TEAL, Introducing Training-Free Activation Sparsity to Improve LLM Efficiency

TEAL (Training-Free Activation Sparsity in LLMs) has emerged as a groundbreaking approach to improve the efficiency of large-scale language models (LLMs) without additional training. According to together.ai, the method achieves 40-50% activation sparsity with minimal degradation by applying size pruning to the hidden state throughout the model. This innovation allows transferring fewer weights to on-chip memory, solving the memory-bound nature of LLM inference and translating into a 1.53-1.8x wall-clock speedup in single-batch decoding.

background

LLM is known for its enormous size, which makes it difficult during inference, mainly due to the speed limitation of transferring parameters from device memory to registers. Various techniques such as quantization, weight sparsity, and speculative decoding have been developed to address this ‘memory wall’. Activation sparsity, which utilizes zero values ​​in the hidden state, is a less explored method that avoids transferring unnecessary weight channels during decoding.

Older models like OPT-175B exhibit high activation sparsity, allowing significant speedups with methods like DejaVu. However, newer models like LLaMA have moved to SwiGLU variants, making these methods difficult to apply. Recent studies have attempted to ‘recover’ models that exhibit activation sparsity, but these models require extensive retraining on large datasets.

Motivational Research: Activation Distribution Characteristics of LLM

Studies have shown that the hidden states of LLM are outliers, zero-centered, and have similar distribution shapes across layers. Specifically, the states before MLP and Attention Blocks are Gaussian in shape, and the intermediate states are Laplacian in shape. This suggests that many low-amplitude activations can be eliminated with negligible model degradation, a notion also observed in other studies such as CATS.

teal

TEAL introduces optimizations by sparsifying all tensors in the model, achieving near-zero degradation at 25% sparsity and minimal degradation at 40% sparsity. At 50% sparsity, the Llama-3 variant shows slightly more degradation than its predecessors Llama-2 and Mistral. TEAL outperforms CATS by sparsifying all tensors and producing lower error by sparsifying the input.

Improved hardware recognition speed

To benchmark real-world speedups, TEAL is integrated with GPT-Fast, achieving significant speedups of up to 1.53x and 1.8x at 40% and 50% sparsity, respectively. The kernel is faster than cuBLAS at 0% sparsity, but there is still room for further optimization.

Compatibility with quantization

TEAL also demonstrates compatibility with quantization, another technique for efficient LLM inference. Combining activation sparsity and quantization opens up a new regime for transferring memory to GPU registers, leading to faster inference speeds.

Application

The most immediate application of TEAL is to accelerate inference in resource-constrained edge settings, especially in single-batch scenarios. It also enables inference providers like Together AI, which hosts over 100 open-source models on large fleets of GPUs, to serve their models more efficiently.

Image source: Shutterstock