Very slow generation on V100 and ViT-B-32::laion2b_e16

[adieyal]

Version 0.7.0

I haven't tested other weights but ViT-B-32::laion2b_e16 is running extremely slowly on 0.7.0

Here is a minimal test:

create(clip_models=[ "ViT-B-32::laion2b_e16"])

I am getting around 32 seconds per iterations whereas with previous versions it would be less than 2 seconds.

[hanxiao]

i see, let me test it

[hanxiao]

seems fine to me

NVIDIA GeForce RTX 3090 (24576 RAM)

[adieyal]

Weird, I'll retry on another gpu

On Tue, 19 Jul 2022, 18:30 Han Xiao, @.***> wrote:

seems fine to me

[image: 071e1aa4f9bf9e1c7d4ede51eb546157] https://user-images.githubusercontent.com/2041322/179801950-0b87e1f5-d961-43e2-ad2a-e8c4dff63417.gif

NVIDIA GeForce RTX 3090 (24576 RAM)

— Reply to this email directly, view it on GitHub https://github.com/jina-ai/discoart/issues/50#issuecomment-1189307331, or unsubscribe https://github.com/notifications/unsubscribe-auth/AADGXMSUY77G2VR35W2LW3LVU3J23ANCNFSM54AQ6MCA . You are receiving this because you authored the thread.Message ID: @.***>

[adieyal]

Tested with A4000 and P100 - they work fine as well. I'll wait until I get another V100 to check in case there is something weird with the instance I'm currently using

[thegenerativegeneration]

Yesterday I had the same problem. Iterations went from 3s to 30s. But only after about 200 iterations. GPU was probably a P100. Diffusion model was custom.

[hanxiao]

closing as cant reproduce, if you find out anything, feel free to reopen

[adieyal]

Happens on another V100. Given that it is only this GPU, perhaps it should just be mentioned as a caveat in the documentation rather than looking for a fix?