OpenClaw (Clawd Bot) with vLLM Running for Free on AMD Developer Cloud

Over the past week, OpenClaw (previously known as Clawd Bot and Molt Bot) has seen a surge of activity. Users are highlighting impressive projects and demonstrating how they use the agent as a personal assistant. However, there has also been ongoing discussion regarding cost and security.

To address these concerns, this guide demonstrates how to use a powerful open-source model that exceeds the limits of typical consumer-grade GPUs by running it on enterprise data center hardware, specifically the AMD Instinct™ MI300X with a massive 192GB of memory, at no cost.

Complimentary AMD Developer Cloud Credits via AMD AI Developer Program 

The AMD AI Developer Program provides members with $100 in free credits, which is enough for approximately 50 hours of usage to get started. Members who showcase useful applications and publicly share their projects may also apply for additional credits.

Beyond the credits, signing up grants you a one-month DeepLearning.AI Premium membership, automatic entry into monthly AMD hardware sweepstakes, and access to free AMD training courses.

Getting Started 

Step 1: Sign up for the AMD AI Developer Program

• Existing Users: If you already have an AMD account, simply sign in and enroll in the program.
• New Users: Select “Create an account” on the login page to set up your AMD account and enroll simultaneously.

Step 2: Activate Your Credits

After you join the program, navigate to the “member-site” to find your free credit activation code.

Step 3: Create a GPU Droplet

Once signed in, you will be directed to the Create a GPU Droplet page, where your credits should appear at the top.

1. Select Hardware: Choose a single MI300X instance.
2. Select Image: Choose the ROCm Software image.
3. Configure Access: Add your SSH key (instructions for creating one are provided in that section of the page).

Once the droplet is created, you can access it via your terminal or by clicking on Web Console button:

Running the Model with vLLM

We will use vLLM, one of the most popular frameworks for LLMs, to run our model. The following steps must be followed in your droplet that you created in the previous step.

Step 1: Environment Setup

In this example, we will use pip in a virtual environment. Run the following commands on your droplet to set up the environment:

Next, install the ROCm-optimized version of vLLM:

Step 2: Configure Firewall and Launch vLLM

In this step, we demonstrate how you can run a suitable model from HuggingFace with vLLM. 

First, open port 8090 to allow traffic to your model: 

Next, run the command below to pull a pruned version of MiniMax-M2.1 model from HuggingFace. This model is 139B parameters in FP8, meaning it fits comfortably within the MI300X 192GB of memory. 

Note: Replace abc-123 in the command below with a secure, unique API Key.

This command downloads the model weights, loads them onto your GPU, and creates an OpenAI-compatible endpoint accessible at:

Running the above should make the model accessible via this URL: http://<your-droplet-ip>:8090/v1

Connecting OpenClaw to Your Local LLM

Now that the model is running, we will configure OpenClaw to use it.

Step 1: Install OpenClaw

If you haven't installed OpenClaw yet, use the following command (for Mac/Linux):

During installation, when asked “How do you want to hatch your bot?”, select “Open the Web UI” and follow the instructions to open the web app.

Step 2: Configure the Provider

In the OpenClaw Web UI, navigate to Settings > Config. We need to add our local LLM information.

• Name: vllm
• API: Select openai-completions
• API Key: Enter the key you defined earlier (e.g., abc-123)
• Base URL: http://<your-droplet-ip>:8090/v1

Step 3: Configure the Model

1. Add a new entry under “Models”: 
   • API: Select openai-completions. 
   • Context Window: Set to 194000 (matching the max-model-len setting).
   • ID: Set to MiniMax-M2.1 (matching the served-model-name).
2. Click Apply.

Step 4: Set the Agent

Finally, connect this model to your main agent:

1. Navigate to the Agents section.
2. Change the Primary Model value to: vllm/MiniMax-M2.1 (this is based on the values we entered under for Models configuration).
3. Click Apply.

Conclusion

You are now ready to start chatting with your agent running on enterprise-grade hardware. You can use this same method to run other open-source models from Hugging Face. Don't forget to share your projects with the community to earn more free GPU credits. Happy building - the world is your Lobster!