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 deploy DeepSeek AI's first-generation frontier model, DeepSeek-R1, in addition to the distilled variations ranging from 1.5 to 70 billion criteria to build, experiment, and responsibly scale your generative AI concepts on AWS.
In this post, we demonstrate how to get going with DeepSeek-R1 on Amazon Bedrock Marketplace and SageMaker JumpStart. You can follow similar actions to release the distilled versions of the models too.
Overview of DeepSeek-R1
DeepSeek-R1 is a large language design (LLM) established by DeepSeek AI that utilizes support learning to improve thinking abilities through a multi-stage training procedure from a DeepSeek-V3-Base foundation. A crucial distinguishing function is its support learning (RL) action, which was used to fine-tune the model's actions beyond the basic pre-training and fine-tuning procedure. By incorporating RL, DeepSeek-R1 can adjust more successfully to user feedback and objectives, eventually boosting both significance and clearness. In addition, DeepSeek-R1 utilizes a chain-of-thought (CoT) method, suggesting it's equipped to break down intricate queries and reason through them in a detailed way. This directed reasoning procedure permits the model to produce more accurate, transparent, and detailed answers. This model combines RL-based fine-tuning with CoT abilities, aiming to produce structured reactions while concentrating on interpretability and user interaction. With its comprehensive abilities DeepSeek-R1 has actually recorded the market's attention as a flexible text-generation design that can be incorporated into different workflows such as representatives, sensible reasoning and data interpretation jobs.
DeepSeek-R1 utilizes a Mix of Experts (MoE) architecture and is 671 billion specifications in size. The MoE architecture allows activation of 37 billion parameters, enabling efficient inference by routing questions to the most pertinent specialist "clusters." This method allows the design to focus on different issue domains while maintaining overall performance. DeepSeek-R1 needs a minimum of 800 GB of HBM memory in FP8 format for reasoning. In this post, we will utilize an ml.p5e.48 xlarge circumstances to release the model. ml.p5e.48 xlarge includes 8 Nvidia H200 GPUs offering 1128 GB of GPU memory.
DeepSeek-R1 distilled models bring the reasoning abilities of the main R1 model to more efficient architectures based on popular open designs like Qwen (1.5 B, 7B, 14B, and 32B) and Llama (8B and 70B). Distillation refers to a procedure of training smaller sized, more effective designs to mimic the behavior and thinking patterns of the larger DeepSeek-R1 design, using it as an instructor design.
You can release DeepSeek-R1 design either through SageMaker JumpStart or Bedrock Marketplace. Because DeepSeek-R1 is an emerging model, we suggest releasing this model with guardrails in location. In this blog site, we will use Amazon Bedrock Guardrails to introduce safeguards, avoid harmful material, and examine models against key security criteria. At the time of composing this blog, for DeepSeek-R1 implementations on SageMaker JumpStart and Bedrock Marketplace, Bedrock Guardrails supports just the ApplyGuardrail API. You can produce numerous guardrails tailored to various use cases and apply them to the DeepSeek-R1 model, enhancing user experiences and standardizing security controls across your generative AI applications.
Prerequisites
To release the DeepSeek-R1 design, you need access to an ml.p5e instance. To check if you have quotas for P5e, open the Service Quotas console and under AWS Services, pick Amazon SageMaker, and verify you're utilizing ml.p5e.48 xlarge for endpoint use. Make certain that you have at least one ml.P5e.48 xlarge circumstances in the AWS Region you are deploying. To request a limit boost, develop a limit boost request and reach out to your account team.
Because you will be deploying this model with Amazon Bedrock Guardrails, make certain you have the appropriate AWS Identity and Gain Access To Management (IAM) approvals to use Amazon Bedrock Guardrails. For instructions, see Set up consents to utilize guardrails for material filtering.
Implementing guardrails with the ApplyGuardrail API
Amazon Bedrock Guardrails permits you to present safeguards, avoid damaging material, and assess designs against key safety criteria. You can execute safety steps for the DeepSeek-R1 model using the Amazon Bedrock ApplyGuardrail API. This allows you to use guardrails to assess user inputs and model actions deployed on Amazon Bedrock Marketplace and SageMaker JumpStart. You can produce a guardrail using the Amazon Bedrock console or the API. For the example code to produce the guardrail, see the GitHub repo.
The general circulation involves the following actions: First, the system receives an input for the model. This input is then processed through the ApplyGuardrail API. If the input passes the guardrail check, it's sent to the design for inference. After receiving the design's 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 suggesting the nature of the intervention and whether it happened at the input or output phase. The examples showcased in the following sections show inference utilizing this API.
Deploy DeepSeek-R1 in Amazon Bedrock Marketplace
Amazon Bedrock Marketplace gives you access to over 100 popular, emerging, and specialized structure models (FMs) through Amazon Bedrock. To gain access to DeepSeek-R1 in Amazon Bedrock, total the following actions:
1. On the Amazon Bedrock console, pick Model catalog under Foundation models in the navigation pane.
At the time of writing 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 necessary details about the model's abilities, pricing structure, and implementation guidelines. You can discover detailed use instructions, including sample API calls and code snippets for integration. The model supports numerous text generation jobs, consisting of material production, code generation, and concern answering, utilizing its support learning optimization and CoT thinking capabilities.
The page also consists of implementation choices and licensing details to help you begin with DeepSeek-R1 in your applications.
3. To start using DeepSeek-R1, select Deploy.
You will be triggered to set up the release details for DeepSeek-R1. The model ID will be pre-populated.
4. For Endpoint name, go into an endpoint name (between 1-50 alphanumeric characters).
5. For Number of circumstances, go into a number of circumstances (between 1-100).
6. For example type, select your instance type. For optimum performance with DeepSeek-R1, a GPU-based circumstances type like ml.p5e.48 xlarge is recommended.
Optionally, you can configure advanced security and infrastructure settings, consisting of virtual personal cloud (VPC) networking, service role consents, and encryption settings. For most use cases, the default settings will work well. However, it-viking.ch for production releases, you might desire to examine these settings to line up with your organization's security and compliance requirements.
7. Choose Deploy to start utilizing the model.
When the implementation is total, you can check DeepSeek-R1's abilities straight in the Amazon Bedrock play ground.
8. Choose Open in play ground to access an interactive user interface where you can experiment with different prompts and change model criteria like temperature level and optimum length.
When using R1 with Bedrock's InvokeModel and Playground Console, use DeepSeek's chat template for optimal results. For example, content for reasoning.
This is an exceptional method to check out the design's thinking and text generation abilities before integrating it into your applications. The play area offers instant feedback, assisting you understand how the design reacts to various inputs and letting you fine-tune your prompts for ideal outcomes.
You can rapidly test the design in the play area 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 deployed DeepSeek-R1 endpoint
The following code example shows how to carry out reasoning utilizing a released DeepSeek-R1 model through Amazon Bedrock utilizing the invoke_model and ApplyGuardrail API. You can create a guardrail utilizing the Amazon Bedrock console or the API. For the example code to create the guardrail, see the GitHub repo. After you have actually produced the guardrail, use the following code to implement guardrails. The script initializes the bedrock_runtime client, configures inference specifications, and sends out a demand to based upon a user prompt.
Deploy DeepSeek-R1 with SageMaker JumpStart
SageMaker JumpStart is an artificial intelligence (ML) hub with FMs, integrated algorithms, and prebuilt ML services that you can release with just a couple of clicks. With SageMaker JumpStart, you can tailor pre-trained models to your use case, with your data, and deploy them into production using either the UI or SDK.
Deploying DeepSeek-R1 model through SageMaker JumpStart offers two convenient approaches: utilizing the user-friendly SageMaker JumpStart UI or executing programmatically through the SageMaker Python SDK. Let's check out both methods to help you pick the method that finest suits your requirements.
Deploy DeepSeek-R1 through SageMaker JumpStart UI
Complete the following actions to deploy DeepSeek-R1 using SageMaker JumpStart:
1. On the SageMaker console, select Studio in the navigation pane.
2. First-time users will be triggered to develop a domain.
3. On the SageMaker Studio console, choose JumpStart in the navigation pane.
The design internet browser shows available designs, with details like the supplier name and design abilities.
4. Search for DeepSeek-R1 to view the DeepSeek-R1 model card.
Each design card shows crucial details, consisting of:
- Model name
- Provider name
- Task classification (for instance, Text Generation).
Bedrock Ready badge (if applicable), suggesting that this design can be registered with Amazon Bedrock, permitting you to use Amazon Bedrock APIs to invoke the model
5. Choose the model card to view the model details page.
The design details page includes the following details:
- The model name and supplier details. Deploy button to release the model. About and Notebooks tabs with detailed details
The About tab includes crucial details, such as:
- Model description. - License details. - Technical specs.
- Usage standards
Before you deploy the design, it's suggested to review the design details and license terms to verify compatibility with your use case.
6. Choose Deploy to proceed with deployment.
7. For Endpoint name, utilize the immediately created name or produce a custom one.
- For example type ¸ choose an instance type (default: ml.p5e.48 xlarge).
- For Initial circumstances count, get in the variety of circumstances (default: 1). Selecting appropriate instance types and counts is important for cost and performance optimization. Monitor your implementation to change these settings as needed.Under Inference type, Real-time reasoning is selected by default. This is enhanced for sustained traffic and low latency.
- Review all configurations for accuracy. For this design, we highly advise adhering to SageMaker JumpStart default settings and making certain that network isolation remains in location.
- Choose Deploy to release the design.
The release procedure can take several minutes to finish.
When implementation is total, 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 implementation progress on the SageMaker console Endpoints page, which will show appropriate metrics and status details. When the implementation is complete, you can conjure up the model using a SageMaker runtime customer and integrate it with your applications.
Deploy DeepSeek-R1 using the SageMaker Python SDK
To get going with DeepSeek-R1 using the SageMaker Python SDK, you will need to set up the SageMaker Python SDK and make certain you have the necessary AWS permissions and environment setup. The following is a detailed code example that shows how to deploy and use DeepSeek-R1 for inference programmatically. The code for deploying the design is offered in the Github here. You can clone the notebook and run from SageMaker Studio.
You can run extra demands against the predictor:
Implement guardrails and run inference with your SageMaker JumpStart predictor
Similar to Amazon Bedrock, you can also utilize the ApplyGuardrail API with your SageMaker JumpStart predictor. You can produce a guardrail utilizing the Amazon Bedrock console or the API, and implement it as shown in the following code:
Clean up
To avoid undesirable charges, finish the steps in this area to tidy up your resources.
Delete the Amazon Bedrock Marketplace deployment
If you released the model utilizing Amazon Bedrock Marketplace, total the following actions:
1. On the Amazon Bedrock console, under Foundation models in the navigation pane, select Marketplace releases. - In the Managed implementations section, locate the endpoint you want to delete.
- Select the endpoint, and on the Actions menu, choose Delete.
- Verify the endpoint details to make certain you're deleting the proper release: 1. Endpoint name.
- Model name.
- Endpoint status
Delete the SageMaker JumpStart predictor
The SageMaker JumpStart design you deployed will sustain costs if you leave it running. Use the following code to erase the endpoint if you wish 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 model using Bedrock Marketplace and SageMaker JumpStart. Visit SageMaker JumpStart in SageMaker Studio or Amazon Bedrock Marketplace now to get going. For more details, describe Use Amazon Bedrock tooling with Amazon SageMaker JumpStart models, SageMaker JumpStart pretrained models, Amazon SageMaker JumpStart Foundation Models, Amazon Bedrock Marketplace, and Beginning with Amazon SageMaker JumpStart.
About the Authors
Vivek Gangasani is a Lead Specialist Solutions Architect for Inference at AWS. He assists emerging generative AI companies build innovative solutions using AWS services and accelerated compute. Currently, he is focused on establishing strategies for fine-tuning and enhancing the inference efficiency of large language designs. In his spare time, Vivek delights in treking, viewing movies, and attempting different foods.
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 technology and Bioinformatics.
Jonathan Evans is a Professional Solutions Architect dealing with generative AI with the Third-Party Model Science group at AWS.
Banu Nagasundaram leads product, engineering, and tactical partnerships for Amazon SageMaker JumpStart, SageMaker's artificial intelligence and generative AI center. She is enthusiastic about constructing options that assist consumers accelerate their AI journey and unlock service value.