Pre-training Transformers in FP4 precision is becoming a promising approach to gain substantial speedup, but it comes with a considerable loss of accuracy. Microscaling (MX) data format provides a fine-grained per-group quantization method to improve the representation ability of the FP4 format and is supported by the next-generation Blackwell GPU architecture. However, training with MXFP4 data format still results in significant degradation and there is a lack of systematic research on the reason. In this work, we propose a novel training method TetraJet for a more accurate FP4 training. We comprehensively evaluate all of the quantizers involved in the training, and identify the weight oscillation problem in the forward pass as the main source of the degradation in MXFP4 training. Therefore, we introduce two novel methods, EMA Quantizer (Q-EMA) and Adaptive Ramping Optimizer (Q-Ramping), to resolve the oscillation problem. Extensive experiments on Vision Transformers demonstrate that TetraJet consistently outperforms the existing 4-bit training methods, and Q-EMA & Q-Ramping can provide additional enhancement by effectively reducing oscillation. We decreased the accuracy degradation by more than $50\%$ compared to the baseline, and can even achieve competitive performance compared to full precision training. The codes are available at https://github.com/thu-ml/TetraJet-MXFP4Training
Visual generative and understanding models typically rely on distinct tokenizers to process images, presenting a key challenge for unifying them within a single framework. Recent studies attempt to address this by connecting the training of VQVAE (for autoregressive generation) and CLIP (for understanding) to build a unified tokenizer. However, directly combining these training objectives has been observed to cause severe loss conflicts. In this paper, we show that reconstruction and semantic supervision do not inherently conflict. Instead, the underlying bottleneck stems from limited representational capacity of discrete token space. Building on these insights, we introduce UniTok, a unified tokenizer featuring a novel multi-codebook quantization mechanism that effectively scales up the vocabulary size and bottleneck dimension. In terms of final performance, UniTok sets a new record of 0.38 rFID and 78.6% zero-shot accuracy on ImageNet. Besides, UniTok can be seamlessly integrated into MLLMs to unlock native visual generation capability, without compromising the understanding performance. Additionally, we show that UniTok favors cfg-free generation, reducing gFID from 14.6 to 2.5 on ImageNet 256$\times$256 benchmark. GitHub: https://github.com/FoundationVision/UniTok.
Low-bit model quantization for image super-resolution (SR) is a longstanding task that is renowned for its surprising compression and acceleration ability. However, accuracy degradation is inevitable when compressing the full-precision (FP) model to ultra-low bit widths (2~4 bits). Experimentally, we observe that the degradation of quantization is mainly attributed to the quantization of activation instead of model weights. In numerical analysis, the condition number of weights could measure how much the output value can change for a small change in the input argument, inherently reflecting the quantization error. Therefore, we propose CondiQuant, a condition number based low-bit post-training quantization for image super-resolution. Specifically, we formulate the quantization error as the condition number of weight metrics. By decoupling the representation ability and the quantization sensitivity, we design an efficient proximal gradient descent algorithm to iteratively minimize the condition number and maintain the output still. With comprehensive experiments, we demonstrate that CondiQuant outperforms existing state-of-the-art post-training quantization methods in accuracy without computation overhead and gains the theoretically optimal compression ratio in model parameters. Our code and model are released at https://github.com/Kai-Liu001/CondiQuant.
Structured pruning and quantization are fundamental techniques used to reduce the size of deep neural networks (DNNs) and typically are applied independently. Applying these techniques jointly via co-optimization has the potential to produce smaller, high-quality models. However, existing joint schemes are not widely used because of (1) engineering difficulties (complicated multi-stage processes), (2) black-box optimization (extensive hyperparameter tuning to control the overall compression), and (3) insufficient architecture generalization. To address these limitations, we present the framework GETA, which automatically and efficiently performs joint structured pruning and quantization-aware training on any DNNs. GETA introduces three key innovations: (i) a quantization-aware dependency graph (QADG) that constructs a pruning search space for generic quantization-aware DNN, (ii) a partially projected stochastic gradient method that guarantees layerwise bit constraints are satisfied, and (iii) a new joint learning strategy that incorporates interpretable relationships between pruning and quantization. We present numerical experiments on both convolutional neural networks and transformer architectures that show that our approach achieves competitive (often superior) performance compared to existing joint pruning and quantization methods.
Large Language Models (LLMs) have emerged as powerful tools for addressing modern challenges and enabling practical applications. However, their computational expense remains a significant barrier to widespread adoption. Quantization has emerged as a promising technique to democratize access and enable low resource device deployment. Despite these advancements, the safety and trustworthiness of quantized models remain underexplored, as prior studies often overlook contemporary architectures and rely on overly simplistic benchmarks and evaluations. To address this gap, we introduce OpenSafetyMini, a novel open-ended safety dataset designed to better distinguish between models. We evaluate 4 state-of-the-art quantization techniques across LLaMA and Mistral models using 4 benchmarks, including human evaluations. Our findings reveal that the optimal quantization method varies for 4-bit precision, while vector quantization techniques deliver the best safety and trustworthiness performance at 2-bit precision, providing foundation for future research.
Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.
Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81$\times$ (16.95$\times$), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B).
Large Language Models (LLMs) suffer severe performance degradation when facing extremely low-bit (sub 2-bit) quantization. Several existing sub 2-bit post-training quantization (PTQ) methods utilize a mix-precision scheme by leveraging an unstructured fine-grained mask to explicitly distinguish salient weights, while which introduces an extra 1-bit or more per weight. To explore the real limit of PTQ, we propose an extremely low-bit PTQ method called PTQ1.61, which enables weight quantization to 1.61-bit for the first time. Specifically, we first introduce a one-dimensional structured mask with negligibly additional 0.0002-bit per weight based on input activations from the perspective of reducing the upper bound of quantization error to allocate corresponding salient weight channels to 4-bit. For non-salient channels binarization, an efficient block-wise scaling factors optimization framework is then presented to take implicit row-wise correlations and angular biases into account. Different from prior works that concentrate on adjusting quantization methodologies, we further propose a novel paradigm called quantization preprocessing, where we argue that transforming the weight distribution of the pretrained model before quantization can alleviate the difficulty in per-channel extremely low-bit PTQ. Extensive experiments indicate our PTQ1.61 achieves state-of-the-art performance in extremely low-bit quantization. Codes are available at https://github.com/zjq0455/PTQ1.61.
Post-training Quantization (PTQ) technique has been extensively adopted for large language models (LLMs) compression owing to its efficiency and low resource requirement. However, current research lacks a in-depth analysis of the superior and applicable scenarios of each PTQ strategy. In addition, existing algorithms focus primarily on performance, overlooking the trade-off among model size, performance, and quantization bitwidth. To mitigate these confusions, we provide a novel benchmark for LLMs PTQ in this paper. Firstly, in order to support our benchmark, we propose a comprehensive taxonomy for existing mainstream methods by scrutinizing their computational strategies (e.g., optimization-based, compensation-based, etc.). Then, we conduct extensive experiments with the baseline within each class, covering models with various sizes (7B-70B), bitwidths, training levels (LLaMA1/2/3/3.1), architectures (Mixtral, DeepSeekMoE and Mamba) and modality (LLaVA1.5 and VILA1.5) on a wide range of evaluation metrics.Through comparative analysis on the results, we summarize the superior of each PTQ strategy and modelsize-bitwidth trade-off considering the performance. For example, our benchmark reveals that compensation-based technique demonstrates outstanding cross-architecture robustness and extremely low-bit PTQ for ultra large models should be reexamined. Finally, we further accordingly claim that a practical combination of compensation and other PTQ strategy can achieve SOTA various robustness. We believe that our benchmark will provide valuable recommendations for the deployment of LLMs and future research on PTQ approaches.We conduct an repository for our benchmark at https://github.com/zjq0455/PTQ_Benchmark.
This work introduces NeuroQuant, a novel post-training quantization (PTQ) approach tailored to non-generalized Implicit Neural Representations for variable-rate Video Coding (INR-VC). Unlike existing methods that require extensive weight retraining for each target bitrate, we hypothesize that variable-rate coding can be achieved by adjusting quantization parameters (QPs) of pre-trained weights. Our study reveals that traditional quantization methods, which assume inter-layer independence, are ineffective for non-generalized INR-VC models due to significant dependencies across layers. To address this, we redefine variable-rate INR-VC as a mixed-precision quantization problem and establish a theoretical framework for sensitivity criteria aimed at simplified, fine-grained rate control. Additionally, we propose network-wise calibration and channel-wise quantization strategies to minimize quantization-induced errors, arriving at a unified formula for representation-oriented PTQ calibration. Our experimental evaluations demonstrate that NeuroQuant significantly outperforms existing techniques in varying bitwidth quantization and compression efficiency, accelerating encoding by up to eight times and enabling quantization down to INT2 with minimal reconstruction loss. This work introduces variable-rate INR-VC for the first time and lays a theoretical foundation for future research in rate-distortion optimization, advancing the field of video coding technology. The materials will be available at https://github.com/Eric-qi/NeuroQuant.
Approximating a probability distribution using a set of particles is a fundamental problem in machine learning and statistics, with applications including clustering and quantization. Formally, we seek a weighted mixture of Dirac measures that best approximates the target distribution. While much existing work relies on the Wasserstein distance to quantify approximation errors, maximum mean discrepancy (MMD) has received comparatively less attention, especially when allowing for variable particle weights. We argue that a Wasserstein-Fisher-Rao gradient flow is well-suited for designing quantizations optimal under MMD. We show that a system of interacting particles satisfying a set of ODEs discretizes this flow. We further derive a new fixed-point algorithm called mean shift interacting particles (MSIP). We show that MSIP extends the classical mean shift algorithm, widely used for identifying modes in kernel density estimators. Moreover, we show that MSIP can be interpreted as preconditioned gradient descent and that it acts as a relaxation of Lloyd's algorithm for clustering. Our unification of gradient flows, mean shift, and MMD-optimal quantization yields algorithms that are more robust than state-of-the-art methods, as demonstrated via high-dimensional and multi-modal numerical experiments.
Large language models (LLMs) offer powerful capabilities but come with significant environmental impact, particularly in carbon emissions. Existing studies benchmark carbon emissions but lack a standardized basis for comparison across different model configurations. To address this, we introduce the concept of functional unit (FU) as a standardized basis and develop FUEL, the first FU-based framework for evaluating LLM serving's environmental impact. Through three case studies, we uncover key insights and trade-offs in reducing carbon emissions by optimizing model size, quantization strategy, and hardware choice, paving the way for more sustainable LLM serving. The code is available at https://github.com/jojacola/FUEL.
We propose a beam codebook design to reduce self-interference (SI) in integrated sensing and communication (ISAC) systems. Our optimization methods, which can be applied to both tapered beamforming and phased arrays, adapt the codebooks to the SI channel such that a certain SI level is achieved. Furthermore, we derive an upper bound on the quantization noise in terms of the achieved SI level, which provides guidelines to pose the optimization problem in order to obtain performance guarantees for sensing. By selecting standard reference codebooks in our simulations, we show substantially improved sensing quality with little impact on 5G-NR communication. Our proposed method is not only less dependent on hyperparameters than other approaches in the literature, but it can also reduce SI further, and thus deliver better sensing and communication performance.
This work introduces Semantically Masked VQ-GAN (SQ-GAN), a novel approach integrating generative models to optimize image compression for semantic/task-oriented communications. SQ-GAN employs off-the-shelf semantic semantic segmentation and a new specifically developed semantic-conditioned adaptive mask module (SAMM) to selectively encode semantically significant features of the images. SQ-GAN outperforms state-of-the-art image compression schemes such as JPEG2000 and BPG across multiple metrics, including perceptual quality and semantic segmentation accuracy on the post-decoding reconstructed image, at extreme low compression rates expressed in bits per pixel.
In this paper, we propose a method to perform empirical analysis of the loss landscape of machine learning (ML) models. The method is applied to two ML models for scientific sensing, which necessitates quantization to be deployed and are subject to noise and perturbations due to experimental conditions. Our method allows assessing the robustness of ML models to such effects as a function of quantization precision and under different regularization techniques -- two crucial concerns that remained underexplored so far. By investigating the interplay between performance, efficiency, and robustness by means of loss landscape analysis, we both established a strong correlation between gently-shaped landscapes and robustness to input and weight perturbations and observed other intriguing and non-obvious phenomena. Our method allows a systematic exploration of such trade-offs a priori, i.e., without training and testing multiple models, leading to more efficient development workflows. This work also highlights the importance of incorporating robustness into the Pareto optimization of ML models, enabling more reliable and adaptive scientific sensing systems.
Compute-in-memory (CIM) is an efficient method for implementing deep neural networks (DNNs) but suffers from substantial overhead from analog-to-digital converters (ADCs), especially as ADC precision increases. Low-precision ADCs can reduce this overhead but introduce partial-sum quantization errors degrading accuracy. Additionally, low-bit weight constraints, imposed by cell limitations and the need for multiple cells for higher-bit weights, present further challenges. While fine-grained partial-sum quantization has been studied to lower ADC resolution effectively, weight granularity, which limits overall partial-sum quantized accuracy, remains underexplored. This work addresses these challenges by aligning weight and partial-sum quantization granularities at the column-wise level. Our method improves accuracy while maintaining dequantization overhead, simplifies training by removing two-stage processes, and ensures robustness to memory cell variations via independent column-wise scale factors. We also propose an open-source CIM-oriented convolution framework to handle fine-grained weights and partial-sums efficiently, incorporating a novel tiling method and group convolution. Experimental results on ResNet-20 (CIFAR-10, CIFAR-100) and ResNet-18 (ImageNet) show accuracy improvements of 0.99%, 2.69%, and 1.01%, respectively, compared to the best-performing related works. Additionally, variation analysis reveals the robustness of our method against memory cell variations. These findings highlight the effectiveness of our quantization scheme in enhancing accuracy and robustness while maintaining hardware efficiency in CIM-based DNN implementations. Our code is available at https://github.com/jiyoonkm/ColumnQuant.
We develop diffusion models for simulating lattice gauge theories, where stochastic quantization is explicitly incorporated as a physical condition for sampling. We demonstrate the applicability of this novel sampler to U(1) gauge theory in two spacetime dimensions and find that a model trained at a small inverse coupling constant can be extrapolated to larger inverse coupling regions without encountering the topological freezing problem. Additionally, the trained model can be employed to sample configurations on different lattice sizes without requiring further training. The exactness of the generated samples is ensured by incorporating Metropolis-adjusted Langevin dynamics into the generation process. Furthermore, we demonstrate that this approach enables more efficient sampling of topological quantities compared to traditional algorithms such as Hybrid Monte Carlo and Langevin simulations.
Recently, large language model (LLM) based text-to-speech (TTS) systems have gradually become the mainstream in the industry due to their high naturalness and powerful zero-shot voice cloning capabilities.Here, we introduce the IndexTTS system, which is mainly based on the XTTS and Tortoise model. We add some novel improvements. Specifically, in Chinese scenarios, we adopt a hybrid modeling method that combines characters and pinyin, making the pronunciations of polyphonic characters and long-tail characters controllable. We also performed a comparative analysis of the Vector Quantization (VQ) with Finite-Scalar Quantization (FSQ) for codebook utilization of acoustic speech tokens. To further enhance the effect and stability of voice cloning, we introduce a conformer-based speech conditional encoder and replace the speechcode decoder with BigVGAN2. Compared with XTTS, it has achieved significant improvements in naturalness, content consistency, and zero-shot voice cloning. As for the popular TTS systems in the open-source, such as Fish-Speech, CosyVoice2, FireRedTTS and F5-TTS, IndexTTS has a relatively simple training process, more controllable usage, and faster inference speed. Moreover, its performance surpasses that of these systems. Our demos are available at https://index-tts.github.io.
Graph Neural Networks (GNNs) are vital for learning from graph-structured data, enabling applications in network analysis, recommendation systems, and speech analytics. Deploying them on edge devices like client PCs and laptops enhances real-time processing, privacy, and cloud independence. GNNs aid Retrieval-Augmented Generation (RAG) for Large Language Models (LLMs) and enable event-based vision tasks. However, irregular memory access, sparsity, and dynamic structures cause high latency and energy overhead on resource-constrained devices. While modern edge processors integrate CPUs, GPUs, and NPUs, NPUs designed for data-parallel tasks struggle with irregular GNN computations. We introduce GraNNite, the first hardware-aware framework optimizing GNN execution on commercial-off-the-shelf (COTS) SOTA DNN accelerators via a structured three-step methodology: (1) enabling NPU execution, (2) optimizing performance, and (3) trading accuracy for efficiency gains. Step 1 employs GraphSplit for workload distribution and StaGr for static aggregation, while GrAd and NodePad handle dynamic graphs. Step 2 boosts performance using EffOp for control-heavy tasks and GraSp for sparsity exploitation. Graph Convolution optimizations PreG, SymG, and CacheG reduce redundancy and memory transfers. Step 3 balances quality versus efficiency, where QuantGr applies INT8 quantization, and GrAx1, GrAx2, and GrAx3 accelerate attention, broadcast-add, and SAGE-max aggregation. On Intel Core Ultra AI PCs, GraNNite achieves 2.6X to 7.6X speedups over default NPU mappings and up to 8.6X energy gains over CPUs and GPUs, delivering 10.8X and 6.7X higher performance than CPUs and GPUs, respectively, across GNN models.
KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we thoroughly analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 38.3% compared with KV8 quantization over various context lengths.