Accelerating decode-heavy LLM inference with speculative decoding on AWS Trainium and vLLM | Modeldex