Lambda-Instructor is an experimental deployment of the text-embedding model Instructor-Large on AWS Lambda.
Instructor-Large is a model built by the NLP Group of The University of Hong Kong under the Apache-2.0 license and performs well on retrieval tasks (i.e. finding related documents for a given sentence). As of June-2023, it seems to be on a level with OpenAI's text-embedding-ada-002 across numerous benchmarks on the Hugging Face leaderboard.
Potential Use Cases:
- Pair with local inference: Generate a large number of embeddings with Instructor-Large on your local machine upfront (vs. paying for commercial APIs such as OpenAI's even for testing).
- Data residency: Deploy the Lambda function in global AWS Regions such including Europe and the US.
- Scalability: AWS Lambda scales virtually unlimited, without the overhead of managing servers.
- Low-cost production inference: with AWS Lambda's per-request pricing vs. running a server 24/7.
- Use with vector databases: such as ChromaDB or Pinecone.
Performance, Cost, and Limitations:
- Cold Start: The Lambda function has a cold start of about 150 seconds (with instructor-base about 50s).
- Inference: Consecutive request are processed in about 6 seconds per request (with instructor-base about 3 seconds).
- Max. Tokens: Instructor-Large seems to be capped at a sequence length of 512 tokens (ca. 380 words), whereas OpenAI's text-embedding-ada-002 supports up to 8191 tokens (ca. 6000 words).
- Cost: The AWS Lambda cost can be estimated at:
- Configuration:
- Region: eu-central-1 (Europe, Frankfurt)
- Arm Price: $0.0000133334 for every GB-second
- Requests: $0.20 per 1M requests
- Memory Size: 10240 MB
- Calculation:
- Duration: $0.0000133334 * 10,24 GB * 6 seconds = 0.000819204096 / request
- Requests: $0.20 / 1M requests = 0.0000002 / requests
- Total = $0.0008194 / request
- As of June 2023, that is pricier than OpenAI's Ada v2 at $0.0001/1K tokens after their 75% price reduction.
- Configuration:
Further improvements: Cost and Cold start could be further improved with AWS Lambda Provisioned Concurrency and AWS Compute Savings Plans. Also check out the AWS Calculator at 1M requests per month for this project.
- General AWS knowledge is helpful.
- You need to have Docker and git-lfs installed locally.
- You need to have aws-sam installed locally and configured with your AWS account.
- An Arm-based local environment is recommended (Mac M1/M2). If you run into issues, try deploying via an Arm-based EC2 machine (guide below).
We will download the instructor model locally, package the app in a Docker container and deploy it on an Arm-based Lambda function. We're using Arm as it is lower-cost with AWS and turned out to be faster.
- Clone the repository
git clone https://github.com/maxsagt/lambda-instructor.git
- Clone the instructor-large model to the ./app/model folder (see https://huggingface.co/hkunlp/instructor-large?clone=true)
git lfs install ./app/model git clone https://huggingface.co/hkunlp/instructor-large ./app/model
- Build the docker container with aws-sam
sam build --cached --parallel
- Test locally with the sample payload in event.json.
sam local invoke -e event.json
- Deploy to AWS. Note that your AWS user or role needs (temporary) IAM permissions for AWS CloudFormation, Elastic Container Registry, S3, Lambda and IAM.
sam deploy
- Done. You will find a Lambda function in AWS that is ready for further configuration. For example:
- Test the Lambda function in the Lambda console.
- Configure a Lambda Function URL to directly use the Lambda function via a URL.
- Add an API Gateway for more advanced API functionalities.
If you do not have an Arm machine at hand, or want to deploy within AWS for faster uploading of the docker container.
- Create an AWS instance:
- Tested on Ubuntu Server 22.04 LTS (HVM), SSD Volume Type, ami-0329d3839379bfd15, Architecture: 64-bit (Arm), Instance Type: t4g.medium, Storage: 20GiB gp3.
- Make sure your VPC is public.
- Start the instance.
- Click on the instance and add an IAM role via Actions -> Security. The IAM Role needs IAM, S3, ECR, Cloudformation and Lambda access.
- Upload repository to remote AWS instance:
git clone https://github.com/maxsagt/lambda-instructor.git cd lambda-instructor zip -r deployment.zip . -x './app/model/*' PUBLIC_IPv4_DNS=ec2-x-xx-xxx-xxx.eu-central-1.compute.amazonaws.com scp -o StrictHostKeyChecking=no -i "../my_key.pem" ./deployment.zip ubuntu@$PUBLIC_IPv4_DNS:deployment.zip
- Log in to your instance and execute the sample deployment script.
ssh -o StrictHostKeyChecking=no -i "../my_key.pem" ubuntu@$PUBLIC_IPv4_DNS sudo apt install unzip unzip -o deployment.zip chmod +x sample_deployment.sh sudo bash -x sample_deployment.sh
- Build and deploy as per above. If the command sam is not recognized, read the sam documentation here.
- Don't forget to terminate the instance after use.
Feedback and contributions are welcome!