Skip to main content


Banana provided serverless GPU inference for AI models, a CI/CD build pipeline and a simple Python framework (Potassium) to server your models.

This page covers how to use the Banana ecosystem within LangChain.

Installation and Setup​

  • Install the python package banana-dev:
pip install banana-dev
  • Get an Banana api key from the dashboard and set it as an environment variable (BANANA_API_KEY)
  • Get your model's key and url slug from the model's details page.

Define your Banana Template​

You'll need to set up a Github repo for your Banana app. You can get started in 5 minutes using this guide.

Alternatively, for a ready-to-go LLM example, you can check out Banana's CodeLlama-7B-Instruct-GPTQ GitHub repository. Just fork it and deploy it within Banana.

Other starter repos are available here.

Build the Banana app​

To use Banana apps within Langchain, you must include the outputs key in the returned json, and the value must be a string.

# Return the results as a dictionary
result = {'outputs': result}

An example inference function would be:

def handler(context: dict, request: Request) -> Response:
"""Handle a request to generate code from a prompt."""
model = context.get("model")
tokenizer = context.get("tokenizer")
max_new_tokens = request.json.get("max_new_tokens", 512)
temperature = request.json.get("temperature", 0.7)
prompt = request.json.get("prompt")
prompt_template=f'''[INST] Write code to solve the following coding problem that obeys the constraints and passes the example test cases. Please wrap your code answer using ```:
input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
output = model.generate(inputs=input_ids, temperature=temperature, max_new_tokens=max_new_tokens)
result = tokenizer.decode(output[0])
return Response(json={"outputs": result}, status=200)

This example is from the file in CodeLlama-7B-Instruct-GPTQ.


from langchain_community.llms import Banana
API Reference:Banana

See a usage example.

Was this page helpful?

You can leave detailed feedback on GitHub.