Skip to content

[webgpu] Unify the present_sequence_length in flash attention #25945

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
guschmue merged 1 commit into from
Sep 15, 2025
Merged

[webgpu] Unify the present_sequence_length in flash attention #25945

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 15 additions & 8 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 @@ -164,7 +164,9 @@ Status FlashAttentionDecodeQKTProgram::GenerateShaderCode(ShaderHelper& shader)

Status ComputeFlashAttentionDecodeQKT(onnxruntime::webgpu::ComputeContext& context, const Tensor* Q,
const Tensor* attention_bias, Tensor* output, Tensor* present_key, Tensor* metadata,
const WebgpuAttentionParameters& parameters, uint32_t num_total_seq_length_tile, uint32_t num_present_sequence_length_tile, uint32_t tile_size) {
const WebgpuAttentionParameters& parameters, uint32_t num_total_seq_length_tile,
uint32_t num_present_sequence_length_tile, uint32_t tile_size,
uint32_t present_sequence_length) {
const float alpha = parameters.scale_ == 0.0f ? 1.f / sqrt(static_cast<float>(parameters.head_size_))
: parameters.scale_;

Expand All @@ -187,8 +189,7 @@ Status ComputeFlashAttentionDecodeQKT(onnxruntime::webgpu::ComputeContext& conte
.AddUniformVariables({{static_cast<uint32_t>(vectorized_head_size)},
{static_cast<uint32_t>(parameters.total_sequence_length_)},
{static_cast<float>(alpha)},
// present_sequence_length is used to index into the KV cache, for static kv cache it is the max sequence length.
{static_cast<uint32_t>(parameters.is_gqa_ ? parameters.seqlen_present_kv_cache_ : parameters.total_sequence_length_)},
present_sequence_length,
{static_cast<uint32_t>(parameters.n_reps)},
{num_total_seq_length_tile},
{num_present_sequence_length_tile},
Expand Down Expand Up @@ -220,7 +221,8 @@ Status ComputeFlashAttentionDecodeSplitVxScore(onnxruntime::webgpu::ComputeConte
const WebgpuAttentionParameters& parameters,
uint32_t num_total_seq_length_tile,
uint32_t num_present_sequence_length_tile,
uint32_t tile_size) {
uint32_t tile_size,
uint32_t present_sequence_length) {
const int components = 4;
int head_size_vec = parameters.v_head_size_ / components;
FlashAttentionDecodeSplitVxProgram program{"FlashAttentionDecodeSplitVx", tile_size, head_size_vec};
Expand All @@ -233,7 +235,7 @@ Status ComputeFlashAttentionDecodeSplitVxScore(onnxruntime::webgpu::ComputeConte
.SetWorkgroupSize(64)
.AddUniformVariables({{static_cast<uint32_t>(parameters.total_sequence_length_)},
{static_cast<uint32_t>(head_size_vec)},
{static_cast<uint32_t>(parameters.is_gqa_ ? parameters.seqlen_present_kv_cache_ : parameters.total_sequence_length_)},
present_sequence_length,
{static_cast<uint32_t>(parameters.n_reps)},
num_total_seq_length_tile,
num_present_sequence_length_tile,
Expand Down Expand Up @@ -279,7 +281,10 @@ Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, co
Tensor* output, const Tensor* past_key, Tensor* present_key, const Tensor* past_value, Tensor* present_value,
const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context) {
ORT_RETURN_IF_ERROR(CopyKVCache(context, parameters, K, past_key, present_key, V, past_value, present_value));
const int present_sequence_length = parameters.is_gqa_ ? parameters.seqlen_present_kv_cache_ : parameters.total_sequence_length_;

// Extract present_sequence_length directly from present_key tensor shape:
// (batch_size, num_heads, total_sequence_length/max_sequence_length, head_size)
const uint32_t present_sequence_length = static_cast<uint32_t>(present_key->Shape()[2]);
if (parameters.sequence_length_ > 1) {
const uint32_t tile_size = 64;
bool has_attention_bias = attention_bias != nullptr;
Expand Down Expand Up @@ -332,12 +337,14 @@ Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, co
const TensorShape metadata_shape(metadata_dims);
Tensor metadata = context.CreateGPUTensor(DataTypeImpl::GetType<float>(), metadata_shape);
ORT_RETURN_IF_ERROR(ComputeFlashAttentionDecodeQKT(context, Q, attention_bias, &qk, present_key, &metadata,
parameters, num_total_seq_length_tile, num_present_sequence_length_tile, tile_size));
parameters, num_total_seq_length_tile, num_present_sequence_length_tile, tile_size,
present_sequence_length));

const TensorShapeVector out_split_vx_dims({parameters.batch_size_, parameters.num_heads_, num_present_sequence_length_tile, parameters.head_size_});
const TensorShape out_split_vx_shape(out_split_vx_dims);
Tensor out_split_vx = context.CreateGPUTensor(Q->DataType(), out_split_vx_shape);
ORT_RETURN_IF_ERROR(ComputeFlashAttentionDecodeSplitVxScore(context, &metadata, &qk, &out_split_vx, present_value, parameters, num_total_seq_length_tile, num_present_sequence_length_tile, tile_size));
ORT_RETURN_IF_ERROR(ComputeFlashAttentionDecodeSplitVxScore(context, &metadata, &qk, &out_split_vx, present_value, parameters,
num_total_seq_length_tile, num_present_sequence_length_tile, tile_size, present_sequence_length));
ORT_RETURN_IF_ERROR(ComputeFlashAttentionDecodeVxReduce(context, &out_split_vx, output, parameters, num_total_seq_length_tile, num_present_sequence_length_tile));

return Status::OK();
Expand Down
Loading