meta-llama-3-yarn 0.1.3

YaRN Implementation compatible with HuggingFace Transformers

meta-llama-3-yarn

Getting Started

Install with pip

pip install meta-llama-3-yarn

Build from source

You can also build and install meta-llama-3-yarn from source:

git clone https://github.com/MeetKai/meta-llama-3-yarn.git
cd meta-llama-3-yarn
pip install -e .

Usage

To use a Llama-3 model regardless whether it is a YaRN-scaled model:

import torch
from transformers import AutoTokenizer
from meta_llama_3_yarn import LlamaForCausalLM

model_name = "meta-llama/Meta-Llama-3-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = LlamaForCausalLM.from_pretrained(
    model_name, device_map="auto", torch_dtype=torch.bfloat16
)

messages = [
    {"role": "user", "content": "Write a piece of quicksort code in C++"}
]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)

result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)

To specifically scale a model using YaRN:

import torch
from transformers import AutoTokenizer
from meta_llama_3_yarn import LlamaForCausalLM, LlamaConfig

model_name = "meta-llama/Meta-Llama-3-8B-Instruct"
context_length = 32768
tokenizer = AutoTokenizer.from_pretrained(model_name)
config = LlamaConfig.from_pretrained(model_name)
config.rope_scaling = {"type": "yarn", "factor": context_length / config.max_position_embeddings}
model = LlamaForCausalLM.from_pretrained(
    model_name, config=config, device_map="auto", torch_dtype=torch.bfloat16
)

messages = [
    {"role": "user", "content": "Write a piece of quicksort code in C++"}
]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)

result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)