Tutorial - Deploy Qwen's QwQ-32B-Preview using Inferless

QwQ-32B-Preview is a large language model developed by the Qwen Team, focusing on enhancing AI's analytical and problem-solving abilities. It employs a dense transformer architecture with 32.5 billion parameters, 64 layers, and a context window of 32,768 tokens, incorporating advanced components such as Rotary Position Embedding (RoPE), SwiGLU activation functions, RMSNorm normalization, and Attention QKV bias.

This design enables the model to process extensive inputs effectively, making it particularly adept at complex reasoning tasks. However, as a preview release, it exhibits certain limitations, including potential language mixing, recursive reasoning loops, and areas requiring improved safety measures.

TL;DR:

• Deployment of QwQ-32B-Preview model using vllm.
• You can expect an average tokens/sec of 21.73 and a latency of 11.71 sec for generating a text of 256 tokens. This setup has an average cold start time of 39.44 sec.
• 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.
• Model import via GitHub with input_schema.py file.
• Recommended GPU: NVIDIA A100 for optimal performance.
• 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-config.yaml file given above. Finally, ensure you save your changes by clicking on the save button.

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 a custom model button.

• Select Github as the method of upload from the Provider list and then select your Github Repository and the branch.
• Choose the type of machine, and specify the minimum and maximum number of replicas for deploying your model.
• Configure Custom Runtime ( If you have pip or apt packages), choose Volume, Secrets and set Environment variables like Inference Timeout / Container Concurrency / Scale Down Timeout
• Once you click “Continue,” click Deploy to start the model import process.

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. It 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 argument to this function inputs, is a dictionary containing all the input parameters. The keys are the same as the name given in inputs. Refer to input for more.

def infer(self, inputs):
    prompts = inputs["prompt"]
    temperature = inputs.get("temperature",0.7)
    top_p = inputs.get("top_p",0.1)
    repetition_penalty = inputs.get("repetition_penalty",1.18)
    top_k = inputs.get("top_k",40)
    max_tokens = inputs.get("max_tokens",256)

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.