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.

CopyQNN: Quantum Neural Network Extraction Attack under Varying Quantum Noise

Quantum Neural Networks (QNNs) have shown significant value across domains, with well-trained QNNs representing critical intellectual property often deployed via cloud-based QNN-as-a-Service (QNNaaS) platforms. Recent work has examined QNN model extraction attacks using classical and emerging quantum strategies. These attacks involve adversaries querying QNNaaS platforms to obtain labeled data for training local substitute QNNs that replicate the functionality of cloud-based models. However, existing approaches have largely overlooked the impact of varying quantum noise inherent in noisy intermediate-scale quantum (NISQ) computers, limiting their effectiveness in real-world settings. To address this limitation, we propose the CopyQNN framework, which employs a three-step data cleaning method to eliminate noisy data based on its noise sensitivity. This is followed by the integration of contrastive and transfer learning within the quantum domain, enabling efficient training of substitute QNNs using a limited but cleaned set of queried data. Experimental results on NISQ computers demonstrate that a practical implementation of CopyQNN significantly outperforms state-of-the-art QNN extraction attacks, achieving an average performance improvement of 8.73% across all tasks while reducing the number of required queries by 90x, with only a modest increase in hardware overhead.

FCoT-VL:Advancing Text-oriented Large Vision-Language Models with Efficient Visual Token Compression

The rapid success of Vision Large Language Models (VLLMs) often depends on the high-resolution images with abundant visual tokens, which hinders training and deployment efficiency. Current training-free visual token compression methods exhibit serious performance degradation in tasks involving high-resolution, text-oriented image understanding and reasoning. In this paper, we propose an efficient visual token compression framework for text-oriented VLLMs in high-resolution scenarios. In particular, we employ a light-weight self-distillation pre-training stage to compress the visual tokens, requiring a limited numbers of image-text pairs and minimal learnable parameters. Afterwards, to mitigate potential performance degradation of token-compressed models, we construct a high-quality post-train stage. To validate the effectiveness of our method, we apply it to an advanced VLLMs, InternVL2. Experimental results show that our approach significantly reduces computational overhead while outperforming the baselines across a range of text-oriented benchmarks. We will release the models and code soon.

Enhancing Web Service Anomaly Detection via Fine-grained Multi-modal Association and Frequency Domain Analysis

Anomaly detection is crucial for ensuring the stability and reliability of web service systems. Logs and metrics contain multiple information that can reflect the system's operational state and potential anomalies. Thus, existing anomaly detection methods use logs and metrics to detect web service systems' anomalies through data fusion approaches. They associate logs and metrics using coarse-grained time window alignment and capture the normal patterns of system operation through reconstruction. However, these methods have two issues that limit their performance in anomaly detection. First, due to asynchrony between logs and metrics, coarse-grained time window alignment cannot achieve a precise association between the two modalities. Second, reconstruction-based methods suffer from severe overgeneralization problems, resulting in anomalies being accurately reconstructed. In this paper, we propose a novel anomaly detection method named FFAD to address these two issues. On the one hand, FFAD employs graph-based alignment to mine and extract associations between the modalities from the constructed log-metric relation graph, achieving precise associations between logs and metrics. On the other hand, we improve the model's fit to normal data distributions through Fourier Frequency Focus, thereby enhancing the effectiveness of anomaly detection. We validated the effectiveness of our model on two real-world industrial datasets and one open-source dataset. The results show that our method achieves an average anomaly detection F1-score of 93.6%, representing an 8.8% improvement over previous state-of-the-art methods.

Meta-Reflection: A Feedback-Free Reflection Learning Framework

Despite the remarkable capabilities of large language models (LLMs) in natural language understanding and reasoning, they often display undesirable behaviors, such as generating hallucinations and unfaithful reasoning. A prevalent strategy to mitigate these issues is the use of reflection, which refines responses through an iterative process. However, while promising, reflection heavily relies on high-quality external feedback and requires iterative multi-agent inference processes, thus hindering its practical application. In this paper, we propose Meta-Reflection, a novel feedback-free reflection mechanism that necessitates only a single inference pass without external feedback. Motivated by the human ability to remember and retrieve reflections from past experiences when encountering similar problems, Meta-Reflection integrates reflective insights into a codebook, allowing the historical insights to be stored, retrieved, and used to guide LLMs in problem-solving. To thoroughly investigate and evaluate the practicality of Meta-Reflection in real-world scenarios, we introduce an industrial e-commerce benchmark named E-commerce Customer Intent Detection (ECID). Extensive experiments conducted on both public datasets and the ECID benchmark highlight the effectiveness and efficiency of our proposed approach.

Empowering 1000 tokens/second on-device LLM prefilling with mllm-NPU

On-device large language models (LLMs) are catalyzing novel mobile applications such as UI task automation and personalized email auto-reply, without giving away users' private data. However, on-device LLMs still suffer from unacceptably long inference latency, especially the time to first token (prefill stage) due to the need of long context for accurate, personalized content generation, as well as the lack of parallel computing capacity of mobile CPU/GPU. To enable practical on-device LLM, we present mllm-NPU, the first-of-its-kind LLM inference system that efficiently leverages on-device Neural Processing Unit (NPU) offloading. Essentially, mllm-NPU is an algorithm-system co-design that tackles a few semantic gaps between the LLM architecture and contemporary NPU design. Specifically, it re-constructs the prompt and model in three levels: (1) At prompt level, it divides variable-length prompts into multiple fixed-sized chunks while maintaining data dependencies; (2) At tensor level, it identifies and extracts significant outliers to run on the CPU/GPU in parallel with minimal overhead; (3) At block level, it schedules Transformer blocks in an out-of-order manner to the CPU/GPU and NPU based on their hardware affinity and sensitivity to accuracy. Compared to competitive baselines, mllm-NPU achieves 22.4x faster prefill speed and 30.7x energy savings on average, and up to 32.8x speedup in an end-to-end real-world application. For the first time, mllm-NPU achieves more than 1,000 tokens/sec prefilling for a billion-sized model (Qwen1.5-1.8B), paving the way towards practical on-device LLM.

LLM-Powered Test Case Generation for Detecting Tricky Bugs

Conventional automated test generation tools struggle to generate test oracles and tricky bug-revealing test inputs. Large Language Models (LLMs) can be prompted to produce test inputs and oracles for a program directly, but the precision of the tests can be very low for complex scenarios (only 6.3% based on our experiments). To fill this gap, this paper proposes AID, which combines LLMs with differential testing to generate fault-revealing test inputs and oracles targeting plausibly correct programs (i.e., programs that have passed all the existing tests). In particular, AID selects test inputs that yield diverse outputs on a set of program variants generated by LLMs, then constructs the test oracle based on the outputs. We evaluate AID on two large-scale datasets with tricky bugs: TrickyBugs and EvalPlus, and compare it with three state-of-the-art baselines. The evaluation results show that the recall, precision, and F1 score of AID outperform the state-of-the-art by up to 1.80x, 2.65x, and 1.66x, respectively.

QuantumLeak: Stealing Quantum Neural Networks from Cloud-based NISQ Machines

Variational quantum circuits (VQCs) have become a powerful tool for implementing Quantum Neural Networks (QNNs), addressing a wide range of complex problems. Well-trained VQCs serve as valuable intellectual assets hosted on cloud-based Noisy Intermediate Scale Quantum (NISQ) computers, making them susceptible to malicious VQC stealing attacks. However, traditional model extraction techniques designed for classical machine learning models encounter challenges when applied to NISQ computers due to significant noise in current devices. In this paper, we introduce QuantumLeak, an effective and accurate QNN model extraction technique from cloud-based NISQ machines. Compared to existing classical model stealing techniques, QuantumLeak improves local VQC accuracy by 4.99\%$\sim$7.35\% across diverse datasets and VQC architectures.

Masked Collaborative Contrast for Weakly Supervised Semantic Segmentation

This study introduces an efficacious approach, Masked Collaborative Contrast (MCC), to emphasize semantic regions in weakly supervised semantic segmentation. MCC adroitly incorporates concepts from masked image modeling and contrastive learning to devise Transformer blocks that induce keys to contract towards semantically pertinent regions. Unlike prevalent techniques that directly eradicate patch regions in the input image when generating masks, we scrutinize the neighborhood relations of patch tokens by exploring masks considering keys on the affinity matrix. Moreover, we generate positive and negative samples in contrastive learning by utilizing the masked local output and contrasting it with the global output. Elaborate experiments on commonly employed datasets evidences that the proposed MCC mechanism effectively aligns global and local perspectives within the image, attaining impressive performance. The source code is available at \url{https://github.com/fwu11/MCC}.

Fast Distributed Inference Serving for Large Language Models

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1$\times$ and 6.4$\times$, respectively.