Goal: Benchmark attention/MLP layers on CPU, GPU, and Apple M-series. Deliverable: Latency, FLOPs utilization, memory bottlenecks. Grading: Benchmark correctness (30%), Comparison depth (40%), Report (30%). Reference: pytorch/examplesLinks to an external site.