Deploy DL/ ML inference pipelines with minimal extra code.
Project description
fastDeploy
easy and performant micro-services for Python Deep Learning inference pipelines
- Deploy any python inference pipeline with minimal extra code
- Auto batching of concurrent inputs is enabled out of the box
- no changes to inference code (unlike tf-serving etc), entire pipeline is run as is
- Promethues metrics (open metrics) are exposed for monitoring
- Auto generates clean dockerfiles and kubernetes health check, scaling friendly APIs
- sequentially chained inference pipelines are supported out of the box
- can be queried from any language via easy to use rest apis
- easy to understand (simple consumer producer arch) and simple code base
Installation:
pip install --upgrade fastdeploy fdclient
# fdclient is optional, only needed if you want to use python client
CLI explained
Start fastDeploy server on a recipe:
# Invoke fastdeploy
python -m fastdeploy --help
# or
fastdeploy --help
# Start prediction "loop" for recipe "echo"
fastdeploy --loop --recipe recipes/echo
# Start rest apis for recipe "echo"
fastdeploy --rest --recipe recipes/echo
Send a request and get predictions:
auto generate dockerfile and build docker image:
# Write the dockerfile for recipe "echo"
# and builds the docker image if docker is installed
# base defaults to python:3.8-slim
fastdeploy --build --recipe recipes/echo
# Run docker image
docker run -it -p8080:8080 fastdeploy_echo
Serving your model (recipe):
Where to use fastDeploy?
- to deploy any non ultra light weight models i.e: most DL models, >50ms inference time per example
- if the model/pipeline benefits from batch inference, fastDeploy is perfect for your use-case
- if you are going to have individual inputs (example, user's search input which needs to be vectorized or image to be classified)
- in the case of individual inputs, requests coming in at close intervals will be batched together and sent to the model as a batch
- perfect for creating internal micro services separating your model, pre and post processing from business logic
- since prediction loop and inference endpoints are separated and are connected via sqlite backed queue, can be scaled independently
Where not to use fastDeploy?
- non cpu/gpu heavy models that are better of running parallely rather than in batch
- if your predictor calls some external API or uploads to s3 etc in a blocking way
- io heavy non batching use cases (eg: query ES or db for each input)
- for these cases better to directly do from rest api code (instead of consumer producer mechanism) so that high concurrency can be achieved
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
fastdeploy-3.0.14.tar.gz
(14.6 kB
view hashes)
Built Distribution
Close
Hashes for fastdeploy-3.0.14-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | a877564f4cd65c3eaab198866a819542b3bc1f209088a34b34cf7d760f97a048 |
|
MD5 | 8741b4a6f5ab7fece968a3e83e485707 |
|
BLAKE2b-256 | 231bf77b3779a1848dd1b842d530e0f26b160f23f637fe60e1ee55689de3c2fb |