ilyahabr/yolo_trocr_bentoml

for reproduce errors

yolo_trocr_bentoml

Virtual env

create env

virtualenv -p python3.8 bentoml_env

install libs

pip install -r requirements.txt

download yolov8 weights

python download_models.py

change permissions (because in docker it won't work without it)

chmod 776 yolov8n-seg.pt
chmod 776 yolov8n.pt

---

RUN WITH SERVE

BENTOML_CONFIG=configuration.yml bentoml serve service.py:svc -p 8995 --development --reload --debug 

Questions for serve:

1. How to allocate gpu: 1 with it?

---

RUN WITH DOCKER

bentoml build

bentoml bentoml containerize ml_pipeline_service:latest

docker run --rm --gpus '"device=1"' -v ./configuration.yml:/home/bentoml/configuration.yml -v $(pwd):/home/bentoml/bento/src/  -p 8995:3000 ml_pipeline_service:(your_tag)

---

Problems:

1. Memory issues arise when executing within a container, witnessing a cumulative increase of 500MB with each inference attempt.

2. While performing a 30-batch operation, an error is encountered, surpassing the 60-second time limit.(can't reproduce in this repo because of different data and pre/post processing)

3. Running with BENTOML_CONFIG=configuration.yml bentoml serve service.py:svc -p 8995 --development --reload --debug consumes significantly less memory than execution within a Docker container. Approximately 3GB for serve of memory is utilized, and for docker 6-8GB, which escalates further with increased batch sizes

4. A RuntimeError is experienced with a batch size of 30, displaying the message: "Unexpected ASGI message 'http.response.start' sent after the response has already been completed." Despite successful functionality and the desired response, this error persists. (can't reproduce in this repo because of different data and pre/post processing)

5. Swagger functionality is impaired when using the Docker container, but it works seamlessly with bentoml serve. Investigating the reason for this disparity.

6. What benefits I get while using custom runner with transformers lib and bentoml.transformers?

7. Can't send defaultdict as input, but send as output is fine.