-
Notifications
You must be signed in to change notification settings - Fork 978
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
FIM implementation question #932
Comments
Still looking for your reply, @haileyschoelkopf. Thanks in advance. |
Hi! Sorry for the delay on this. At the time I did not thoroughly compare the different variants implemented in the paper. My rationale for not implementing joint PSM-SPM was that at inference time, one's interface to infilling (if e.g. serving a model that has FIM capabilities) can be controlled behind the scenes--there's no reason that pops out to me that explains why one would want to sometimes do PSM mode, sometimes SPM mode, at inference time. I suppose it might be the case that one mode allows more attention to the prefix vs. suffix compared to the other, but I am not certain this is likely or desirable. Does this make sense to you? Some other information that may interest you:
|
Thank you very much for your detailed and informative response! It has greatly helped me in understanding the issue and resolving my problem. |
Hello, maintainers and contributors,
I hope you are all doing well. I am writing this issue to inquire about the FIM implementation in this repository. I noticed that the implementation differs from the one described in the original paper.
The FIM paper mentioned that the joint 50-50 SPM+PSM mixture may be the most efficient and flexible than the solely SPM mode. However, this FIM-clean branch only implemented the SPM mode, with
<SUF> ◦ Enc(suffix) ◦ <PRE> ◦ Enc(prefix) ◦ <MID> ◦ Enc(middle) ◦ <EOT>
instead of
<PRE> ◦ <SUF> ◦ Enc(suffix) ◦ <MID> ◦ Enc(prefix) ◦ Enc(middle) ◦ <EOT>
Is there a specific reason for these changes, such as improvements in performance or stability?
I appreciate your time and work on this project, and I'm looking forward to understanding the rationale behind the differences in the implementation.
Thank you in advance for your help and support.
Best regards,
The text was updated successfully, but these errors were encountered: