Skip to content

Address Windows CUDA build issue #24149

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yuslepukhin merged 2 commits into from
Mar 24, 2025
Merged

Address Windows CUDA build issue #24149

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cee5231
Fix debug CUDA build issue on Windows
yuslepukhin Mar 24, 2025
63c1eed
Adjust CopyQK() specialiazation
yuslepukhin Mar 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 21 additions & 17 deletions onnxruntime/contrib_ops/cuda/bert/attention_qk.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,26 +37,17 @@ Status CopyQK(cudaStream_t stream,
const int qk_size,
const T* input,
QK* output) {
if constexpr (std::is_same<T, QK>::value) {
CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(output, input, qk_size * sizeof(QK), cudaMemcpyDeviceToDevice, stream));
return Status::OK();
}
const bool half2float = std::is_same<T, half>::value && std::is_same<QK, float>::value;
const bool float2half = std::is_same<T, float>::value && std::is_same<QK, half>::value;
ORT_ENFORCE(half2float || float2half);
constexpr const bool half2float = std::is_same<T, half>::value && std::is_same<QK, float>::value;
constexpr const bool float2half = std::is_same<T, float>::value && std::is_same<QK, half>::value;
static_assert(half2float || float2half, "This function supports either <float,half> or <half,float>");

int block_size = 256;
constexpr const int block_size = 256;
int num_blocks = (qk_size + block_size - 1) / block_size;
ConvertAndCopyQK<<<num_blocks, block_size, 0, stream>>>(qk_size, input, output);

return CUDA_CALL(cudaGetLastError());
}

template Status CopyQK<float, float>(cudaStream_t stream,
const int qk_size,
const float* input,
float* output);

template Status CopyQK<float, half>(cudaStream_t stream,
const int qk_size,
const float* input,
Expand All @@ -67,10 +58,23 @@ template Status CopyQK<half, float>(cudaStream_t stream,
const half* input,
float* output);

template Status CopyQK<half, half>(cudaStream_t stream,
const int qk_size,
const half* input,
half* output);
template <>
Status CopyQK(cudaStream_t stream,
const int qk_size,
const float* input,
float* output) {
CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(output, input, qk_size * sizeof(float), cudaMemcpyDeviceToDevice, stream));
return Status::OK();
}

template <>
Status CopyQK(cudaStream_t stream,
const int qk_size,
const half* input,
half* output) {
CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(output, input, qk_size * sizeof(half), cudaMemcpyDeviceToDevice, stream));
return Status::OK();
}

} // namespace cuda
} // namespace contrib
Expand Down
Loading