Quantization
-
The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs
2026-07-09 17:35:02Baha Rababah, Cuneyt Gurcan Akcora, Carson K. LeungAbstract
Post-training quantization is widely used to deploy large language models in resource-constrained settings, yet its evaluation relies almost exclusively on accuracy and perplexity. We show that these metrics fail to capture behavioral changes induced by quantization. We introduce correctness agreement, a decision-level metric that measures overlap in correct predictions between a base model and its quantized variants, independent of absolute accuracy. Across multiple models and quantization schemes from 8-bit to 2-bit, we find that behavioral divergence emerges under moderate quantization even when task performance appears preserved. To explain this effect, we analyze quantization as a structural operator on attention weights and quantify layer-wise distortions using statistical and distributional measures. Our results reveal non-linear breakpoints at low bit-widths and show that query and key projections are consistently more sensitive than value and output projections. These findings expose an illusion of equivalence between base and quantized models and motivate behavioral evaluation beyond conventional performance metrics.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08734
PDF
-
It Takes Few to TANGO: A Quantized Distributed Model for Binaural Speech Enhancement
2026-07-09 16:17:42Zahra Benslimane, Pierre Chouteau, Martyna Poreba, Fabrice Auzanneau, Michal Szczepanski, Fabian Chersi, Romain SerizelAbstract
Neural network-based multichannel speech enhancement systems achieve strong enhancement performance, but their computational and memory requirements limit deployment on resource-constrained devices. This paper investigates low-precision inference for TANGO, a hybrid distributed binaural speech enhancement system combining neural mask estimation with spatial filtering. We evaluate post-training quantization and quantization-aware training for the neural components, and analyze how quantization errors in the mask estimators propagate through the downstream spatial filtering stage. Our analysis shows that, although quantization degrades intermediate mask estimates, the spatial filtering stage compensates for most quantization-induced errors. Leveraging this robustness, we simplify TANGO into MN-TANGO, reducing both model size and computational complexity while maintaining comparable final performance. By combining INT8 weight-and-activation quantization with ERB compression and grouped recurrent layers, the most compact MN-TANGO reaches 4.65 MMAC/s and 0.177 MB.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08645
PDF
-
A Quantized Native Runtime for On-Device Semantic Audio Generation
Abstract
Semantic audio applications increasingly require controllable generation on commodity and embedded hardware rather than through framework-heavy datacenter stacks. We present \textit{aria}, a dependency-free native runtime that runs the complete text-to-music pipeline of Stable Audio~3 (SA3) on ordinary GPUs, CPU-only machines, and a Raspberry~Pi~5, with no Python or deep-learning framework underneath. Our main contribution is a study of quantization: running the model at lower numerical precision to fit tight memory budgets, saving memory in place rather than adding to it. Because the runtime owns every internal tensor, it also exposes activation steering, a low-cost way to steer what the model generates. We judge the quality cost with three independent measures of the output (prompt adherence, overall audio quality, taste preservation), each compared against the ordinary variation between random seeds. Eight-bit precision shows no measurable quality loss on any measure while sharply cutting memory, and it is the fastest mode on the GPU; four-bit adds a small, bounded cost but shrinks the footprint enough to run the $1.2$-billion-parameter model on an $8$\,GB Pi. Against the official implementation, aria matches or exceeds generation speed and starts about seven times faster. A case study of the steering interface generates music carrying taste associations (\emph{sonic seasoning}), with genuine but bounded control for a subset of attributes. These results make a compact, quantized runtime with built-in control a practical basis for on-device semantic audio in Internet-of-Sounds settings. The \textit{aria} runtime is released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08526
PDF
-
Progression as Latent Drift: Generative Forecasting of Slow-Evolving Pathologies
2026-07-09 09:13:03Yuxiang Feng, Juncheng Wang, Chao Xu, Wenlong Hou, Huihan Wang, Yijie Qian, Yang Liu, Baigui Sun, Yong Liu, Shujun WanAbstract
Forecasting the future anatomy of slow-evolving neurodegenerative diseases could enable earlier, more targeted intervention and improve clinical trial design, but it remains challenging because true progression signals are subtle in longitudinal MRI. In this low-signal regime, transferring modern generative sequence models directly is unreliable: training is dominated by stable baseline anatomy and confounded by dense, sample-specific nuisance variation. We first provide a theoretical analysis that explains these failures through two modes. Identity collapse occurs when optimization is driven toward reproducing the current anatomy, which prevents the model from learning faint temporal change. The continuous interpolation trap arises when standard smooth networks cannot separate localized biological drift from pervasive noise, which leads to spurious changes that diffuse across the volume. To address both issues, we propose Latent Drift, a progressive generative framework that learns change in a compressed semantic representation rather than synthesizing full-resolution anatomy. This design removes pixel-level identity from the prediction target and concentrates model capacity on progression-relevant dynamics. We further apply Finite Scalar Quantization to the learned change representation, which suppresses small, high-frequency nuisance fluctuations while preserving consistent structural drift. Experiments on longitudinal 3D brain MRI show that Latent Drift improves patient-specific neuro-forecasting over diffusion and autoregressive transformer baselines across generative fidelity and clinically relevant evaluation metrics. Project page: \href{this https URL}{this https URL}.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08270
PDF
-
Closing the Null Space: Guidance-Aware Quantization for Classifier-Free Diffusion
2026-07-09 08:42:07Abdullah Al Shafi, Sumaiya Rahim SumaAbstract
Deploying classifier-free guidance (CFG) diffusion models under real-world compute budgets requires quantization, yet existing post-training quantization (PTQ) methods treat CFG models as single-branch networks, ignoring the paired conditional/unconditional structure that CFG inference fundamentally relies on. This structural blind spot has two consequences. At the system level, the two-pass CFG execution pattern imposes a latency overhead that parameter-count and bit-operation metrics conceal entirely, and commodity INT8 inference stacks fail to realize the theoretical efficiency gains that BOPs calculations promise. At the algorithmic level, calibrating against the guidance gap alone admits an exact null space: a quantized model can achieve perfect gap-fidelity diagnostics while the unconditional branch drifts arbitrarily, corrupting every guided prediction at inference time. This paper terms this the branch-drift trap, proves its existence analytically, and confirms it empirically through a false-positive result in which the best-calibrated model by standard diagnostics simultaneously produces the worst sample quality. To close the trap, Guidance-Aware Mixed Precision (GAMP) is proposed, which calibrates directly on the guided prediction, derives per-layer activation-bit sensitivity from guided-output degradation, and allocates bits via a greedy knapsack -- provably preventing unconditional branch drift by construction.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08241
PDF
-
Latency-Constrained DNN Architecture Learning for Edge Systems using Zerorized Batch Normalization
2026-07-08 02:35:25Shuo Huai, Di Liu, Hao Kong, Weichen Liu, Ravi Subramaniam, Christian Makaya, Qian LinAbstract
Deep learning applications have been widely adopted on edge devices, to mitigate the privacy and latency issues of accessing cloud servers. Deciding the number of neurons during the design of a deep neural network to maximize performance is not intuitive. Particularly, many application scenarios are real-time and have a strict latency constraint, while conventional neural network optimization methods do not directly change the temporal cost of model inference for latency-critical edge systems. In this work, we propose a latency-oriented neural network learning method to optimize models for high accuracy while fulfilling the latency constraint. For efficiency, we also introduce a universal hardware-customized latency predictor to optimize this procedure to learn a model that satisfies the latency constraint by only a one-shot training process. The experiment results reveal that, compared to state-of-the-art methods, our approach can well-fit the 'hard' latency constraint and achieve high accuracy. Under the same training settings as the original model and satisfying a 34 ms latency constraint on the ImageNet-100 dataset, we reduce GoogLeNet's latency from 40.32 ms to 34 ms with a 0.14% accuracy reduction on the NVIDIA Jetson Nano. When coupled with quantization, our method can be further improved to only 0.04% drop for GoogLeNet. On the NVIDIA Jetson TX2, we compress VGG-19 from 119.98 ms to 34 ms and even improve its accuracy by 0.5%, and we scale GoogLeNet up from 20.27 ms to 34 ms and achieve higher accuracy by 0.78%. We also open source this framework at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2607.06922
PDF
-
Hardware-aware Graph Neural Networks prunning for embedded event-based vision
2026-07-07 19:09:03Piotr Wzorek, Kamil Jeziorek, Tomasz KryjakAbstract
Event-based cameras are gaining popularity as the sensor of choice for mobile robotics, due to their high performance in dynamic environments. However, these applications require efficient real-time data processing with low latency and power consumption. One strategy to meet these stringent requirements is hardware acceleration of efficient algorithms that preserve the temporal sparsity of event data. In this work, we propose an optimization strategy for Graph Convolutional Neural Networks models aimed at adapting their architecture to the limited resources of embedded heterogeneous FPGA platforms. Our method incorporates hardware-aware pruning and quantization, taking into account the trade-off between on-chip memory savings and inference accuracy. Strategic exploration of the design space with Fine Grid Search and Greedy layer-wise Iterative Deepening Search methods enables flexible adaptation of the model architecture to the target platform. Our approach was evaluated across various network configurations and multiple datasets, resulting in BRAM memory reductions of 28.8% for CIFAR-10 (with a 1.65% decrease in accuracy), 31.4% for MNIST-DVS (accuracy drop of 3.55%), and 26.5% for N-Caltech101 (with a 5.18% accuracy reduction).
Abstract (translated)
URL
https://arxiv.org/abs/2607.06739
PDF
-
DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression
2026-07-07 17:29:01Anna Cordoba, Adam Puente Tercero, Nerea Angulo Hijo, Mar Linares Tercero, Julia Barrientos, Ainhoa Miranda, Jesus OliveraAbstract
Long-context language model inference is increasingly limited by the memory bandwidth and capacity required to store key-value caches, yet existing compression methods often apply uniform budgets across layers or tokens and degrade retrieval when lexical cues and semantic states require different preservation. We introduce DepthWeave-KV, a token-adaptive cache compression method that factorizes key and value states across neighboring transformer layers using shared low-rank channel bases while retaining lightweight token-specific residuals where attention behavior is sensitive. DepthWeave-KV combines cross-depth residual factorization with a token-conditional depth router that allocates higher reconstruction rank to instruction-bearing and retrieval-critical tokens, and uses calibration-free online error tracking from attention-output probes to adapt compression during generation without retraining the base model. A fused CUDA implementation jointly performs basis lookup, residual dequantization, and attention projection to reduce decode-time memory traffic. Across LongBench, Needle-in-a-Haystack, L-Eval, and long-form QA and summarization benchmarks, DepthWeave-KV achieves near-full-cache task quality with substantially lower memory use, improving average score and retrieval accuracy over prior compressed caches while reaching 8.3x KV memory reduction and 72.8 tokens per second at 64K context.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06523
PDF
-
From Textural Counterpoint to Feature Encoding: A Multi-Dimensional Machine Representation Study of Haydn's 'The Lark' Integrating Electroacoustic Analysis
2026-07-07 06:58:46Yakun Liu, Zhiyu Jin, Hai Luan, Dong Liu, Xiaonan LiAbstract
Chamber music, as a highly precise multi-part interactive system, contains a logic of "role assignment and dynamic interaction" that provides an extremely valuable blueprint for exploring human-computer collaborative composition paradigms. Addressing the lack of role perception capabilities in existing deep music generation models during polyphonic interactions, this paper conducts an interdisciplinary analysis of Haydn's String Quartet in D Major, The Lark (Op. 64, No. 5). We propose a novel research path: "Classical Morphology Qualitative Analysis-Electroacoustic Quantitative Measurement-Machine Representation Reconstruction." The study first utilizes auditory analysis to dissect the counterpoint morphology of the leading voice and the underlying groove in the first movement. Subsequently, it introduces spectrum and dynamic feature analysis tools from a Digital Audio Workstation (DAW) to translate subjective auditory perception into objective, measurable physical parameters. Building on this, the paper introduces a fundamentally new approach to low-level computer feature extraction: completely abandoning the traditional mechanical quantization grid, introducing Event-based Timestamps to record the duration of micro-timing, and transforming acoustic features into an independent "Role-Aware Encoding" as an aesthetic heuristic mechanism (a phenomenological anchor). This study not only completes the logical loop spanning classical analysis, electronic music mapping, and AI symbolic generation but also establishes a profound theoretical foundation-from the perspectives of interactive aesthetics and media philosophy-for constructing human-computer collaborative music systems imbued with "social attributes" and "otherness awareness."
Abstract (translated)
URL
https://arxiv.org/abs/2607.05902
PDF
-
FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models
2026-07-07 00:34:30Bowen Xue, Zihan Min, Xingyang Li, Zhekai Zhang, Haocheng Xi, Lvmin Zhang, Maneesh Agrawala, Jun-Yan Zhu, Song Han, Yujun Lin, Muyang LiAbstract
Diffusion models have become a dominant paradigm for high-quality generative modeling, while post-training is essential for adapting them to diverse downstream applications. However, post-training of large diffusion models is still challenging due to the prohibitive memory footprints and slow training speed, which existing parameter-efficient fine-tuning methods only partially address. To overcome these limitations, we propose FourTune, an efficient post-training framework for diffusion models based on an end-to-end W4A4G4 paradigm. FourTune introduces a triple-branch hybrid pipeline that augments the standard LoRA architecture with a frozen numerical stabilizer to isolate quantization-sensitive outliers, enabling stable training under native 4-bit computation. In addition, FourTune employs hardware-efficient block-wise quantization and customized fused kernels to support efficient quantized backpropagation and reduce memory bandwidth overhead. Across customization, reinforcement learning, and distillation tasks, FourTune matches the quality of full-precision fine-tuning. On FLUX.1-dev (12B), FourTune reduces memory overhead by 2.25$\times$ and increases end-to-end training throughput by 2.27$\times$ compared to BF16 LoRA.
Abstract (translated)
URL
https://arxiv.org/abs/2607.05711
PDF
-
TriRoute: Unified Learned Routing for Joint Adaptive Attention, Experts, and KV-Cache Allocation
2026-07-07 00:12:46Andrii Balashov, Olena PonomarovaAbstract
Conditional computation can decouple language model quality from per-token inference cost, yet leading techniques act on a single axis in isolation: Mixture-of-Experts (MoE) sparsifies the FFN, Mixture-of-Depths (MoD) skips whole transformer blocks, and KV-cache quantization compresses attention memory. We argue these three decisions (attention resolution, expert selection, and cache bit-width) are strongly coupled and should be made jointly: a token rare enough to warrant full attention may also need high-precision caching regardless of which expert processes it. We introduce TriRoute, a single lightweight controller shared across all three axes that, for every token at every layer, emits a coordinated policy: (i) an attention mode (skip/local/full), (ii) a sparse set of FFN experts (with a null expert recovering MoD), and (iii) a KV-cache bit-width. The controller trains end-to-end via a heterogeneous relaxation (Gumbel-Softmax with straight-through estimation for categorical decisions and load-balanced top-k gating for experts) under a Lagrangian budget constraint that turns the average compute and memory cost into a controllable knob. We identify a cross-axis routing-collapse cascade in naive joint training, where collapse on one axis propagates to the others, and address it with per-axis normalization and a coupling-aware balancing loss. On decoder-only models from 160M to 1.3B parameters at compute-optimal token counts, TriRoute Pareto-dominates the best independent MoD+MoE+KV-quantization combination at matched inference FLOPs and memory, while better preserving tail-case robustness on rare entities, code, and arithmetic that pure perplexity optimization erodes. Post-hoc analysis reveals interpretable structure: the controller allocates full attention and high-precision cache to sentence-initial positions, rare subwords, and named entities, while cheaply routing function words.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06601
PDF
-
MiLSD: A Micro Line-Segment Detector for Resource-Constrained Devices
2026-07-07 00:04:50Parsa Hassani Shariat Panahi, Amir Hossein Jalilvand, M. Hassan NajafiAbstract
Line segment detection is a key building block in visual SLAM, 3D reconstruction, and industrial inspection. Recent deep learning methods have greatly improved accuracy, yet even the smallest models require several megabytes of memory, exceeding low-cost MCU capacity. This work investigates the maximum achievable accuracy under a sub-megabyte budget. We propose MiLSD, a detector tailored for MCU-level constraints, and systematically compare three output representations within a compact fully-convolutional backbone. Our study shows that the proposed F-Clip center-with-length-and-angle formulation learns most effectively at small model sizes. We find that 8-bit quantization preserves full-precision performance, while 4-bit quantization causes significant degradation, particularly in angle regression, with quantization-aware training recovering only part of the loss. With a one-megabyte activation budget and inference enhancements including sub-pixel decoding, test-time augmentation, and a lightweight verifier, MiLSD improves sAP10 on ShanghaiTech Wireframe from 10.6 (25k parameters, 0.25 MB) to 24.1 within 1 MB. Rather than competing with GPU-scale parsers, we map the accuracy memory trade-off across representations, bit-widths, capacities, and post-processing strategies for embedded vision systems.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06600
PDF
-
Clustered Codebook Quantization for 2D Gaussian-based Image Compression
2026-07-06 22:16:37Runze Cheng, Yicheng Zhan, Josef Spjut, Kaan Ak\c{s}itAbstract
Gaussian-based image representations effectively model image content using compact parametric primitives while preserving high visual fidelity, yet storing a large number of floating-point parameters per primitive degrades rate-distortion efficiency at higher fidelity targets. To improve the rate-distortion performance in Gaussian representation, we present our Cluster-Guided Vector Quantization (CGVQ), a Gaussian primitive based image compression method. Our key idea is to partition Gaussian parameters further into homogeneous groups prior to quantization, enabling higher compression efficiency and accurate parameter reconstruction. In practice, our extensive experiments show that CGVQ decreases the bpp by 20% with respect to our baseline, while maintaining on-par visual quality
Abstract (translated)
URL
https://arxiv.org/abs/2607.05667
PDF
-
Is Your NPU Ready for LLMs? Dissecting the Hidden Efficiency Bottlenecks in Mobile LLM Inference
2026-07-06 09:42:34Guanyu Cai, Ruiming Tian, Lang Yang, Zhouhong Ren, Jinliang Yuan, Lingkun Li, Jiliang WangAbstract
Deploying Large Language Models (LLMs) on mobile devices enhances privacy and reduces latency, but is severely bottlenecked by hardware inefficiency. We present the first comprehensive, cross-layer measurement study of mobile LLM inference, uniquely spanning five mainstream frameworks (e.g., this http URL, GENIE) and three hardware backends (CPU, GPU, NPU). To enable this analysis, we develop PowerBench, a fine-grained profiling tool that provides the first backend-specific energy attribution, moving beyond traditional device-level measurements. Our study yields three critical insights: (1) Framework-induced performance gaps are substantially amplified on NPUs, reaching up to 10x using custom operators due to divergent offloading and quantization strategies. (2) We identify a distinct phase split where NPUs excel at compute-bound prefilling, while CPUs outperform all other backends in memory-bound decoding. This is driven by the NPU's preference for large, fixed-shape workloads, which conflicts with the small-kernel, dynamic nature of decoding. (3) Backend-specific profiling uncovers substantial scheduling headroom missed by prior work. Suboptimal thread configurations, uncoordinated NPU sleep latencies, and CPU polling intervals result in up to 40% energy waste. Leveraging these findings, we present an energy-oriented best-practice configuration for mobile LLM inference. We estimate that this configuration could reduce energy consumption by up to 54.8% on the NPU backend across three datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2607.05475
PDF
-
Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models
Abstract
We study how correctness is assembled inside aligned language models, not only whether the final answer is right. Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip: in mid layers (25-90% depth), internal preference transiently commits to the incorrect answer and is rescued only by late-layer correction. We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B). Four findings follow. (1) Alignment amplification of the causal wrong-dip is recipe-specific and emergent: it emerges at 3B in Qwen2.5, remains high, and peaks at 32B (paired t up to 9.7), reverses in Llama-3-8B (t=-2.31), and sits between for Mistral-7B. (2) The dip predicts real compression failures: high-dip items are 3-7x more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind, a double dissociation confirmed by late-layer ablation. (3) The dip is trainable: a LoRA fine-tune with a mid-layer wrong-margin penalty matches output-only SFT accuracy while cutting the causal dip by 67-70% and improving compression robustness; output-only SFT worsens the causal dip by up to 2.8x at perfect surface accuracy. (4) With controlled readouts, the phenomenon survives natural-language I/O: dip stratification of structural-damage failures is significant on naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer. Together, output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.
Abstract (translated)
URL
https://arxiv.org/abs/2607.04640
PDF
-
Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention
2026-07-05 17:31:36Siyu Ding, Mingchuan Ma, Jiabo Tong, Xingrun Xing, Ziming Wang, Guoqi LiAbstract
Recent NVFP4 pretraining methods mainly target transformer linear layers, leaving optimizer states, optimizer arithmetic and attention underexplored in 4-bit pipelines. This critical gap blocks stable full-stack 4-bit pretraining, as the three core modules exhibit unique numerical failure patterns: linear layers hit hard quantization noise limits with dimension-propagated error amplification; AdamW second moments are heavy-tailed non-negative values fragile to low-precision denominators; attention carries error-prone computation paths demanding strict forward-backward quantization consistency. We propose Full-Stack FP4, the first complete NVFP4 pretraining framework resolving all three stability bottlenecks via module-wise precision strategies. For linear projections, LoRA-SVD lightweight decomposition suppresses quantization noise and breaks the direct-quantization error ceiling, shrinking the linear-only loss gap from 1.40% to 0.61%. For optimizers, we design AdamW second-moment transformation for robust NVFP4 storage and fully support native NVFP4 Newton-Schulz iterations for the Root (Muon) optimizer. For attention, a mixed-precision scheme quantizes Q/K/V and backward dS while guarding vulnerable paths in BF16, paired with unified tensor reuse to sustain forward-backward alignment. We further analyze fast error accumulation in naive low-bit matrix multiplication and the extreme sensitivity of PV / dOV^T attention branches. All modules are plug-and-play with cumulative stability and efficiency improvements. Our 3B/64B-token pretraining validates near-BF16 performance with merely 1.47% loss gap, verifying feasible stable end-to-end NVFP4 LLM pretraining.
Abstract (translated)
URL
https://arxiv.org/abs/2607.04422
PDF
-
Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs
2026-07-05 16:00:29Akhiad Bercovich, Talor Abramovich, Daniel Afrimi, Shay Aharon, Nir Ailon, Vladimir Anisimov, Omer Ullman Argov, Maor Ashkenazi, Tomer Asida, Nave Assaf, Tomer Bar Natan, Alexander Bukharin, Grzegorz Chlebus, Marcin Chochowski, Eric Chung, Mohammad Dabbah, Carlo del Mundo, Ewa Dobrowolska, Ido Galil, Yaniv Galron, Amnon Geifman, Yonatan Geifman, Izik Golan, Alex Gronskiy, Tomasz Grzegorzek, Netanel Haber, Lior Kadoch, Grzegorz Karch, Tomer Keren, Abhinav Khattar, Amir Klein, Tugrul Konuk, Roi Koren, Daniel Korzekwa, Shaun Kotek, Konstantinos Krommydas, Itay Levy, Ofri Masad, Yoav Miron, Pavlo Molchanov, Shahar Mor, Zach Moshe, Saurav Muralidharan, Najeeb Nabwani, Besmira Nushi, Mostofa Patwary, Omri Puny, Johannes Rausch, Tomer Ronen, Sepehr Sameni, Itamar Schen, Elad Segal, Daniel Serebrenik, Ido Shahaf, Soumye Singhal, Daniil Sorokin, Sharath Turuvekere Sreenivas, Marta Stepniewska-Dziubinska, Ali Taghibakhshi, Nima Tajbakhsh, Oren Tropp, Dor Tzur, Anna Warno, Yi-Fu Wu, Michal Zawalski, Jiaqi Zeng, Yian Zhang, Ran Zilberstein, Amit Zuker, Ran El-YanivAbstract
We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints. In interactive serving workloads on a single 8xB200 node, Puzzle-75B-A9B achieves approximately 2x higher server throughput than Nemotron-3-Super at matched user throughput constraints. In ultra-long-context deployment on a single H100 GPU, the compressed model increases 1M-token concurrency from 1 request to 8 requests. Puzzle-75B-A9B is constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and a Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. We evaluate Puzzle-75B-A9B on a broad suite of reasoning, coding, multilingual, long-context, and agentic benchmarks. Despite substantial compression, the model retains strong downstream accuracy relative to the parent model across a wide range of tasks. These results demonstrate that large hybrid MoE models can be substantially optimized for deployment efficiency while maintaining strong downstream capability.
Abstract (translated)
URL
https://arxiv.org/abs/2607.04371
PDF
-
HiFA4: Training-Free 4-bit FlashAttention on Ascend HIF4 NPUs for LLM Inference
2026-07-05 13:39:52Hui Dong, Yanzhao Li, Jie Gao, Chunlu Li, Zhiyuan Zhang, Yupeng Sun, Zhenyuan Chen, Zhiqiang ZouAbstract
We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16. To our knowledge, HiFA4 is the first Ascend-HIF4-targeted design of this kind evaluated on standard NLP benchmarks. HiFA4 combines two mechanisms. Smooth-QK applies a calibration-static per-channel equivalent rescaling to Q and K after RoPE, transferring quantization difficulty from K to Q without per-tile online reduction at inference. P-Reordering accumulates the softmax normalizer from the same quantized attention weights P_hat used in the PV GEMM, rather than from a higher-precision reconstruction. We show that this inconsistent formulation introduces a coherent output-scaling error, and validate the effect on a Qwen3-8B Layer-0 MMLU trace, where all 3.6M measured attention tiles exhibit net probability-mass loss with median epsilon_bar = -0.064. P-Reordering also allows the normalizer to be fused into the PV Cube GEMM. Across five LLMs, HiFA4 reduces quantization-induced decision drift. On Qwen3-8B, it recovers 37.5% of the accuracy gap introduced by direct HIF4 quantization, narrows the sample-weighted accuracy loss from 1.12 pp to 0.70 pp, reduces BF16-inconsistent MMLU predictions from 16.3% to 8.2%, and cuts MMLU accuracy regressions by 57% (1071 to 465). On Gemma2-9B, mild smoothing keeps HiFA4 within 0.7 pp of BF16 while reducing MMLU regressions by 27%. On LLaMA3.1-8B, Mistral-7B, and Phi-4B, where Smooth-QK is disabled, P-Reordering with the adopted Q-Mean auxiliary still reduces full-set MMLU regressions by 41-52%. A preliminary instruction-scheduling analysis projects a 35.4% critical-path latency reduction relative to BF16 by fusing the softmax normalizer into the PV Cube GEMM; on-hardware validation is left to future work.
Abstract (translated)
URL
https://arxiv.org/abs/2607.04302
PDF
-
TokAN: Accent Normalization Using Self-Supervised Speech Tokens
2026-07-04 15:45:38Qibing Bai, Shuai Wang, Yuhan Du, Bohan Li, Yannan Wang, Haizhou LiAbstract
Accent normalization (AN) seeks to convert non-native (L2) accented speech into standard (L1) speech while preserving speaker identity. The current techniques either require naturally recorded parallel L1-L2 speech for training, or suffer from quality degradation when supervised by synthesized targets. In this paper, we present TokAN, a token-based accent normalization framework that operates on self-supervised discrete speech tokens extracted from a L1-L2 jointly trained vector-quantization (VQ) tokenizer, without the need of synthetic supervisory speech. An autoregressive encoder-decoder model performs token-to-token conversion, translating L2-accented token sequences into the tokens of standard voice. We also introduce reinforcement learning (RL) post-training based on Group Relative Policy Optimization (GRPO), using word error rate and accent classifier confidence as complementary rewards. A non-autoregressive flow-matching synthesizer recovers the Mel-spectrogram from the converted tokens, conditioned on the source speaker embedding. We also develop a flow-matching duration predictor that supports total-duration-aware synthesis, making TokAN applicable to duration-critical tasks such as voice dubbing and live casting. Experiments on seven English accents demonstrate that TokAN reduced the word error rate from 12.40% to 9.89% after supervised fine-tuning, and further to 9.23% after RL post-training, consistently outperforming frame-to-frame, direct flow-matching, and prompt-based token-conversion baselines in terms of accent reduction and intelligibility.
Abstract (translated)
URL
https://arxiv.org/abs/2607.03928
PDF
-
CineMobile: On-Device Image-to-Video Diffusion for Cinematic Camera Motion Generation
2026-07-04 10:24:35Xuyao Huang, Zelai Deng, Xu Wang, Xizhong Xiao, Zhijie DengAbstract
The growing demand for image-to-video creation on mobile devices has increasingly focused on cinematic motion effects like bullet time, dolly zoom, slow motion, etc. While Diffusion Transformers (DiTs) exhibit strong performance in video generation, their large parameter sizes and multi-step iterative denoising processes lead to substantial computational overhead, making efficient generation on mobile devices challenging. We propose CineMobile to bridge the gap. In particular, CineMobile adopts a three-fold optimization strategy: (1) leveraging a distillation-guided pruning approach to derive a compact yet efficient model that retains the essential video generation capabilities required for cinematic effects; (2) optimizing the compressed model into a 4-step generator via a combination of diffusion distillation and reinforcement learning; (3) employing a hybrid post-training quantization strategy to compress the model footprint to under 1 GB. Experimental results show that compared to the teacher model with the Wan 2.1 architecture, CineMobile achieves a 40x speedup in generation while maintaining comparable visual quality. Specifically, CineMobile generates 49-frame 480p videos with a per-step denoising latency of 0.6s on an NVIDIA H200 GPU and 20s on the MediaTek Dimensity 8400 Ultimate 5G platform, with a peak memory usage of 1.8 GB, demonstrating its practical applicability for mobile-based image-to-video creation.
Abstract (translated)
URL
https://arxiv.org/abs/2607.03803
PDF
