Mengtao Yuan

@l3utterfly Could you share the command you used in torchtune, as well as the export_llama?

@l3utterfly I can see two options, 1. Have you tried directly safe_open your .safetensors file, and use the codes @mergennachin suggested? Then you could use torch.save to save the resulting...

@l3utterfly Let me try to convert the safetensor files and let you know if there's a way to workaround. > I tried to fine tune with torchtune, but it appears...

@l3utterfly I took a deeper look into the state dict (`sd`) in your code. There are two issues. 1. The difference is that, the checkpoint returned by `checkpointer.load_checkpoint()` is a...

Update: After chatting with @kartikayk , we need another convert from torchtune to meta's llama3 format. So the code below should work, ```Python from torchtune.utils import FullModelHFCheckpointer from torchtune.models import...

@l3utterfly It's a great idea. Let me put up a PR for this with documentations.

cc @manuelcandales @SS-JIA

@l3utterfly to clarify, if we can do a "stack" style on the kv cache. For example, if there's original prompt and outputs. Later we can feed the model with the...

@mhs4670go Yes, ExecuTorch is pinned to a specific PyTorch version. You may run ./install_requirements to set it up.

@salykova Thank you for the list! We may picked Phi3 as it's relatively new and popular, but are definitely considering enabling other models. The long term goal is that we...