-
-
Notifications
You must be signed in to change notification settings - Fork 780
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
Version 2.3.6 of flash-attn Is Incompatible With axolotl #911
Comments
Did you mean flash-attn instead of accelerate? |
Apologies - you are correct. I'll edit the issue accordingly. |
I experienced this too. Downgrading to 2.3.3 solved it for now. |
FYI, I hit this in the container ( |
I got the same. I think the issue may be that flash attention wheels are built using the nightly build of PyTorch. |
Until this is fixed, it's easy to install a working version:
|
I am experiencing this issue in a colab pro+ enviroment, despite using flash-attn 2.3.3, for me, flash-attn 2.3.2 worked. Output of ´pip list´
|
I am getting this with flash attention 2.3.3. |
Installing from source worked for me:
|
I have this same issue |
Same issue with flash_attn 2.5.9.post1 and 2.3.3. |
@nctu6 what version of PyTorch? |
The incompatible problem occurs with flash_attn versions 2.5.9.post1 and 2.3.3 on H100. axolotl 0.4.0
|
Please check that this issue hasn't been reported before.
Expected Behavior
Installation via the git instructions in a clean conda environment was successful. I should be able to execute scripts using
axolotl.cli.inference
.Current behaviour
Executing scripts results in an
ImportError
.This error has been recently reported by "Oobabooga users".
Steps to reproduce
Config yaml
Possible solution
Rolling back to flash-attn==2.3.3 resolves the issue.
Which Operating Systems are you using?
Python Version
3.9
axolotl branch-commit
main/4d6490b
Acknowledgements
The text was updated successfully, but these errors were encountered: