Is there any more flexible way to set up the pretrained_model folder loading instead of downloading from huggingface or put into .cache/huggingface

[rflai]

like pretrained_model = "/mnt/Databases/evo-1-8k-base"

[KPHippe]

The internal code bootstraps from the HF model

evo/evo/models.py

Line 71 in 9242312

def load_checkpoint(

So you can set your HF_HOME environment variable and it will download it once to that location, and every other time will load from that folder (assuming you set the same path)

[meiyiyeshi]

My server cannot connect to the internet to download models. Can I download the model offline on another machine, then upload it to a specific directory on the server (for example, /data/model/evo-1-131k-base)?
Later, when loading the model with the following code:

from evo import Evo
import torch

device = 'cuda:0'

evo_model = Evo('evo-1-131k-base')
model, tokenizer = evo_model.model, evo_model.tokenizer
model.to(device)
model.eval()

can I simply change it to

evo_model = Evo('/data/model/evo-1-131k-base')

?
I’ve tried that, but it keeps failing and says “invalid model.”

[meiyiyeshi]

The internal code bootstraps from the HF model

evo/evo/models.py

Line 71 in 9242312

def load_checkpoint(
So you can set your HF_HOME environment variable and it will download it once to that location, and every other time will load from that folder (assuming you set the same path)