Sieve functions run on a Debian-based OS with Python 3.8 by default. Typically, you’ll want to customize this to your own liking, which Sieve let’s you do in a few ways.

One think you’ll typically want to do is add custom python packages, like opencv. To do this, you can add it to a list of packages as follows under the python_packages tag.

@sieve.function(
	name="hello",
	python_packages=[
		"opencv-python"
	]
)
def hello(img: sieve.Image) -> sieve.Image:
	import cv2
	...

In the same way, you can also specify custom system packages that will get apt-install-ed.

@sieve.function(
	name="hello",
	python_packages=[
		"opencv-python"
	],
	system_packages=[
		"ffmpeg"
	]
)
def hello(img: sieve.Image) -> sieve.Image:
	import subprocess
	subprocess.run(["ffmpeg", "-h"])
	...

Sieve also let’s you specify custom shell commands for example if you want to download model weights or anything else to the container before it runs.

@sieve.function(
	name="hello",
	python_packages=[
		"opencv-python"
	],
	system_packages=[
		"ffmpeg"
	],
	run_commands=[
		"mkdir -p /root/.cache/models/",
    "wget -c 'https://whisperx.s3.eu-west-2.amazonaws.com/model_weights/segmentation/0b5b3216d60a2d32fc086b47ea8c67589aaeb26b7e07fcbe620d6d0b83e209ea/pytorch_model.bin' -P /root/.cache/models/"
	]
)
def hello(img: sieve.Image) -> sieve.Image:
	import subprocess
	subprocess.run(["ffmpeg", "-h"])
	...

There are loads of other things you can customize including CUDA versions, python versions, and more which you can read more about here

Example: Video DubbingGPU Acceleration