Difference between Build for Inferencing and Build for training?

[lamikr]

I have build the onnxruntime 17.3 now for AMD rocm by using the documentation how to do the build for inferencing and training.
https://onnxruntime.ai/docs/build/inferencing.html
https://onnxruntime.ai/docs/build/training.html

Documentation does not however explain what is the main difference between these builds. If I do the build for training, does it also include the inferencing features? Basically it training build just adds the "--enable_training" in top of the inferencing build.

Then there is also 2 AMD specific builds described in
https://onnxruntime.ai/docs/build/eps.html

and it's little unclear can I also enable both the "-use-rocm --rocm_home " and
"--use_migraphx --migraphx_home" flags on same build or does one flag remove some features from another flag?

[LaurentBerger]

I'm not an expert in ONNX and I'm facing same issue with ONNX documentation
About training it's here https://onnxruntime.ai/training and https://onnxruntime.ai/docs/get-started/training-pytorch.html and https://onnxruntime.ai/docs/api/c/training_c_cpp_api.html
To train a new model you need

In order to train a model with onnxruntime, the following training artifacts must be generated:

The training onnx model
The checkpoint file
The optimizer onnx model
The eval onnx model model (optional)

These training artifacts can be generated as part of an offline step using the python utilities made available in the onnxruntime-training python package.

[carzh]

Regarding training build vs inference build:
Inference build is the default build. Using the --enable_training build flag will build for training, which includes the inferencing capabilities with the training capabilities added. In other words, the training build contains both inferencing and training functionality.

Regarding large model training vs on-device training:
ONNXRuntime supports two types of training capabilities: large model training, which will accelerate a PyTorch Python training script with minimal code changes, and on-device training, which enables training in other platforms / languages. On-device training requires the extra step of generating 3-4 training artifacts.

We do need better documentation to make clear what ORT offerings are. Let me know if you have any more questions about training.