Skip to content

Fix attention QK linkage error #24134

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
kunal-vaishnavi merged 4 commits into from
Mar 24, 2025
Merged

Fix attention QK linkage error #24134

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
9d94517
Update attention_qk.cu
kunal-vaishnavi Mar 21, 2025
dca5c5b
Update attention_impl.cu
kunal-vaishnavi Mar 21, 2025
72d15aa
Add missing kernel
kunal-vaishnavi Mar 22, 2025
586f534
Fix ROCm build
kunal-vaishnavi Mar 22, 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
8 changes: 2 additions & 6 deletions onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -762,12 +762,8 @@ Status UnfusedAttention(
} else { // no mask
if (nullptr != data.output_qk) {
int64_t qk_size = (int64_t)batch_size * num_heads * sequence_length * total_sequence_length;
if (std::is_same<T, QK>::value) {
cudaMemcpyAsync(data.output_qk, data.scratch, qk_size * sizeof(QK), cudaMemcpyDeviceToDevice, stream);
} else {
ORT_RETURN_IF_ERROR(
(CopyQK<T, QK>(stream, static_cast<int>(qk_size), data.scratch, reinterpret_cast<QK*>(data.output_qk))));
}
ORT_RETURN_IF_ERROR(
(CopyQK<T, QK>(stream, static_cast<int>(qk_size), data.scratch, reinterpret_cast<QK*>(data.output_qk))));
}
ORT_RETURN_IF_ERROR(
ComputeSoftmax<T>(
Expand Down
22 changes: 22 additions & 0 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 @@ -24,11 +24,23 @@ __global__ void ConvertAndCopyQK(const int count, const half* input, float* outp
}
}

template <typename T>
__global__ void ConvertAndCopyQK(const int count, const T* input, T* output) {
int idx = threadIdx.x + blockIdx.x * blockDim.x;
if (idx < count) {
output[idx] = input[idx];
}
}

template <typename T, typename QK>
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);
Expand All @@ -40,6 +52,11 @@ Status CopyQK(cudaStream_t stream,
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 @@ -50,6 +67,11 @@ 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);

} // namespace cuda
} // namespace contrib
} // namespace onnxruntime
Loading