Zach Anderson
Sep 01, 2024 08:34
TEAL offers a training-free approach to activation sparsity, significantly enhancing the efficiency of large language models (LLMs) with minimal degradation.
TEAL (Training-Free Activation Sparsity in LLMs) has emerged as a groundbreaking approach to improve the efficiency of large language models (LLMs) without requiring additional training. According to together.ai, this method applies magnitude pruning to hidden states throughout the model, achieving 40-50% activation sparsity with minimal degradation. This innovation allows for the transfer of fewer weights to on-chip memory, addressing the memory-bound nature of LLM inference and translating into 1.53-1.8x wall-clock speedups in single-batch decoding.
Background
LLMs are known for their massive size, which poses challenges during inference, primarily due to the speed limitations of transferring parameters from device memory to registers. Various techniques such as quantization, weight sparsity, and speculative decoding have been developed to tackle this ‘memory wall’. Activation sparsity, which leverages zero values in hidden states, is a less explored method that avoids transferring unnecessary weight channels during decoding.
Older models like OPT-175B show high activation sparsity, enabling methods like DejaVu to achieve significant speedups. However, newer models like LLaMA have moved to SwiGLU variants, making it harder to apply such methods. Recent research has attempted to ‘recover’ models that exhibit activation sparsity, but these require extensive retraining on massive datasets.
Motivating Study: Distributional Properties of Activations in LLMs
Research has shown that hidden states in LLMs exhibit outliers and are zero-centered with similar distributional shapes across layers. Specifically, states before MLP and Attention Blocks are Gaussian-shaped, while intermediate states are Laplacian-shaped. This suggests that many low-magnitude activations can be pruned with negligible model degradation, a concept also observed in other studies like CATS.
TEAL
TEAL introduces an optimization by sparsifying every tensor in the model, achieving near-zero degradation at 25% sparsity and minimal degradation at 40% sparsity. At 50% sparsity, Llama-3 variants show slightly more degradation compared to older Llama-2 and Mistral variants. TEAL outperforms CATS by sparsifying every tensor and choosing to sparsify through input, yielding lower error.
Hardware-Aware Speed-up
To benchmark real-world speedups, TEAL was integrated with GPT-Fast, achieving significant speedups of up to 1.53x and 1.8x at 40% and 50% sparsity, respectively. While the kernel is faster than cuBLAS at 0% sparsity, 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 unlocks new regimes for transferring memory to GPU registers, allowing for higher inference speed-ups.
Applications
TEAL’s most immediate application is accelerating inference in resource-constrained edge settings, particularly in single-batch scenarios. It also aids inference providers like Together AI, which hosts over 100 open-source models across a large fleet of GPUs, by serving models more efficiently.
Image source: Shutterstock
Credit: Source link
