Speculative Decoding in vLLM on AMD GPUs

Speculative decoding lets vLLM verify multiple drafted tokens in one target-model pass, boosting output-token throughput. Tests on AMD Instinct MI300X and MI355X GPUs with ROCm covered five drafting methods: native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark. Performance varied by model family, draft checkpoint, workload, and acceptance behavior.