upload miqu ckpt to hf

miqu-upload-hf.py

from transformers import LlamaConfig as LC, LlamaForCausalLM as LLM, LlamaTokenizer as LT
from accelerate import init_empty_weights, load_checkpoint_and_dispatch
import torch
lt = LT.from_pretrained("NousResearch/Llama-2-7b-hf")
c = LC.from_pretrained("NousResearch/Llama-2-70b-hf")
c.max_position_embeddings = 32764
c.rope_theta = 1000000
with init_empty_weights(): m = LLM(c)
m = m.half().eval()
m.requires_grad_(False)

ckpt = 'hf_pm2.pt'
print(f'loading {ckpt=}')
model = load_checkpoint_and_dispatch(m, checkpoint=ckpt, device_map='auto', no_split_module_classes=["LlamaDecoderLayer"])
model.save_pretrained("./miqu-1-70b-sf", push_to_hub=True, safe_serialization=True)