Skip to content

Conversation

Green-Sky
Copy link
Contributor

No description provided.

@wbruna
Copy link
Contributor

wbruna commented Aug 23, 2025

Related: #760 .

@Green-Sky
Copy link
Contributor Author

@wbruna right, you where first. 😄

@rmatif
Copy link
Contributor

rmatif commented Aug 23, 2025

This will solve this issue also #757

@wbruna
Copy link
Contributor

wbruna commented Aug 23, 2025

@wbruna right, you where first. 😄

Oh, my PR doesn't actually bump SD_TYPE_COUNT, it just makes it easier to upgrade/downgrade ggml without dealing with that :-)

@Green-Sky
Copy link
Contributor Author

Green-Sky commented Sep 5, 2025

CUDA fp16 sd1 performance degraded ggml-org/ggml@323951f...5fdc78f
diffusion from ~1.6it/s to ~1.4it/s
vae from ~0.7s to ~0.8s
(diffusion fa on/off sees similar degradation)

Nevermind, I think my gpu just got hot. 😅

@wbruna
Copy link
Contributor

wbruna commented Sep 5, 2025

CUDA fp16 sd1 performance degraded ggml-org/[email protected]
diffusion from ~1.6it/s to ~1.4it/s
vae from ~0.7s to ~0.8s
(diffusion fa on/off sees similar degradation)

Nevermind, I think my gpu just got hot. 😅

Well, this did help me notice FA on the wan branch makes full-black images for SD1.5 🙁

I also noticed a consistent slowdown for the Conv2D VAE, but it's very small (~3%).

@wbruna
Copy link
Contributor

wbruna commented Sep 6, 2025

Well, this did help me notice FA on the wan branch makes full-black images for SD1.5 🙁

Huh, except it's working on master now - and giving a nice 20% speed boost. Chroma is now working with FA, too.

(btw, this PR was superseded by #778 )

@Green-Sky Green-Sky closed this Sep 6, 2025
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.

3 participants