grazier
1.0.0
Grazier is a Python library for easily calling large language models from a unified API.
From OpenAI:
From Anthropic:
From Google/GCP:
From Huggingface
From Facebook (via Huggingface)
From Stanford (via Huggingface)
From Berkeley (via Huggingface)
From StabilityAI (via Huggingface)
From AllenAI (via Huggingface)
From AI21
Image/Language Models:
Grazier can easily be installed using pip:
pip install grazierEach of the LLMs may need additional setup, which you can find in the engine setup section below.
For completion engines, it's as simple as:
import grazier
grazier.list_models()
['gptj-6B', 'gpt2', 'gpt2-med', 'gpt2-lg', 'gpt2-xl', 'distilgpt2', 'gptneo-125M', 'gptneo-1.3B', 'gptneo-2.7B', 'stablelm-3B', 'stablelm-7B', 'opt-125M', 'opt-350M', 'opt-1.3b', 'opt-2.7b', 'opt-6.7b', 'opt-13b', 'opt-30b', 'opt-66b', 'llama-7B', 'llama-13B', 'llama-30B', 'llama-65B', 'gpt3-davinci3', 'gpt3-davinci2', 'gpt3-curie', 'gpt3-babbage', 'gpt3-ada', 'palm']
gpt2 = grazier.get("gpt2")
completion = gpt2("I enjoy walking with my cute dog, but sometimes he gets scared and")
print(completion)For chat engines, all you need to do is add the type="chat" parameter:
from grazier import Conversation, Speaker, get, list_models
conversation = Conversation()
conversation.add_turn("You are a funny person.", speaker=Speaker.SYSTEM)
conversation.add_turn("Hi, how are you?", speaker=Speaker.USER)
conversation.add_turn("I am doing well, how about you?", speaker=Speaker.AI)
conversation.add_turn("What are you planning to do today?", speaker=Speaker.USER)
list_models(type="chat")
['claude', 'claude-100k', 'claude-instant', 'claude-instant-100k', 'bard', 'koala-7b', 'koala-13b-v1', 'koala-13b-v2', 'vicuna-7b', 'vicuna-13b', 'alpaca-13b', 'chat-gpt', 'gpt4', 'gpt4-32k', 'stablelm-3b', 'stablelm-7b', 'palm']
gpt4 = get("gpt4", type="chat")
next_turn = gpt4(conversation)
print(next_turn)For vision-augmented (image) engines, use type="image"
import grazier
from PIL import Image
grazier.list_models(type="image")
['blip2-opt-2.7b', 'blip2-opt-6.7b', 'blip2-opt-2.7b-coco', 'blip2-opt-6.7b-coco', 'blip2-flan-t5-xl', 'blip2-flan-t5-xxl', 'blip2-flan-t5-xl-coco', 'openflamingo-3b-vitl-mpt1b', 'openflamingo-3b-vitl-mpt1b-dolly', 'openflamingo-9b-vitl-mpt7b', 'openflamingo-4b-vitl-rpj3b']
blip2 = grazier.get("blip2-opt-2.7b", type="image")
image = Image.open('test_data/dog.jpg')
completion = blip2(image, "A photo of")
print(completion)Each engine may require some specific details to be passed in. For example, OpenAI engines require an API key. These details are generally set up with environment variables.
For OpenAI engines, you will need to set the OPENAI_API_KEY and OPENAI_API_ORG environment variables. You can find
your API key and organization ID on the OpenAI dashboard. You can set these environment
variables in your shell or in a .env file in the root of your project. For example, in a .env file, you would have:
OPENAI_API_KEY=<your key>
OPENAI_API_ORG=<your org id>or on the command line:
export OPENAI_API_KEY=<your key>
export OPENAI_API_ORG=<your org id>For Anthropic engines, you will need to set the ANTHROPIC_API_KEY environment variable. You can find your API key at
the Anthropic dashboard. You can set this environment variable in
your shell or in a .env file in the root of your project. For example, in a .env file, you would have:
ANTHROPIC_API_KEY=<your key>or on the command line:
export ANTHROPIC_API_KEY=<your key>For Google engines, we use the Vertex cloud API, which requires a Google Cloud Platform (GCP) project. You can create a
GCP project at the GCP console. You will also need to enable the Vertex AI API for
your project, set up a services account, and download the account JSON credentials. You can find instructions for that
following steps 1 to 6 of the tutorial here.
Finally, you will need to set the GOOGLE_APPLICATION_CREDENTIALS environment variable to the path of the JSON file.
You can set this environment variable in your shell or in a .env file in the root of your project. For example, in a
.env file, you would have:
GOOGLE_APPLICATION_CREDENTIALS=<path to your JSON file>or on the command line:
export GOOGLE_APPLICATION_CREDENTIALS=<path to your JSON file>For the Bard engine, you will need to get your Bard __Secure-1PSID and __Secure-1PSIDTS tokens. Get the value of this variable by first going to https://bard.google.com/, then log in, press F12 for console, and go to the "Application" tab, then "Cookies", then copy the value of the "__Secure-1PSID" and "__Secure-1PSIDTS" cookies. You can then set the environment variables:
BARD__Secure_1PSID=<your session id>
BARD__Secure_1PSIDTS=<your session id timestamp>Most of the huggingface engines require no additional setup, however, some of the larger models require a GPU to run with any kind of efficiency (and some require multiple GPUs with large amounts of memory). You can find more details about the requirements for each model on the Huggingface model hub.
For these engines, you will need to obtain and postprocess the weights yourself (due to Facebook's licensing). You can find the instructions for doing so on each model page:
Once the weights have been downloaded and processed, you can set the following environment variables to the root
directory containing a folder for each variant (The format is, {root_dir}/{model-prefix}/weights.bin, the root directory would
be root_dir, and the model-prefix is the name of the model, e.g. tulu-65b):
LLAMA_WEIGHTS_ROOT=<path to the llama weights>
ALPACA_WEIGHTS_ROOT=<path to the alpaca weights>
KOALA_WEIGHTS_ROOT=<path to the koala weights>
VICUNA_WEIGHTS_ROOT=<path to the vicuna weights>
ALLENAI_WEIGHTS_ROOT=<path to the allenai weights>For AI21 models, you will need to set the AI21_API_KEY environment variable. You can find your API key at
the AI21 Studio Dashboard. You can set this environment variable in
your shell or in a .env file in the root of your project. For example, in a .env file, you would have:
AI21_API_KEY=<your key>or on the command line:
export AI21_API_KEY=<your key>If you use grazier in your work, please cite:
@misc{grazier,
author = {David Chan},
title = {grazier: Easily call Large Language Models from a unified API},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {url{
https://github.com/DavidMChan/grazier
}}
}
grazier is licensed under the terms of the MIT license. See LICENSE for more information.
