# Banana

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

This page covers how to use the [Banana](https://www.banana.dev) ecosystem within LangChain.

It is broken into two parts: 
* installation and setup, 
* and then references to specific Banana wrappers.

## Installation and Setup

- Install with `pip install banana-dev`
- Get an Banana api key from the [Banana.dev dashboard](https://app.banana.dev) 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](https://docs.banana.dev/banana-docs/).

Alternatively, for a ready-to-go LLM example, you can check out Banana's [CodeLlama-7B-Instruct-GPTQ](https://github.com/bananaml/demo-codellama-7b-instruct-gptq) GitHub repository. Just fork it and deploy it within Banana.

Other starter repos are available [here](https://github.com/orgs/bananaml/repositories?q=demo-&type=all&language=&sort=).

## Build the Banana app

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

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

An example inference function would be:

```python
@app.handler("/")
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 ```:
    {prompt}
    [/INST]
    '''
    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 `app.py` file in [CodeLlama-7B-Instruct-GPTQ](https://github.com/bananaml/demo-codellama-7b-instruct-gptq).

## Wrappers

### LLM

Within Langchain, there exists a Banana LLM wrapper, which you can access with

```python
from langchain.llms import Banana
```

You need to provide a model key and model url slug, which you can get from the model's details page in the [Banana.dev dashboard](https://app.banana.dev).

```python
llm = Banana(model_key="YOUR_MODEL_KEY", model_url_slug="YOUR_MODEL_URL_SLUG")
```