Inference optimization techniques for 10x throughput gains

Post-training optimization involves quantization, speculative decoding, and custom GPU kernel rewrites to improve model speed. These methods address throughput bottlenecks and quadratic attention scaling issues in video generation models.