Slackbot
06/06/2023, 4:09 PMChaoyu
06/06/2023, 4:57 PMMatthieu Vanhoutte
06/06/2023, 7:34 PMservice: "bentoml_gpu_ct2_service:cfct5s" # Same as the argument passed to `bentoml serve`
labels:
owner: tinycoaching-ml-team
stage: dev
include:
- "*.py" # A pattern for matching which files to include in the bento
exclude:
- "__pycache__/"
- "transformers_save_model.py"
python:
packages: # Additional pip packages required by the service
- "transformers==4.29.2"
- "tokenizers==0.13.3"
- "sentencepiece==0.1.99"
- "sacremoses==0.0.53"
- "pydantic==1.10.8"
- "six==1.16.0"
- "websocket-client==1.5.2"
- "python-dateutil==2.8.2"
- "tzlocal==5.0.1"
- "attrs==23.1.0"
- "protobuf==3.20.2"
- "ctranslate2==3.14.0"
- "starlette==0.26.1"
docker:
distro: debian
python_version: "3.8.16"
cuda_version: "11.6"
This generates in the Dockerfile FROM nvidia/cuda:11.6.2-cudnn8-runtime-ubuntu20.04 as base-debian
which is not the version I want but nvidia/cuda:11.6.0-cudnn8-runtime-ubuntu18.04 as base-debian
So I modified the Dockerfile as explained above, which functioned with previous version of bentoML (v.1.0.7). But with version v1.0.21, whatever I specify in the Dockerfile, it downloads nvidia/cuda:11.6.2-cudnn8-runtime-ubuntu20.04
Any idead @Chaoyu?Chaoyu
06/06/2023, 7:40 PMChaoyu
06/06/2023, 7:40 PMChaoyu
06/06/2023, 7:42 PM--enable-features
flag, bentoml containerize
now will regenerate Dockerfile based on user provided argumentsChaoyu
06/06/2023, 7:44 PMdocker build
command and provide the corresponding Dockerfile path and set build context to the bento directory - this is a bit hacky and rely on BentoML’s internal implementation details. But hopefully can be a quick short-term fix for your case.Matthieu Vanhoutte
06/07/2023, 7:50 AM