Hu Liu

Mehran Taghian, Yunke Peng, Xing Huang et al.

Large foundation models have become central to modern machine learning, with performance scaling predictably with model size and data. However, training and deploying such models incur substantial computational and memory costs, motivating the development of low-precision training techniques. Recent work has demonstrated that 4-bit floating-point (FP4) formats--such as MXFP4 and NVFP4--can be successfully applied to linear GEMM operations in large language models (LLMs), achieving up to 4x improvements in compute throughput and memory efficiency compared to higher-precision baselines. In this work, we investigate the recently proposed HiFloat4 FP4 format for Huawei Ascend NPUs and systematically compare it with MXFP4 in large-scale training settings. All experiments are conducted on Ascend NPU clusters, with linear and expert GEMM operations performed entirely in FP4 precision. We evaluate both dense architectures (e.g., Pangu and LLaMA-style models) and mixture-of-experts (MoE) models, where both standard linear layers and expert-specific GEMMs operate in FP4. Furthermore, we explore stabilization techniques tailored to FP4 training that significantly reduce numerical degradation, maintaining relative error within 1% of full-precision baselines while preserving the efficiency benefits of 4-bit computation. Our results provide a comprehensive empirical study of FP4 training on NPUs and highlight the practical trade-offs between FP4 formats in large-scale dense and MoE models.

Yuanyong Luo, Zhongxing Zhang, Richard Wu et al.

This preliminary white paper proposes a novel 8-bit floating-point data format HiFloat8 (abbreviated as HiF8) for deep learning. HiF8 features tapered precision. For normal value encoding, it provides 7 exponent values with 3-bit mantissa, 8 exponent values with 2-bit mantissa, and 16 exponent values with 1-bit mantissa. For denormal value encoding, it extends the dynamic range by 7 extra powers of 2, from 31 to 38 binades (notice that FP16 covers 40 binades). Meanwhile, HiF8 encodes all the special values except that positive zero and negative zero are represented by only one bit-pattern. Thanks to the better balance between precision and dynamic range, HiF8 can be simultaneously used in both forward and backward passes of AI training. In this paper, we will describe the definition and rounding methods of HiF8, as well as the tentative training and inference solutions. To demonstrate the efficacy of HiF8, massive simulation results on various neural networks, including traditional neural networks and large language models (LLMs), will also be presented.

Yuguang Liu, Luyao Xia, Hu Liu et al.

Consumption Drives Production (CDP) on social platforms aims to deliver interpretable incentive signals for creator ecosystem building and resource utilization improvement, which strongly relies on attribution. In large-scale and complex recommendation systems, the absence of accurate labels together with unobserved confounding renders backdoor adjustments alone insufficient for reliable attribution. To address these problems, we propose Adversarial Learning Mediator based Multi-Touch Attribution (ALM-MTA), an extensible causal framework that leverages front-door identification with an adversarially learned mediator: a proxy trained to distill outcome information to strengthen the causal pathway from treatment to outcome and eliminate shortcut leakage. We then introduce contrastive learning that conditions front-door marginalization on high-match consumption-upload pairs to ensure positivity in large treatment spaces. To assess causality from non-RCT logs, we also incorporate a non-personalized bucketed protocol, estimating grouped uplift and computing AUUC over treatment clusters. Finally, we evaluate ALM-MTA using a real-world recommendation system with 400 million DAU and 30 billion samples. ALM-MTA increases DAU by 0.04% and daily active creators by 0.6%, with unit exposure efficiency increased by 670%. On causal utility, ALM-MTA achieves higher grouped AUUC than the SOTA in every propensity bucket, with a maximum gain of 0.070. In terms of accuracy, ALM-MTA improves upload AUC by 40% compared to SOTA. These results demonstrate that front-door deconfounding with adversarial mediator learning provides accurate, personalized, and operationally efficient attribution for creator ecosystem optimization.

Pengfei Zuo, Huimin Lin, Junbo Deng et al.

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

Zisheng Ye, Xiaoyu He, Maoyuan Song et al.

As the performance gains from accelerating quantized matrix multiplication plateau, the softmax operation becomes the critical bottleneck in Transformer inference. This bottleneck stems from two hardware limitations: (1) limited data bandwidth between matrix and vector compute cores, and (2) the significant area cost of high-precision (FP32/16) exponentiation units (EXP2). To address these issues, we introduce a novel low-precision workflow that employs a specific 8-bit floating-point format (HiF8) and block-aware precision rescaling for softmax. Crucially, our algorithmic innovations make low-precision softmax feasible without the significant model accuracy loss that hampers direct low-precision approaches. Specifically, our design (i) halves the required data movement bandwidth by enabling matrix multiplication outputs constrained to 8-bit, and (ii) substantially reduces the EXP2 unit area by computing exponentiations in low (8-bit) precision. Extensive evaluation on language models and multi-modal models confirms the validity of our method. By alleviating the vector computation bottleneck, our work paves the way for doubling end-to-end inference throughput without increasing chip area, and offers a concrete co-design path for future low-precision hardware and software.

Yuanyong Luo, Jing Huang, Yu Cheng et al.

This paper introduces HiFloat4 (HiF4), a block floating-point data format tailored for deep learning. Each HiF4 unit packs 64 4-bit elements with 32 bits of shared scaling metadata, averaging 4.5 bits per value. The metadata specifies a three-level scaling hierarchy, capturing inter- and intra-group dynamic range while improving the utilization of the representational space. In addition, the large 64-element group size enables matrix multiplications to be executed in a highly fixed-point manner, significantly reducing hardware area and power consumption. To evaluate the proposed format, we conducted inference experiments on several language models, including LLaMA, Qwen, Mistral, DeepSeek-V3.1 and LongCat. Results show that HiF4 achieves higher average accuracy than the state-of-the-art NVFP4 format across multiple models and diverse downstream tasks.

Shunyu Zhang, Hu Liu, Wentian Bao et al.

Scale-calibrated ranking systems are ubiquitous in real-world applications nowadays, which pursue accurate ranking quality and calibrated probabilistic predictions simultaneously. For instance, in the advertising ranking system, the predicted click-through rate (CTR) is utilized for ranking and required to be calibrated for the downstream cost-per-click ads bidding. Recently, multi-objective based methods have been wildly adopted as a standard approach for Calibrated Ranking, which incorporates the combination of two loss functions: a pointwise loss that focuses on calibrated absolute values and a ranking loss that emphasizes relative orderings. However, when applied to industrial online applications, existing multi-objective CR approaches still suffer from two crucial limitations. First, previous methods need to aggregate the full candidate list within a single mini-batch to compute the ranking loss. Such aggregation strategy violates extensive data shuffling which has long been proven beneficial for preventing overfitting, and thus degrades the training effectiveness. Second, existing multi-objective methods apply the two inherently conflicting loss functions on a single probabilistic prediction, which results in a sub-optimal trade-off between calibration and ranking. To tackle the two limitations, we propose a Self-Boosted framework for Calibrated Ranking (SBCR).

Yaofang Liu, Xinyue Zhang, Wenlong Wan et al.

Coronary angiography is the "gold standard" for diagnosing coronary artery disease (CAD). At present, the methods for detecting and evaluating coronary artery stenosis cannot satisfy the clinical needs, e.g., there is no prior study of detecting stenoses in prespecified vessel segments, which is necessary in clinical practice. Two vascular stenosis detection methods are proposed to assist the diagnosis. The first one is an automatic method, which can automatically extract the entire coronary artery tree and mark all the possible stenoses. The second one is an interactive method. With this method, the user can choose any vessel segment to do further analysis of its stenoses. Experiments show that the proposed methods are robust for angiograms with various vessel structures. The precision, sensitivity, and $F_1$ score of the automatic stenosis detection method are 0.821, 0.757, and 0.788, respectively. Further investigation proves that the interactive method can provide a more precise outcome of stenosis detection, and our quantitative analysis is closer to reality. The proposed automatic method and interactive method are effective and can complement each other in clinical practice. The first method can be used for preliminary screening, and the second method can be used for further quantitative analysis. We believe the proposed solution is more suitable for the clinical diagnosis of CAD.

Yaofang Liu, Xinyue Zhang, Wenlong Wan et al.

Coronary angiography is the "gold standard" for diagnosing coronary artery disease (CAD). At present, the methods for detecting and evaluating coronary artery stenosis cannot satisfy the clinical needs, e.g., there is no prior study of detecting stenoses in prespecified vessel segments, which is necessary in clinical practice. Two vascular stenosis detection methods are proposed to assist the diagnosis. The first one is an automatic method, which can automatically extract the entire coronary artery tree and mark all the possible stenoses. The second one is an interactive method. With this method, the user can choose any vessel segment to do further analysis of its stenoses. Experiments show that the proposed methods are robust for angiograms with various vessel structures. The precision, sensitivity, and $F_1$ score of the automatic stenosis detection method are 0.821, 0.757, and 0.788, respectively. Further investigation proves that the interactive method can provide a more precise outcome of stenosis detection, and our quantitative analysis is closer to reality. The proposed automatic method and interactive method are effective and can complement each other in clinical practice. The first method can be used for preliminary screening, and the second method can be used for further quantitative analysis. We believe the proposed solution is more suitable for the clinical diagnosis of CAD.

Hu Liu, Jing Lu, Xiwei Zhao et al.

Click-through rate (CTR) prediction is one of the fundamental tasks for e-commerce search engines. As search becomes more personalized, it is necessary to capture the user interest from rich behavior data. Existing user behavior modeling algorithms develop different attention mechanisms to emphasize query-relevant behaviors and suppress irrelevant ones. Despite being extensively studied, these attentions still suffer from two limitations. First, conventional attentions mostly limit the attention field only to a single user's behaviors, which is not suitable in e-commerce where users often hunt for new demands that are irrelevant to any historical behaviors. Second, these attentions are usually biased towards frequent behaviors, which is unreasonable since high frequency does not necessarily indicate great importance. To tackle the two limitations, we propose a novel attention mechanism, termed Kalman Filtering Attention (KFAtt), that considers the weighted pooling in attention as a maximum a posteriori (MAP) estimation. By incorporating a priori, KFAtt resorts to global statistics when few user behaviors are relevant. Moreover, a frequency capping mechanism is incorporated to correct the bias towards frequent behaviors. Offline experiments on both benchmark and a 10 billion scale real production dataset, together with an Online A/B test, show that KFAtt outperforms all compared state-of-the-arts. KFAtt has been deployed in the ranking system of a leading e commerce website, serving the main traffic of hundreds of millions of active users everyday.

Hu Liu, Jing Lu, Hao Yang et al.

As one of the largest B2C e-commerce platforms in China, JD com also powers a leading advertising system, serving millions of advertisers with fingertip connection to hundreds of millions of customers. In our system, as well as most e-commerce scenarios, ads are displayed with images.This makes visual-aware Click Through Rate (CTR) prediction of crucial importance to both business effectiveness and user experience. Existing algorithms usually extract visual features using off-the-shelf Convolutional Neural Networks (CNNs) and late fuse the visual and non-visual features for the finally predicted CTR. Despite being extensively studied, this field still face two key challenges. First, although encouraging progress has been made in offline studies, applying CNNs in real systems remains non-trivial, due to the strict requirements for efficient end-to-end training and low-latency online serving. Second, the off-the-shelf CNNs and late fusion architectures are suboptimal. Specifically, off-the-shelf CNNs were designed for classification thus never take categories as input features. While in e-commerce, categories are precisely labeled and contain abundant visual priors that will help the visual modeling. Unaware of the ad category, these CNNs may extract some unnecessary category-unrelated features, wasting CNN's limited expression ability. To overcome the two challenges, we propose Category-specific CNN (CSCNN) specially for CTR prediction. CSCNN early incorporates the category knowledge with a light-weighted attention-module on each convolutional layer. This enables CSCNN to extract expressive category-specific visual patterns that benefit the CTR prediction. Offline experiments on benchmark and a 10 billion scale real production dataset from JD, together with an Online A/B test show that CSCNN outperforms all compared state-of-the-art algorithms.