FlagAttention is a library implementing memory-efficient attention operators in the Triton language. It provides optimized implementations of flash attention, piecewise attention, paged attention, and gated delta rule operations, designed for use in large language model training and inference. The library features modular architecture with separate implementations for each attention variant, comprehensive testing, and example usage patterns. It leverages Triton for GPU-accelerated computations while maintaining compatibility with PyTorch's autograd system.
hotpdf is a Python library for fast and memory-efficient PDF text extraction and search. It wraps pdfminer.six to load PDF pages into memory using a sparse matrix representation, a trie for text search, and a span map for grouping characters. The library provides a HotPdf class with methods to find text (with options for case sensitivity and span extraction), extract text from bounding boxes, extract spans, and merge multiple PDF documents. The architecture separates concerns into processing (processor.py), memory mapping (memory_map.py), data structures (data/classes.py), and utilities (utils.py).
nanoexpress is a professional backend framework for Node.js built on top of uWebSockets.js for high performance. It provides HTTP server functionality with routing, middleware support, WebSocket capabilities, request validation using AJV, Swagger/OpenAPI documentation generation, and utilities for file serving, compression, and cookie handling. The framework follows a modular architecture with core components including App (application wrapper), Config (configuration handling), and Route (routing system with middleware chain). It emphasizes performance while maintaining Express-like usability patterns.