Skip to content

[MLAS] Add an AVX-512 fused kernel for LinearAttention #31674

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
mirounga merged 8 commits into from
Aug 12, 2026
Merged

[MLAS] Add an AVX-512 fused kernel for LinearAttention #31674

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
24f9b88
Fix LinearAttention output shape inference for standard GQA (q > kv)
mirounga Jul 27, 2026
a6acfff
[Webgpu] Fix LinearAttention output shape for standard GQA (q > kv)
mirounga Jul 28, 2026
188b06a
[MLAS] Move the LinearAttention recurrence into MLAS
mirounga Aug 4, 2026
2482e83
[MLAS] Add an AVX-512 fused kernel for LinearAttention
mirounga Aug 4, 2026
3189073
Merge remote-tracking branch 'origin/main' into rs_buf72
mirounga Aug 5, 2026
b351482
[MLAS] Address review feedback on the LinearAttention AVX-512 kernel
mirounga Aug 6, 2026
72a141b
Merge branch 'main' into rs_buf72
mirounga Aug 10, 2026
39de2a4
[MLAS] Second round of review feedback on LinearAttention
mirounga Aug 10, 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
4 changes: 4 additions & 0 deletions cmake/onnxruntime_mlas.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,6 +59,8 @@ onnxruntime_add_static_library(onnxruntime_mlas
${MLAS_SRC_DIR}/flashattn_qkv.cpp
${MLAS_SRC_DIR}/flashattn_gqa.cpp
${MLAS_SRC_DIR}/qkv_quant.cpp
${MLAS_SRC_DIR}/linear_attention.h
${MLAS_SRC_DIR}/linear_attention.cpp
${MLAS_SRC_DIR}/cast.cpp
${MLAS_SRC_DIR}/layernorm.cpp
${MLAS_SRC_DIR}/rotary_embedding.h
Expand Down Expand Up @@ -245,6 +247,7 @@ function(setup_mlas_source_for_windows)
${MLAS_SRC_DIR}/intrinsics/avx512/silu_avx512f.cpp
${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
${MLAS_SRC_DIR}/intrinsics/avx512/sconv_nchw_depthwise_multiplier_greater_than_1_avx512f.cpp
${MLAS_SRC_DIR}/linear_attention_kernel_avx512f.cpp
)

set_source_files_properties(${mlas_platform_srcs_avx512} PROPERTIES COMPILE_FLAGS "/arch:AVX512")
Expand Down Expand Up @@ -928,6 +931,7 @@ else()
${MLAS_SRC_DIR}/intrinsics/avx512/silu_avx512f.cpp
${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
${MLAS_SRC_DIR}/intrinsics/avx512/sconv_nchw_depthwise_multiplier_greater_than_1_avx512f.cpp
${MLAS_SRC_DIR}/linear_attention_kernel_avx512f.cpp
)
set_source_files_properties(${mlas_platform_srcs_avx512f} PROPERTIES COMPILE_FLAGS "-mavx512f")

Expand Down
2 changes: 1 addition & 1 deletion docs/ContribOperators.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2895,7 +2895,7 @@ This version of the operator has been available since version 1 of the 'com.micr

<dl>
<dt><tt>output</tt> : T</dt>
<dd>Attention output with 3D packed shape (B, T, H_q * d_v).</dd>
<dd>Attention output with 3D packed shape (B, T, max(H_q, H_kv) * d_v). Standard GQA emits one output per query head; inverse GQA, where H_kv exceeds H_q, emits one per KV head.</dd>
<dt><tt>present_state</tt> : S</dt>
<dd>Updated recurrent state with shape (B, H_kv, d_k, d_v), or (W, B, H_kv, d_k, d_v) when state_window = W > 0. Slot W-1 is the state after the last token; slot j is the state after token (T - W + j).</dd>
</dl>
Expand Down
Loading
Loading