Efficient Speculative Decoding for Llama at Scale: Challenges and Solutions
This work addresses efficiency problems for deploying large language models in production environments, though it is incremental as it optimizes an existing method.
The paper tackles engineering challenges in scaling speculative decoding for Llama models in production, achieving a new state-of-the-art inference latency of about 4 ms per token (10% faster than previous methods) and speed-ups of 1.4x to 2.0x for large batch sizes.
Speculative decoding is a standard method for accelerating the inference speed of large language models. However, scaling it for production environments poses several engineering challenges, including efficiently implementing different operations (e.g., tree attention and multi-round speculative decoding) on GPU. In this paper, we detail the training and inference optimization techniques that we have implemented to enable EAGLE-based speculative decoding at a production scale for Llama models. With these changes, we achieve a new state-of-the-art inference latency for Llama models. For example, Llama4 Maverick decodes at a speed of about 4 ms per token (with a batch size of one) on 8 NVIDIA H100 GPUs, which is 10% faster than the previously best known method. Furthermore, for EAGLE-based speculative decoding, our optimizations enable us to achieve a speed-up for large batch sizes between 1.4x and 2.0x at production scale.