One post tagged with "DeepSeek"

Deploying large language models (LLMs) like DeepSeek-V3 and DeepSeek-R1 has become more accessible with Alibaba Cloud's Platform for AI (PAI). This guide provides a step-by-step walkthrough to help you set up your Alibaba Cloud account, configure necessary resources, and deploy DeepSeek models efficiently.

Creating an Alibaba Cloud Account​

Before leveraging Alibaba Cloud services, you need to create an account:

1. Visit the Official Website: Navigate to Alibaba Cloud

2. Initiate Sign-Up: Click on the Free Trial button located at the top-right corner.

3. Choose Account Type: Select between Personal or Enterprise account types based on your need.

4. Enter Account Information:

   1. Email Address Provide a valid email address.
   2. Password: Set a secure password.

5. Verify Your Account:

   1. Email Verification: Click on "Send" to receive a verification code in your email. Enter the code to verify.
   2. Phone Verification: Alternatively, provide your mobile number to receive a verification code via SMS.

6. Agree to Terms: Read and accept the Membership Agreement, Privacy Policy, Product Terms, and Terms of Use.

7. Complete Sign-Up: Click on Sign Up to finalize the registration.

8. Add Billing Information: Before purchasing services, add your billing and payment details. Note that if you intend to use services in mainland China, real-name verification is required.

Selecting the Appropriate Region​

Choosing the right region is crucial for performance and compliance:

1. Geographical Proximity: Select a region closer to your target audience to reduce latency.

2. Service Availability: Ensure the desired services are available in the selected region.

3. Pricing Considerations: Be aware that resource pricing may vary between region.

4. Internal Communication: If your services need to communicate over an internal network, deploy them within the same region.

Note: Once a resource is created in a region, its region cannot be changed.

Understanding DeepSeek Models​

Alibaba Cloud's PAI Model Gallery offers access to advanced LLM:

1. DeepSeek-V3: A 671B parameter mixture of experts (MoE) model designed for complex language tasks.

2. DeepSeek-R1: Built upon DeepSeek-V3-Base, this model excels in reasoning and inference tasks.

3. Distilled Versions: For resource-constrained environments, distilled models like DeepSeek-R1-Distill-Qwen-32B offer a balance between performance and resource usage.

Deployment Methods on PAI​

Alibaba Cloud PAI provides multiple deployment option:

1. BladeLLM Accelerated Deployment: High-performance inference framework optimized for large-scale models.

2. SGLang Accelerated Deployment: Fast service framework tailored for large and visual language models, compatible with OpenAI APIs.

3. vLLM Accelerated Deployment: Widely-used library for LLM inference acceleration, offering efficient deployment capabilities.

4. Transformers Standard Deployment: Standard deployment method without inference acceleration, supporting both API and WebUI interfaces.

Step-by-Step Deployment Guide​

1. Access the PAI Console: Log in to your Alibaba Cloud account and navigate to the PAI Console.

2. Select a Workspace: In the upper-left corner, choose a region that aligns with your business requirement. From the left-side navigation pane, select Workspaces and click on your desired workspace.

3. Navigate to Model Gallery: Within your workspace, go to QuickStart > *Model Gallery.

4. Choose a Model: Browse the available models and select one that fits your needs, such as DeepSeek-R1-Distill-Qwen-32B.

5. Configure Deployment: **Click Deploy in the upper right corner. Choose your preferred deployment method and adjust resource settings as necessary.

6. Monitor Deployment: After initiating deployment, monitor the progress via Model Gallery > Job Management > *Deployment Jobs. Here, you can view deployment status and access call information.

Usage Recommendations​

To optimize model performance:

1. Temperature Settings: Set the temperature between 0.5 and 0.7, ideally at 0.6. This controls the randomness of responses — lower values produce more predictable output, higher values are more creative.
2. Prompt Formatting: Always include the full instruction within the user prompt. Do not use the system prompt field, as it is ignored by most DeepSeek deployments.
3. Math & Reasoning Tasks: For structured tasks like solving math problems or logical questions, append this phrase to your prompt:

Accessing the Model Interface​

After deployment, interacting with your model can be done via WebUI or API, depending on the method:

✅ For Transformers Standard Deployment​

1. Supports both API and WebUI.
2. Once deployed, click View Web App on the service page in the PAI console to open a simple chatbot-style interface.

⚡ For Accelerated Deployments (BladeLLM, vLLM, SGLang)​

1. These support API access only for higher performance.
2. Alibaba Cloud provides Gradio-based WebUI code in the deployment page, which you can use to launch a custom interface locally.

Resource Considerations​

Deploying full-scale models requires substantial cloud resources:

1. DeepSeek-V3 or full R1 models may require:

2. 8 GPU cores

3. 96 GB GPU memory

4. Up to 2 TB of system memory for optimal inference

5. Distilled models like DeepSeek-R1-Distill-Qwen-32B offer:

6. Lower resource consumption

7. Faster response times

8. Reduced costs

📌 You can choose the instance size in the deployment UI. For most users, we recommend starting with a distilled model and upgrading based on need.

🧠 Deploying DeepSeek-R1-Distill-Qwen-32B on Alibaba Cloud​

Once you're inside the Alibaba Cloud PAI Model Gallery, follow the steps below to deploy the DeepSeek-R1-Distill-Qwen-32B model — a powerful yet resource-efficient distilled version designed for high-performance inference at lower cost.

🔍 Step 1: Locate the Model​

1. Navigate to your PAI workspace.
2. Click on Quick Start > Model Gallery.
3. In the search bar, type Qwen-32B or DeepSeek.
4. Locate the model: deepseek-r1-distill-qwen-32b.
5. Click on the model card to open its detail page.

⚙️ Step 2: Click “Deploy”​

On the model detail page:

1. Click the blue “Deploy” button in the top-right.

2. Choose your deployment method:

   1. For standard usage with a WebUI, choose Transformers Standard Deployment.
   2. For faster performance, choose BladeLLM, vLLM, or SGLang depending on your preference and API-only usage.

3. In the deployment configuration panel:

   1. Service Name: Auto-generated or give it a custom name like deepseek-qwen32b-service.
   2. Resource Type: Select your desired GPU/CPU resource. For distilled models, fewer resources like ecs.gn7i-c8g1.2xlarge may be sufficient.
   3. Click Next and then Deploy.

Within a few minutes, your service will be provisioned.

🌐 Step 3: Get the API Endpoint​

Once deployed:

1. Go to Job Management > Deployment Jobs.
2. Click on the deployed service.
3. Copy the API endpoint URL provided — you'll use this in Postman or any HTTP client.

📬 Step 4: Test the Model in Postman​

Here's how to use Postman to make your first inference call:

🧾 Request Type: POST​

📍 URL:​

Use the endpoint you copied, for example:

https://your-service-name.pai-eas.aliyuncs.com/api/predict

🧠 Headers:​

Content-Type: application/json
Authorization: Bearer <your-access-token>  # Optional depending on your setup

📦 Body (raw JSON):​

For deepseek-r1-distill-qwen-32b, use the following format:

{
  "model": "deepseek-r1-distill-qwen-32b",
  "inputs": [
    {
      "role": "user",
      "content": "What is the capital of France?"
    }
  ],
  "parameters": {
    "temperature": 0.6,
    "max_tokens": 512
  }
}

📝 You can modify:

1. "content": to ask any question.
2. "temperature": to control randomness (0.6 is ideal).
3. "max_tokens": controls how long the output will be.

▶️ Click “Send”​

You'll receive a JSON response containing the model's answer under content.

✅ Tips for Better Results​

1. For math or reasoning problems, include:

"content": "Solve this step by step: 248 + 382. Put the final answer in \\boxed{}"

1. Use "stream": false unless your deployment supports streaming output.

🧠 What Happens Behind the Scenes​

By deploying Qwen-32B, you're spinning up an inference-optimized container hosted on Alibaba Cloud's GPU infrastructure, preloaded with model weights and optimized runtime. It's production-grade and ready to scale with API Gateway, logging, and access control if needed.

Integrating with Alibaba Cloud API Gateway​

For secure and scalable integration, expose your model's inference endpoint through Cloud-Native API Gateway:

Step 1: Create an AI Service​

1. Go to API Gateway Console > AI Services > Create AI Service
2. Fill in:
3. Service Source: PAI-EAS
4. Model Supplier: (select appropriate deployment method)
5. Endpoint URL: Found in your model's service page

Step 2: Publish an AI API​

1. Navigate to APIs > Create API
2. Attach the AI Service you created
3. Customize method (POST), headers, and body parameters

Step 3: Monitor & Secure​

1. Use the API Gateway's tools to:
2. Throttle requests
3. Apply authentication
4. View logs and response times

This ensures stable public access while protecting backend models from abuse or excessive load.

Conclusion​

Alibaba Cloud's Platform for AI (PAI) has made advanced LLM deployment accessible to everyone — from developers to enterprises. With just a few clicks, you can spin up state-of-the-art models like DeepSeek-R1 and DeepSeek-V3 and integrate them into your apps or websites securely and efficiently.

By combining one-click deployment, high-performance inference frameworks, and secure API exposure, Alibaba Cloud offers a best-in-class experience for anyone looking to harness the power of large language models.

Want help building your first chatbot, virtual assistant, or generative app with DeepSeek? Reach out to us at Arina Technologies — we specialize in AI-based cloud solutions.