-
Notifications
You must be signed in to change notification settings - Fork 55
do not zero out sink tokens during recache #21
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
|
Thanks for your modification! However, I believe there’s a logical issue with the change. The cache zeroing occurs under During the subsequent cache update process, whether the region In my view, the |
|
Good point! Looking at this more closely, I actually think that the zero'ing out is unnecessary since as you mention the cache entries will be overwritten during the subsequent recompute. I think that the proper fix is to actually modify https://github.com/NVlabs/LongLive/pull/21/files#diff-4ed558c68c6ec9a180de74af9d37373fdf8fff84ac40cbf34a6669ad03599a18R260 to account for whether we are using a global sink or not. |
|
This is what I get with the model shared on hugging face before this MR is applied rank0-0-0_lora.mp4And then with the fix, this is what we get with the global sink enabled: rank0-0-0_lora.mp4And this is with the fix + the global sink disabled: rank0-0-0_lora.mp4At least on this example, it seems that keeping the global sink is important for maintaining the appearance of the boy. |
|
Yeah! That’s great! In fact, based on the experimental setup described in the LongLive paper, the That said, your modification effectively fixes the behavior when |
2681bcb to
f5597f1
Compare
f5597f1 to
d69a1fa
Compare
fixes #20