armed-gpt / gpt-blazing

gpt-blazing

This project draws inspiration from gpt-fast and applies the same performance optimization strategy to MORE models. Unlike gpt-fast, this project aims to be a “framework” or “library”.

Installation

pip install torch>=2.2.0 --index-url https://download.pytorch.org/whl/cu118
pip install gpt-blazing

Usage

Download a gpt-blazing converted model.

Original model	👇👇 gpt-blazing converted model
🤗 baichuan-inc/Baichuan2-13B-Chat	🤗 gpt-blazing/baichuan2-13b-chat
more to be supported	...

Run the following demo.

from datetime import datetime

from gpt_blazing.engine import Engine
from gpt_blazing.model.interface import Role
from gpt_blazing.model.baichuan2.inference import (
    Baichuan2ModelInferenceConfig,
    Baichuan2ModelInference,
)


init_dt_begin = datetime.now()
engine = Engine(
    Baichuan2ModelInference(
        Baichuan2ModelInferenceConfig(
            model_folder='the path of model folder you just downloaded.',
            device='cuda:0',
        )
    )
)
init_dt_end = datetime.now()
print('init:', (init_dt_end - init_dt_begin).total_seconds())

generate_dt_begin = datetime.now()
response = engine.generate([(Role.USER, "帮我写一篇与A股主题相关的作文，800字左右")])
generate_dt_end = datetime.now()
generate_total_seconds = (generate_dt_end - generate_dt_begin).total_seconds()
print('generate:', generate_total_seconds, response.num_tokens / generate_total_seconds)

print(response.content)

Performance

GPU: 3090

Model	Technique	Tokens/Second
Baichuan2 13b	INT8 (this project)	50.1
Baichuan2 13b	INT8 (huggingface)	7.9
Llama2 13b	INT8 (gpt-fast)	55.5