A high-speed AI learning and inference engine developed in Rust can now be easily and flexibly controlled from Python. The key point is that **"the computationally intensive parts are processed quickly in Rust, while the parts requiring flexibility, such as data preparation and control, can be written in Python."** The three main advancements are as follows: 1. Building a "learning loop" solely in Python (PyTrainer) The control of the learning progress, which was previously hidden within Rust, can now be operated step by step from Python. Optimized role distribution: Data loading and processing can be done using familiar Python libraries (like Pandas), while the subsequent computational processing can be left to Rust. Simple operation: By just calling the command train_step, complex calculations such as forward propagation, backward propagation, and parameter updates are processed collectively on the Rust side. High parallel performance: During computation, Python's limitations (GIL/exclusive locks) are lifted, allowing for smooth operation even when data loading or GUI updates are happening in the background without interrupting the learning process. Incorporation of the latest technology: The latest optimizers implemented in Rust (like ScheduleFreeOptimizer) can be utilized directly. 2. Direct use of a high-speed inference engine (BitLlama) It is now possible to directly load the lightweight and fast BitNet model from Python for text generation. Elimination of overhead: By generating text (token sequences) all at once on the Rust side without using Python's loop processing, the speed reduction specific to Python does not occur. Practical applications: It can be used immediately for checking whether learning is progressing (text output) or as the backend for a lightweight chatbot. 3. Complete state saving and restoration It is now possible to save and restore the learning progress, including detailed internal states, to a file. Complete backup: Not only the model's "weights" but also the optimizer's "internal states (parameters like momentum)" are saved together. Safe interruption and resumption: It can be saved with just the command save_checkpoint, making it easy to manage interruptions and resumptions of learning from Python scripts. Future prospects With this foundation in place, it will be easier to create a "system to manage the entire learning process" on the Python side or to integrate with existing learning GUI tools.