Views
No views yet
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from transformers.models.llama import modeling_llama
3
4
5def nope_monkey_patch(q, k, cos, sin, position_ids, unsqueeze_dim=1):
6 return q, k
7
8
9modeling_llama.apply_rotary_pos_emb = nope_monkey_patch
10
11model_path = "AntNLP/TinyLlama-NoPE-1.1B"
12tokenizer = AutoTokenizer.from_pretrained(model_path)
13model = AutoModelForCausalLM.from_pretrained(model_path).cuda()
14
15input_ids = tokenizer("Hello, TinyLlama-NoPE", return_tensors="pt").input_ids.cuda()
16output = model.generate(input_ids, do_sample=True, max_length=50)
17print(tokenizer.decode(output[0], skip_special_tokens=True))@misc{wang2024length,
title={Length Generalization of Causal Transformers without Position Encoding},
author={Jie Wang and Tao Ji and Yuanbin Wu and Hang Yan and Tao Gui and Qi Zhang and Xuanjing Huang and Xiaoling Wang},
year={2024},
eprint={2404.12224},
archivePrefix={arXiv},
primaryClass={cs.CL}
}