Large language models (LLMs) are increasingly utilized for complex tasks requiring longer context lengths, with some models supporting up to 128K or 1M tokens. This trend, however, presents significant challenges in inference speed and memory management. The primary bottleneck in long-context LLM inference is the quadratic computational complexity of attention mechanisms, causing substantial slowdowns as sequence length increases. KV cache mechanism alleviates this issue by storing pre-computed data, but introduces memory requirements that scale linearly with context length, hindering efficient LLM deployment. Quantization emerges as a promising approach to address the widening gap between LLM size and memory capacity. However, traditional quantization schemes often yield suboptimal compression results for KV caches due to two key factors: i) On-the-fly quantization and de-quantization, causing significant performance overhead; ii) Prevalence of outliers in KV values, challenging low-bitwidth uniform quantization. To this end, we propose EVASION, a novel quantization framework achieving low-bitwidth KV cache through product quantization. First, we conduct a thorough analysis of KV cache distribution, revealing the limitations of existing quantization schemes. Second, we introduce a non-uniform quantization algorithm based on product quantization, which efficiently compresses data while preserving accuracy. Third, we develop a high-performance GPU inference framework for EVASION that leverages sparse computation and asynchronous quantization, significantly enhancing inference speed. Comprehensive evaluation results demonstrate that EVASION can achieve 4 bits quantization trivial perplexity and accuracy loss.

Transformer-based models have demonstrated remarkable performance in computer vision tasks. However, their increasing model size leads to substantial memory demands and higher latency, hindering practical deployment. This paper presents an adaptive dynamic layer-skipping framework based on Markov Decision Process, which determines optimal computational paths based on the current state of input samples. We introduce a Temporal Importance Difference Reward mechanism to address the credit assignment problem in layer-skipping decisions, and develop a knowledge distillation strategy using learnable cognitive tokens to compensate for information loss. Experiments on various models demonstrate that our method significantly reduces computational costs while maintaining accuracy, offering a practical solution for deploying high-performance Transformer models in resource-constrained environments. The code is available at https://github.com/wjjkhl/ASTER

Large-scale deep neural networks (DNNs), such as large language models (LLMs), have revolutionized the artificial intelligence (AI) field and become increasingly popular. However, training or fine-tuning such models requires substantial computational power and resources, where the memory capacity of a single acceleration device like a GPU is one of the most important bottlenecks. Owing to the prohibitively large overhead (e.g., 10×) of GPUs' native memory allocator, DNN frameworks like PyTorch and TensorFlow adopt a caching allocator that maintains a memory pool with a splitting mechanism for fast memory (de)allocation. Unfortunately, the caching allocator's efficiency degrades quickly for popular memory reduction techniques such as re-computation, offloading, distributed training, and low-rank adaptation. The primary reason is that those memory reduction techniques introduce frequent and irregular memory (de)allocation requests, leading to severe fragmentation problems for the splitting-based caching allocator. To mitigate this fragmentation problem, we propose a novel memory allocation framework based on low-level GPU virtual memory management called GPU memory lake (GMLake). GMLake employs a novel virtual memory stitching (VMS) mechanism, which can fuse or combine non-contiguous memory blocks with a virtual memory address mapping. GMLake can reduce average of 9.2 GB (up to 25 GB) GPU memory usage and 15% (up to 33%) fragmentation among eight LLM models on GPU A100 with 80 GB memory. GMLake is completely transparent to the DNN models and memory reduction techniques and ensures the seamless execution of resource-intensive deep-learning tasks. We have open-sourced GMLake at https://github.com/intelligent-machine-learning/glake/tree/main/GMLake.