We propose Rotate, Clip, and Partition (RCP), a quantization-aware training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4(2-bit weight, 4-bit activation, and 4-bit KV cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design, by quantitatively analyzing the impact of random rotation on 2-bit weight quantization. Our weight quantizer features Learnable Direct Partitioning (LDP), which introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a specialized GPU kernel that supports GEMV on non-uniform W2A4. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 ppl and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code will be made available at blind_review.
We introduce Slam, a recipe for training high-quality Speech Language Models (SLMs) on a single academic GPU in 24 hours. We do so through empirical analysis of model initialisation and architecture, synthetic training data, preference optimisation with synthetic data and tweaking all other components. We empirically demonstrate that this training recipe also scales well with more compute getting results on par with leading SLMs in a fraction of the compute cost. We hope these insights will make SLM training and research more accessible. In the context of SLM scaling laws, our results far outperform predicted compute optimal performance, giving an optimistic view to SLM feasibility. See code, data, models, samples at - https://pages.cs.huji.ac.il/adiyoss-lab/slamming .
Although being widely adopted for evaluating generated audio signals, the Fr\'echet Audio Distance (FAD) suffers from significant limitations, including reliance on Gaussian assumptions, sensitivity to sample size, and high computational complexity. As an alternative, we introduce the Kernel Audio Distance (KAD), a novel, distribution-free, unbiased, and computationally efficient metric based on Maximum Mean Discrepancy (MMD). Through analysis and empirical validation, we demonstrate KAD's advantages: (1) faster convergence with smaller sample sizes, enabling reliable evaluation with limited data; (2) lower computational cost, with scalable GPU acceleration; and (3) stronger alignment with human perceptual judgments. By leveraging advanced embeddings and characteristic kernels, KAD captures nuanced differences between real and generated audio. Open-sourced in the kadtk toolkit, KAD provides an efficient, reliable, and perceptually aligned benchmark for evaluating generative audio models.
It is critical for robots to explore Federated Learning (FL) settings where several robots, deployed in parallel, can learn independently while also sharing their learning with each other. This collaborative learning in real-world environments requires social robots to adapt dynamically to changing and unpredictable situations and varying task settings. Our work contributes to addressing these challenges by exploring a simulated living room environment where robots need to learn the social appropriateness of their actions. First, we propose Federated Root (FedRoot) averaging, a novel weight aggregation strategy which disentangles feature learning across clients from individual task-based learning. Second, to adapt to challenging environments, we extend FedRoot to Federated Latent Generative Replay (FedLGR), a novel Federated Continual Learning (FCL) strategy that uses FedRoot-based weight aggregation and embeds each client with a generator model for pseudo-rehearsal of learnt feature embeddings to mitigate forgetting in a resource-efficient manner. Our results show that FedRoot-based methods offer competitive performance while also resulting in a sizeable reduction in resource consumption (up to 86% for CPU usage and up to 72% for GPU usage). Additionally, our results demonstrate that FedRoot-based FCL methods outperform other methods while also offering an efficient solution (up to 84% CPU and 92% GPU usage reduction), with FedLGR providing the best results across evaluations.
Until now, in the wake of the COVID-19 pandemic in 2019, lung diseases, especially diseases such as lung cancer and chronic obstructive pulmonary disease (COPD), have become an urgent global health issue. In order to mitigate the goal problem, early detection and accurate diagnosis of these conditions are critical for effective treatment and improved patient outcomes. To further research and reduce the error rate of hospital diagnoses, this comprehensive study explored the potential of computer-aided design (CAD) systems, especially utilizing advanced deep learning models such as U-Net. And compared with the literature content of other authors, this study explores the capabilities of U-Net in detail, and enhances the ability to simulate CAD systems through the VGG16 algorithm. An extensive dataset consisting of lung CT images and corresponding segmentation masks, curated collaboratively by multiple academic institutions, serves as the basis for empirical validation. In this paper, the efficiency of U-Net model is evaluated rigorously and precisely under multiple hardware configurations, such as single CPU, single GPU, distributed GPU and federated learning, and the effectiveness and development of the method in the segmentation task of lung disease are demonstrated. Empirical results clearly affirm the robust performance of the U-Net model, most effectively utilizing four GPUs for distributed learning, and these results highlight the potential of U-Net-based CAD systems for accurate and timely lung disease detection and diagnosis huge potential.
Despite the efficacy of network sparsity in alleviating the deployment strain of Large Language Models (LLMs), it endures significant performance degradation. Applying Low-Rank Adaptation (LoRA) to fine-tune the sparse LLMs offers an intuitive approach to counter this predicament, while it holds shortcomings include: 1) The inability to integrate LoRA weights into sparse LLMs post-training, and 2) Insufficient performance recovery at high sparsity ratios. In this paper, we introduce dynamic Low-rank Sparse Adaptation (LoSA), a novel method that seamlessly integrates low-rank adaptation into LLM sparsity within a unified framework, thereby enhancing the performance of sparse LLMs without increasing the inference latency. In particular, LoSA dynamically sparsifies the LoRA outcomes based on the corresponding sparse weights during fine-tuning, thus guaranteeing that the LoRA module can be integrated into the sparse LLMs post-training. Besides, LoSA leverages Representation Mutual Information (RMI) as an indicator to determine the importance of layers, thereby efficiently determining the layer-wise sparsity rates during fine-tuning. Predicated on this, LoSA adjusts the rank of the LoRA module based on the variability in layer-wise reconstruction errors, allocating an appropriate fine-tuning for each layer to reduce the output discrepancies between dense and sparse LLMs. Extensive experiments tell that LoSA can efficiently boost the efficacy of sparse LLMs within a few hours, without introducing any additional inferential burden. For example, LoSA reduced the perplexity of sparse LLaMA-2-7B by 68.73 and increased zero-shot accuracy by 16.32$\%$, achieving a 2.60$\times$ speedup on CPU and 2.23$\times$ speedup on GPU, requiring only 45 minutes of fine-tuning on a single NVIDIA A100 80GB GPU. Code is available at https://github.com/wzhuang-xmu/LoSA.
In this paper, we address the challenge of determining the layer-wise sparsity rates of large language models (LLMs) through a theoretical perspective. Specifically, we identify a critical issue of ''$\textbf{reconstruction error explosion}$'' in existing LLMs sparsification methods. This refers to the cumulative effect of reconstruction errors throughout the sparsification process, where errors from earlier layers propagate and amplify in subsequent layers. As a result, the overall reconstruction error increases significantly, leading to a substantial degradation in model performance. Through theoretical analysis, we derive a simple yet effective approach to layer-wise sparsity allocation that mitigates this issue. Our method uses a monotonically increasing arithmetic progression, reducing the process of determining sparsity rates for multiple layers to the determination of a single common difference hyperparameter. Remarkably, this allows for the optimal layer-wise sparsity rates to be identified with just a few trials. Both our theoretical analysis and experimental results demonstrate that this sparsity allocation scheme is near optimal. Extensive experiments show that our method significantly improves the performance of sparse LLMs across various architectures, outperforming existing layer-wise sparsity methods. Furthermore, it enhances the performance of various compression techniques and is applicable to vision and multimodal models. Notably, our method achieves a reduction of 52.10 in perplexity for the 70$\%$ sparse LLaMA2-7B model obtained via Wanda, improves average zero-shot accuracy by 10.50$\%$, and delivers speedups of 2.63$\times$ and 2.23$\times$ on CPU and GPU, respectively.
Triton, a high-level Python-like language designed for building efficient GPU kernels, is widely adopted in deep learning frameworks due to its portability, flexibility, and accessibility. However, programming and parallel optimization still require considerable trial and error from Triton developers. Despite advances in large language models (LLMs) for conventional code generation, these models struggle to generate accurate, performance-optimized Triton code, as they lack awareness of its specifications and the complexities of GPU programming. More critically, there is an urgent need for systematic evaluations tailored to Triton. In this work, we introduce TritonBench, the first comprehensive benchmark for Triton operator generation. TritonBench features two evaluation channels: a curated set of 184 real-world operators from GitHub and a collection of operators aligned with PyTorch interfaces. Unlike conventional code benchmarks prioritizing functional correctness, TritonBench also profiles efficiency performance on widely deployed GPUs aligned with industry applications. Our study reveals that current state-of-the-art code LLMs struggle to generate efficient Triton operators, highlighting a significant gap in high-performance code generation. TritonBench will be available at https://github.com/thunlp/TritonBench.
Despite exceptional capabilities, Large Language Models (LLMs) still face deployment challenges due to their enormous size. Post-training structured pruning is a promising solution that prunes LLMs without the need for retraining, reducing computational overhead, and it is hardware-deployment friendly. However, the training-free nature of post-training structured pruning leads to significant performance degradation. We argue that the key to mitigating this issue lies in accurately determining the pruning rate for each layer. Meanwhile, we find that LLMs may have prior knowledge about their own redundancy. Based on this insight, we introduce $\textbf{Self-Pruner}$ an end-to-end automatic self-pruning framework for LLMs, which efficiently search layer-wise pruning rates. Specifically, $\textbf{Self-Pruner}$ leverages LLMs to autonomously execute the entire evolutionary search process to search for pruning rate configurations. In this process, LLMs are used to generate populations, select parent solutions from the current population, and perform crossover and mutation operations to produce offspring solutions. In this way, LLMs automatically generate and evaluate a large number of candidate solutions, effectively converging to find the pruning rate configurations with minimal human intervention. Extensive experiments demonstrate $\textbf{Self-Pruner}$'s better performance compared to existing state-of-the-art methods. Notably, $\textbf{Self-Pruner}$ prunes LLaMA-2-70B to 49B level with only 0.80$\%$ drop in accuracy across seven commonsense reasoning tasks, achieving a 1.39$\times$ speedup on NVIDIA A100 80GB GPU. Further pruning to 35B level resulted in only a 3.80$\%$ decrease in accuracy while obtaining a 1.70$\times$ speedup.
Bytes form the basis of the digital world and thus are a promising building block for multimodal foundation models. Recently, Byte Language Models (BLMs) have emerged to overcome tokenization, yet the excessive length of bytestreams requires new architectural paradigms. Therefore, we present the Multiscale Byte Language Model (MBLM), a model-agnostic hierarchical decoder stack that allows training with context windows of $5$M bytes on single GPU in full model precision. We thoroughly examine MBLM's performance with Transformer and Mamba blocks on both unimodal and multimodal tasks. Our experiments demonstrate that hybrid architectures are efficient in handling extremely long byte sequences during training while achieving near-linear generational efficiency. To the best of our knowledge, we present the first evaluation of BLMs on visual Q\&A tasks and find that, despite serializing images and the absence of an encoder, a MBLM with pure next token prediction can match custom CNN-LSTM architectures with designated classification heads. We show that MBLMs exhibit strong adaptability in integrating diverse data representations, including pixel and image filestream bytes, underlining their potential toward omnimodal foundation models. Source code is publicly available at: https://github.com/ai4sd/multiscale-byte-lm
Efficiently handling long contexts is crucial for large language models (LLMs). While rotary position embeddings (RoPEs) enhance length generalization, effective length extrapolation remains challenging and often requires costly fine-tuning. In contrast, recent training-free approaches suffer from the attention sink phenomenon, leading to severe performance degradation. In this paper, we introduce ParallelComp, a novel training-free method for long-context extrapolation that extends LLMs' context length from 4K to 128K while maintaining high throughput and preserving perplexity, and integrates seamlessly with Flash Attention. Our analysis offers new insights into attention biases in parallel attention mechanisms and provides practical solutions to tackle these challenges. To mitigate the attention sink issue, we propose an attention calibration strategy that reduces biases, ensuring more stable long-range attention. Additionally, we introduce a chunk eviction strategy to efficiently manage ultra-long contexts on a single A100 80GB GPU. To further enhance efficiency, we propose a parallel KV cache eviction technique, which improves chunk throughput by 1.76x, thereby achieving a 23.50x acceleration in the prefilling stage with negligible performance loss due to attention calibration. Furthermore, ParallelComp achieves 91.17% of GPT-4's performance on long-context tasks using an 8B model trained on 8K-length context, outperforming powerful closed-source models such as Claude-2 and Kimi-Chat.
Transformer-based Large Language Models rely critically on KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy designed specifically to reduce both memory bandwidth and capacity demand of KV cache during the decode phase. RocketKV contains two consecutive stages. In the first stage, it performs coarse-grain KV cache eviction on the input sequence tokens with SnapKV++, a method improved upon SnapKV by introducing adaptive pooling size and full compatibility with grouped-query attention. In the second stage, it adopts a hybrid attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensional reductions. Combining these two stages, RocketKV achieves significant KV cache fetching bandwidth and storage savings while maintaining comparable accuracy to full KV cache attention. We show that RocketKV provides end-to-end speedup by up to 3$\times$ as well as peak memory reduction by up to 31% in the decode phase on an NVIDIA H100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks.
Simulation agents are essential for designing and testing systems that interact with humans, such as autonomous vehicles (AVs). These agents serve various purposes, from benchmarking AV performance to stress-testing system limits, but all applications share one key requirement: reliability. To enable sound experimentation, a simulation agent must behave as intended. It should minimize actions that may lead to undesired outcomes, such as collisions, which can distort the signal-to-noise ratio in analyses. As a foundation for reliable sim agents, we propose scaling self-play to thousands of scenarios on the Waymo Open Motion Dataset under semi-realistic limits on human perception and control. Training from scratch on a single GPU, our agents solve almost the full training set within a day. They generalize to unseen test scenes, achieving a 99.8% goal completion rate with less than 0.8% combined collision and off-road incidents across 10,000 held-out scenarios. Beyond in-distribution generalization, our agents show partial robustness to out-of-distribution scenes and can be fine-tuned in minutes to reach near-perfect performance in such cases. We open-source the pre-trained agents and integrate them with a batched multi-agent simulator. Demonstrations of agent behaviors can be viewed at https://sites.google.com/view/reliable-sim-agents, and we open-source our agents at https://github.com/Emerge-Lab/gpudrive.
This paper investigates the efficacy of RWKV, a novel language model architecture known for its linear attention mechanism, for generating sentence embeddings in a zero-shot setting. I conduct a layer-wise analysis to evaluate the semantic similarity captured by embeddings from different hidden layers of a pre-trained RWKV model. The performance is assessed on the Microsoft Research Paraphrase Corpus (MRPC) dataset using Spearman correlation and compared against a GloVe-based baseline. My results indicate that while RWKV embeddings capture some semantic relatedness, they underperform compared to the GloVe baseline in terms of Spearman correlation. I also analyze the inference time and GPU memory usage, highlighting the computational trade-offs associated with RWKV embeddings. The findings suggest that while RWKV offers potential advantages in terms of linear scaling, its zero-shot sentence embedding quality for semantic similarity tasks requires further investigation and potential task-specific fine-tuning to match or exceed simpler baselines.
Recent Multimodal Large Language Models (MLLMs) have achieved remarkable performance but face deployment challenges due to their quadratic computational complexity, growing Key-Value cache requirements, and reliance on separate vision encoders. We propose mmMamba, a framework for developing linear-complexity native multimodal state space models through progressive distillation from existing MLLMs using moderate academic computational resources. Our approach enables the direct conversion of trained decoder-only MLLMs to linear-complexity architectures without requiring pre-trained RNN-based LLM or vision encoders. We propose an seeding strategy to carve Mamba from trained Transformer and a three-stage distillation recipe, which can effectively transfer the knowledge from Transformer to Mamba while preserving multimodal capabilities. Our method also supports flexible hybrid architectures that combine Transformer and Mamba layers for customizable efficiency-performance trade-offs. Distilled from the Transformer-based decoder-only HoVLE, mmMamba-linear achieves competitive performance against existing linear and quadratic-complexity VLMs, while mmMamba-hybrid further improves performance significantly, approaching HoVLE's capabilities. At 103K tokens, mmMamba-linear demonstrates 20.6$\times$ speedup and 75.8% GPU memory reduction compared to HoVLE, while mmMamba-hybrid achieves 13.5$\times$ speedup and 60.2% memory savings. Code and models are released at https://github.com/hustvl/mmMamba
Point cloud segmentation (PCS) is to classify each point in point clouds. The task enables robots to parse their 3D surroundings and run autonomously. According to different point cloud representations, existing PCS models can be roughly divided into point-, voxel-, and range image-based models. However, no work has been found to report comprehensive comparisons among the state-of-the-art point-, voxel-, and range image-based models from an application perspective, bringing difficulty in utilizing these models for real-world scenarios. In this paper, we provide thorough comparisons among the models by considering the LiDAR data motion compensation and the metrics of model parameters, max GPU memory allocated during testing, inference latency, frames per second, intersection-over-union (IoU) and mean IoU (mIoU) scores. The experimental results benefit engineers when choosing a reasonable PCS model for an application and inspire researchers in the PCS field to design more practical models for a real-world scenario.
Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.
Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.
Imposing an effective structural assumption on neural network weight matrices has been the major paradigm for designing Parameter-Efficient Fine-Tuning (PEFT) systems for adapting modern large pre-trained models to various downstream tasks. However, low rank based adaptation has become increasingly challenging due to the sheer scale of modern large language models. In this paper, we propose an effective kernelization to further reduce the number of parameters required for adaptation tasks. Specifically, from the classical idea in numerical analysis regarding matrix Low-Separation-Rank (LSR) representations, we develop a kernel using this representation for the low rank adapter matrices of the linear layers from large networks, named the Low Separation Rank Adaptation (LSR-Adapt) kernel. With the ultra-efficient kernel representation of the low rank adapter matrices, we manage to achieve state-of-the-art performance with even higher accuracy with almost half the number of parameters as compared to conventional low rank based methods. This structural assumption also opens the door to further GPU-side optimizations due to the highly parallelizable nature of Kronecker computations.
Referring Multi-Object Tracking (RMOT) is a relatively new concept that has rapidly gained traction as a promising research direction at the intersection of computer vision and natural language processing. Unlike traditional multi-object tracking, RMOT identifies and tracks objects and incorporates textual descriptions for object class names, making the approach more intuitive. Various techniques have been proposed to address this challenging problem; however, most require the training of the entire network due to their end-to-end nature. Among these methods, iKUN has emerged as a particularly promising solution. Therefore, we further explore its pipeline and enhance its performance. In this paper, we introduce a practical module dubbed Memory-Efficient Cross-modality -- MEX. This memory-efficient technique can be directly applied to off-the-shelf trackers like iKUN, resulting in significant architectural improvements. Our method proves effective during inference on a single GPU with 4 GB of memory. Among the various benchmarks, the Refer-KITTI dataset, which offers diverse autonomous driving scenes with relevant language expressions, is particularly useful for studying this problem. Empirically, our method demonstrates effectiveness and efficiency regarding HOTA tracking scores, substantially improving memory allocation and processing speed.