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
I am starting to play with the relatively recent scheduler TCDScheduler and encountered a bug yesterday.
The thing is, I'm embarrassed to say I am unable to provide the exact details of the bug. Right now I am unable to reproduce it again but I'd like to make sure it gets eradicated.
Here is however information from my memory that may help reproduce it:
diffusers version : it was either 0.27.2 or 0.28.0
ubuntu 20.04, CUDA 12.2
where it crashed: here, indicating that I somehow called TCDScheduler.step with a too high eta (it was a float around 7 if I recall well). I did not call .step directly, it was called under the hood when calling my pipeline
what pipeline used: a standard txt2img sd15 or sdxl pipeline
doing what: switching schedulers randomly in the below list, to verify that configs work well and that I can infer smoothly while changing scheduler at every inference
I had a cursory glance at diffusers code and couldn't see how I managed to somehow get an invalid eta value. I saw it defaults to 0.3 in TCDScheduler.step
Do you have a hint? A way to produce the bug when switching schedulers, maybe by being clumsy about the specific config params to use?
This is perhaps best dealt with by either @mhh0318 or @vanakema as they're the main contributors to this scheduler's code. (I am aware that the code originates from an external repo (jabir-zheng)).
cc @yi
Reproduction
as stated above, no MRE unfortunately but some level of detail to help narrow down the problem
Logs
again unfortunately all I know is I ran into the `AssertionError "gamma must be less than or equal to 1.0"`
Oh man haha, I only appear like one of the main contributors because I did a huge refactoring of the Tensor type used across the repo. Someone else would definitely be better equip to answer this
Update: the error came from our implem, not diffusers'.
(FWIW, I had generated a random eta value for DDIMScheduler that was > 1 and then the scheduler was changed to TCDScheduler with the same eta value, triggering the assertionerror I linked above)
Describe the bug
Hi,
I am starting to play with the relatively recent scheduler TCDScheduler and encountered a bug yesterday.
The thing is, I'm embarrassed to say I am unable to provide the exact details of the bug. Right now I am unable to reproduce it again but I'd like to make sure it gets eradicated.
Here is however information from my memory that may help reproduce it:
TCDScheduler.step
with a too high eta (it was a float around 7 if I recall well). I did not call .step directly, it was called under the hood when calling my pipelinetxt2img
sd15 or sdxl pipelineI had a cursory glance at diffusers code and couldn't see how I managed to somehow get an invalid
eta
value. I saw it defaults to 0.3 inTCDScheduler.step
Do you have a hint? A way to produce the bug when switching schedulers, maybe by being clumsy about the specific config params to use?
This is perhaps best dealt with by either @mhh0318 or @vanakema as they're the main contributors to this scheduler's code. (I am aware that the code originates from an external repo (jabir-zheng)).
cc @yi
Reproduction
as stated above, no MRE unfortunately but some level of detail to help narrow down the problem
Logs
again unfortunately all I know is I ran into the `AssertionError "gamma must be less than or equal to 1.0"`
System Info
diffusers
version: 0.27.2 or 0.28.0Who can help?
@yiyixuxu @vanakema @mhh
The text was updated successfully, but these errors were encountered: