Unintended error caused by compiling fused_kernels about megatron-lm HOT 4 CLOSED

It's on the list but pretty low priority since we seldom run into it. If you can figure out why the lock mechanism in torch.utils.cpp_extension.load doesn't always work as I think it should I'm all ears. :) Might be we are expecting more from it than it is meant to handle, might be we are using it wrong, might be a bug in torch, etc.

from megatron-lm.

We've had reports of this happening internally as well. The problem with only compiling on the master rank is we need to support the use case that the code lives locally on several nodes (i.e. a multinode environment and the code isn't on a shared disk). So the kernels need to be compiled on several ranks.

torch.utils.cpp_extension.load is supposed to lock the directory so that only one rank runs the build at a time (and only the first run actually does anything, subsequent runs just see that it is already built and returns). Apparently that lock isn't properly deleted/released in some instances.

from megatron-lm.

Thanks for your response! I'm glad that your group already had reports of this happening.
Is your group internally planning to resolve this issue?

from megatron-lm.

I will close this issue, and if there is an additional issue then I will re-open it. Thank you :)

from megatron-lm.