The global scarcity of GPUs necessitates more sophisticated strategies for Deep Learning jobs in shared cluster environments. Accurate estimation of how much GPU memory a job will require is fundamental to enabling advanced scheduling and GPU sharing, which helps prevent out-of-memory (OOM) errors and resource underutilization. However, existing estimation methods have limitations. Approaches relying on static analysis or historical data with machine learning often fail to accurately capture runtime dynamics. Furthermore, direct GPU analysis consumes scarce resources, and some techniques require intrusive code modifications. Thus, the key challenge lies in precisely estimating dynamic memory requirements, including memory allocator nuances, without consuming GPU resources and non-intrusive code changes. To address this challenge, we propose xMem, a novel framework that leverages CPU-only dynamic analysis to accurately estimate peak GPU memory requirements a priori. We conducted a thorough evaluation of xMem against state-of-the-art solutions using workloads from 25 different models, including architectures like Convolutional Neural Networks and Transformers. The analysis of 5209 runs, which includes ANOVA and Monte Carlo results, highlights xMem's benefits: it decreases the median relative error by 91% and significantly reduces the probability of estimation failure as safe OOM thresholds by 75%, meaning that the estimated value can often be used directly without causing OOM. Ultimately, these improvements lead to a 368% increase in memory conservation potential over current solutions.

However, the model size and corresponding computational complexity are constantly scaled up in pursuit of higher performance.

Agent-based simulation is an indispensable paradigm for studying complex systems. These systems can comprise billions of agents, requiring the computing resources of multiple servers to simulate. Unfortunately, the state-of-the-art platform, BioDynaMo, does not scale out across servers due to its shared-memory-based implementation. To overcome this key limitation, we introduce TeraAgent, a distributed agent-based simulation engine. A critical challenge in distributed execution is the exchange of agent information across servers, which we identify as a major performance bottleneck. We propose two solutions: 1) a tailored serialization mechanism that allows agents to be accessed and mutated directly from the receive buffer, and 2) leveraging the iterative nature of agent-based simulations to reduce data transfer with delta encoding. Built on our solutions, TeraAgent enables extreme-scale simulations with half a trillion agents (an 84x improvement), reduces time-to-result with additional compute nodes, improves interoperability with third-party tools, and provides users with more hardware flexibility.

Scaling test-time compute has emerged as a key ingredient for enabling large language models (LLMs) to solve difficult problems, but comes with high latency and inference cost. We introduce sleep-time compute, which allows models to "think" offline about contexts before queries are presented: by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time. To demonstrate the efficacy of our method, we create modified versions of two reasoning tasks - Stateful GSM-Symbolic and Stateful AIME. We find that sleep-time compute can reduce the amount of test-time compute needed to achieve the same accuracy by ~ 5x on Stateful GSM-Symbolic and Stateful AIME and that by scaling sleep-time compute we can further increase accuracy by up to 13% on Stateful GSM-Symbolic and 18% on Stateful AIME. Furthermore, we introduce Multi-Query GSM-Symbolic, which extends GSM-Symbolic by including multiple related queries per context. By amortizing sleep-time compute across related queries about the same context using Multi-Query GSM-Symbolic, we can decrease the average cost per query by 2.5x. We then conduct additional analysis to understand when sleep-time compute is most effective, finding the predictability of the user query to be well correlated with the efficacy of sleep-time compute. Finally, we conduct a case-study of applying sleep-time compute to a realistic agentic SWE task.

Autonomous Vehicles (AVs) have entered the commercialization stage, but their limited ability to interact and express intentions still poses challenges in interactions with Human-driven Vehicles (HVs). Recent advances in large language models (LLMs) enable bidirectional human-machine communication, but the conflict between slow inference speed and the need for real-time decision-making challenges practical deployment. To address these issues, this paper introduces a parallel Actor-Reasoner framework designed to enable explicit bidirectional AV-HV interactions across multiple scenarios. First, by facilitating interactions between the LLM-driven Reasoner and heterogeneous simulated HVs during training, an interaction memory database, referred to as the Actor, is established. Then, by introducing the memory partition module and the two-layer memory retrieval module, the Actor's ability to handle heterogeneous HVs is significantly enhanced. Ablation studies and comparisons with other decision-making methods demonstrate that the proposed Actor-Reasoner framework significantly improves safety and efficiency. Finally, with the combination of the external Human-Machine Interface (eHMI) information derived from Reasoner's reasoning and the feasible action solutions retrieved from the Actor, the effectiveness of the proposed Actor-Reasoner is confirmed in multi-scenario field interactions. Our code is available at https://github.com/FanGShiYuu/Actor-Reasoner.

The "state" of State Space Models (SSMs) represents their memory, which fades exponentially over an unbounded span. By contrast, Attention-based models have "eidetic" (i.e., verbatim, or photographic) memory over a finite span (context size). Hybrid architectures combine State Space layers with Attention, but still cannot recall the distant past and can access only the most recent tokens eidetically. Unlike current methods of combining SSM and Attention layers, we allow the state to be allocated based on relevancy rather than recency. In this way, for every new set of query tokens, our models can "eidetically" access tokens from beyond the Attention span of current Hybrid SSMs without requiring extra hardware resources. We describe a method to expand the memory span of the hybrid state by "reserving" a fraction of the Attention context for tokens retrieved from arbitrarily distant in the past, thus expanding the eidetic memory span of the overall state. We call this reserved fraction of tokens the "expansion span," and the mechanism to retrieve and aggregate it "Span-Expanded Attention" (SE-Attn). To adapt Hybrid models to using SE-Attn, we propose a novel fine-tuning method that extends LoRA to Hybrid models (HyLoRA) and allows efficient adaptation on long spans of tokens. We show that SE-Attn enables us to efficiently adapt pre-trained Hybrid models on sequences of tokens up to 8 times longer than the ones used for pre-training. We show that HyLoRA with SE-Attn is cheaper and more performant than alternatives like LongLoRA when applied to Hybrid models on natural language benchmarks with long-range dependencies, such as PG-19, RULER, and other common natural language downstream tasks.

In order to reduce the computational complexity of large language models, great efforts have been made to to improve the efficiency of transformer models such as linear attention and flash-attention. However, the model size and corresponding computational complexity are constantly scaled up in pursuit of higher performance. In this work, we present MemoryFormer, a novel transformer architecture which significantly reduces the computational complexity (FLOPs) from a new perspective. We eliminate nearly all the computations of the transformer model except for the necessary computation required by the multi-head attention operation. This is made possible by utilizing an alternative method for feature transformation to replace the linear projection of fully-connected layers. Specifically, we first construct a group of in-memory lookup tables that store a large amount of discrete vectors to replace the weight matrix used in linear projection. We then use a hash algorithm to retrieve a correlated subset of vectors dynamically based on the input embedding. The retrieved vectors combined together will form the output embedding, which provides an estimation of the result of matrix multiplication operation in a fully-connected layer. Compared to conducting matrix multiplication, retrieving data blocks from memory is a much cheaper operation which requires little computations. We train MemoryFormer from scratch and conduct extensive experiments on various benchmarks to demonstrate the effectiveness of the proposed model.

Convolutional Neural Networks (CNNs) are important for many machine learning tasks. They are built with different types of layers: convolutional layers that detect features, dropout layers that help to avoid over-reliance on any single neuron, and residual layers that allow the reuse of features. However, CNNs lack a dynamic feature retention mechanism similar to the human brain's memory, limiting their ability to use learned information in new contexts. To bridge this gap, we introduce the "Squeeze-and-Remember" (SR) block, a novel architectural unit that gives CNNs dynamic memory-like functionalities. The SR block selectively memorizes important features during training, and then adaptively re-applies these features during inference. This improves the network's ability to make contextually informed predictions. Empirical results on ImageNet and Cityscapes datasets demonstrate the SR block's efficacy: integration into ResNet50 improved top-1 validation accuracy on ImageNet by 0.52% over dropout2d alone, and its application in DeepLab v3 increased mean Intersection over Union in Cityscapes by 0.20%. These improvements are achieved with minimal computational overhead. This show the SR block's potential to enhance the capabilities of CNNs in image processing tasks.

In order to follow the ever-growing computational complexity and data intensity of state-of-the-art AI models, new computing paradigms are being proposed. These paradigms aim at achieving high energy efficiency, by mitigating the Von Neumann bottleneck that relates to the energy cost of moving data between the processing cores and the memory. Convolutional Neural Networks (CNNs) are particularly susceptible to this bottleneck, given the massive data they have to manage. Systolic Arrays (SAs) are promising architectures to mitigate the data transmission cost, thanks to high data utilization carried out by an array of Processing Elements (PEs). These PEs continuously exchange and process data locally based on specific dataflows (like weight stationary and row stationary), in turn reducing the number of memory accesses to the main memory. The hardware specialization of SAs can meet different workloads, ranging from matrix multiplications to multi-dimensional convolutions. In this paper, we propose TrIM: a novel dataflow for SAs based on a Triangular Input Movement and compatible with CNN computing. When compared to state-of-the-art SA dataflows, like weight stationary and row stationary, the high data utilization offered by TrIM guarantees ~10x less memory access. Furthermore, considering that PEs continuously overlap multiplications and accumulations, TrIM achieves high throughput (up to 81.8% higher than row stationary), other than requiring a limited number of registers (up to 15.6x fewer registers than row stationary).