Speculative Decoding in vLLM on AMD GPUs