Today, we are delighted to announce that DeepSeek R1 distilled Llama and Qwen designs are available through Amazon Bedrock Marketplace and Amazon SageMaker JumpStart. With this launch, you can now release DeepSeek AI's first-generation frontier design, DeepSeek-R1, together with the distilled versions ranging from 1.5 to 70 billion criteria to construct, experiment, and properly scale your generative AI ideas on AWS.
In this post, we show how to get started with DeepSeek-R1 on Amazon Bedrock Marketplace and SageMaker JumpStart. You can follow comparable actions to release the distilled versions of the models too.
Overview of DeepSeek-R1
DeepSeek-R1 is a big language design (LLM) established by DeepSeek AI that utilizes reinforcement learning to improve thinking capabilities through a multi-stage training process from a DeepSeek-V3-Base foundation. A key identifying function is its support learning (RL) action, which was utilized to refine the design's reactions beyond the basic pre-training and fine-tuning process. By integrating RL, DeepSeek-R1 can adjust better to user feedback and goals, ultimately improving both relevance and clarity. In addition, DeepSeek-R1 employs a chain-of-thought (CoT) technique, indicating it's geared up to break down complex inquiries and reason through them in a detailed manner. This assisted thinking process allows the design to produce more precise, transparent, and detailed answers. This model combines RL-based fine-tuning with CoT abilities, aiming to create structured actions while concentrating on interpretability and user interaction. With its extensive abilities DeepSeek-R1 has actually captured the market's attention as a versatile text-generation design that can be integrated into various workflows such as representatives, rational thinking and information analysis jobs.
DeepSeek-R1 uses a Mix of Experts (MoE) architecture and is 671 billion criteria in size. The MoE architecture permits activation of 37 billion parameters, allowing efficient reasoning by routing questions to the most pertinent professional "clusters." This approach enables the model to focus on various problem domains while maintaining total performance. DeepSeek-R1 requires at least 800 GB of HBM memory in FP8 format for reasoning. In this post, we will utilize an ml.p5e.48 xlarge instance to release the design. ml.p5e.48 xlarge includes 8 Nvidia H200 GPUs supplying 1128 GB of GPU memory.
DeepSeek-R1 distilled designs bring the reasoning abilities of the main R1 design to more efficient architectures based upon popular open models like Qwen (1.5 B, 7B, 14B, and 32B) and Llama (8B and 70B). Distillation refers to a procedure of training smaller, more effective models to mimic the habits and reasoning patterns of the bigger DeepSeek-R1 model, utilizing it as an instructor model.
You can release DeepSeek-R1 design either through SageMaker JumpStart or Bedrock Marketplace. Because DeepSeek-R1 is an emerging design, we recommend deploying this model with guardrails in location. In this blog, we will utilize Amazon Bedrock Guardrails to present safeguards, avoid harmful content, and evaluate designs against crucial safety criteria. At the time of writing this blog site, for DeepSeek-R1 deployments on SageMaker JumpStart and Bedrock Marketplace, Bedrock Guardrails supports just the ApplyGuardrail API. You can create numerous guardrails tailored to various use cases and apply them to the DeepSeek-R1 model, enhancing user experiences and standardizing security controls throughout your generative AI applications.
Prerequisites
To deploy the DeepSeek-R1 design, you require access to an ml.p5e instance. To examine if you have quotas for P5e, open the Service Quotas console and under AWS Services, choose Amazon SageMaker, and confirm you're utilizing ml.p5e.48 xlarge for endpoint use. Make certain that you have at least one ml.P5e.48 xlarge instance in the AWS Region you are releasing. To request a limit increase, produce a limit increase demand and reach out to your account group.
Because you will be deploying this model with Amazon Bedrock Guardrails, make certain you have the proper AWS Identity and Gain Access To Management (IAM) permissions to use Amazon Bedrock Guardrails. For instructions, see Set up permissions to utilize guardrails for content filtering.
Implementing guardrails with the ApplyGuardrail API
Amazon Bedrock Guardrails permits you to introduce safeguards, prevent harmful content, and examine models against crucial safety criteria. You can implement safety steps for the DeepSeek-R1 model using the Amazon Bedrock ApplyGuardrail API. This allows you to apply guardrails to examine user inputs and model responses deployed on Amazon Bedrock Marketplace and SageMaker JumpStart. You can develop a guardrail using the Amazon Bedrock console or the API. For the example code to develop the guardrail, see the GitHub repo.
The basic circulation includes the following steps: First, the system receives an input for the design. This input is then processed through the ApplyGuardrail API. If the input passes the guardrail check, it's sent out to the model for inference. After receiving the output, another guardrail check is used. If the output passes this final check, it's returned as the result. However, if either the input or output is intervened by the guardrail, a message is returned indicating the nature of the intervention and whether it happened at the input or output stage. The examples showcased in the following areas demonstrate inference using this API.
Deploy DeepSeek-R1 in Amazon Bedrock Marketplace
Amazon Bedrock Marketplace offers you access to over 100 popular, emerging, and specialized foundation designs (FMs) through Amazon Bedrock. To gain access to DeepSeek-R1 in Amazon Bedrock, complete the following actions:
1. On the Amazon Bedrock console, choose Model brochure under Foundation designs in the navigation pane.
At the time of composing this post, you can utilize the InvokeModel API to invoke the model. It does not support Converse APIs and other Amazon Bedrock tooling.
2. Filter for DeepSeek as a company and choose the DeepSeek-R1 design.
The model detail page offers essential details about the design's capabilities, prices structure, and execution standards. You can discover detailed usage instructions, including sample API calls and code bits for combination. The design supports numerous text generation jobs, consisting of material development, code generation, and concern answering, utilizing its support discovering optimization and CoT reasoning capabilities.
The page likewise includes deployment options and licensing details to assist you get started with DeepSeek-R1 in your applications.
3. To start utilizing DeepSeek-R1, pick Deploy.
You will be prompted to set up the release details for DeepSeek-R1. The design ID will be pre-populated.
4. For Endpoint name, get in an endpoint name (in between 1-50 alphanumeric characters).
5. For Number of circumstances, enter a number of circumstances (in between 1-100).
6. For Instance type, it-viking.ch choose your instance type. For optimal efficiency with DeepSeek-R1, a GPU-based instance type like ml.p5e.48 xlarge is recommended.
Optionally, you can set up advanced security and infrastructure settings, consisting of virtual personal cloud (VPC) networking, service role authorizations, and file encryption settings. For most use cases, the default settings will work well. However, for production implementations, you might wish to evaluate these settings to align with your company's security and compliance requirements.
7. Choose Deploy to begin utilizing the design.
When the deployment is total, you can evaluate DeepSeek-R1's abilities straight in the Amazon Bedrock playground.
8. Choose Open in play ground to access an interactive interface where you can experiment with different prompts and adjust model criteria like temperature and maximum length.
When using R1 with Bedrock's InvokeModel and Playground Console, utilize DeepSeek's chat template for optimal results. For example, material for inference.
This is an excellent method to check out the design's thinking and text generation abilities before incorporating it into your applications. The playground offers instant feedback, assisting you comprehend how the design reacts to numerous inputs and letting you tweak your prompts for optimum outcomes.
You can rapidly evaluate the design in the playground through the UI. However, to conjure up the released design programmatically with any Amazon Bedrock APIs, you require to get the endpoint ARN.
Run inference using guardrails with the released DeepSeek-R1 endpoint
The following code example demonstrates how to perform reasoning using a released DeepSeek-R1 model through Amazon Bedrock using the invoke_model and ApplyGuardrail API. You can create a guardrail utilizing the Amazon Bedrock console or the API. For the example code to produce the guardrail, see the GitHub repo. After you have actually developed the guardrail, use the following code to implement guardrails. The script initializes the bedrock_runtime client, configures inference criteria, and sends a demand to generate text based on a user prompt.
Deploy DeepSeek-R1 with SageMaker JumpStart
SageMaker JumpStart is an artificial intelligence (ML) hub with FMs, integrated algorithms, and prebuilt ML options that you can deploy with just a couple of clicks. With SageMaker JumpStart, you can tailor pre-trained designs to your usage case, with your data, and deploy them into production using either the UI or SDK.
Deploying DeepSeek-R1 model through SageMaker JumpStart provides 2 convenient methods: utilizing the user-friendly SageMaker JumpStart UI or executing programmatically through the SageMaker Python SDK. Let's explore both methods to help you pick the technique that best fits your needs.
Deploy DeepSeek-R1 through SageMaker JumpStart UI
Complete the following actions to deploy DeepSeek-R1 utilizing SageMaker JumpStart:
1. On the SageMaker console, select Studio in the navigation pane.
2. First-time users will be triggered to produce a domain.
3. On the SageMaker Studio console, choose JumpStart in the navigation pane.
The design browser displays available designs, with details like the company name and design capabilities.
4. Search for DeepSeek-R1 to see the DeepSeek-R1 model card.
Each model card reveals essential details, including:
- Model name
- Provider name
- Task category (for instance, Text Generation).
Bedrock Ready badge (if appropriate), showing that this model can be signed up with Amazon Bedrock, enabling you to use Amazon Bedrock APIs to invoke the design
5. Choose the design card to see the design details page.
The model details page includes the following details:
- The model name and company details. Deploy button to deploy the design. About and Notebooks tabs with detailed details
The About tab includes important details, such as:
- Model description. - License details.
- Technical specifications.
- Usage guidelines
Before you deploy the model, it's advised to review the design details and license terms to validate compatibility with your usage case.
6. Choose Deploy to proceed with release.
7. For Endpoint name, use the immediately produced name or develop a custom one.
- For Instance type ¸ pick a circumstances type (default: ml.p5e.48 xlarge).
- For Initial instance count, enter the variety of instances (default: 1). Selecting appropriate instance types and counts is crucial for expense and performance optimization. Monitor your deployment to adjust these settings as needed.Under Inference type, Real-time reasoning is chosen by default. This is optimized for sustained traffic and low latency.
- Review all setups for accuracy. For this model, we strongly recommend sticking to SageMaker JumpStart default settings and making certain that network isolation remains in location.
- Choose Deploy to deploy the design.
The implementation procedure can take several minutes to finish.
When implementation is complete, your endpoint status will change to InService. At this moment, the design is ready to accept reasoning demands through the endpoint. You can keep an eye on the deployment development on the SageMaker console Endpoints page, which will show relevant metrics and status details. When the release is total, you can conjure up the design using a SageMaker runtime client and incorporate it with your applications.
Deploy DeepSeek-R1 using the SageMaker Python SDK
To get begun with DeepSeek-R1 utilizing the SageMaker Python SDK, you will need to install the SageMaker Python SDK and make certain you have the essential AWS approvals and environment setup. The following is a detailed code example that shows how to release and utilize DeepSeek-R1 for inference programmatically. The code for releasing the design is offered in the Github here. You can clone the note pad and run from SageMaker Studio.
You can run extra demands against the predictor:
Implement guardrails and run reasoning with your SageMaker JumpStart predictor
Similar to Amazon Bedrock, you can likewise use the ApplyGuardrail API with your SageMaker JumpStart predictor. You can develop a guardrail using the Amazon Bedrock console or the API, and implement it as displayed in the following code:
Clean up
To avoid unwanted charges, complete the steps in this section to tidy up your resources.
Delete the Amazon Bedrock Marketplace deployment
If you deployed the model utilizing Amazon Bedrock Marketplace, complete the following actions:
1. On the Amazon Bedrock console, under Foundation designs in the navigation pane, select Marketplace implementations. - In the Managed implementations area, locate the endpoint you want to erase.
- Select the endpoint, and on the Actions menu, choose Delete.
- Verify the endpoint details to make certain you're erasing the proper implementation: 1. Endpoint name.
- Model name.
- Endpoint status
Delete the SageMaker JumpStart predictor
The SageMaker JumpStart design you deployed will sustain expenses if you leave it running. Use the following code to erase the endpoint if you want to stop sustaining charges. For more details, see Delete Endpoints and Resources.
Conclusion
In this post, we checked out how you can access and deploy the DeepSeek-R1 design utilizing Bedrock Marketplace and SageMaker JumpStart. Visit SageMaker JumpStart in SageMaker Studio or Amazon Bedrock Marketplace now to begin. For more details, refer to Use Amazon Bedrock tooling with Amazon SageMaker JumpStart models, SageMaker JumpStart pretrained models, Amazon SageMaker JumpStart Foundation Models, Amazon Bedrock Marketplace, and Starting with Amazon SageMaker JumpStart.
About the Authors
Vivek Gangasani is a Lead Specialist Solutions Architect for Inference at AWS. He helps emerging generative AI companies construct ingenious services using AWS services and sped up compute. Currently, he is focused on developing techniques for fine-tuning and optimizing the reasoning performance of large language designs. In his spare time, Vivek enjoys treking, enjoying movies, and trying various cuisines.
Niithiyn Vijeaswaran is a Generative AI Specialist Solutions Architect with the Third-Party Model Science group at AWS. His area of focus is AWS AI accelerators (AWS Neuron). He holds a Bachelor's degree in Computer Science and Bioinformatics.
Jonathan Evans is a Professional Solutions Architect dealing with generative AI with the Third-Party Model Science team at AWS.
Banu Nagasundaram leads product, engineering, and tactical collaborations for Amazon SageMaker JumpStart, SageMaker's artificial intelligence and generative AI hub. She is enthusiastic about developing services that assist customers accelerate their AI journey and unlock business value.