Neuron Software Classification
Contents
This document is relevant for: Inf1
, Inf2
, Trn1
, Trn1n
Neuron Software Classification#
Table of Contents
Overview#
This document explains the Neuron software classification for APIs, libraries, packages, features, and Neuron supported model classes mentioned in the Neuron documentation.
Note
For APIs, libraries, packages, features and model classes, only Alpha and Beta software classifications will be mentioned. Otherwise, they should be considered as “Stable.”
Note
APIs, libraries, packages, features, and model classes at Alpha or Beta classification should not be used in production environments, and are meant for early access and feedback purposes only. Alpha and Beta releases are Developer Preview releases under the AWS SDK policy.
APIs Software Classification#
This section details the classification of APIs supported in any Neuron
Components, in addition to environment variables and flags (e.g.
compiler flags). Examples of APIs supported by Neuron are Neuron APIs like
torch_neuron.trace()
, Neuron Environment variables like
PyTorch NeuronX Environment Variables, and Neuron flags like Neuron compiler flags.
Note
Alpha and Beta classified APIs are APIs in a Developer Preview release (see AWS SDK policy) that should not be used in production environments and are meant for early access and feedback purposes only.
API Contract |
API Backward Compatibility |
|
---|---|---|
Alpha |
Major changes may happen |
No |
Beta |
Minor changes may happen |
No |
Stable |
Incremental changes in new releases (without breaking the API contract)* |
Yes* |
Note
*In case when a new Neuron version of a Stable release will break backwards compatibility, AWS will notify customers of the breaking change at least one month before the change.
Packages / Libraries Software Classification#
This section details the classification of Neuron packages or libraries such as Neuron Runtime, PyTorch Neuron or Neuron Distributed.
Note
Alpha and Beta classified packages/libraries are packages/libraries in a Developer Preview release (see AWS SDK policy) that should not be used in production environments and are meant for early access and feedback purposes only.
Testing |
Features |
Performance |
|
---|---|---|---|
Alpha |
Basic |
Basic |
|
Beta |
Basic |
Minimal Viable Product (MVP)* |
|
Stable |
Standard Product Testing |
Incremental additions/changes in new releases |
Tested |
Note
*A minimum viable product (MVP) for a package/library contains just enough features to be usable by early customers who can then provide feedback for future development. MVP can be different per use case and depends on the specific package/library of interest. Please note that in many cases, an MVP can also represent an advanced level of features.
Features Software Classification#
This section details the classification for Neuron features. An example of a Neuron feature is Neuron Persistent Cache in the Transformers Neuron library.
Note
Alpha and Beta classified features are features in a Developer Preview release (see AWS SDK policy) that should not be used in production environments and are meant for early access and feedback purposes only.
Testing |
Functionality |
Performance |
|
---|---|---|---|
Alpha |
Basic |
Basic |
|
Beta |
Basic |
Minimal Viable Product (MVP)* |
|
Stable |
Standard Product Testing |
Incremental additions/changes in new releases |
Tested |
Note
*A minimum viable product (MVP) for a feature contains just enough functionality to be usable by early customers who can then provide feedback for future development. MVP can be different per use case and depends on the specific feature of interest. Please note that in many cases, an MVP can also represent an advanced level of functionality.
Neuron Model Classes Software Classification#
This section details the classification for Neuron model classes which mainly refers throughput/latency and accuracy for both training and inference.
Note
A Neuron supported model class is tightly coupled with a specific supported ML Framework (e.g. PyTorch Neuron), specific ML library (e.g. NeuronX Distributed) and the workload type (e.g. Training or Inference). For example a model can be supported at Beta level in PyTorch Neuron for training and Stable level in PyTorch Neuron for inference.
Note
Alpha and Beta classified model classes are model classes in a Developer Preview release (see AWS SDK policy) that should not be used in production environments and are meant for early access and feedback purposes only.
Accuracy / Convergence |
Throughput / Latency |
|
---|---|---|
Beta |
Validated |
|
Stable |
Validated |
Tested |
This document is relevant for: Inf1
, Inf2
, Trn1
, Trn1n