What is the format of weights for the input model to convert.py?

[RonanKMcGovern]

Here is a comment from the ReadMe regarding usage of convert.py:

# obtain the original LLaMA model weights and place them in ./models

However, it's not clear what format those weights need to be in...

1. In what format do I need to have the original weights? consolidated.00.pth or .bin?
2. If .bin, do the weights need to be float 32 or can they be bf16?
3. If .bin, is there a way to handle sharded weights? e.g. Llama 2 7B sharded into ~10 GB and ~3 GB? Can I specify each shard as a command line argument here:

# convert the 7B model to ggml FP16 format
python3 convert.py models/7B/

[RonanKMcGovern]

Ok, I should have just looked at convert.py before in llama ccp.

To answer my own Qs:

• convert.py supports input models in many formats (including .pth and .bin)
• convert.py supports multiple data types, including float 32 and bf16
• convert.py even supports shards and finds them automatically (very nice!)

[Kushagrapuro]

thnx