Symbolic regression (SR) is an emerging branch of machine learning focused on discovering simple and interpretable mathematical expressions from data. Although a wide-variety of SR methods have been developed, they often face challenges such as high computational cost, poor scalability with respect to the number of input dimensions, fragility to noise, and an inability to balance accuracy and complexity. This work introduces SyMANTIC, a novel SR algorithm that addresses these challenges. SyMANTIC efficiently identifies (potentially several) low-dimensional descriptors from a large set of candidates (from $\sim 10^5$ to $\sim 10^{10}$ or more) through a unique combination of mutual information-based feature selection, adaptive feature expansion, and recursively applied $\ell_0$-based sparse regression. In addition, it employs an information-theoretic measure to produce an approximate set of Pareto-optimal equations, each offering the best-found accuracy for a given complexity. Furthermore, our open-source implementation of SyMANTIC, built on the PyTorch ecosystem, facilitates easy installation and GPU acceleration. We demonstrate the effectiveness of SyMANTIC across a range of problems, including synthetic examples, scientific benchmarks, real-world material property predictions, and chaotic dynamical system identification from small datasets. Extensive comparisons show that SyMANTIC uncovers similar or more accurate models at a fraction of the cost of existing SR methods.
Genetic programming is an optimization algorithm inspired by evolution which automatically evolves the structure of interpretable computer programs. The fitness evaluation in genetic programming suffers from high computational requirements, limiting the performance on difficult problems. Consequently, there is no efficient genetic programming framework that is usable for a wide range of tasks. To this end, we developed Kozax, a genetic programming framework that evolves symbolic expressions for arbitrary problems. We implemented Kozax using JAX, a framework for high-performance and scalable machine learning, which allows the fitness evaluation to scale efficiently to large populations or datasets on GPU. Furthermore, Kozax offers constant optimization, custom operator definition and simultaneous evolution of multiple trees. We demonstrate successful applications of Kozax to discover equations of natural laws, recover equations of hidden dynamic variables, evolve a control policy and optimize an objective function. Overall, Kozax provides a general, fast, and scalable library to optimize white-box solutions in the realm of scientific computing.
Equipping large language models (LLMs) with latent-space memory has attracted increasing attention as they can extend the context window of existing language models. However, retaining information from the distant past remains a challenge. For example, MemoryLLM (Wang et al., 2024a), as a representative work with latent-space memory, compresses past information into hidden states across all layers, forming a memory pool of 1B parameters. While effective for sequence lengths up to 16k tokens, it struggles to retain knowledge beyond 20k tokens. In this work, we address this limitation by introducing M+, a memory-augmented model based on MemoryLLM that significantly enhances long-term information retention. M+ integrates a long-term memory mechanism with a co-trained retriever, dynamically retrieving relevant information during text generation. We evaluate M+ on diverse benchmarks, including long-context understanding and knowledge retention tasks. Experimental results show that M+ significantly outperforms MemoryLLM and recent strong baselines, extending knowledge retention from under 20k to over 160k tokens with similar GPU memory overhead.
Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem.
Establishing the long-context capability of large vision-language models is crucial for video understanding, high-resolution image understanding, multi-modal agents and reasoning. We introduce Long-VITA, a simple yet effective large multi-modal model for long-context visual-language understanding tasks. It is adept at concurrently processing and analyzing modalities of image, video, and text over 4K frames or 1M tokens while delivering advanced performances on short-context multi-modal tasks. We propose an effective multi-modal training schema that starts with large language models and proceeds through vision-language alignment, general knowledge learning, and two sequential stages of long-sequence fine-tuning. We further implement context-parallelism distributed inference and logits-masked language modeling head to scale Long-VITA to infinitely long inputs of images and texts during model inference. Regarding training data, Long-VITA is built on a mix of $17$M samples from public datasets only and demonstrates the state-of-the-art performance on various multi-modal benchmarks, compared against recent cutting-edge models with internal data. Long-VITA is fully reproducible and supports both NPU and GPU platforms for training and testing. We hope Long-VITA can serve as a competitive baseline and offer valuable insights for the open-source community in advancing long-context multi-modal understanding.
Knowledge of the primordial matter density field from which the large-scale structure of the Universe emerged over cosmic time is of fundamental importance for cosmology. However, reconstructing these cosmological initial conditions from late-time observations is a notoriously difficult task, which requires advanced cosmological simulators and sophisticated statistical methods to explore a multi-million-dimensional parameter space. We show how simulation-based inference (SBI) can be used to tackle this problem and to obtain data-constrained realisations of the primordial dark matter density field in a simulation-efficient way with general non-differentiable simulators. Our method is applicable to full high-resolution dark matter $N$-body simulations and is based on modelling the posterior distribution of the constrained initial conditions to be Gaussian with a diagonal covariance matrix in Fourier space. As a result, we can generate thousands of posterior samples within seconds on a single GPU, orders of magnitude faster than existing methods, paving the way for sequential SBI for cosmological fields. Furthermore, we perform an analytical fit of the estimated dependence of the covariance on the wavenumber, effectively transforming any point-estimator of initial conditions into a fast sampler. We test the validity of our obtained samples by comparing them to the true values with summary statistics and performing a Bayesian consistency test.
Self-play has powered breakthroughs in two-player and multi-player games. Here we show that self-play is a surprisingly effective strategy in another domain. We show that robust and naturalistic driving emerges entirely from self-play in simulation at unprecedented scale -- 1.6~billion~km of driving. This is enabled by Gigaflow, a batched simulator that can synthesize and train on 42 years of subjective driving experience per hour on a single 8-GPU node. The resulting policy achieves state-of-the-art performance on three independent autonomous driving benchmarks. The policy outperforms the prior state of the art when tested on recorded real-world scenarios, amidst human drivers, without ever seeing human data during training. The policy is realistic when assessed against human references and achieves unprecedented robustness, averaging 17.5 years of continuous driving between incidents in simulation.
Scaling Large Language Model (LLM) training relies on multi-dimensional parallelism, where High-Bandwidth Domains (HBDs) are critical for communication-intensive parallelism like Tensor Parallelism (TP) and Expert Parallelism (EP). However, existing HBD architectures face fundamental limitations in scalability, cost, and fault resiliency: switch-centric HBDs (e.g., NVL-72) incur prohibitive scaling costs, while GPU-centric HBDs (e.g., TPUv3/Dojo) suffer from severe fault propagation. Switch-GPU hybrid HBDs such as TPUv4 takes a middle-ground approach by leveraging Optical Circuit Switches, but the fault explosion radius remains large at the cube level (e.g., 64 TPUs). We propose InfiniteHBD, a novel transceiver-centric HBD architecture that unifies connectivity and dynamic switching at the transceiver level using Optical Circuit Switching (OCS). By embedding OCS within each transceiver, InfiniteHBD achieves reconfigurable point-to-multipoint connectivity, allowing the topology to adapt into variable-size rings. This design provides: i) datacenter-wide scalability without cost explosion; ii) fault resilience by isolating failures to a single node, and iii) full bandwidth utilization for fault-free GPUs. Key innovations include a Silicon Photonic (SiPh) based low-cost OCS transceiver (OCSTrx), a reconfigurable k-hop ring topology co-designed with intra-/inter-node communication, and an HBD-DCN orchestration algorithm maximizing GPU utilization while minimizing cross-ToR datacenter network traffic. The evaluation demonstrates that InfiniteHBD achieves 31% of the cost of NVL-72, near-zero GPU waste ratio (over one order of magnitude lower than NVL-72 and TPUv4), near-zero cross-ToR traffic when node fault ratios under 7%, and improves Model FLOPs Utilization by 3.37x compared to NVIDIA DGX (8 GPUs per Node).
AI methods, such as generative models and reinforcement learning, have recently been applied to combinatorial optimization (CO) problems, especially NP-hard ones. This paper compares such GPU-based methods with classical CPU-based methods on Maximum Independent Set (MIS). Experiments on standard graph families show that AI-based algorithms fail to outperform and, in many cases, to match the solution quality of the state-of-art classical solver KaMIS running on a single CPU. Some GPU-based methods even perform similarly to the simplest heuristic, degree-based greedy. Even with post-processing techniques like local search, AI-based methods still perform worse than CPU-based solvers. We develop a new mode of analysis to reveal that non-backtracking AI methods, e.g. LTFT (which is based on GFlowNets), end up reasoning similarly to the simplest degree-based greedy approach, and thus worse than KaMIS. We also find that CPU-based algorithms, notably KaMIS, have strong performance on sparse random graphs, which appears to refute a well-known conjectured upper bound for efficient algorithms from Coja-Oghlan & Efthymiou (2015).
This technical report demonstrates that large language models (LLMs) can maintain their learning capacity while reducing their non-embedding parameters by up to 77%. We achieve this by adapting a parameter reduction technique originally developed for computer vision, replacing dense layers with an optimized subnetwork that contains grouped pointwise convolutions. Using Microsoft's phi-3-mini-4k-instruct as our baseline, we show that our optimized model (kphi-3) achieves comparable validation loss while using only 15-23% of the original non-embedding parameters. All experiments were conducted on a single NVIDIA L2 GPU within a 3-day timeframe, supporting the democratization of AI research. Our findings suggest that current LLM architectures may be substantially overparameterized, opening possibilities for more efficient model training and deployment.
One approach to reducing the massive costs of large language models (LLMs) is the use of quantized or sparse representations for training or deployment. While post-training compression methods are very popular, the question of obtaining even more accurate compressed models by directly training over such representations, i.e., Quantization-Aware Training (QAT), is still open: for example, a recent study (arXiv:2411.04330v2) put the "optimal" bit-width at which models can be trained using QAT, while staying accuracy-competitive with standard FP16/BF16 precision, at 8-bits weights and activations. We advance this state-of-the-art via a new method called QuEST, which is Pareto-competitive with FP16, i.e., it provides better accuracy at lower model size, while training models with weights and activations in 4-bits or less. Moreover, QuEST allows stable training with 1-bit weights and activations. QuEST achieves this by improving two key aspects of QAT methods: (1) accurate and fast quantization of the (continuous) distributions of weights and activations via Hadamard normalization and MSE-optimal fitting; (2) a new trust gradient estimator based on the idea of explicitly minimizing the error between the noisy gradient computed over quantized states and the "true" (but unknown) full-precision gradient. Experiments on Llama-type architectures show that QuEST induces stable scaling laws across the entire range of hardware-supported precisions, and can be extended to sparse representations. We provide GPU kernel support showing that models produced by QuEST can be executed efficiently. Our code is available at https://github.com/IST-DASLab/QuEST.
Emerging AI accelerators increasingly adopt wafer-scale manufacturing technologies, integrating hundreds of thousands of AI cores in a mesh architecture with large distributed on-chip memory (tens of GB in total) and ultra-high on-chip memory bandwidth (tens of PB/s). However, current LLM inference systems, optimized for shared memory architectures like GPUs, fail to exploit these accelerators fully. We introduce WaferLLM, the first wafer-scale LLM inference system. WaferLLM is guided by a novel PLMR model (pronounced as "Plummer") that captures the unique hardware characteristics of wafer-scale architectures. Leveraging this model, WaferLLM pioneers wafer-scale LLM parallelism, optimizing the utilization of hundreds of thousands of on-chip cores. It also introduces MeshGEMM and MeshGEMV, the first GEMM and GEMV implementations designed to scale effectively on wafer-scale accelerators. Evaluations show that WaferLLM achieves up to 200$\times$ higher accelerator utilization than state-of-the-art methods. Leveraging a wafer-scale accelerator (Cerebras WSE2), WaferLLM delivers GEMV operations 606$\times$ faster and 16$\times$ more energy-efficient than on an NVIDIA A100 GPU. For full LLM inference, WaferLLM achieves 10-20$\times$ speedups over A100 GPU clusters running SGLang and vLLM. These advantages are expected to grow as wafer-scale AI models, software, and hardware continue to mature. WaferLLM is open-sourced at https://github.com/MeshInfra/WaferLLM.
In late January 2025, DeepSeek released their new reasoning model (DeepSeek R1); which was developed at a fraction of the cost yet remains competitive with OpenAI's models, despite the US's GPU export ban. This report discusses the model, and what its release means for the field of Generative AI more widely. We briefly discuss other models released from China in recent weeks, their similarities; innovative use of Mixture of Experts (MoE), Reinforcement Learning (RL) and clever engineering appear to be key factors in the capabilities of these models. This think piece has been written to a tight time-scale, providing broad coverage of the topic, and serves as introductory material for those looking to understand the model's technical advancements, as well as it's place in the ecosystem. Several further areas of research are identified.
Speculative decoding is an effective and lossless method for Large Language Model (LLM) inference acceleration. It employs a smaller model to generate a draft token sequence, which is then verified by the original base model. In multi-GPU systems, inference latency can be further reduced through tensor parallelism (TP), while the optimal TP size of the draft model is typically smaller than that of the base model, leading to GPU idling during the drafting stage. To solve this problem, we propose EasySpec, a layer-parallel speculation strategy that optimizes the efficiency of multi-GPU utilization.EasySpec breaks the sequential execution order of layers in the drafting model, enabling multi-layer parallelization across devices, albeit with some induced approximation errors. After each drafting-and-verification iteration, the draft model's key-value (KV) cache is calibrated in a single forward pass, preventing long-term error accumulation at minimal additional latency. We evaluated EasySpec on several mainstream open-source LLMs, using smaller versions of models from the same series as drafters. The results demonstrate that EasySpec can achieve a peak speedup of 4.17x compared to vanilla decoding, while preserving the original distribution of the base LLMs. Specifically, the drafting stage can be accelerated by up to 1.62x with a maximum accuracy drop of only 7%, requiring no training or fine-tuning on the draft models.
Although deep models have been widely explored in solving partial differential equations (PDEs), previous works are primarily limited to data only with up to tens of thousands of mesh points, far from the million-point scale required by industrial simulations that involve complex geometries. In the spirit of advancing neural PDE solvers to real industrial applications, we present Transolver++, a highly parallel and efficient neural solver that can accurately solve PDEs on million-scale geometries. Building upon previous advancements in solving PDEs by learning physical states via Transolver, Transolver++ is further equipped with an extremely optimized parallelism framework and a local adaptive mechanism to efficiently capture eidetic physical states from massive mesh points, successfully tackling the thorny challenges in computation and physics learning when scaling up input mesh size. Transolver++ increases the single-GPU input capacity to million-scale points for the first time and is capable of continuously scaling input size in linear complexity by increasing GPUs. Experimentally, Transolver++ yields 13% relative promotion across six standard PDE benchmarks and achieves over 20% performance gain in million-scale high-fidelity industrial simulations, whose sizes are 100$\times$ larger than previous benchmarks, covering car and 3D aircraft designs.
We present PixelBrax, a set of continuous control tasks with pixel observations. We combine the Brax physics engine with a pure JAX renderer, allowing reinforcement learning (RL) experiments to run end-to-end on the GPU. PixelBrax can render observations over thousands of parallel environments and can run two orders of magnitude faster than existing benchmarks that rely on CPU-based rendering. Additionally, PixelBrax supports fully reproducible experiments through its explicit handling of any stochasticity within the environments and supports color and video distractors for benchmarking generalization. We open-source PixelBrax alongside JAX implementations of several RL algorithms at github.com/trevormcinroe/pixelbrax.
Cross-attention is commonly adopted in multimodal large language models (MLLMs) for integrating visual information into the language backbone. However, in applications with large visual inputs, such as video understanding, processing a large number of visual tokens in cross-attention layers leads to high memory demands and often necessitates distributed computation across multiple GPUs. Existing distributed attention mechanisms face significant communication overheads, making cross-attention layers a critical bottleneck for efficient training and inference of MLLMs. To address this, we propose LV-XAttn, a distributed, exact cross-attention mechanism with minimal communication overhead. We observe that in applications involving large visual inputs, the size of the query block is typically much smaller than that of the key-value blocks. Thus, in LV-XAttn we keep the large key-value blocks locally on each GPU and exchange smaller query blocks across GPUs. We also introduce an efficient activation recomputation technique to support longer visual context. We theoretically analyze the communication benefits of LV-XAttn and show that it can achieve speedups for a wide range of models. Our evaluations with Llama 3-V, mPLUG-Owl3 and OpenFlamingo models find that LV-XAttn achieves up to 10.62$\times$ end-to-end speedup compared to existing approaches.
In high-energy physics, the increasing luminosity and detector granularity at the Large Hadron Collider are driving the need for more efficient data processing solutions. Machine Learning has emerged as a promising tool for reconstructing charged particle tracks, due to its potentially linear computational scaling with detector hits. The recent implementation of a graph neural network-based track reconstruction pipeline in the first level trigger of the LHCb experiment on GPUs serves as a platform for comparative studies between computational architectures in the context of high-energy physics. This paper presents a novel comparison of the throughput of ML model inference between FPGAs and GPUs, focusing on the first step of the track reconstruction pipeline$\unicode{x2013}$an implementation of a multilayer perceptron. Using HLS4ML for FPGA deployment, we benchmark its performance against the GPU implementation and demonstrate the potential of FPGAs for high-throughput, low-latency inference without the need for an expertise in FPGA development and while consuming significantly less power.
The machine learning (ML) training over disparate data sources traditionally involves materialization, which can impose substantial time and space overhead due to data movement and replication. Factorized learning, which leverages direct computation on disparate sources through linear algebra (LA) rewriting, has emerged as a viable alternative to improve computational efficiency. However, the adaptation of factorized learning to leverage the full capabilities of modern LA-friendly hardware like GPUs has been limited, often requiring manual intervention for algorithm compatibility. This paper introduces Ilargi, a novel factorized learning framework that utilizes matrix-represented data integration (DI) metadata to facilitate automatic factorization across CPU and GPU environments without the need for costly relational joins. Ilargi incorporates an ML-based cost estimator to intelligently selects between factorization and materialization based on data properties, algorithm complexity, hardware environments, and their interactions. This strategy ensures up to 8.9x speedups on GPUs and achieves over 20% acceleration in batch ML training workloads, thereby enhancing the practicability of ML training across diverse data integration scenarios and hardware platforms. To our knowledge, this work is the very first effort in GPU-compatible factorized learning.
Transformers have demonstrated great success in numerous domains including natural language processing and bioinformatics. This success stems from the use of the attention mechanism by these models in order to represent and propagate pairwise interactions between individual tokens of sequential data. However, the primary limitation of this operation is its quadratic memory and time complexity in relation to the input's context length - the length of a sequence over which the interactions need to be captured. This significantly limits the length of sequences that can be inferred upon by these models. Extensive research has been conducted to reduce the number of pairwise interactions to sub-quadratic in relation to the context length by introducing sparsity into the attention mechanism through the development of sparse attention masks. However, efficient implementations that achieve "true sparsity" are lacking. In this work, we address this issue by proposing a graph computing view of attention where tokens are perceived as nodes of the graph and the attention mask determines the edges of the graph. Using this view, we develop graph processing algorithms to implement the attention mechanism. Both theoretically and empirically, we demonstrate that our algorithms only perform the needed computations, i.e., they are work optimal. We also perform extensive experimentation using popular attention masks to explore the impact of sparsity on execution time and achievable context length. Our experiments demonstrate significant speedups in execution times compared to state-of-the-art attention implementations such as FlashAttention for large sequence lengths. We also demonstrate that our algorithms are able to achieve extremely long sequence lengths of as high as 160 million on a single NVIDIA A100 GPU (SXM4 80GB).