MLflow VS onnxruntime

One day I was hopping around a few popular ML libraries in Python, including MLflow. While glancing at its source code, one class attracted my interest, LazyLoader in __init__.py (well, this actually mirrors from the wandb project, but the original code has changed from what MLflow is using now, as you can see).

Tools: Implement logging using tools like MLFlow or Weights & Biases (W&B), which provide a structured way to track experiments, compare them visually, and share findings with your team. These tools integrate seamlessly with most machine learning frameworks, making it easier to adopt them in your existing workflows.

CometML and mlMLFlow are popular development and experimentation tools, although some express concerns about their proprietary and weak data storage with its lack of tamper-proof guarantees.

MLflow is an open source MLOps tool that allows users to manage the entire life cycle of machine learning models. It has four key components:

Experiment tracking tools like MLflow, Weights and Biases, and Neptune.ai provide a pipeline that automatically tracks meta-data and artifacts generated from each experiment you run. Although they have varying features and functionalities, experiment tracking tools provide a systematic structure that handles the iterative model development approach.

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.

First time I see this, by the first impression I think the output quality of this one is better than mine and my code is only based on one model

In my understanding, It'll possible if the model's author build to https://onnxruntime.ai ONNX Runtime. And maybe the downside is user will need to download ton of data to their device, currently it's ~100-200mb

ONNX Runtime is a high-performance inference engine for executing ONNX (Open Neural Network Exchange) models. It provides a simple way to run large language models like Llama, Phi, Gemma, and Mistral via the onnxruntime-genai API.

Starting from version 1.5.1 the backend integrates changes borrowed from sam_onnx_full_export, to support OnnxRuntime 1.17.x and later versions. Please note that on MacOS directly running the project from the command line suffers from memory leaks, making inference operations slower than normal. It's best therefore running the project inside a docker container, unless in case of development or debugging activities.

I've found a suitable for my idea ONNX example. I'm going to use this example as a strong foundation for the project. But to make things more interesting I'll add just a few enhancements:

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.