-
Notifications
You must be signed in to change notification settings - Fork 1.1k
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
[BUG] Crash in mlx.core.fast.scaled_dot_product_attention with matrix attention kernel #1643
Comments
It's ok if you can't write a test case. But can you provide some code or a pointer to a repo which reproduces the issue with some consistency? |
Yep, it should be reproducible this way. It downloads a small dataset (~1.5gb) and then should crash shortly after the training loop begins.
|
It crashed for me with this message:
|
Seems like it depends on the Python runtime version. I see that error in 3.11 but it crashes with this in 3.9:
|
Thank you! 🙏 |
Describe the bug
I'm seeing a crash in mlx.core.fast.scaled_dot_product_attention after the change in #1610, which is in the latest version. Rolling that specific change out fixes it, but it's large so I can't say what the specific issue is.
I tried to write a reduced test case, but it doesn't seem to trigger the issue from a single invocation. I can consistently reproduce it in the context of a larger codebase when running a training loop.
The captured inputs are here in case that's helpful.
Here's a backtrace from lldb when this happens:
Desktop (please complete the following information):
The text was updated successfully, but these errors were encountered: