GitHub Trends

#python#agent#llm#llm_agent#llm_reasoning#machine_learning_systems#mlsys#reinforcement_learning#rl AReaL is a free, open-source system for fast asynchronous reinforcement learning to train large AI models in math, coding, search, and agents. It decouples generation and training for up to 2.77x speedup, stable performance, and easy setup on single or 1000+ GPUs with algorithms like GRPO/PPO. Install via git/pip, run examples like GSM8K math instantly. You benefit by building top AI agents affordably and quickly, reproducing results with shared data/models, saving time/money vs. slow synchronous tools. https://github.com/inclusionAI/AReaL

#python#artificial_intelligence#cloud_ml#computer_systems#courseware#deep_learning#edge_machine_learning#embedded_ml#machine_learning#machine_learning_systems#mobile_ml#textbook#tinyml You can learn how to build real-world AI systems from start to finish with an open-source textbook originally from Harvard University. It teaches you not just how to train AI models but how to design scalable systems, manage data pipelines, deploy models in production, monitor them continuously, and optimize for devices like phones or IoT gadgets. This helps you become an engineer who can create efficient, reliable, and sustainable AI systems that work well in practice. The book offers hands-on labs, community support, and free online access, making it easier to gain practical skills in machine learning systems engineering. https://github.com/harvard-edge/cs249r_book

#python#large_language_models#machine_learning_systems#natural_language_processing Flash Linear Attention (FLA) is a fast, memory-efficient library for advanced linear attention models used in transformers, written in PyTorch and Triton, and compatible with NVIDIA, AMD, and Intel GPUs. It offers many state-of-the-art linear attention models and fused modules that speed up training and reduce memory use. You can easily replace standard attention layers in your models with FLA’s efficient versions, improving training and inference speed, especially for long sequences. FLA supports hybrid models mixing linear and standard attention, and integrates with Hugging Face Transformers for easy use and evaluation. This helps you train and run large language models faster and with less memory, making your AI projects more efficient and scalable. https://github.com/fla-org/flash-linear-attention