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.

DSPO: Direct Semantic Preference Optimization for Real-World Image Super-Resolution

Recent advances in diffusion models have improved Real-World Image Super-Resolution (Real-ISR), but existing methods lack human feedback integration, risking misalignment with human preference and may leading to artifacts, hallucinations and harmful content generation. To this end, we are the first to introduce human preference alignment into Real-ISR, a technique that has been successfully applied in Large Language Models and Text-to-Image tasks to effectively enhance the alignment of generated outputs with human preferences. Specifically, we introduce Direct Preference Optimization (DPO) into Real-ISR to achieve alignment, where DPO serves as a general alignment technique that directly learns from the human preference dataset. Nevertheless, unlike high-level tasks, the pixel-level reconstruction objectives of Real-ISR are difficult to reconcile with the image-level preferences of DPO, which can lead to the DPO being overly sensitive to local anomalies, leading to reduced generation quality. To resolve this dichotomy, we propose Direct Semantic Preference Optimization (DSPO) to align instance-level human preferences by incorporating semantic guidance, which is through two strategies: (a) semantic instance alignment strategy, implementing instance-level alignment to ensure fine-grained perceptual consistency, and (b) user description feedback strategy, mitigating hallucinations through semantic textual feedback on instance-level images. As a plug-and-play solution, DSPO proves highly effective in both one-step and multi-step SR frameworks.

Transferable text data distillation by trajectory matching

In the realm of large language model (LLM), as the size of large models increases, it also brings higher training costs. There is a urgent need to minimize the data size in LLM training. Compared with data selection method, the data distillation method aims to synthesize a small number of data samples to achieve the training effect of the full data set and has better flexibility. Despite its successes in computer vision, the discreteness of text data has hitherto stymied its exploration in natural language processing (NLP). In this work, we proposed a method that involves learning pseudo prompt data based on trajectory matching and finding its nearest neighbor ID to achieve cross-architecture transfer. During the distillation process, we introduce a regularization loss to improve the robustness of our distilled data. To our best knowledge, this is the first data distillation work suitable for text generation tasks such as instruction tuning. Evaluations on two benchmarks, including ARC-Easy and MMLU instruction tuning datasets, established the superiority of our distillation approach over the SOTA data selection method LESS. Furthermore, our method demonstrates a good transferability over LLM structures (i.e., OPT to Llama).

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.

Saliency-driven Dynamic Token Pruning for Large Language Models

Despite the recent success of large language models (LLMs), LLMs are particularly challenging in long-sequence inference scenarios due to the quadratic computational complexity of the attention mechanism. Inspired by the interpretability theory of feature attribution in neural network models, we observe that not all tokens have the same contribution. Based on this observation, we propose a novel token pruning framework, namely Saliency-driven Dynamic Token Pruning (SDTP), to gradually and dynamically prune redundant tokens based on the input context. Specifically, a lightweight saliency-driven prediction module is designed to estimate the importance score of each token with its hidden state, which is added to different layers of the LLM to hierarchically prune redundant tokens. Furthermore, a ranking-based optimization strategy is proposed to minimize the ranking divergence of the saliency score and the predicted importance score. Extensive experiments have shown that our framework is generalizable to various models and datasets. By hierarchically pruning 65\% of the input tokens, our method greatly reduces 33\% $\sim$ 47\% FLOPs and achieves speedup up to 1.75$\times$ during inference, while maintaining comparable performance. We further demonstrate that SDTP can be combined with KV cache compression method for further compression.

Mixture of Lookup Experts

Mixture-of-Experts (MoE) activates only a subset of experts during inference, allowing the model to maintain low inference FLOPs and latency even as the parameter count scales up. However, since MoE dynamically selects the experts, all the experts need to be loaded into VRAM. Their large parameter size still limits deployment, and offloading, which load experts into VRAM only when needed, significantly increase inference latency. To address this, we propose Mixture of Lookup Experts (MoLE), a new MoE architecture that is efficient in both communication and VRAM usage. In MoLE, the experts are Feed-Forward Networks (FFNs) during training, taking the output of the embedding layer as input. Before inference, these experts can be re-parameterized as lookup tables (LUTs) that retrieves expert outputs based on input ids, and offloaded to storage devices. Therefore, we do not need to perform expert computations during inference. Instead, we directly retrieve the expert's computation results based on input ids and load them into VRAM, and thus the resulting communication overhead is negligible. Experiments show that, with the same FLOPs and VRAM usage, MoLE achieves inference speeds comparable to dense models and significantly faster than MoE with experts offloading, while maintaining performance on par with MoE.

Unshackling Context Length: An Efficient Selective Attention Approach through Query-Key Compression

Handling long-context sequences efficiently remains a significant challenge in large language models (LLMs). Existing methods for token selection in sequence extrapolation either employ a permanent eviction strategy or select tokens by chunk, which may lead to the loss of critical information. We propose Efficient Selective Attention (ESA), a novel approach that extends context length by efficiently selecting the most critical tokens at the token level to compute attention. ESA reduces the computational complexity of token selection by compressing query and key vectors into lower-dimensional representations. We evaluate ESA on long sequence benchmarks with maximum lengths up to 256k using open-source LLMs with context lengths of 8k and 32k. ESA outperforms other selective attention methods, especially in tasks requiring the retrieval of multiple pieces of information, achieving comparable performance to full-attention extrapolation methods across various tasks, with superior results in certain tasks.

GenVidBench: A Challenging Benchmark for Detecting AI-Generated Video

The rapid advancement of video generation models has made it increasingly challenging to distinguish AI-generated videos from real ones. This issue underscores the urgent need for effective AI-generated video detectors to prevent the dissemination of false information through such videos. However, the development of high-performance generative video detectors is currently impeded by the lack of large-scale, high-quality datasets specifically designed for generative video detection. To this end, we introduce GenVidBench, a challenging AI-generated video detection dataset with several key advantages: 1) Cross Source and Cross Generator: The cross-generation source mitigates the interference of video content on the detection. The cross-generator ensures diversity in video attributes between the training and test sets, preventing them from being overly similar. 2) State-of-the-Art Video Generators: The dataset includes videos from 8 state-of-the-art AI video generators, ensuring that it covers the latest advancements in the field of video generation. 3) Rich Semantics: The videos in GenVidBench are analyzed from multiple dimensions and classified into various semantic categories based on their content. This classification ensures that the dataset is not only large but also diverse, aiding in the development of more generalized and effective detection models. We conduct a comprehensive evaluation of different advanced video generators and present a challenging setting. Additionally, we present rich experimental results including advanced video classification models as baselines. With the GenVidBench, researchers can efficiently develop and evaluate AI-generated video detection models. Datasets and code are available at https://genvidbench.github.io.

Eve: Efficient Multimodal Vision Language Models with Elastic Visual Experts

Multimodal vision language models (VLMs) have made significant progress with the support of continuously increasing model sizes and data volumes. Running VLMs on edge devices has become a challenge for their widespread application. There are several efficient VLM efforts, but they often sacrifice linguistic capabilities to enhance multimodal abilities, or require extensive training. To address this quandary,we introduce the innovative framework of Efficient Vision Language Models with Elastic Visual Experts (Eve). By strategically incorporating adaptable visual expertise at multiple stages of training, Eve strikes a balance between preserving linguistic abilities and augmenting multimodal capabilities. This balanced approach results in a versatile model with only 1.8B parameters that delivers significant improvements in both multimodal and linguistic tasks. Notably, in configurations below 3B parameters, Eve distinctly outperforms in language benchmarks and achieves state-of-the-art results 68.87% in VLM Benchmarks. Additionally, its multimodal accuracy outstrips that of the larger 7B LLaVA-1.5 model.

Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning

Large Language Models (LLMs) have shown remarkable abilities across various language tasks, but solving complex reasoning problems remains a challenge. While existing methods like Chain-of-Thought (CoT) and Tree-of-Thought (ToT) enhance reasoning by decomposing problems or structuring prompts, they typically perform a single pass of reasoning and may fail to revisit flawed paths, compromising accuracy. To address this, we propose a novel reasoning framework called Forest-of-Thought (FoT), which integrates multiple reasoning trees to leverage collective decision-making for solving complex logical problems. FoT utilizes sparse activation strategies to select the most relevant reasoning paths, improving both efficiency and accuracy. Additionally, we introduce a dynamic self-correction strategy that enables real-time error correction and learning from past mistakes, as well as consensus-guided decision making strategies to optimize correctness and computational resources. Experimental results demonstrate that the FoT framework, combined with these strategies, significantly enhances the reasoning capabilities of LLMs, enabling them to solve complex tasks with greater precision and efficiency.