Large Language Models (LLMs) and other large foundation models have achieved impressive results, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model selector to choose from an ensemble of models for query processing.Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model selector, we achieve optimal rates in both offline and online settings. Empirically, simulations show that our caching and model selection algorithm greatly improves over the baselines, with up to $50\times$ improvement over the baseline when the ratio between the maximum cost and minimum cost is $100$. Experiments on real datasets show a $4.3\times$ improvement in FLOPs over the baseline when the ratio for FLOPs is $10$, and a $1.8\times$ improvement in latency when the ratio for average latency is $1.85$.

Low-power microcontroller (MCU) hardware is currently evolving from single-core architectures to predominantly multi-core architectures. In parallel, new embedded software building blocks are more and more written in Rust, while C/C++ dominance fades in this domain. On the other hand, small artificial neural networks (ANN) of various kinds are increasingly deployed in edge AI use cases, thus deployed and executed directly on low-power MCUs. In this context, both incremental improvements and novel innovative services will have to be continuously retrofitted using ANNs execution in software embedded on sensing/actuating systems already deployed in the field. However, there was so far no Rust embedded software platform automating parallelization for inference computation on multi-core MCUs executing arbitrary TinyML models. This paper thus fills this gap by introducing Ariel-ML, a novel toolkit we designed combining a generic TinyML pipeline and an embedded Rust software platform which can take full advantage of multi-core capabilities of various 32bit microcontroller families (Arm Cortex-M, RISC-V, ESP-32). We published the full open source code of its implementation, which we used to benchmark its capabilities using a zoo of various TinyML models. We show that Ariel-ML outperforms prior art in terms of inference latency as expected, and we show that, compared to pre-existing toolkits using embedded C/C++, Ariel-ML achieves comparable memory footprints. Ariel-ML thus provides a useful basis for TinyML practitioners and resource-constrained embedded Rust developers.

In this paper, we show how to run pi0-level multi-view VLA at 30Hz frame rate and at most 480Hz trajectory frequency using a single consumer GPU. This enables dynamic and real-time tasks that were previously believed to be unattainable by large VLA models. To achieve it, we introduce a bag of strategies to eliminate the overheads in model inference. The real-world experiment shows that the pi0 policy with our strategy achieves a 100% success rate in grasping a falling pen task. Based on the results, we further propose a full streaming inference framework for real-time robot control of VLA. Code is available at https://github.com/Dexmal/realtime-vla.

We present FlagEvalMM, an open-source evaluation framework designed to comprehensively assess multimodal models across a diverse range of vision-language understanding and generation tasks, such as visual question answering, text-to-image/video generation, and image-text retrieval. We decouple model inference from evaluation through an independent evaluation service, thus enabling flexible resource allocation and seamless integration of new tasks and models. Moreover, FlagEvalMM utilizes advanced inference acceleration tools (e.g., vLLM, SGLang) and asynchronous data loading to significantly enhance evaluation efficiency. Extensive experiments show that FlagEvalMM offers accurate and efficient insights into model strengths and limitations, making it a valuable tool for advancing multimodal research. The framework is publicly accessible at https://github.com/flageval-baai/FlagEvalMM.

Large Language Models (LLMs) and other large foundation models have achieved impressive results, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model selector to choose from an ensemble of models for query processing.Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model selector, we achieve optimal rates in both offline and online settings. Empirically, simulations show that our caching and model selection algorithm greatly improves over the baselines, with up to 50\times improvement over the baseline when the ratio between the maximum cost and minimum cost is 100 .

Vision Transformer models exhibit immense power yet remain opaque to human understanding, posing challenges and risks for practical applications. While prior research has attempted to demystify these models through input attribution and neuron role analysis, there's been a notable gap in considering layer-level information and the holistic path of information flow across layers. In this paper, we investigate the significance of influential neuron paths within vision Transformers, which is a path of neurons from the model input to output that impacts the model inference most significantly. We first propose a joint influence measure to assess the contribution of a set of neurons to the model outcome. And we further provide a layer-progressive neuron locating approach that efficiently selects the most influential neuron at each layer trying to discover the crucial neuron path from input to output within the target model. Our experiments demonstrate the superiority of our method finding the most influential neuron path along which the information flows, over the existing baseline solutions. Additionally, the neuron paths have illustrated that vision Transformers exhibit some specific inner working mechanism for processing the visual information within the same image category. We further analyze the key effects of these neurons on the image classification task, showcasing that the found neuron paths have already preserved the model capability on downstream tasks, which may also shed some lights on real-world applications like model pruning. Transformer (V aswani et al., 2017) models in the vision domain, such as supervised Vision Transformers (Dosovitskiy et al., 2021) (ViT) or self-supervised pretrained models (He et al., 2022; Oquab et al., 2023), have showcased remarkable performance in various real-world tasks like image classification (Dosovitskiy et al., 2021) and image synthesis (Peebles & Xie, 2023). However, the inner workings of these vision Transformer models remain elusive, despite their impressive achievements. Understanding the internal mechanisms of vision models is crucial for both research and practical applications. When confronted with the model decision outputs, one may raise some questions that, how is the vision Transformer model processing the input information by layer, and which part of the model is significant to derive the final outcome? Unraveling the synergy within these models is essential for comprehending machine learning systems.

--With the rapid advent of generative models, efficiently deploying these models on specialized hardware has become critical. T ensor Processing Units (TPUs) are designed to accelerate AI workloads, but their high power consumption necessitates innovations for improving efficiency. Compute-in-memory (CIM) has emerged as a promising paradigm with superior area and energy efficiency. In this work, we present a TPU architecture that integrates digital CIM to replace conventional digital systolic arrays in matrix multiply units (MXUs). We first establish a CIM-based TPU architecture model and simulator to evaluate the benefits of CIM for diverse generative model inference. Building upon the observed design insights, we further explore various CIM-based TPU architectural design choices. Up to 44.2% and 33.8% performance improvement for large language model and diffusion transformer inference, and 27.3 reduction in MXU energy consumption can be achieved with different design choices, compared to the baseline TPUv4i architecture. Generative models, such as large language models (LLMs) and diffusion models (DMs), have exhibited exceptional performance in generating content across various modalities. For example, LLMs have dominated NLP tasks, powering applications like ChatGPT [1].

We present Entropy Adaptive Decoding (EAD), a novel approach for efficient language model inference that dynamically switches between different-sized models based on prediction uncertainty. By monitoring rolling entropy in model logit distributions, our method identifies text regions where a smaller model suffices and switches to a larger model only when prediction uncertainty exceeds a threshold. Unlike speculative decoding approaches that maintain perfect output fidelity through verification, EAD accepts controlled output divergence in exchange for computational efficiency. Our experiments on the MATH benchmark demonstrate remarkable efficiency gains across different model families. Using the LLaMA family, we maintain 96.7\% of the 11B model's performance (50.4\% vs 52.1\%) while using it for only 43\% of tokens, decreasing computational cost by 41.5\%. These gains become more pronounced with larger size differentials in the Qwen family, where we achieve 92.9\% of the 14B model's performance (74.3\% vs 80.0\%) while using it for just 25\% of tokens, decreasing computational cost by 67\%. The consistency of these results across model pairs suggests that language model computation can be significantly optimized by selectively deploying model capacity based on local generation complexity. Our findings indicate that current approaches to model inference may be unnecessarily conservative in their pursuit of perfect output fidelity, and that accepting minor performance trade-offs can enable dramatic reductions in computational costs.

Large scale deep learning model, such as modern language models and diffusion architectures, have revolutionized applications ranging from natural language processing to computer vision. However, their deployment in distributed or decentralized environments raises significant privacy concerns, as sensitive data may be exposed during inference. Traditional techniques like secure multi-party computation, homomorphic encryption, and differential privacy offer partial remedies but often incur substantial computational overhead, latency penalties, or limited compatibility with non-linear network operations. In this work, we introduce Equivariant Encryption (EE), a novel paradigm designed to enable secure, "blind" inference on encrypted data with near zero performance overhead. Unlike fully homomorphic approaches that encrypt the entire computational graph, EE selectively obfuscates critical internal representations within neural network layers while preserving the exact functionality of both linear and a prescribed set of non-linear operations. This targeted encryption ensures that raw inputs, intermediate activations, and outputs remain confidential, even when processed on untrusted infrastructure. We detail the theoretical foundations of EE, compare its performance and integration complexity against conventional privacy preserving techniques, and demonstrate its applicability across a range of architectures, from convolutional networks to large language models. Furthermore, our work provides a comprehensive threat analysis, outlining potential attack vectors and baseline strategies, and benchmarks EE against standard inference pipelines in decentralized settings. The results confirm that EE maintains high fidelity and throughput, effectively bridging the gap between robust data confidentiality and the stringent efficiency requirements of modern, large scale model inference.

Resource utilization is one of the key factors in fully exploiting the computing power of massively parallel devices, including GPUs. As a common method to improve utilization and reduce overhead, the benefit of the batching technique should never be underestimated [7, 8, 11]. In most cases, it is handy to batch regular workloads that share the same type and size, which also have similar amounts of computation and memory access. For example, in the CUDA programming model, this kind of regular workloads can be conveniently batched along an additional thread block or grid dimension [15]. However, irregular workloads do not naturally fit into this scheme. Irregular workloads may show one or more of the following characteristics that prevent regular batching[1]: variable amounts of computation, special memory access patterns, control flow divergence, etc. Moreover, heterogeneous workloads almost raise the difficulty of batching to an unreachable level. Here, by heterogeneous, we refer to workloads of different types of operations, e.g., some of the workloads are reduction, while others are element-wise operations. Irregular workloads are often managed in a task-parallel fashion instead of batching, where an individual workload is regarded as a task, and all tasks are dynamically scheduled [1, 19].