stopes
stopes copied to clipboard
open source publishing code
Why ?
Wikipedia expressed interest into hosting the model themselves, so I'm sharing the code to push the model to AWS
How ?
The deploy.py script will:
- create a docker image with a copy of fairseq nllb branch (the git clone isn't part of the script will add that to the readme), inference code, tokenizer model (sentencepiece)
- push the docker image to AWS image registry
- push the model weights to an S3 bucket
- create a "sagemaker model" which is basically a tuple (docker image, model weights)
- deploy the model to a "staging" Sagemaker endpoint
Test plan
Actually not tested. I used a slightly different version internally, but I've removed some hardcoded path/names from the script and I need to check if the Docker build still works