onnxruntime VS MLflow

ONNX Runtime: ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator

Embedding generation uses the ONNX runtime under the hood. This is a cross-platform inferencing library that supports multiple execution providers from CPU to specialized GPUs.

tfjs is dead, looking at the commit history. The standard now is to convert PyTorch to onnx, then use onnxruntime (https://github.com/microsoft/onnxruntime/tree/main/js/web) to run the model on the browsdr.

- slang[2] has the potential, but the meta programming part is not as strong as C++, existing libraries cannot be used.

The above conclusion is drawn from my work https://github.com/microsoft/onnxruntime/tree/dev/opencl, purely nightmare to work with thoes drivers and jit compilers. Hopefully Vcc can take compute shader more seriously.

[1]: https://www.circle-lang.org/

I did. It depends what you want, for an overview of how ONNX Runtime works then Microsoft have a bunch of things on https://onnxruntime.ai, but the Java content is a bit lacking on there as I've not had time to write much. Eventually I'll probably write something similar to the C# SD tutorial they have on there but for the Java API.

For writing ONNX models from Java we added an ONNX export system to Tribuo in 2022 which can be used by anything on the JVM to export ONNX models in an easier way than writing a protobuf directly. Tribuo doesn't have full coverage of the ONNX spec, but we're happy to accept PRs to expand it, otherwise it'll fill out as we need it.

What about models besides GPT? Most of the popular vector encoding models aren't using this architecture.

If you really didn't want PyTorch/Transformers, you could consider exporting your models to ONNX (https://github.com/microsoft/onnxruntime).

How can this be? The current state of practice in AI/ML work requires adaptivity, which is uncommon in classical computational fields. There are myriad tools that capture the work across the many instances of the AI/ML lifecycle. The idea that any one tool could sufficiently capture the dynamic work is unrealistic. Take, for example, an experiment tracking tool like W&B or MLFlow; some form of experiment tracking is necessary in typical model training lifecycles. Such a tool requires some notion of a dataset. However, a tool focusing on experiment tracking is orthogonal to the needs of analyzing model performance at the data sample level, which is critical to understanding the failure modes of models. The way one does this depends on the type of data and the AI/ML task at hand. In other words, MLOps is inherently an intricate mosaic, as the capabilities and best practices of AI/ML work evolve.

MLflow is an open-source platform for managing the end-to-end machine learning lifecycle. It includes features for experiment tracking, model versioning, and deployment, enabling developers to track and compare experiments, package models into reproducible runs, and manage model deployment across multiple environments.

Platforms such as MLflow monitor the development stages of machine learning models. In parallel, Data Version Control (DVC) brings version control system-like functions to the realm of data sets and models.

MLFlow - http://mlflow.org

MLflow:

In search for a tool that needs comparable configuration I looked into mlflow and found this. https://github.com/mlflow/mlflow/blob/master/mlflow/environment_variables.py There they define a class _EnvironmentVariable and create many objects out of it, for any variable they need. The get method of this class is in principle a decorated os.getenv. Maybe that is something I can take as orientation.

I have been using DVC and MLflow since then DVC had only data tracking and MLflow only model tracking. I can say both are awesome now and maybe the only factor I would like to mention is that IMO, MLflow is a bit harder to learn while DVC is just a git practically.