How to Serve PyTorch Models with TorchServe
Hamid Shojanazeri is a Partner Engineer at PyTorch, here to demonstrate the basics of using TorchServe. As the preferred model serving solution for PyTorch, TorchServe allows you to expose a web API for your model that may be accessed directly or via your application. With default model handlers that perform basic data transforms, TorchServe can be a very effective tool for those participating in our Hackathon.
What is new?
New examples for serving HuggingFace Transformers, MMF: https://github.com/pytorch/serve/tree...
Ensemble model support with examples for Neural Machine Translation
Model interpretability using Captum.ai - https://github.com/pytorch/serve/tree...
Kubeflow Pipelines support : for open source Kubeflow pipelines and Google Vertex AI (samples https://github.com/kubeflow/pipelines...)
KFServing integrations with Canary rollouts and auto-scaling (samples https://github.com/kubeflow/kfserving...) (blog https://blog.kubeflow.org/release/off...)
MLFlow integrations with examples -- Open source library for MLOps - https://github.com/mlflow/mlflow-torc...
TorchServe on the cloud
AWS: https://torchserve-on-aws.workshop.aw...
Google Cloud: https://cloud.google.com/ai-platform/...
Microsoft: https://techcommunity.microsoft.com/t...
Dynabench and Flores
https://dynabench.org/tasks/3#overall
https://github.com/facebookresearch/d...
http://www.statmt.org/wmt21/large-sca...
PyTorch Github repo: github.com/pytorch/serve
To register for the PyTorch 2021 Annual Hackathon, please visit https://pytorch2021.devpost.com
