Modal
This page covers how to use the Modal ecosystem to run LangChain custom LLMs. It is broken into two parts:
- Modal installation and web endpoint deployment
- Using deployed web endpoint with
LLM
wrapper class.
Installation and Setup
- Install with
pip install modal
- Run
modal token new
Define your Modal Functions and Webhooks
You must include a prompt. There is a rigid response structure:
class Item(BaseModel):
prompt: str
@stub.function()
@modal.web_endpoint(method="POST")
def get_text(item: Item):
return {"prompt": run_gpt2.call(item.prompt)}
The following is an example with the GPT2 model:
from pydantic import BaseModel
import modal
CACHE_PATH = "/root/model_cache"
class Item(BaseModel):
prompt: str
stub = modal.Stub(name="example-get-started-with-langchain")
def download_model():
from transformers import GPT2Tokenizer, GPT2LMHeadModel
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2LMHeadModel.from_pretrained('gpt2')
tokenizer.save_pretrained(CACHE_PATH)
model.save_pretrained(CACHE_PATH)
# Define a container image for the LLM function below, which
# downloads and stores the GPT-2 model.
image = modal.Image.debian_slim().pip_install(
"tokenizers", "transformers", "torch", "accelerate"
).run_function(download_model)
@stub.function(
gpu="any",
image=image,
retries=3,
)
def run_gpt2(text: str):
from transformers import GPT2Tokenizer, GPT2LMHeadModel
tokenizer = GPT2Tokenizer.from_pretrained(CACHE_PATH)
model = GPT2LMHeadModel.from_pretrained(CACHE_PATH)
encoded_input = tokenizer(text, return_tensors='pt').input_ids
output = model.generate(encoded_input, max_length=50, do_sample=True)
return tokenizer.decode(output[0], skip_special_tokens=True)
@stub.function()
@modal.web_endpoint(method="POST")
def get_text(item: Item):
return {"prompt": run_gpt2.call(item.prompt)}
Deploy the web endpoint
Deploy the web endpoint to Modal cloud with the modal deploy
CLI command.
Your web endpoint will acquire a persistent URL under the modal.run
domain.
LLM wrapper around Modal web endpoint
The Modal
LLM wrapper class which will accept your deployed web endpoint's URL.
from langchain_community.llms import Modal
endpoint_url = "https://ecorp--custom-llm-endpoint.modal.run" # REPLACE ME with your deployed Modal web endpoint's URL
llm = Modal(endpoint_url=endpoint_url)
llm_chain = LLMChain(prompt=prompt, llm=llm)
question = "What NFL team won the Super Bowl in the year Justin Beiber was born?"
llm_chain.run(question)
API Reference:Modal