⚡️ Beyond Transformers with Power Retention скачать в хорошем качестве

⚡️ Beyond Transformers with Power Retention

Jacob Buckman, CEO of Manifest AI, joins us to discuss their solution to one of AI's most expensive computational bottlenecks: the ever-growing KV cache for long context inference. Power Retention replaces Attention in the transformer architecture, and moves memory requirements from growing to fixed memory size. Show Notes: https://huggingface.co/manifestai/pow... https://github.com/m-a-n-i-f-e-s-t/re... https://github.com/m-a-n-i-f-e-s-t/vi... https://manifestai.com/articles/relea... 00:00 Introduction and Manifest AI's Mission 01:36 The Core Problem: Growing KV Cache 04:12 Power Retention Architecture Explained 05:55 Inference Infrastructure Benefits 08:35 Vidrial: Just-in-Time CUDA Kernel Framework 14:44 Metamorphosis: Converting Transformers to Power Retention 19:35 Live Demo: PowerCoder Performance 26:04 Open Source Strategy and Community Building 30:07 The Data Challenge for Long Context