Skip to content
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

DMMHA: add unit tests; fix CPU, CUDA kernel #22567

Merged
merged 16 commits into from
Nov 2, 2024
Merged

DMMHA: add unit tests; fix CPU, CUDA kernel #22567

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
41ad318
DecoderMaskedSelfAttention: Merge code for float and float16
mindest Oct 21, 2024
c294331
Merge more functions for float and float16
mindest Oct 21, 2024
5f171e7
Fix errors; refactor
mindest Oct 22, 2024
d79c3c3
Update doc.
mindest Oct 22, 2024
9df4782
Add unit tests for DecoderMaskedMultiHeadAttention.
mindest Oct 22, 2024
50dfa85
Fix scale order; revert doc change
mindest Oct 23, 2024
c8ab330
Revert test change
mindest Oct 24, 2024
3d64c5c
Enable qk output for cross-attn only
mindest Oct 24, 2024
5558436
Fix scale order in reference code
mindest Oct 24, 2024
f8bb900
Fix mask in CPU and CUDA attn; fix offset for out_qk;
mindest Oct 28, 2024
69db052
Set macro USE_CUDA; float output qk.
mindest Oct 28, 2024
cef311c
Add missed changes.
mindest Oct 28, 2024
601b26d
Enable mask, attn_bias; reorder key in CUDA cross-attn
mindest Oct 28, 2024
94cb346
Fix error: type, softmax lowest
mindest Oct 29, 2024
8302d56
Loosen tol for fp16.
mindest Oct 29, 2024
6fedb82
Fix: rm unused arg
mindest Oct 29, 2024
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
10 changes: 6 additions & 4 deletions onnxruntime/contrib_ops/cpu/bert/decoder_masked_multihead_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -339,6 +339,7 @@ void DecoderMaskedMultiHeadAttention<T>::ComputeAttentionProbsWithBeams(
T* attention_probs_ptr = reinterpret_cast<T*>(attention_probs) + last_offset;
math::Dot<float, CPUMathUtil>(head_size, q_vec, K + i * head_size, attention_probs_ptr, nullptr);

*attention_probs_ptr *= scale;
// Apply the attention bias and mask
if (attn_bias_data != nullptr) {
*attention_probs_ptr += attn_bias_data[attn_bias_base_offset + past_sequence_length];
Expand All @@ -348,7 +349,6 @@ void DecoderMaskedMultiHeadAttention<T>::ComputeAttentionProbsWithBeams(
if (is_masked) {
*attention_probs_ptr += mask_filter_value_;
}
*attention_probs_ptr *= scale;
}

{
Expand All @@ -362,6 +362,8 @@ void DecoderMaskedMultiHeadAttention<T>::ComputeAttentionProbsWithBeams(
const T* past_k_vec = past_key_data + beam_batch_offset + beam_offset + j * head_size;
T* output = reinterpret_cast<T*>(attention_probs) + j + i * probs_matrix_size;
math::Dot<float, CPUMathUtil>(head_size, q_vec, past_k_vec, output, nullptr);

*output *= scale;
// Apply the attention bias and mask
if (attn_bias_data != nullptr) {
*output += attn_bias_data[attn_bias_base_offset + j];
Expand All @@ -371,11 +373,11 @@ void DecoderMaskedMultiHeadAttention<T>::ComputeAttentionProbsWithBeams(
if (is_masked) {
*output += mask_filter_value_;
}
*output *= scale;
}
}
// Append current key to present key (past_present_share_buffer_ is true)
memcpy(present_key_data + i * max_sequence_length * head_size, K + i * head_size, head_size * sizeof(T));
memcpy(present_key_data + (i * max_sequence_length + past_sequence_length) * head_size,
K + i * head_size, head_size * sizeof(T));
}
});

Expand Down Expand Up @@ -460,7 +462,7 @@ void DecoderMaskedMultiHeadAttention<T>::ComputeVxAttentionScoreWithBeams(
}
}
// Append current value to present value (past_present_share_buffer_ is true)
memcpy(present_value_data + i * max_sequence_length * v_head_size,
memcpy(present_value_data + (i * max_sequence_length + past_sequence_length) * v_head_size,
V + i * v_head_size,
v_head_size * sizeof(T));
}
Expand Down
4 changes: 2 additions & 2 deletions onnxruntime/contrib_ops/cpu/bert/decoder_masked_multihead_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ class DecoderMaskedMultiHeadAttention final : public OpKernel, public AttentionC
const Tensor* cache_indir,
OpKernelContext* context,
int beam_width,
Tensor* scaled_qk = nullptr) const;
Tensor* output_qk = nullptr) const;
void ComputeAttentionProbsWithBeams(T* attention_probs,
const T* Q,
const T* K,
Expand All @@ -50,7 +50,7 @@ class DecoderMaskedMultiHeadAttention final : public OpKernel, public AttentionC
bool broadcast_attn_bias_dim_1,
const int32_t* cache_indir_data,
int beam_width,
T* scaled_qk_data = nullptr) const;
T* output_qk_data = nullptr) const;
void ComputeVxAttentionScoreWithBeams(T* output,
T* tmp_buffer,
const T* attention_probs,
Expand Down
1 change: 0 additions & 1 deletion onnxruntime/core/graph/contrib_ops/bert_defs.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -908,7 +908,6 @@ ONNX_MS_OPERATOR_SET_SCHEMA(
OpSchema::Optional)
.Input(9,
"cache_indirection",
// This input is useful for CUDA EP only.
"A buffer of shape [batch_size, beam_width, max_output_length] where an `[i, j, k]` entry specifies "
"which beam the `k`-th token came from for the `j`-th beam for batch `i` in the current iteration",
"M",
Expand Down
Loading
Loading