[CPU] Implement X-Attention for intel CPU#32086

Open

mangguo321 wants to merge 28 commits intoopenvinotoolkit:masterfrom

mangguo321:mang/xatt

Commits on Nov 20, 2025

Enable xattention for intel CPU
mangguo321
committed
Integrate mask to PA
mangguo321
committed
Support multiple sequences for xattention
mangguo321
committed
Fix find block issue
mangguo321
committed
Copy key buffer before executing GEMM and bf16 precision support.
mangguo321
committed
revert pa conflict changes with sparse attetnion execution

liubo-intel
authored and
mangguo321
committed
enable sparse attetnion execution

liubo-intel
authored and
mangguo321
committed
Use optimized softmax kernel in X-Attention
mangguo321
committed
Optimize block sum using hardware instructions
mangguo321
committed
Add macro definications to not support ARM platform
mangguo321
committed
modify xattention_block_size and PA block_size check and scale calculation places

liubo-intel
authored and
mangguo321
committed
optimize sparse softmax mask in sparse mask attention: extend attn_softmax_kernel(float) to support sparse mask

liubo-intel
authored and
mangguo321
committed
Add const to unchanged variable
mangguo321
committed
Minor changes in find blocks logic and remove useless parameter.
mangguo321
committed
Fix clang-format issue
mangguo321
committed
optimize use_softmax_sparse_mask check

liubo-intel
authored and
mangguo321
committed
Apply suggestions from code review

liubo-intel
authored and
mangguo321
committed
Add test case.
mangguo321
committed
Fix Clang-format issue
mangguo321
committed
Avoid repeated brgemm kernel creation and temp buffer creation
mangguo321
committed
Fix ARM cross compile
mangguo321
committed
Fix CI failure
mangguo321
committed
Fix arm compile error
mangguo321
committed
Fix RISCV64 error
mangguo321
committed
Change sum_block logic to handle column num.
mangguo321
committed
Add const to unchanged variables. Set brgemm output buffer.
mangguo321
committed
fix arm testcase fail issue: Fix ARM testcase failure: Retain the C++ implementation of scale_add2_reduce_max not only as a reference but also as a tail processing step (primarily for ARM).

liubo-intel
authored and
mangguo321
committed
Add test case compare with reference implementation
mangguo321
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[CPU] Implement X-Attention for intel CPU#32086

[CPU] Implement X-Attention for intel CPU#32086
mangguo321 wants to merge 28 commits intoopenvinotoolkit:masterfrom
mangguo321:mang/xatt

Commits on Nov 20, 2025

Enable xattention for intel CPU

Integrate mask to PA

Support multiple sequences for xattention

Fix find block issue

Copy key buffer before executing GEMM and bf16 precision support.

revert pa conflict changes with sparse attetnion execution

enable sparse attetnion execution

Use optimized softmax kernel in X-Attention

Optimize block sum using hardware instructions

Add macro definications to not support ARM platform

modify xattention_block_size and PA block_size check and scale calculation places

optimize sparse softmax mask in sparse mask attention: extend attn_softmax_kernel(float) to support sparse mask

Add const to unchanged variable

Minor changes in find blocks logic and remove useless parameter.

Fix clang-format issue

optimize use_softmax_sparse_mask check

Apply suggestions from code review

Add test case.

Fix Clang-format issue

Avoid repeated brgemm kernel creation and temp buffer creation

Fix ARM cross compile

Fix CI failure

Fix arm compile error

Fix RISCV64 error

Change sum_block logic to handle column num.

Add const to unchanged variables. Set brgemm output buffer.

fix arm testcase fail issue: Fix ARM testcase failure: Retain the C++ implementation of scale_add2_reduce_max not only as a reference but also as a tail processing step (primarily for ARM).

Add test case compare with reference implementation

[CPU] Implement X-Attention for intel CPU#32086mangguo321 wants to merge 28 commits intoopenvinotoolkit:masteropenvinotoolkit/openvino:masterfrom mangguo321:mang/xattmangguo321/openvino:mang/xattCopy head branch name to clipboard

Commits

Commits on Nov 20, 2025

[CPU] Implement X-Attention for intel CPU#32086
mangguo321 wants to merge 28 commits intoopenvinotoolkit:masterfrom
mangguo321:mang/xatt