Efficiency: FlashAttention, quantization, speculative decoding

A · 10-second intuition Make each token cheaper by moving fewer bytes (fused kernels, smaller numbers) or by guessing several tokens and verifying them at once.