Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models (ARs). However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KVcache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKVCache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference.

Multi-agent large language model (LLM) systems are increasingly adopted for complex language processing tasks that require communication and coordination among agents. However, these systems often suffer substantial overhead from repeated reprocessing of overlapping contexts across agents. In typical pipelines, once an agent receives a message from its predecessor, the full context-including prior turns-must be reprocessed from scratch, leading to inefficient processing. While key-value (KV) caching is an effective solution for avoiding redundant computation in single-agent settings where prefixes remain unchanged, it cannot be directly reused in multi-agent scenarios due to diverging prefixes introduced by agent-specific context extensions. We identify that the core challenge lies in the offset variance of KV-caches across agents.

This paper presents MoE-Gen, a high-throughput MoE inference system optimized for single-GPU execution. Existing inference systems rely on model-based or continuous batching strategies, originally designed for interactive inference, which result in excessively small batches for MoE's key modules-attention and expert modules-leading to poor throughput. To address this, we introduce module-based batching, which accumulates tokens in host memory and dynamically launches large batches on GPUs to maximize utilization. Additionally, we optimize the choice of batch sizes for each module in an MoE to fully overlap GPU computation and communication, maximizing throughput. Evaluation demonstrates that MoE-Gen achieves 8-31x higher throughput compared to state-of-the-art systems employing model-based batching (FlexGen, MoE-Lightning, DeepSpeed), and offers even greater throughput improvements over continuous batching systems (e.g., vLLM and Ollama) on popular MoE models (DeepSeek and Mixtral) across offline inference tasks. MoE-Gen's source code is publicly available at https://github.com/EfficientMoE/MoE-Gen

Slim attention shrinks the context memory size by 2x for transformer models with MHA (multi-head attention), which can speed up inference by up to 2x for large context windows. Slim attention is an exact, mathematically identical implementation of the standard attention mechanism and therefore does not compromise model accuracy. In other words, slim attention losslessly compresses the context memory by a factor of 2. For encoder-decoder transformers, the context memory size can be reduced even further: For the Whisper models for example, slim attention reduces the context memory by 8x, which can speed up token generation by 5x for batch size 64 for example. And for rare cases where the MHA projection dimension is larger than the embedding dimension, the memory can be reduced by a factor of 32 for the T5-11B model for example. See https://github.com/OpenMachine-ai/transformer-tricks for code and more transformer tricks, and https://www.youtube.com/watch?v=uVtk3B6YO4Y for a video about this paper.

Large language models (LLMs) are widely used across various applications, but their substantial computational requirements pose significant challenges, particularly in terms of HBM bandwidth bottlenecks and inter-device communication overhead. In this paper, we present PRESERVE, a novel prefetching framework designed to optimize LLM inference by overlapping memory reads for model weights and KV-cache with collective communication operations. Through extensive experiments conducted on commercial AI accelerators, we demonstrate up to 1.6x end-to-end speedup on state-of-the-art, open-source LLMs. Additionally, we perform a design space exploration that identifies the optimal hardware configuration for the proposed method, showing a further 1.25x improvement in performance per cost by selecting the optimal L2 cache size. Our results show that PRESERVE has the potential to mitigate the memory bottlenecks and communication overheads, offering a solution to improve the performance and scalability of the LLM inference systems.

Transformers, driven by attention mechanisms, form the foundation of large language models (LLMs). As these models scale up, efficient GPU attention kernels become essential for high-throughput and low-latency inference. Diverse LLM applications demand flexible and high-performance attention solutions. We present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneity using block-sparse format and composable formats to optimize memory access and reduce redundancy. It also offers a customizable attention template, enabling adaptation to various settings through Just-In-Time (JIT) compilation. Additionally, FlashInfer's load-balanced scheduling algorithm adjusts to dynamism of user requests while maintaining compatibility with CUDAGraph which requires static configuration. FlashInfer have been integrated into leading LLM serving frameworks like SGLang, vLLM and MLC-Engine. Comprehensive kernel-level and end-to-end evaluations demonstrate FlashInfer's ability to significantly boost kernel performance across diverse inference scenarios: compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.

As the use of large language models (LLMs) expands rapidly, so does the range of knowledge needed to supplement various LLM queries. Thus, enabling flexible and efficient injection of new knowledge in LLM inference is critical. Three high-level options exist: (i) embedding the knowledge in LLM's weights (i.e., fine-tuning), (ii) including the knowledge as a part of LLM's text input (i.e., in-context learning), or (iii) injecting the KV caches of the new knowledge to LLM during prefill. This paper argues that, although fine-tuning and in-context learning are popular, using KV caches as the medium of knowledge could simultaneously enable more modular management of knowledge injection and more efficient LLM serving with low cost and fast response. To realize these benefits, we envision a Knowledge Delivery Network (KDN), a new system component in LLM services that dynamically optimizes the storage, transfer, and composition of KV cache across LLM engines and other compute and storage resources. We believe that, just like content delivery networks (CDNs), such as Akamai, enabled the success of the Internet ecosystem through their efficient data delivery, KDNs will be critical to the success of LLM applications through their efficient knowledge delivery. We have open-sourced a KDN prototype at https://github.com/LMCache/LMCache.

Model quantization has become a crucial technique to address the issues of large memory consumption and long inference times associated with LLMs. Mixed-precision quantization, which distinguishes between important and unimportant parameters, stands out among numerous quantization schemes as it achieves a balance between precision and compression rate. However, existing approaches can only identify important parameters through qualitative analysis and manual experiments without quantitatively analyzing how their importance is determined. We propose a new criterion, so-called 'precision alignment', to build a quantitative framework to holistically evaluate the importance of parameters in mixed-precision quantization. Our observations on floating point addition under various real-world scenarios suggest that two addends should have identical precision, otherwise the information in the higher-precision number will be wasted. Such an observation offers an essential principle to determine the precision of each parameter in matrix multiplication operation. As the first step towards applying the above discovery to large model inference, we develop a dynamic KV-Cache quantization technique to effectively reduce memory access latency. Different from existing quantization approaches that focus on memory saving, this work directly aims to accelerate LLM inference through quantifying floating numbers. The proposed technique attains a 25% saving of memory access and delivers up to 1.3x speedup in the computation of attention in the decoding phase of LLM, with almost no loss of precision.

The KV-Cache technique has become the standard for the inference of large language models (LLMs). It caches states of self-attention to avoid recomputation. Yet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system, especially when confronted with ultra-large models and long-context queries. A natural remedy is to discard the KV-Cache for less important tokens, with StreamingLLM as an example, but the used static eviction strategies cannot flexibly adapt to varying contexts. Remedies like H2O leverage accumulative attention scores to perform dynamic eviction but suffer from the attention bias issue in capturing contextual information. This paper bridges this gap by devising a parameterized KV-Cache eviction mechanism, dubbed as Attention-Gate, which accepts the whole context as input and yields eviction flags for each token to realize in-context eviction. The subsequent self-attention module proceeds according to the flags and only the KV states for the remaining tokens need to be cached. The Attention-Gates can vary among different heads and layers and be trivially plugged into pre-trained LLMs, tuned by cost-effective continual pre-training or supervised fine-tuning objectives to acquire what to discard. The computational and memory overhead introduced by Attention-Gates is minimal. Our method is validated across multiple tasks, demonstrating both efficiency and adaptability. After a highly efficient continual pre-training, it achieves higher average accuracy and evicts more tokens compared to traditional training-free methods. In supervised fine-tuning, it not only evicts many tokens but also outperforms LoRA-finetuned LLMs on some datasets, such as RTE, where it improves accuracy by 13.9% while evicting 62.8% of tokens, showing that effective eviction of redundant tokens can even enhance performance.

KV-Cache compression methods generally sample a KV-Cache of effectual tokens or quantize it into lower bits. However, these methods cannot exploit the redundancy of the hidden dimension of KV tensors. This paper investigates a unique hidden dimension approach called Palu, a novel KV-Cache compression framework that utilizes low-rank projection. Palu decomposes the linear layers into low-rank matrices, caches the smaller intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) a low-rank-aware quantization algorithm, and (4) matrix fusion with optimized GPU kernels. Our extensive experiments with popular LLMs show that Palu can compress KV-Cache by more than 91.25% while maintaining a significantly better accuracy (up to 1.19 lower perplexity) than state-of-the-art KV-Cache quantization methods at a similar or even higher memory usage. When compressing KV-Cache for 50%, Palu delivers up to 1.61x end-to-end speedup for the attention module. Our code is publicly available at https://github.com/shadowpa0327/Palu.