When deploying DeepSeek-R1 on AWS, the storage requirements are not explicitly detailed in terms of specific AWS storage services. However, the model itself requires significant storage space for its weights. For instance, the full DeepSeek-R1 model originally has a footprint of about 720GB, though optimizations can reduce this size significantly, such as down to 131GB through quantization and compression techniques[3].
For AWS deployments, you typically need to consider the following aspects:
1. Model Storage: The model weights need to be stored in a location accessible by your AWS instances. This could be in an S3 bucket, which is commonly used for storing large datasets and model files. The cost of storing the model in S3 depends on the region and the storage class used (e.g., Standard, Standard-IA, One Zone-IA, etc.).
2. Instance Storage: When running the model on EC2 instances, you'll need sufficient storage for the operating system, any additional software required, and potentially temporary storage for model outputs or intermediate results. This storage is typically provided by the instance's root volume or additional EBS volumes.
3. Data Storage for Inference: If you're using the model for inference tasks, you might need additional storage for input data and output results. This could also be managed through S3 or other AWS storage services like EBS or instance store volumes.
In terms of specific AWS services, Amazon S3 is often used for storing model weights due to its scalability and reliability. For running the model, Amazon EC2 instances are commonly used, with storage needs depending on the instance type and configuration. For large-scale deployments, Amazon SageMaker can also be utilized, which provides a managed environment for machine learning tasks, including model storage and inference.
For cost considerations, the storage costs in AWS depend on the specific services used. For example, S3 storage costs vary based on the storage class and region, while EC2 instance costs include both compute and storage components. Additionally, using services like SageMaker can simplify the management of these costs by providing a managed environment for ML tasks.
Citations:
[1] https://dev.to/askyt/deepseek-r1-architecture-training-local-deployment-and-hardware-requirements-3mf8
[2] https://www.byteplus.com/en/topic/383522
[3] https://www.reddit.com/r/selfhosted/comments/1ic8zil/yes_you_can_run_deepseekr1_locally_on_your_device/
[4] https://aws.amazon.com/blogs/machine-learning/deploy-deepseek-r1-distilled-models-on-amazon-sagemaker-using-a-large-model-inference-container/
[5] https://www.aboutamazon.com/news/aws/aws-deepseek-r1-fully-managed-generally-available
[6] https://aws.amazon.com/blogs/machine-learning/optimize-hosting-deepseek-r1-distilled-models-with-hugging-face-tgi-on-amazon-sagemaker-ai/
[7] https://repost.aws/questions/QU-hcixrtFSaSoKH8GL-KogA/pricing-model-of-deepseek-r1-distilled-llama-models-with-amazon-bedrock-custom-model-import
[8] https://community.aws/content/2Z6DlAohx12yuNoEAs7qb5YTH0q/leveraging-deepseek-r1-on-aws?lang=en
[9] https://community.aws/content/2sEuHQlpyIFSwCkzmx585JckSgN/deploying-deepseek-r1-14b-on-amazon-ec2?lang=en
[10] https://aws.amazon.com/blogs/aws/deepseek-r1-now-available-as-a-fully-managed-serverless-model-in-amazon-bedrock/
[11] https://vagon.io/blog/a-step-by-step-guide-to-running-deepseek-r1-on-vagon-cloud-desktops