Alexander Borzunov

Hi @arikon, Apparently, the sequence manager contains a thread that is not restarted in a child process after fork. I'd recommend to recreate the sequence manager in all children processes...

@arikon, your last code works well for me (see the full snippet [here](https://gist.github.com/borzunov/42a4b60233500d12ce43c9a4b2eb5c40)). Note that you don't have to call `.share_memory()` since fork uses copy-on-write (it already "shares" the memory...

Hi @SephReed, Can you share your OS and if you're using Anaconda or Docker? If it's Anaconda, can you please try Docker instead? I believe it's a CUDA/PyTorch installation issue,...

@SephReed, can you please share what this command prints? ```bash python -c "import torch; print(torch.cuda.is_available())" ```

I have just shipped the new PyPI release (v1.1.2), the notebooks should now work as is.

...and one more fix was merged with #248.

Hi @ryanshrott, We don't have convenient methods for this yet but you can save tensors directly, then load them to a newly created model, as described here: https://github.com/bigscience-workshop/petals/issues/268#issuecomment-1439610005 Methods for...

Hi @ryanshrott, That code is for saving to a file, but you can create a repo and upload it to model hub using the `huggingface_hub` library. The methods for uploading...

Hi @HuYang719! It seems that you're using the latest version of hivemind (1.0.0 or newer), while this tutorial was created for an old version (0.9.9.post1, the latest version at the...

I've drafted the updated [NAT traversal tutorial](https://github.com/yandex-research/DeDLOC/blob/hivemind-1.0.0/p2p/NAT-traversal.md) for the latest version of hivemind (1.0.0). Later, I'll update the circuit relay tutorial too and integrate them into our main [docs &...