arXiv subjects

Yonghan Dong

Publications and source records attributed to Yonghan Dong.

7 recordsLinked to original sources

arXiv2026

Activation Outliers Matter: Robust Recovery for Quantized Multimodal LLMs

Low-bit quantization offers a promising avenue for reducing the computational and memory demands of Multimodal Large Language Models (MLLMs). Recent hardware support for low-precision formats, ranging from MXFP8 to ultra-low-bit formats such as MXFP4 and HiF4, has accelerated research into efficient MLLM training and deployment. In this work, we present a systematic study of these quantization schemes in representative MLLMs that span both video generation and reasoning tasks. Our analysis shows that MXFP8 achieves near-lossless performance, whereas aggressive 4-bit quantization leads to significant degradation. Through extensive ablations, we identify activation quantization as the primary source of this performance loss, contributing substantially more than weight quantization. Motivated by this observation, we propose Residual Fallback Quantization (RFQ), a lightweight activation reconstruction framework that supplements the primary ulta-low-bit activation representation with an auxiliary quantized residual pathway. By explicitly modeling and compensating for quantization errors, RFQ improves activation fidelity while preserving the efficiency advantages of ultra-low-bit computation. RFQ requires no architectural modifications and incurs negligible computational overhead. Extensive experiments on Wan2.2 and Qwen3-VL demonstrate that RFQ consistently recovers a substantial portion of the performance lost under the quantization of MXFP4 and HiF4, significantly narrowing the gap to BF16 baselines across both generation and 4 reasoning benchmarks. Our findings establish activation quantization as the dominant bottleneck in ultra-low-bit MLLMs and highlight residual-based activation reconstruction as an effective and practical strategy for robust 4-bit deployment.

cs.LG↗

arXiv2026

GNMR: Runtime Stability Control for Low-Precision Large Language Model Training

Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $Δ$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\mathrm{maxO}$ budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.

cs.LG↗

arXiv2025

VA-MoE: Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting

This paper presents Variables Adaptive Mixture of Experts (VAMoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VAMoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct subpatterns of atmospheric variables (temperature, humidity, wind speed). Moreover, the proposed method employs a variable adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on real world ERA5 dataset demonstrate that VAMoE performs comparable against SoTA models in both short term (1 days) and long term (5 days) forecasting tasks, with only about 25% of trainable parameters and 50% of the initial training data.

cs.LG↗

arXiv2025

Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

cs.CL↗

arXiv2025

Pangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUs

We present Pangu Ultra, a Large Language Model (LLM) with 135 billion parameters and dense Transformer modules trained on Ascend Neural Processing Units (NPUs). Although the field of LLM has been witnessing unprecedented advances in pushing the scale and capability of LLM in recent years, training such a large-scale model still involves significant optimization and system challenges. To stabilize the training process, we propose depth-scaled sandwich normalization, which effectively eliminates loss spikes during the training process of deep models. We pre-train our model on 13.2 trillion diverse and high-quality tokens and further enhance its reasoning capabilities during post-training. To perform such large-scale training efficiently, we utilize 8,192 Ascend NPUs with a series of system optimizations. Evaluations on multiple diverse benchmarks indicate that Pangu Ultra significantly advances the state-of-the-art capabilities of dense LLMs such as Llama 405B and Mistral Large 2, and even achieves competitive results with DeepSeek-R1, whose sparse model structure contains much more parameters. Our exploration demonstrates that Ascend NPUs are capable of efficiently and effectively training dense models with more than 100 billion parameters. Our model and system will be available for our commercial customers.

cs.CL↗

arXiv2023

Dense Affinity Matching for Few-Shot Segmentation

Few-Shot Segmentation (FSS) aims to segment the novel class images with a few annotated samples. In this paper, we propose a dense affinity matching (DAM) framework to exploit the support-query interaction by densely capturing both the pixel-to-pixel and pixel-to-patch relations in each support-query pair with the bidirectional 3D convolutions. Different from the existing methods that remove the support background, we design a hysteretic spatial filtering module (HSFM) to filter the background-related query features and retain the foreground-related query features with the assistance of the support background, which is beneficial for eliminating interference objects in the query background. We comprehensively evaluate our DAM on ten benchmarks under cross-category, cross-dataset, and cross-domain FSS tasks. Experimental results demonstrate that DAM performs very competitively under different settings with only 0.68M parameters, especially under cross-domain FSS tasks, showing its effectiveness and efficiency.

cs.CV↗

arXiv2023

Multi-Content Interaction Network for Few-Shot Segmentation

Few-Shot Segmentation (FSS) is challenging for limited support images and large intra-class appearance discrepancies. Most existing approaches focus on extracting high-level representations of the same layers for support-query correlations, neglecting the shift issue between different layers and scales, due to the huge difference between support and query samples. In this paper, we propose a Multi-Content Interaction Network (MCINet) to remedy this issue by fully exploiting and interacting with the multi-scale contextual information contained in the support-query pairs to supplement the same-layer correlations. Specifically, MCINet improves FSS from the perspectives of boosting the query representations by incorporating the low-level structural information from another query branch into the high-level semantic features, enhancing the support-query correlations by exploiting both the same-layer and adjacent-layer features, and refining the predicted results by a multi-scale mask prediction strategy, with which the different scale contents have bidirectionally interacted. Experiments on two benchmarks demonstrate that our approach reaches SOTA performances and outperforms the best competitors with many desirable advantages, especially on the challenging COCO dataset.

cs.CV↗