FFN Fusion: Rethinking Sequential Computation in Large Language Models
This addresses the issue of slow and costly inference for users of large language models, though it is incremental as it builds on existing optimization techniques like quantization and pruning.
The paper tackles the problem of high sequential computation in large language models by introducing FFN Fusion, an architectural optimization technique that parallelizes sequences of Feed-Forward Network layers, resulting in a 1.71X speedup in inference latency and 35X lower per-token cost for a 253B parameter model while maintaining strong benchmark performance.
We introduce FFN Fusion, an architectural optimization technique that reduces sequential computation in large language models by identifying and exploiting natural opportunities for parallelization. Our key insight is that sequences of Feed-Forward Network (FFN) layers, particularly those remaining after the removal of specific attention layers, can often be parallelized with minimal accuracy impact. We develop a principled methodology for identifying and fusing such sequences, transforming them into parallel operations that significantly reduce inference latency while preserving model behavior. Applying these techniques to Llama-3.1-405B-Instruct, we create Llama-Nemotron-Ultra-253B-Base (Ultra-253B-Base), an efficient and soon-to-be publicly available model that achieves a 1.71X speedup in inference latency and 35X lower per-token cost while maintaining strong performance across benchmarks. Through extensive experiments on models from 49B to 253B parameters, we demonstrate that FFN Fusion becomes increasingly effective at larger scales and can complement existing optimization techniques like quantization and pruning. Most intriguingly, we find that even full transformer blocks containing both attention and FFN layers can sometimes be parallelized, suggesting new directions for neural architecture design.