RFX (Random Forests X), where X stands for compression or quantization, presents a production-ready implementation of Breiman and Cutler's Random Forest classification methodology in Python. RFX v1.0 provides complete classification: out-of-bag error estimation, overall and local importance measures, proximity matrices with QLORA compression, case-wise analysis, and interactive visualization (rfviz)--all with CPU and GPU acceleration. Regression, unsupervised learning, CLIQUE importance, and RF-GAP proximity are planned for v2.0. This work introduces four solutions addressing the proximity matrix memory bottleneck limiting Random Forest analysis to ~60,000 samples: (1) QLORA (Quantized Low-Rank Adaptation) compression for GPU proximity matrices, reducing memory from 80GB to 6.4MB for 100k samples (12,500x compression with INT8 quantization) while maintaining 99% geometric structure preservation, (2) CPU TriBlock proximity--combining upper-triangle storage with block-sparse thresholding--achieving 2.7x memory reduction with lossless quality, (3) SM-aware GPU batch sizing achieving 95% GPU utilization, and (4) GPU-accelerated 3D MDS visualization computing embeddings directly from low-rank factors using power iteration. Validation across four implementation modes (GPU/CPU x case-wise/non-case-wise) demonstrates correct implementation. GPU achieves 1.4x speedup over CPU for overall importance with 500+ trees. Proximity computation scales from 1,000 to 200,000+ samples (requiring GPU QLORA), with CPU TriBlock filling the gap for medium-scale datasets (10K-50K samples). RFX v1.0 eliminates the proximity memory bottleneck, enabling proximity-based Random Forest analysis on datasets orders of magnitude larger than previously feasible. Open-source production-ready classification following Breiman and Cutler's original methodology.

Sparse Matrix-Vector Multiplication (SpMV) is a fundamental operation in the inference of sparse Large Language Models (LLMs). Because existing SpMV methods perform poorly under the low and unstructured sparsity (30-90%) commonly observed in pruned LLMs, unstructured pruning provided only limited memory reduction and speedup. We propose MACKO-SpMV, a GPU-optimized format and kernel co-designed to reduce storage overhead while preserving compatibility with the GPU's execution model. This enables efficient SpMV for unstructured sparsity without specialized hardware units (e.g., tensor cores) or format-specific precomputation. Empirical results show that at sparsity 50%, MACKO is the first approach with significant 1.5x memory reduction and 1.2-1.5x speedup over dense representation. Speedups over other SpMV baselines: 2.8-13.0x over cuSPARSE, 1.9-2.6x over Sputnik, and 2.2-2.5x over DASP. Applied to Llama2-7B pruned with Wanda to sparsity 50%, it delivers 1.5x memory reduction and 1.5x faster inference at fp16 precision. Thanks to MACKO, unstructured pruning at 50% sparsity is now justified in real-world LLM workloads.

The scaling of Generative AI (GenAI) models into the hundreds of billions of parameters makes low-precision computation indispensable for efficient deployment. We argue that the fundamental solution lies in developing low-precision floating-point formats, which inherently provide numerical stability, memory savings, and hardware efficiency without dequantization overhead. In this paper, we present a theoretical and empirical study of an exponent concentration phenomenon in GenAI weights: exponents consistently exhibit low entropy across architectures and modalities. We show that this arises naturally from $α$-stable distributions induced by stochastic gradient descent, and we prove tight bounds on the entropy of exponents. Our analysis establishes a theoretical compression limit near FP4.67, which motivates the design of a practical FP8 format. Building on these insights, we propose Exponent-Concentrated FP8 (ECF8), a lossless compression framework with entropy-aware encoding and GPU-optimized decoding. Experiments on LLMs and DiTs up to 671B parameters demonstrate up to 26.9% memory savings and 177.1% throughput acceleration, with perfectly lossless computations, i.e., no deviation in model outputs. Our results establish exponent concentration as a statistical law of trained models and open a principled path for lossless low-precision floating-point design in the FP8 era.

Neural Information Processing Systems http://nips.cc/

Training large neural networks with end-to-end backpropagation creates significant memory bottlenecks, limiting accessibility to state-of-the-art AI research. We propose $\textit{DiffusionBlocks}$, a novel training framework that interprets neural network blocks as performing denoising operations in a continuous-time diffusion process. By partitioning the network into independently trainable blocks and optimizing noise level assignments based on equal cumulative probability mass, our approach achieves significant memory efficiency while maintaining competitive performance compared to traditional backpropagation in generative tasks. Experiments on image generation and language modeling tasks demonstrate memory reduction proportional to the number of blocks while achieving superior performance. DiffusionBlocks provides a promising pathway for democratizing access to large-scale neural network training with limited computational resources.

Hyperdimensional Computing (HDC) is emerging as a promising approach for edge AI, offering a balance between accuracy and efficiency. However, current HDC-based applications often rely on high-precision models and/or encoding matrices to achieve competitive performance, which imposes significant computational and memory demands, especially for ultra-low power devices. While recent efforts use techniques like precision reduction and pruning to increase the efficiency, most require retraining to maintain performance, making them expensive and impractical. To address this issue, we propose a novel Post Training Compression algorithm, Decomposition-Pruning-Quantization (DPQ-HD), which aims at compressing the end-to-end HDC system, achieving near floating point performance without the need of retraining. DPQ-HD reduces computational and memory overhead by uniquely combining the above three compression techniques and efficiently adapts to hardware constraints. Additionally, we introduce an energy-efficient inference approach that progressively evaluates similarity scores such as cosine similarity and performs early exit to reduce the computation, accelerating prediction inference while maintaining accuracy. We demonstrate that DPQ-HD achieves up to 20-100x reduction in memory for image and graph classification tasks with only a 1-2% drop in accuracy compared to uncompressed workloads. Lastly, we show that DPQ-HD outperforms the existing post-training compression methods and performs better or at par with retraining-based state-of-the-art techniques, requiring significantly less overall optimization time (up to 100x) and faster inference (up to 56x) on a microcontroller

The paper proposes a method that reduces memory consumption of back prop, and as a result allows the use of larger batch sizes. The authors state that this is significant e.g. with batch norm, where batch size matters. In my own experience, I believe this is also significant for improved GPU utilization and data-parallel training. The paper is original in that I have not seen a similar treatment (although the actual solution may have been relatively straight-forward; asking the right question was an important part of this). The paper is well written and easy to follow.

The advancements in large language models (LLMs) have propelled the improvement of video understanding tasks by incorporating LLMs with visual models. However, most existing LLM-based models (e.g., VideoLLaMA, VideoChat) are constrained to processing short-duration videos. Recent attempts to understand long-term videos by extracting and compressing visual features into a fixed memory size. Nevertheless, those methods leverage only visual modality to merge video tokens and overlook the correlation between visual and textual queries, leading to difficulties in effectively handling complex question-answering tasks. To address the challenges of long videos and complex prompts, we propose AdaCM$^2$, which, for the first time, introduces an adaptive cross-modality memory reduction approach to video-text alignment in an auto-regressive manner on video streams. Our extensive experiments on various video understanding tasks, such as video captioning, video question answering, and video classification, demonstrate that AdaCM$^2$ achieves state-of-the-art performance across multiple datasets while significantly reducing memory usage. Notably, it achieves a 4.5% improvement across multiple tasks in the LVU dataset with a GPU memory consumption reduction of up to 65%.

The ubiquity of vision transformers (ViTs) for various edge applications, including personalized learning, has created the demand for on-device fine-tuning. However, training with the limited memory and computation power of edge devices remains a significant challenge. In particular, the memory required for training is much higher than that needed for inference, primarily due to the need to store activations across all layers in order to compute the gradients needed for weight updates. Previous works have explored reducing this memory requirement via frozen-weight training as well storing the activations in a compressed format. However, these methods are deemed inefficient due to their inability to provide training or inference speedup. In this paper, we first investigate the limitations of existing on-device training methods aimed at reducing memory and compute requirements. We then present block selective reprogramming (BSR) in which we fine-tune only a fraction of total blocks of a pre-trained model and selectively drop tokens based on self-attention scores of the frozen layers. To show the efficacy of BSR, we present extensive evaluations on ViT-B and DeiT-S with five different datasets. Compared to the existing alternatives, our approach simultaneously reduces training memory by up to 1.4x and compute cost by up to 2x while maintaining similar accuracy. We also showcase results for Mixture-of-Expert (MoE) models, demonstrating the effectiveness of our approach in multitask learning scenarios.

In the last decade, Convolutional Neural Network with a multi-layer architecture has advanced rapidly. However, training its complex network is very space-consuming, since a lot of intermediate data are preserved across layers, especially when processing high-dimension inputs with a big batch size. That poses great challenges to the limited memory capacity of current accelerators (e.g., GPUs). Existing efforts mitigate such bottleneck by external auxiliary solutions with additional hardware costs, and internal modifications with potential accuracy penalty. Differently, our analysis reveals that computations intra- and inter-layers exhibit the spatial-temporal weak dependency and even complete independency features. That inspires us to break the traditional layer-by-layer (column) dataflow rule. Now operations are novelly re-organized into rows throughout all convolution layers. This lightweight design allows a majority of intermediate data to be removed without any loss of accuracy. We particularly study the weak dependency between two consecutive rows. For the resulting skewed memory consumption, we give two solutions with different favorite scenarios. Evaluations on two representative networks confirm the effectiveness. We also validate that our middle dataflow optimization can be smoothly embraced by existing works for better memory reduction.