config.json file missing
#1
by
Ashmal
- opened
zrthxn/vicuna-embedding does not appear to have a file named config.json. Checkout 'https://huggingface.co/zrthxn/vicuna-embedding/main' for available files.
Config file for this doesn't exist because this doesnt work with a existing HF model. You have to manually load the model and tokenizer.
Something like...
from torch import nn
# ...
tokenizer = LlamaTokenizer.from_pretrained("...")
# ...
embed = nn.Embedding(config.vocab_size, config.hidden_size, config.padding_idx)
with open(Path(base_path).joinpath("llama_embed.bin"), "rb") as f:
state_dict = torch.load(f, map_location=device)
embed.load_state_dict(state_dict, strict=True)