Tutorial - Deploy Llama-3.2-3B-Instruct using Inferless

Llama-3.2-3B-Instruct is Meta’s streamlined, 3B small language model fine-tuned for instruction following and conversational applications. This model is optimized for multilingual dialogue, retrieval, summarization, and prompt rewriting, delivering impressive performance despite its modest size. Its efficient, auto-regressive transformer architecture with innovations like Grouped-Query Attention ensures low latency and ease of deployment, making it ideal for local inference on a wide range of hardware, from edge devices to personal laptops. Whether used as a base for further fine-tuning or as a ready-to-use conversational agent, Llama-3.2-3B-Instruct offers a compelling balance of speed, capability, and accessibility.

TL;DR:

• Deployment of Llama-3.2-3B-Instruct model using vLLM.
• Dependencies defined in inferless-runtime-config.yaml.
• GitHub/GitLab template creation with app.py, inferless-runtime-config.yaml and inferless.yaml.
• Model class in app.py with initialize, infer, and finalize functions.
• Custom runtime creation with necessary system and Python packages.
• Recommended GPU: NVIDIA A100.
• Custom runtime selection in advanced configuration.
• Final review and deployment on the Inferless platform.

Fork the Repository

Get started by forking the repository. You can do this by clicking on the fork button in the top right corner of the repository page.

This will create a copy of the repository in your own GitHub account, allowing you to make changes and customize it according to your needs.

Create a Custom Runtime in Inferless

To access the custom runtime window in Inferless, simply navigate to the sidebar and click on the Create new Runtime button. A pop-up will appear.

Next, provide a suitable name for your custom runtime and proceed by uploading the inferless-runtime.yaml file given above. Finally, ensure you save your changes by clicking on the save button.

Add Your Hugging Face Access Token

Go into the inferless.yaml and replace <hugging_face_token> with your hugging face access token. Make sure to check the repo is private to protect your hugging face token.

Import the Model in Inferless

Log in to your inferless account, select the workspace you want the model to be imported into and click the Add Model button.

Select the PyTorch as framework and choose Repo(custom code) as your model source and select your provider, and use the forked repo URL as the Model URL.

Enter all the required details to Import your model. Refer this link for more information on model import.

---

Curl Command

Following is an example of the curl command you can use to make inference. You can find the exact curl command in the Model's API page in Inferless.

curl --location '<your_inference_url>' \
    --header 'Content-Type: application/json' \
    --header 'Authorization: Bearer <your_api_key>' \
    --data '{
      "inputs": [
        {
          "name": "prompt",
          "shape": [1],
          "data": ["What is deep learning?"],
          "datatype": "BYTES"
        },
        {
          "name": "temperature",
          "optional": true,
          "shape": [1],
          "data": [0.7],
          "datatype": "FP32"
        },
        {
          "name": "top_p",
          "optional": true,
          "shape": [1],
          "data": [0.1],
          "datatype": "FP32"
        },
        {
          "name": "repetition_penalty",
          "optional": true,
          "shape": [1],
          "data": [1.18],
          "datatype": "FP32"
        },
        {
          "name": "max_tokens",
          "optional": true,
          "shape": [1],
          "data": [512],
          "datatype": "INT16"
        },
        {
          "name": "top_k",
          "optional": true,
          "shape": [1],
          "data": [40],
          "datatype": "INT8"
        }
      ]
    }'

---

Customizing the Code

Open the app.py file. This contains the main code for inference. The InferlessPythonModel has three main functions, initialize, infer and finalize.

Initialize - This function is executed during the cold start and is used to initialize the model. If you have any custom configurations or settings that need to be applied during the initialization, make sure to add them in this function.

Infer - This function is where the inference happens. The infer function leverages both RequestObjects and ResponseObjects to handle inputs and outputs in a structured and maintainable way.

• RequestObjects: Defines the input schema, validating and parsing the input data.
• ResponseObjects: Encapsulates the output data, ensuring consistent and structured API responses.

  def infer(self, request: RequestObjects) -> ResponseObjects:
    sampling_params = SamplingParams(temperature=request.temperature,top_p=request.top_p,repetition_penalty=request.repetition_penalty,
                     top_k=request.top_k,max_tokens=request.max_tokens)
    input_text = self.tokenizer.apply_chat_template([{"role": "user", "content": request.prompt}], tokenize=False)
    result = self.llm.generate(input_text, sampling_params)
    result_output = [output.outputs[0].text for output in result]

    generateObject = ResponseObjects(generated_text = result_output[0])        
    return generateObject

Finalize - This function is used to perform any cleanup activity for example you can unload the model from the gpu by setting to None.

def finalize(self):
    self.llm = None

For more information refer to the Inferless docs.