End2End Serverless Transformers On AWS Lambda for NLP π You need no servers
Deploy transformers with ease πββοΈ
Go through this video and slide deck for full info.
Current available pipelines
- classification
- sentence encoding
- translation (coming soon)
- token classification
- text generation
- zero shot classification
What you get with this?
- ability to run transformers without servers
- complete CI/CD
- concurrency upto 1000 (default AWS limit)
How to use this?
- clone the repo
- keep the pipeline folder you want to use
- modify the source and tests
- keep the corresponding github action in
.github/workflows
- modify directory, registry and lambda function name in workflow
- create repository in AWS ECR
- update ECR path in the workflow
- set up secrets in repo (needed for access to AWS; this creds should have access to ECR and Lambda)
- AWS_ACCESS_KEY_ID
- AWS_SECRET_ACCESS_KEY
- push the code
- create PR
- this will build the container
- run all the tests
- push container to ECR registry
- update lambda with the new container (this will not happen when you push the first time)
- create lambda function if it does not exist
- give appropriate IAM role
- set timeout and RAM
- create API in API gateway and link to lambda
Done! Now you can call the lambda using the API