prohor33/qa-roberta-ru-saas

Question answering on russian with XLMRobertaLarge as a service

QA Roberta Ru SaaS

Question answering on russian with XLMRobertaLarge as a service. Thanks for the model to Alexander Kaigorodov.

Stack

• Flask
• Gunicorn

Build image

sudo docker build . --tag qa-roberta-ru-saas

Run on CPU and predict

sudo docker run --rm -p 8080:8080 --name qa-roberta-ru-saas qa-roberta-ru-saas

curl -H "Content-Type: application/json" --data @tests/app/data/test_input.json 0.0.0.0:8080/predict

Run on GPU

Change device to cuda:0 in config before docker build:

device: cuda:0

After build:

sudo docker run --rm --gpus 0 -p 8080:8080 --name qa-roberta-ru-saas qa-roberta-ru-saas

To run with restart:

sudo docker run --gpus 0 -p 8080:8080 --restart always --name qa-roberta-ru-saas qa-roberta-ru-saas

To stop it later:

docker update --restart unless-stopped qa-roberta-ru-saas

To run tests:

pytest tests/

To run app without docker container

PYTHONPATH=. python app/app_main.py

TODO:

• GPU/CPU support
• Support of context longer than 512 bpe
• Predict on long context with sliding window