Skip to content

feat:add mmdit flash attn support#1269

Merged
SWivid merged 2 commits into
SWivid:mainfrom
ZhikangNiu:main
Feb 26, 2026
Merged

feat:add mmdit flash attn support#1269
SWivid merged 2 commits into
SWivid:mainfrom
ZhikangNiu:main

Conversation

@ZhikangNiu
Copy link
Copy Markdown
Collaborator

No description provided.

@ZhikangNiu ZhikangNiu marked this pull request as draft February 23, 2026 12:40
@ZhikangNiu ZhikangNiu changed the title feat:add mmdit flash attn support [WIP]feat:add mmdit flash attn support Feb 23, 2026
@ZhikangNiu ZhikangNiu marked this pull request as ready for review February 26, 2026 08:57
@ZhikangNiu ZhikangNiu changed the title [WIP]feat:add mmdit flash attn support feat:add mmdit flash attn support Feb 26, 2026
@ZhikangNiu
Copy link
Copy Markdown
Collaborator Author

ZhikangNiu commented Feb 26, 2026

flash_attn 2.8.3

Training avg iter time(8*H200)

  • Torch backend Avg iter time:1.8476s
  • Flash Attention Backend + FusedAdamW Avg iter time: 0.6533s

Results

trained on Emilia and updated 100k (wer/sim/utmos), evaluated three different seeds

seedtts en

  • torch 11.319/0.525/3.548
  • flash attn 10.750/0.530/3.601

librispeech pc test-clean

  • torch 11.878/0.506/3.537
  • flash attn 9.647/0.518/3.634

seedtts zh

  • torch 7.684/0.668/3.026
  • flash attn 5.863/0.670/3.044

@ZhikangNiu ZhikangNiu requested a review from SWivid February 26, 2026 11:14
@ZhikangNiu
Copy link
Copy Markdown
Collaborator Author

@SWivid

@SWivid SWivid merged commit 097772c into SWivid:main Feb 26, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants