Closed vikhyat closed 6 months ago
Loading the older version from the hugging face transformers commit seems to work to load the weights. Would this be sufficient or were you thinking of mapping the keys in the older model to the newer model format?
import torch
from transformers import AutoConfig, AutoModelForCausalLM
config = AutoConfig.from_pretrained('microsoft/phi-1_5',
revision='24f9ea14df973a49a0d87c16d04df88d90067468',
trust_remote_code=True)
model = AutoModelForCausalLM.from_config(config,
code_revision='24f9ea14df973a49a0d87c16d04df88d90067468',
trust_remote_code=True)
model.load_state_dict(torch.load('text_model.pt'))
Could potentially modify the config file in the model cards and the text_encoder to remove the explicit state dict loading as well.
I still think there's value in mapping the keys, that version doesn't support generation options like k-beam search IIRC
The text model weights provided are for the old version of Phi before it was integrated into the Huggingface transformers library. Need to write a script to convert the weight keys into the new format and switch to using the Huggingface implementation directly.