You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
When I was doing some optimization for my pipeline, i found that the BrownianTree somehow took a bit more time.
Describe the solution you'd like.
I further dig into torchsde document, and found that they encouraged to use BrownianInterval to have best benefits for underlying structure utilization. The BrownianTree is actually just an abstraction layer of the BrownianInterval and as we all know, python function calls take time!
Code:
#diffusers/src/diffusers/schedulers/scheduling_dpmsolver_sde.py:41
self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
# Modified
self.trees = [torchsde.BrownianInterval(t0, t1, size=w0.shape, dtype=w0.dtype, device=w0.device, cache_size=None, entropy=s, **kwargs) for s in seed]
Is your feature request related to a problem? Please describe.
When I was doing some optimization for my pipeline, i found that the BrownianTree somehow took a bit more time.
Describe the solution you'd like.
I further dig into torchsde document, and found that they encouraged to use
BrownianInterval
to have best benefits for underlying structure utilization. TheBrownianTree
is actually just an abstraction layer of theBrownianInterval
and as we all know, python function calls take time!Code:
Additional context.
torchsde doc link
The text was updated successfully, but these errors were encountered: