DeepSeek-R1 on Amazon Bedrock

This repository will guide you through the process of importing and using the distilled DeepSeek-R1-Distill-Llama-8B model based on Llama-3.1-8B as the base model from Hugging Face on Amazon Bedrock.

To learn more about DeepSeek-R1, please visit DeepSeek.

For detailed paper walkthrough on DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, check out this paper read on DeepSeek-R1 by Umar Jamil

Prerequisites

• AWS Account with Bedrock access
• Python environment with the following packages:
  • huggingface_hub
  • boto3

Setup Process

1. Download Model Weights

   • The model weights are downloaded from Hugging Face Hub
   • Model used: deepseek-ai/DeepSeek-R1-Distill-Llama-8B

2. Upload to S3

   • Model weights are uploaded to an S3 bucket
   • Target path: s3://[your-bucket]/models/DeepSeek-R1-Distill-Llama-8B/

3. Import to Amazon Bedrock

   • Navigate to AWS Console > Bedrock > Foundation Models > Imported Models
   • Click "Import Model"
   • Name the model (e.g., my-DeepSeek-R1-Distill-Llama-8B)
   • Provide the S3 location of the model weights
   • Wait for successful import
   • Note down the Model ARN for API calls

Usage

Run the Jupyter notebook deepseek-bedrock.ipynb for detailed implementation.

License

DeepSeek-R1 series support commercial use, allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs. Please note that:

• DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-14B and DeepSeek-R1-Distill-Qwen-32B are derived from Qwen-2.5 series, which are originally licensed under Apache 2.0 License, and now finetuned with 800k samples curated with DeepSeek-R1.
• DeepSeek-R1-Distill-Llama-8B is derived from Llama3.1-8B-Base and is originally licensed under llama3.1 license.
• DeepSeek-R1-Distill-Llama-70B is derived from Llama3.3-70B-Instruct and is originally licensed under llama3.3 license.