SCBench: A KV Cache-Centric Analysis of Long-Context Methods

Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.

MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention

The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparsethat can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of long-context LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.

Local Model Explanations and Uncertainty Without Model Access

We present a model-agnostic algorithm for generating post-hoc explanations and uncertainty intervals for a machine learning model when only a sample of inputs and outputs from the model is available, rather than direct access to the model itself. This situation may arise when model evaluations are expensive; when privacy, security and bandwidth constraints are imposed; or when there is a need for real-time, on-device explanations. Our algorithm constructs explanations using local polynomial regression and quantifies the uncertainty of the explanations using a bootstrapping approach. Through a simulation study, we show that the uncertainty intervals generated by our algorithm exhibit a favorable trade-off between interval width and coverage probability compared to the naive confidence intervals from classical regression analysis. We further demonstrate the capabilities of our method by applying it to black-box models trained on two real datasets.

Global Multiclass Classification from Heterogeneous Local Models

Multiclass classification problems are most often solved by either training a single centralized classifier on all $K$ classes, or by reducing the problem to multiple binary classification tasks. This paper explores the uncharted region between these two extremes: How can we solve the $K$-class classification problem by combining the predictions of smaller classifiers, each trained on an arbitrary number of classes $R \in \{2, 3, \ldots, K\}$? We present a mathematical framework for answering this question, and derive bounds on the number of classifiers (in terms of $K$ and $R$) needed to accurately predict the true class of an unlabeled sample under both adversarial and stochastic assumptions. By exploiting a connection to the classical set cover problem in combinatorics, we produce an efficient, near-optimal scheme (with respect to the number of classifiers) for designing such configurations of classifiers, which recovers the well-known one-vs.-one strategy as a special case when $R=2$. Experiments with the MNIST and CIFAR-10 datasets show that our scheme is capable of matching the performance of centralized classifiers in practice. The results suggest that our approach offers a promising direction for solving the problem of data heterogeneity which plagues current federated learning methods.