The Universal Model Adapter

This package acts as an adapter between Huggingface Transformers and several different APIs. As of now, these are the Huggingface Inference API and the OpenAI Inference API.

This works by mock transformers.PreTrainedModel classes that share the same generate() method, but make API calls on the backend. Several dev models are also available for mocking generation or performing debugging tasks.

Use Case

This package is best used in projects that use multiple different model sources interchangeably. In these kinds of projects, a unified generation interface greatly simplifies a lot of code. For example, a project that uses text generated from both Huggingface models and GPT models from OpenAI's API.

Fine-Gained Source Control

An advantage of this package is that it can either automatically resolve the source of a model from its name, or you can specify the source (OpenAI, Huggingface, etc.) manually. This can be done through an extra parameter to the pretrained_from_...() methods. For example:

from universalmodels.interface import pretrained_from_name, ModelSrc

model_name = "mistralai/Mistral-7B-v0.1"
# This will automatically resolve the model's source to 
# a local Huggingface transformers model (ModelSrc.HF_LOCAL)
local_model, tokenizer = pretrained_from_name(model_name, model_src=ModelSrc.AUTO)

# This will attempt to start the FastChat service and run 
# a local instance of the OpenAI API to run optimized generation
fschat_model, tokenizer = pretrained_from_name(model_name, model_src=ModelSrc.OPENAI_API)

# This will create a mock model without any generation logic attached.
# This is useful for when the shell of a model is needed as a reference.
# This option does not load any local models into memory or activate FastChat.
mock_model, tokenizer = pretrained_from_name(model_name, model_src=ModelSrc.NO_LOAD)

Quick Start

Installing from PyPI

pip3 install "universalmodels[fastchat]"

Installing from Source

git clone https://github.com/matthew-pisano/UniversalModels
cd UniversalModels
pip3 install -e ".[fastchat]"

Installing the fastchat extra enables support for using fastchat on compatible locally installed huggingface models. See FastChat supported models for more information on which models are supported.

Example Usage

In the following example, note that the interfaces for the Huggingface and OpenAI modles are the same. This is the primary benefit of using this package.

import torch
from universalmodels import pretrained_from_name
from universalmodels.constants import set_seed

# Set the global seed to encourage deterministic generation 
# NOTE: DOES NOT affect OpenAI API models
set_seed(42)

# Huggingface model example
hf_model_name = "mistralai/Mixtral-8x7B-Instruct-v0.1"
hf_model, hf_tokenizer = pretrained_from_name(hf_model_name)

hf_tokens = hf_tokenizer.encode("Repeat the following: 'Hello there from a huggingface model'")
hf_resp_tokens = hf_model.generate(torch.Tensor([hf_tokens]).int())[0]
hf_response = hf_tokenizer.decode(hf_resp_tokens)
print(hf_response)

# OpenAI model example
oai_model_name = "openai/gpt-3.5"
oai_model, oai_tokenizer = pretrained_from_name(oai_model_name)

oai_tokens = oai_tokenizer.encode("Repeat the following: 'Hello there from an openai model'")
oai_resp_tokens = oai_model.generate(torch.Tensor([oai_tokens]).int())[0]
oai_response = oai_tokenizer.decode(oai_resp_tokens)
print(oai_response)

Important

Make sure your API keys are set for OpenAI and Huggingface before using models that require them!