Fix CUDA build with contrib ops disabled (#28554)
### 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.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>
Co-authored-by: Tianlei Wu <tlwu@microsoft.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
