention.cc) depends on contrib_ops internals (flash attention, memory efficient attention, unfused attention helpers) but was compiled unconditionally. When building with --disable_contrib_ops, GetAttentionKernelOptions() is unavailable (guarded by #ifndef DISABLE_CONTRIB_OPS in cuda_kernel.h), causing a compile error.

Changes:

• cmake/onnxruntime_providers_cuda.cmake — When contrib ops are disabled (and not in CUDA minimal mode), include the contrib_ops/cuda/bert/ attention infrastructure files (flash attention, memory efficient attention, unfused attention helpers, etc.) so the ONNX domain Attention kernel can compile and link. Uses elseif(onnxruntime_DISABLE_CONTRIB_OPS AND NOT onnxruntime_CUDA_MINIMAL) to avoid including these files in CUDA minimal builds where llm/attention.cc isn't compiled and cudnn_frontend.h isn't available.
• onnxruntime/core/providers/cuda/cuda_execution_provider.h — Remove #ifndef DISABLE_CONTRIB_OPS guards from the AttentionKernelOptions include, GetAttentionKernelOptions() method, and attention_kernel_options_ member variable
• onnxruntime/core/providers/cuda/cuda_kernel.h — Remove #ifndef DISABLE_CONTRIB_OPS guard from GetAttentionKernelOptions()

The CUDA Attention kernel and its underlying attention backends (flash, memory efficient, unfused) are now always available in full CUDA builds regardless of whether contrib ops are enabled. No changes are needed in cuda_execution_provider.cc since the Attention kernel registrations remain unconditional.

Motivation and Context

Building onnxruntime with CUDA enabled and --disable_contrib_ops fails:

error C2039: 'GetAttentionKernelOptions': is not a member of 'onnxruntime::cuda::Attention<float>'

This is a valid build configuration (useful for reducing compile time) that should be supported. Rather than excluding the CUDA Attention kernel when contrib ops are disabled, the necessary attention infrastructure from contrib_ops/cuda/bert/ is included in the build so the ONNX domain Attention op retains full CUDA acceleration. The fix is scoped to non-minimal CUDA builds only, since CUDA minimal builds use a non-recursive glob that doesn't include llm/attention.cc and don't have cudnn_frontend available.

[github-actions]

You can commit the suggested changes from lintrunner.

[tianleiwu]

There is CI build failure for Build Linux TensorRT CUDA Minimal x64 Release

@copilot, please update ‎cmake/onnxruntime_providers_cuda.cmake to handle cuda minimal build nicely.

[Copilot]

There is CI build failure for Build Linux TensorRT CUDA Minimal x64 Release

python3 tools/ci_build/...

Fixed in 8b79ac7. The issue was that the else() block also triggered for CUDA minimal builds (where cudnn_frontend.h isn't available). Changed to elseif(onnxruntime_DISABLE_CONTRIB_OPS AND NOT onnxruntime_CUDA_MINIMAL) — since CUDA minimal uses a non-recursive glob that doesn't include llm/attention.cc, the bert/ infrastructure isn't needed there.