Skip to content

Use concat to ONNX attention cache update #27612

Description

Follow up to #27542

(1) If we do "temp = Concat(past_key, new_key)", then this Concat will copy the past_key to temp. But we end up doing a second copy when we do a mem-cpy from temp to present_key. If we "present_key = Concat(past_key, new_key)", we will avoid the second copy (while we still do the first copy in Concat itself).

// Copy past KV (BNSH) into present buffers (BNSH). Strided copy because
// past has [B, N_kv, past_seq, H] and present has [B, N_kv, total_seq, H].
const size_t past_k_row_bytes = static_cast<size_t>(parameters.past_sequence_length) *
parameters.head_size * sizeof(T);
const size_t present_k_row_bytes = static_cast<size_t>(parameters.total_sequence_length) *
parameters.head_size * sizeof(T);
CUDA_RETURN_IF_ERROR(cudaMemcpy2DAsync(
present_key->MutableData<T>(), present_k_row_bytes,
past_key->Data<T>(), past_k_row_bytes,
past_k_row_bytes, num_kv_rows,
cudaMemcpyDeviceToDevice, cuda_stream));
const size_t past_v_row_bytes = static_cast<size_t>(parameters.past_sequence_length) *
parameters.v_head_size * sizeof(T);
const size_t present_v_row_bytes = static_cast<size_t>(parameters.total_sequence_length) *
parameters.v_head_size * sizeof(T);
CUDA_RETURN_IF_ERROR(cudaMemcpy2DAsync(
present_value->MutableData<T>(), present_v_row_bytes,
past_value->Data<T>(), past_v_row_bytes,
past_v_row_bytes, num_kv_rows,
cudaMemcpyDeviceToDevice, cuda_stream));

(2) CUDA_KERNEL_ASSERT is noop in release build. Check the convention in onnxruntime to see an official way to raise for invalid.

CUDA_KERNEL_ASSERT(false); // mask must be contiguous (no True after False)

#27542 (comment)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions