This is a technical deep-dive into speculative decoding, a production optimization where a lightweight "draft" model proposes multiple candidate next tokens and the main model verifies them in one batch pass. The source covers five methods (native MTP, Gemma 4 MTP, EAGLE-3, DFlash, DSpark) tested on AMD MI300X/MI355X GPUs with models like Gemma 4 and Qwen, reporting throughput gains ranging from 1.08× to 2.87× depending on model, workload, and proposal length. The practical takeaway is concrete: speculative decoding works, but optimal token-proposal depth varies by task and checkpoint—you need to measure on your own workload. Configuration is straightforward via `--speculative-config`, though separate checkpoints are required for all methods except native MTP. The source also covers how to train your own speculator and what acceptance metrics matter.