8000 upload 2nd version of sdpa backward by yuankuns · Pull Request #546 · intel/sycl-tla · GitHub
[go: up one dir, main page]
Skip to content

upload 2nd version of sdpa backward#546

Open
yuankuns wants to merge 71 commits intointel:mainfrom
yuankuns:sdpabackward
Open

upload 2nd version of sdpa backward#546
yuankuns wants to merge 71 commits intointel:mainfrom
yuankuns:sdpabackward

Conversation

@yuankuns
Copy link
@yuankuns yuankuns commented Oct 3, 2025 •
edited
Loading
  1. Enable flash attention 2 backward
  2. layout bhsd/bshd
  3. dtype fp16/bf16
  4. is_causal done,
  5. GQA done
  6. NUM_HEAD=1 done
  7. atomic add done

@Antonyvance Antonyvance added the redesign required Implementation require a redesign label Oct 17, 2025
yuankuns added 20 commits October 24, 2025 13:40
@yuankuns
refactor the dq dk dv
0b140 8000 c8
@yuankuns
fix error in host compilation
3492d71
@yuankuns
fix acc issue in loading lse when m is not even
94d8abc
@yuankuns
enable is causal
9afbe85
@yuankuns
enable tri-dao style gqa/mqa
4d48a5b
@yuankuns
enable parallel over seqlen_kv
78b3652
@yuankuns
align n_block calculation to m_block
aba5590
@yuankuns
fix tile shape for 128 headdim
18a8441
@yuankuns
Revert "fix tile shape for 128 headdim"
3bc283e 
This reverts commit 18a8441.
@yuankuns
fix oneapi 2025.3 warning; enable 64x64 tileing for 128 and converter…
b0deafd 
… issue
@yuankuns
enable double buffer of p/ds in gmem
dbc3290
@yuankuns
fix 2d load gap in bmg & pvc
34ca80e
@yuankuns
remove useless buff
1f49712
@yuankuns
reduce dq atomic add operation
4e1ba37
@yuankuns
enable bottom right mask
9262749
@yuankuns
set stride 1 for n_block pickup
f26515e
@yuankuns
isolate each block
360041b
@yuankuns
upload 2nd version of sdpa backward
88bc40a
@yuankuns
upload cutlass dot_do_o
e3b12eb
@yuankuns
fix dp save issue
1e0889a
@yuankuns
fix oneapi 2025.3 warning; enable 64x64 tileing for 128 and converter…
a8626c8 
… issue
@yuankuns
enable double buffer of p/ds in gmem
dcc5395
@yuankuns
fix 2d load gap in bmg & pvc
2393e51
@yuankuns
remove useless buff
dc09692
@yuankuns
reduce dq atomic add operation
c24e792
@yuankuns
enable bottom right mask
2961e29
@yuankuns
set stride 1 for n_block pickup
3c0aed0
@yuankuns
isolate each block
c600aa7
@yuankuns
Merge branch 'sdpabackward' of github.com:yuankuns/cutlass-sycl into …
17448d7 
…sdpabackward
@yuankuns
move to reorder API
337b570
@yuankuns
fix register spill issues
9837bdb
@yuankuns
remove register spill when no causal
3134c78
@yuankuns
optimize o dot do
ed06e28
@yuankuns
rearrange gemm sequence
3a3fdad
@yuankuns
fix row pad of P/dS on pvc/bmg
3da1719
@yuankuns
fix row number not dividable by 16
93eaa3b
@yuankuns
optimize dq conversion
a6b5a9a
@yuankuns
isolate kernel into header files
37531a3
@yuankuns
add benchmark script
bd5cb10
@yuankuns
add reference code
64a3a6e
@yuankuns
upload gpu version reference code
682aa62
@yuankuns
optimize gpu perf of reference code
9d13f0f
@yuankuns
fix lse/out generation with calculation in reference
2176788
@yuankuns
fix is_causal nan in refence code
d938b01
@yuankuns
fix reference with host fwd
C349
bd244b8
@yuankuns
optimize scale_apply_exp2 with inline asm
0406af7
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

redesign required Implementation require a redesign

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@yuankuns @Antonyvance
0