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.

NL2KQL: From Natural Language to Kusto Query

Data is growing rapidly in volume and complexity. Proficiency in database query languages is pivotal for crafting effective queries. As coding assistants become more prevalent, there is significant opportunity to enhance database query languages. The Kusto Query Language (KQL) is a widely used query language for large semi-structured data such as logs, telemetries, and time-series for big data analytics platforms. This paper introduces NL2KQL an innovative framework that uses large language models (LLMs) to convert natural language queries (NLQs) to KQL queries. The proposed NL2KQL framework includes several key components: Schema Refiner which narrows down the schema to its most pertinent elements; the Few-shot Selector which dynamically selects relevant examples from a few-shot dataset; and the Query Refiner which repairs syntactic and semantic errors in KQL queries. Additionally, this study outlines a method for generating large datasets of synthetic NLQ-KQL pairs which are valid within a specific database contexts. To validate NL2KQL's performance, we utilize an array of online (based on query execution) and offline (based on query parsing) metrics. Through ablation studies, the significance of each framework component is examined, and the datasets used for benchmarking are made publicly available. This work is the first of its kind and is compared with available baselines to demonstrate its effectiveness.

TD-GEN: Graph Generation With Tree Decomposition

We propose TD-GEN, a graph generation framework based on tree decomposition, and introduce a reduced upper bound on the maximum number of decisions needed for graph generation. The framework includes a permutation invariant tree generation model which forms the backbone of graph generation. Tree nodes are supernodes, each representing a cluster of nodes in the graph. Graph nodes and edges are incrementally generated inside the clusters by traversing the tree supernodes, respecting the structure of the tree decomposition, and following node sharing decisions between the clusters. Finally, we discuss the shortcomings of standard evaluation criteria based on statistical properties of the generated graphs as performance measures. We propose to compare the performance of models based on likelihood. Empirical results on a variety of standard graph generation datasets demonstrate the superior performance of our method.

Variational Learning with Disentanglement-PyTorch

Unsupervised learning of disentangled representations is an open problem in machine learning. The Disentanglement-PyTorch library is developed to facilitate research, implementation, and testing of new variational algorithms. In this modular library, neural architectures, dimensionality of the latent space, and the training algorithms are fully decoupled, allowing for independent and consistent experiments across variational methods. The library handles the training scheduling, logging, and visualizations of reconstructions and latent space traversals. It also evaluates the encodings based on various disentanglement metrics. The library, so far, includes implementations of the following unsupervised algorithms VAE, Beta-VAE, Factor-VAE, DIP-I-VAE, DIP-II-VAE, Info-VAE, and Beta-TCVAE, as well as conditional approaches such as CVAE and IFCVAE. The library is compatible with the Disentanglement Challenge of NeurIPS 2019, hosted on AICrowd, and achieved the 3rd rank in both the first and second stages of the challenge.

A Study into Echocardiography View Conversion

Transthoracic echo is one of the most common means of cardiac studies in the clinical routines. During the echo exam, the sonographer captures a set of standard cross sections (echo views) of the heart. Each 2D echo view cuts through the 3D cardiac geometry via a unique plane. Consequently, different views share some limited information. In this work, we investigate the feasibility of generating a 2D echo view using another view based on adversarial generative models. The objective optimized to train the view-conversion model is based on the ideas introduced by LSGAN, PatchGAN and Conditional GAN (cGAN). The size and length of the left ventricle in the generated target echo view is compared against that of the target ground-truth to assess the validity of the echo view conversion. Results show that there is a correlation of 0.50 between the LV areas and 0.49 between the LV lengths of the generated target frames and the real target frames.

A Preliminary Study of Disentanglement With Insights on the Inadequacy of Metrics

Disentangled encoding is an important step towards a better representation learning. However, despite the numerous efforts, there still is no clear winner that captures the independent features of the data in an unsupervised fashion. In this work we empirically evaluate the performance of six unsupervised disentanglement approaches on the mpi3d toy dataset curated and released for the NeurIPS 2019 Disentanglement Challenge. The methods investigated in this work are Beta-VAE, Factor-VAE, DIP-I-VAE, DIP-II-VAE, Info-VAE, and Beta-TCVAE. The capacities of all models were progressively increased throughout the training and the hyper-parameters were kept intact across experiments. The methods were evaluated based on five disentanglement metrics, namely, DCI, Factor-VAE, IRS, MIG, and SAP-Score. Within the limitations of this study, the Beta-TCVAE approach was found to outperform its alternatives with respect to the normalized sum of metrics. However, a qualitative study of the encoded latents reveal that there is not a consistent correlation between the reported metrics and the disentanglement potential of the model.

GAN-enhanced Conditional Echocardiogram Generation

Echocardiography (echo) is a common means of evaluating cardiac conditions. Due to the label scarcity, semi-supervised paradigms in automated echo analysis are getting traction. One of the most sought-after problems in echo is the segmentation of cardiac structures (e.g. chambers). Accordingly, we propose an echocardiogram generation approach using generative adversarial networks with a conditional patch-based discriminator. In this work, we validate the feasibility of GAN-enhanced echo generation with different conditions (segmentation masks), namely, the left ventricle, ventricular myocardium, and atrium. Results show that the proposed adversarial algorithm can generate high-quality echo frames whose cardiac structures match the given segmentation masks. This method is expected to facilitate the training of other machine learning models in a semi-supervised fashion as suggested in similar researches.

Variational Shape Completion for Virtual Planning of Jaw Reconstructive Surgery

The premorbid geometry of the mandible is of significant relevance in jaw reconstructive surgeries and occasionally unknown to the surgical team. In this paper, an optimization framework is introduced to train deep models for completion (reconstruction) of the missing segments of the bone based on the remaining healthy structure. To leverage the contextual information of the surroundings of the dissected region, the voxel-weighted Dice loss is introduced. To address the non-deterministic nature of the shape completion problem, we leverage a weighted multi-target probabilistic solution which is an extension to the conditional variational autoencoder (CVAE). This approach considers multiple targets as acceptable reconstructions, each weighted according to their conformity with the original shape. We quantify the performance gain of the proposed method against similar algorithms, including CVAE, where we report statistically significant improvements in both deterministic and probabilistic paradigms. The probabilistic model is also evaluated on its ability to generate anatomically relevant variations for the missing bone. As a unique aspect of this work, the model is tested on real surgical cases where the clinical relevancy of its reconstructions and their compliance with surgeon's virtual plan are demonstrated as necessary steps towards clinical adoption.

Muscle Excitation Estimation in Biomechanical Simulation Using NAF Reinforcement Learning

Motor control is a set of time-varying muscle excitations which generate desired motions for a biomechanical system. Muscle excitations cannot be directly measured from live subjects. An alternative approach is to estimate muscle activations using inverse motion-driven simulation. In this article, we propose a deep reinforcement learning method to estimate the muscle excitations in simulated biomechanical systems. Here, we introduce a custom-made reward function which incentivizes faster point-to-point tracking of target motion. Moreover, we deploy two new techniques, namely, episode-based hard update and dual buffer experience replay, to avoid feedback training loops. The proposed method is tested in four simulated 2D and 3D environments with 6 to 24 axial muscles. The results show that the models were able to learn muscle excitations for given motions after nearly 100,000 simulated steps. Moreover, the root mean square error in point-to-point reaching of the target across experiments was less than 1% of the length of the domain of motion. Our reinforcement learning method is far from the conventional dynamic approaches as the muscle control is derived functionally by a set of distributed neurons. This can open paths for neural activity interpretation of this phenomenon.