This is open_llama_13b sharded into 2 GB shards, and in 8-bit precision using bitsandbytes==0.38.0. Please refer to the original model card for details.

loading

pip install -U -q sentencepiece transformers accelerate bitsandbytes

load the model and tokenizer:

python

import torch
from transformers import LlamaTokenizer, LlamaForCausalLM

model_name = "ethzanalytics/open_llama_13b-sharded-8bit"
tokenizer = LlamaTokenizer.from_pretrained(model_name, use_fast=False)
model = LlamaForCausalLM.from_pretrained(
    model_name,
    load_in_8bit=True,
    device_map="auto",
)