Skip to content

webgpu: Fuse FlashAttention decode kernels and extend to any sequence length #28389

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
qjia7 merged 17 commits into from
Jun 11, 2026
Merged

webgpu: Fuse FlashAttention decode kernels and extend to any sequence length #28389

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
17 commits
Select commit Hold shift + click to select a range
435c445
webgpu: Extend FlashAttention decode path to support any sequence length
qjia7 May 7, 2026
39ffe45
webgpu: Add m_tile optimization and remove Subgroups requirement for …
qjia7 May 11, 2026
ecfc56a
webgpu: Fuse QKT and SplitVx into single QKV kernel for FlashAttentio…
qjia7 May 12, 2026
934ed2a
webgpu: Route between prefill and split-reduce paths in FlashAttention
qjia7 May 12, 2026
35b35f4
webgpu: Fix clang-format in FlashAttentionDecodeQKVProgram constructor
qjia7 May 12, 2026
9b8dadf
Merge branch 'main' into webgpu-flash-attention-relax-subgroups
qjia7 May 19, 2026
b7e6592
Remove subgroups feature check from use_split_reduce routing
qjia7 May 19, 2026
46fd59a
Merge branch 'main' into webgpu-flash-attention-relax-subgroups
qjia7 May 21, 2026
07ef10d
Remove use_seqlen_k extension from decode kernels, keep only use_indi…
qjia7 May 21, 2026
b690d0f
Extend indirect dispatch to support sequence_length > 1 in split-redu…
qjia7 May 21, 2026
4241a91
Extract shared write_indirect_dispatch WGSL function for dispatch nor…
qjia7 May 21, 2026
42419a6
Merge branch 'main' into webgpu-flash-attention-relax-subgroups
qjia7 Jun 2, 2026
a16023e
Merge branch 'main' into webgpu-flash-attention-relax-subgroups
qjia7 Jun 8, 2026
10ef54d
Tune flash attention split-reduce threshold to seq_len < 32
qjia7 Jun 8, 2026
b0dcedf
Use getByOffset/setByOffset helpers in flash attention decode shaders
qjia7 Jun 9, 2026
ea60286
Normalize indirect-dispatch helper to accept (x, y, z) and consolidat…
qjia7 Jun 9, 2026
13ea0e4
Simplify flash attention split-reduce routing to seq_len < 32
qjia7 Jun 9, 2026
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
203 changes: 84 additions & 119 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention.cc
View file
Open in desktop

Large diffs are not rendered by default.

54 changes: 24 additions & 30 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,11 +117,17 @@ class FlashAttentionProgram final : public Program<FlashAttentionProgram> {
bool has_head_sink_;
};

class FlashAttentionDecodeQKTProgram final : public Program<FlashAttentionDecodeQKTProgram> {
class FlashAttentionDecodeQKVProgram final : public Program<FlashAttentionDecodeQKVProgram> {
public:
FlashAttentionDecodeQKTProgram(const std::string& kernel_name,
bool has_attention_bias, uint32_t tile_size, bool use_indirect_dispatch)
: Program{kernel_name}, has_attention_bias_(has_attention_bias), tile_size_(tile_size), use_indirect_dispatch_(use_indirect_dispatch) {
FlashAttentionDecodeQKVProgram(const std::string& kernel_name,
bool has_attention_bias, uint32_t tile_size, int head_size_vec,
bool use_indirect_dispatch, bool q_BNSH = false,
bool is_unidirectional = false, bool use_seqlen_k = false,
uint32_t m_tile = 1)
: Program{kernel_name}, has_attention_bias_(has_attention_bias), tile_size_(tile_size),
head_size_vec_(head_size_vec), use_indirect_dispatch_(use_indirect_dispatch),
q_BNSH_(q_BNSH), is_unidirectional_(is_unidirectional), use_seqlen_k_(use_seqlen_k),
m_tile_(m_tile) {
Comment thread
qjia7 marked this conversation as resolved.
Outdated
}

Status GenerateShaderCode(ShaderHelper& sh) const override;
Expand All @@ -135,41 +141,24 @@ class FlashAttentionDecodeQKTProgram final : public Program<FlashAttentionDecode
{"num_heads", ProgramUniformVariableDataType::Uint32},
{"batch_size", ProgramUniformVariableDataType::Uint32},
{"attn_bias_dim0", ProgramUniformVariableDataType::Uint32},
{"attn_bias_dim1", ProgramUniformVariableDataType::Uint32});
{"attn_bias_dim1", ProgramUniformVariableDataType::Uint32},
{"new_sequence_length", ProgramUniformVariableDataType::Uint32});

private:
bool has_attention_bias_;
uint32_t tile_size_;
bool use_indirect_dispatch_;
};

class FlashAttentionDecodeSplitVxProgram final : public Program<FlashAttentionDecodeSplitVxProgram> {
public:
FlashAttentionDecodeSplitVxProgram(const std::string& kernel_name, uint32_t tile_size, int head_size_vec, bool use_indirect_dispatch, bool has_head_sink = false)
: Program{kernel_name}, tile_size_(tile_size), head_size_vec_(head_size_vec), use_indirect_dispatch_(use_indirect_dispatch), has_head_sink_(has_head_sink) {
}

Status GenerateShaderCode(ShaderHelper& sh) const override;

WEBGPU_PROGRAM_DEFINE_UNIFORM_VARIABLES({"total_sequence_length", ProgramUniformVariableDataType::Uint32},
{"head_size_vec", ProgramUniformVariableDataType::Uint32},
{"present_sequence_length", ProgramUniformVariableDataType::Uint32},
{"n_reps", ProgramUniformVariableDataType::Uint32},
{"num_present_sequence_length_tile", ProgramUniformVariableDataType::Uint32},
{"batch_heads", ProgramUniformVariableDataType::Uint32},
{"num_heads", ProgramUniformVariableDataType::Uint32});

private:
uint32_t tile_size_;
int head_size_vec_;
bool use_indirect_dispatch_;
bool has_head_sink_;
bool q_BNSH_;
bool is_unidirectional_;
bool use_seqlen_k_;
uint32_t m_tile_;
};

class FlashAttentionDecodeVxReduceProgram final : public Program<FlashAttentionDecodeVxReduceProgram> {
public:
FlashAttentionDecodeVxReduceProgram(const std::string& kernel_name, uint32_t tile_size, uint32_t seq_tile_size, bool use_indirect_dispatch)
: Program{kernel_name}, tile_size_(tile_size), seq_tile_size_(seq_tile_size), use_indirect_dispatch_(use_indirect_dispatch) {
FlashAttentionDecodeVxReduceProgram(const std::string& kernel_name, uint32_t tile_size, uint32_t seq_tile_size, bool use_indirect_dispatch, bool use_seqlen_k = false, bool has_head_sink = false, uint32_t m_tile = 1)
: Program{kernel_name}, tile_size_(tile_size), seq_tile_size_(seq_tile_size), use_indirect_dispatch_(use_indirect_dispatch), use_seqlen_k_(use_seqlen_k), has_head_sink_(has_head_sink), m_tile_(m_tile) {
}

Status GenerateShaderCode(ShaderHelper& sh) const override;
Expand All @@ -178,12 +167,17 @@ class FlashAttentionDecodeVxReduceProgram final : public Program<FlashAttentionD
{"num_total_seq_length_tile", ProgramUniformVariableDataType::Uint32},
{"num_present_sequence_length_tile", ProgramUniformVariableDataType::Uint32},
{"num_head_size_tile", ProgramUniformVariableDataType::Uint32},
{"batch_heads", ProgramUniformVariableDataType::Uint32});
{"batch_heads", ProgramUniformVariableDataType::Uint32},
{"new_sequence_length", ProgramUniformVariableDataType::Uint32},
{"num_heads", ProgramUniformVariableDataType::Uint32});

private:
uint32_t tile_size_;
uint32_t seq_tile_size_;
bool use_indirect_dispatch_;
bool use_seqlen_k_;
bool has_head_sink_;
uint32_t m_tile_;
};

Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, const Tensor* attention_bias,
Expand Down
118 changes: 0 additions & 118 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention_decode_qkt.wgsl.template
View file
Open in desktop

This file was deleted.

Loading
Loading