[BUG] RuntimeError, an illegal memory access was encountered

[giacomopas]

Hello,
just switched from 14900K to an AMD Ryzen 9 9950X.
I tried running a batch of profiles with NAM (like I usually did like till yesterday with the 14900K machine) and I was stopped at the first profile.

Never happened in the other machine with Intel processor.
Could it be related? How?
Any solution?

Attaching Screenshots

Desktop (please complete the following information):

• OS / Windows 11 23H2
• Local

[sdatkinson]

Hi!

Did you re-install after making the switch? If not, can you give that a try and report back?

Also, can you post the errors as copy-pasted text instead of as screencaps? That makes it easier to navigate.

[giacomopas]

Here it is attached.

I reinstalled PyTorch, then reinstalled NAM.
Nothing changed.

Thanks for your time!

unitlted1.txt

[sdatkinson]

Hmm. This isn't a bug with NAM. It looks like PyTorch's fault.

I'd recommend opening up the bug there, because there's not much I can do for you to fix that.

In the meantime, I'd maybe just recommend re-installing deeper, e.g.

• A fresh environment
• Clear Anaconda's caches
• Reinstall Anaconda (not sure if it caches (now-incorrect) information about the CPU in your computer.)

As the message says, you could also use CUDA_LAUNCH_BLOCKING=1 or compile PyTorch with TORCH_USE_CUDA_DSA--but as this implies, it's a bit low-level (hence why I'm pretty confident it's not an issue with NAM's Python code).

E.g. it doesn't seem right that there's MKL things being installed--if I'm not mistaken, those aren't supported by non-Intel CPUs.

[giacomopas]

Ok, I’ll open a bug report there. In the meantime, I tried completely uninstalling Anaconda and reinstalling everything from scratch, but now I'm getting a different error.

unitlted2.txt