$ show google/praxis
Praxis is a JAX-based deep learning library designed for building and training large-scale machine learning models. It provides a comprehensive suite of neural network layers (transformers, convolutions, normalizations, etc.) with advanced features like quantization-aware training, sparsity techniques, and video processing components. The library emphasizes modularity through Fiddle-based configuration and includes utilities for beam search decoding, input processing, and model parallelism. Praxis is actively maintained with extensive test coverage and demonstrates strong engineering practices in its abstraction layers and hardware-specific optimizations.
The LLM identified 3 files worth highlighting for reuse:
Implements a flexible beam search algorithm with advanced features such as suffix alignment, beam masking, and early exit conditions. The component is self-contained, depending only on JAX and a few internal utility modules, making it reusable in any sequence generation project. It demonstrates skill in managing beam search state efficiently and integrating with sampling and scoring functions.
Provides a `QuantizationLayer` mixin that adds quantization-aware training capabilities to any layer via op injection. This isolates quantization concerns, allowing layers to remain agnostic to quantization details. It supports multiple quantization modes (AQT, PTQ, FQ) and precisions, demonstrating non-obvious skill in creating a flexible abstraction for model compression.
Implements a `Repeat` layer that stacks another layer a configurable number of times, with options for unrolling, checkpointing, and summary handling. This generic utility is useful for building deep architectures (e.g., transformer blocks) and only depends on the base layer and internal utilities. It demonstrates skill in providing flexible repetition strategies that optimize for memory and computation.