Make xformers optional even if it is available#1753
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
|
cc @patil-suraj @pcuenca, I tend to agree here with @kn . If we make it fully optional we should however not wrap it into a "try .... else ..." statement, but just let it fail if it's not available |
pcuenca
left a comment
pcuenca
left a comment
There was a problem hiding this comment.
Thanks, I think this makes sense!
|
good idea on raising an exception when xformers is used but not available! PTAL |
|
I'm in favor of this as it better follows the "no optimizations by default" API of PyTorch and should be "less surprising" for the user long-term. This change might suddenly slow down some notebooks, so also kindly pinging @apolinario and @patil-suraj here. |
patil-suraj
left a comment
patil-suraj
left a comment
There was a problem hiding this comment.
Thanks a lot for the PR, agree with it, makes sense to add an explicit argument here to enable exformers. Just left a nit. Also, let's add a note in the examples readme about this new argument.
| ), | ||
| ) | ||
| parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") | ||
| parser.add_argument("--use_xformers", action="store_true", help="Whether or not to use xformers.") |
There was a problem hiding this comment.
(nit) could we call this arg enable_xformers_memory_efficient_attention to align with the method name?
kn
force-pushed
the
make-xformers-optional
branch
from
465c1d6 to
8dc93ae
Compare
|
renamed the flag and added a note in READMEs - PTAL |
|
Thanks for the PR @kn ! |
5 participants
Problem:
xformers being available doesn't always mean it should be used. For example, I install xformers because it makes inference fast but I find xformers make my dreambooth training slower.
Solution:
Require explicit flag to enable xformers when running training scripts.
Related recent change: #1640