Hey. BentoML 1.2.5 isn't locking my Python dependencies correctly. I have the following `bentoml.yaml`:

service: "src.api.bert_classifier.api:svc"
labels:
  owner: delta
  project: event-tagging-svc
include:
  - "src/**/*.py"
  - "src/**/*.yaml"
  - "src/**/*.sh"
python:
  packages:
    - bentoml
    - torch
    - pytorch-lightning
    - torchmetrics
    - transformers
    - pydantic<2
    - dotmap
    - nltk
    - numpy
    - pandas
    - scikit-learn
    - beautifulsoup4
    - lxml
  extra_index_url:
    - "<https://download.pytorch.org/whl/cu118>"
docker:
  base_image: "<http://gcr.io/klue-1265/cuda-miniconda3:cuda119|gcr.io/klue-1265/cuda-miniconda3:cuda119>"
  setup_script: "src/api/bert_classifier/setup.sh"

And I have use

pipenv

to deal with my dependencies. My environment has

transformers = "==4.27.3"

, for example, but the generated

requirements.lock.txt

contains

transformers==4.38.2

. I'm building it with:

pipenv run bentoml build -f src/api/bert_classifier/bentofile.yaml

So, I'm sure it's using the correct environment during the build.

Might be an issue with pipenv and BentoML interaction, were you able to do it this way with older versions of BentoML? We didn’t make any changes there recently

I was reading the code and noticed that you guys use

piptools compile

. After further investigation, it seems they're not compatible with pipenv.

Hi Fernando, sorry for messaging you directly. I’m new to bentoml and I’m trying to build a bento with python dependencies using the following config. But when I did

bentoml build

, the resulting bento is very small (~20KB) and apparently it does not have the dependencies. It seems you were able to build bentos with dependencies included. Could you share how you did it? Was it simply configure a bentofile.yaml with dependencies and run `bentoml build`…? Thanks!

python:
  packages:
  - bentoml
  - open_clip_torch
  - opencv-python

@Fernando Camargo yes if you’re already using pipenv to manage your dependency versions, I’d recommend putting the generated requirements.txt directly in bentofile.yaml and set lock_packages to false so BentoML won’t try to pin dependencies again

What does “during deployment time” mean…? If I want to use bentoml as a library and use Ah ok got it. But if I wanted to use

bentoml.batch.run_in_spark

, how do I make sure Spark workers have the required dependencies?

In the case of spark, if you want to run Bento in-place on the spark workers, currently users will need to ensure the libraries are installed themselves

Ok sounds like the only way to decouple client logic from server dependencies is to use BentoCloud?

Using BentoCloud would definitely be easier to setup. It is possible to deploy OSS BentoML to your own infrastructure and connect to it in your spark job a well. You may need to manually manage the deployment lifecycle to match your batch jobs’ cadence as well as doing autoscaling

Ok got it, thank you so much Chaoyu. Sorry for hijacking your thread Fernando.