MemoryLLM: Plug-n-Play Interpretable Feed-Forward Memory for Transformers

TL;DR

MemoryLLM introduces a novel approach to interpret and enhance transformer feed-forward networks by decoupling them from self-attention, enabling token-wise neural retrieval memory and improving inference efficiency.

Contribution

The paper presents MemoryLLM, a method to make FFNs context-free and interpretable, and introduces Flex-MemoryLLM to bridge performance gaps in transformer models.

Findings

MemoryLLM enables token-wise lookups for FFNs, improving interpretability.

Decoupled FFNs can be pre-computed for efficient inference.

Flex-MemoryLLM reduces performance gap in transformer architectures.

Abstract

Understanding how transformer components operate in LLMs is important, as it is at the core of recent technological advances in artificial intelligence. In this work, we revisit the challenges associated with interpretability of feed-forward modules (FFNs) and propose MemoryLLM, which aims to decouple FFNs from self-attention and enables us to study the decoupled FFNs as context-free token-wise neural retrieval memory. In detail, we investigate how input tokens access memory locations within FFN parameters and the importance of FFN memory across different downstream tasks. MemoryLLM achieves context-free FFNs by training them in isolation from self-attention directly using the token embeddings. This approach allows FFNs to be pre-computed as token-wise lookups (ToLs), enabling on-demand transfer between VRAM and storage, additionally enhancing inference efficiency. We also introduce Flex-MemoryLLM, positioning it between a conventional transformer design and MemoryLLM. This architecture bridges the performance gap caused by training FFNs with context-free token-wise embeddings.