architecture-optimizer-codesign

Introduces Parallax, a parameterized local linear attention mechanism with hardware-aware optimization that improves LLM pretraining efficiency and performance, achieving Pareto improvements at 0.6B and 1.7B scales.