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

WIP: Enable FA for GQA #23630

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

WIP: Enable FA for GQA #23630

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0a0a5ca
use correct total length to fix static kv_cache performance
guschmue Feb 7, 2025
8f808c2
Simple version of enabling GQA
sushraja-msft Feb 10, 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
2 changes: 1 addition & 1 deletion onnxruntime/contrib_ops/webgpu/bert/attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -439,7 +439,7 @@ Status ApplyAttention(const Tensor* Q, const Tensor* K, const Tensor* V, const T
WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context, const Tensor* seqlen_k) {
const int output_count = std::min({context.OutputCount(), 1 + (past_key != nullptr ? 1 : 0) + (past_value != nullptr ? 1 : 0)});
const int past_sequence_length = output_count > 1 ? parameters.past_sequence_length_ : 0;
const int total_sequence_length = past_sequence_length + parameters.kv_sequence_length_;
const int total_sequence_length = parameters.total_sequence_length_;

const TensorShapeVector probs_dims({parameters.batch_size_, parameters.num_heads_,
parameters.sequence_length_, total_sequence_length});
Expand Down
7 changes: 5 additions & 2 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -420,8 +420,10 @@ Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, co
FlashAttentionProgram program{"FlashAttention", has_attention_bias, parameters.head_size_, parameters.num_heads_};
program.AddInputs({{Q, ProgramTensorMetadataDependency::TypeAndRank, 4},
{present_key, ProgramTensorMetadataDependency::TypeAndRank, 4},
{present_value, ProgramTensorMetadataDependency::TypeAndRank, 4},
{attention_bias, ProgramTensorMetadataDependency::TypeAndRank}});
{present_value, ProgramTensorMetadataDependency::TypeAndRank, 4}});
if (has_attention_bias) {
program.AddInput({attention_bias, ProgramTensorMetadataDependency::TypeAndRank});
}
program.AddOutputs({{output, ProgramTensorMetadataDependency::TypeAndRank, 4}});
const float alpha = parameters.scale_ == 0.0f ? 1.f / sqrt(static_cast<float>(parameters.head_size_))
: parameters.scale_;
Expand All @@ -443,6 +445,7 @@ bool CanApplyFlashAttention(const Tensor* bias, const Tensor* present_key, const
return parameters.batch_size_ == 1 &&
bias == nullptr &&
parameters.sequence_length_ > 1 &&
parameters.qkv_format_ == Q_K_V_BSNH &&
context.Device().HasFeature(wgpu::FeatureName::Subgroups) &&
present_key != nullptr && present_value != nullptr && present_key->SizeInBytes() > 0 &&
present_value->SizeInBytes() > 0 && parameters.head_size_ % 4 == 0;
Expand Down
6 changes: 6 additions & 0 deletions onnxruntime/contrib_ops/webgpu/bert/group_query_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
#include "contrib_ops/webgpu/bert/attention_common.h"
#include "contrib_ops/webgpu/bert/group_query_attention.h"
#include "contrib_ops/webgpu/webgpu_contrib_kernels.h"
#include "contrib_ops/webgpu/bert/flash_attention.h"

#include "core/providers/webgpu/webgpu_supported_types.h"

Expand Down Expand Up @@ -74,6 +75,11 @@ Status GroupQueryAttention::ComputeInternal(onnxruntime::webgpu::ComputeContext&
Tensor* present_value = context.Output(2, present_kv_shape);
parameters.past_present_share_buffer_ = present_key != nullptr && present_value != nullptr && past_key != nullptr && past_value != nullptr && past_key->DataRaw() == present_key->DataRaw() && past_value->DataRaw() == present_value->DataRaw();

if (CanApplyFlashAttention(/*bias*/ nullptr, present_key, present_value, parameters, context)) {
return ApplyFlashAttention(query, key, value, /*attention_bias*/ nullptr, output, past_key, present_key, past_value,
present_value, parameters, context);
}

TensorShapeVector q_new_dims({parameters.batch_size_, parameters.num_heads_,
parameters.sequence_length_, parameters.head_size_});
TensorShape q_new_shape(q_new_dims);
Expand Down
Loading