Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

This work presents Pangu Embedded, an efficient Large Language Model (LLM) reasoner developed on Ascend Neural Processing Units (NPUs), featuring flexible fast and slow thinking capabilities. Pangu Embedded addresses the significant computational costs and inference latency challenges prevalent in existing reasoning-optimized LLMs. We propose a two-stage training framework for its construction. In Stage 1, the model is finetuned via an iterative distillation process, incorporating inter-iteration model merging to effectively aggregate complementary knowledge. This is followed by reinforcement learning on Ascend clusters, optimized by a latency-tolerant scheduler that combines stale synchronous parallelism with prioritized data queues. The RL process is guided by a Multi-source Adaptive Reward System (MARS), which generates dynamic, task-specific reward signals using deterministic metrics and lightweight LLM evaluators for mathematics, coding, and general problem-solving tasks. Stage 2 introduces a dual-system framework, endowing Pangu Embedded with a "fast" mode for routine queries and a deeper "slow" mode for complex inference. This framework offers both manual mode switching for user control and an automatic, complexity-aware mode selection mechanism that dynamically allocates computational resources to balance latency and reasoning depth. Experimental results on benchmarks including AIME 2024, GPQA, and LiveCodeBench demonstrate that Pangu Embedded with 7B parameters, outperforms similar-size models like Qwen3-8B and GLM4-9B. It delivers rapid responses and state-of-the-art reasoning quality within a single, unified model architecture, highlighting a promising direction for developing powerful yet practically deployable LLM reasoners.

Pangu Pro MoE: Mixture of Grouped Experts for Efficient Sparsity

The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.

SlimLLM: Accurate Structured Pruning for Large Language Models

Large language models(LLMs) have garnered significant attention and demonstrated impressive capabilities in a wide range of applications. However, due to their enormous computational costs, the deployment and application of LLMs are often severely limited. To address this issue, structured pruning is an effective solution to compress the parameters of LLMs. Determining the importance of each sub-module in LLMs and minimizing performance loss are critical issues that need to be carefully addressed in structured pruning. In this paper, we propose an effective and fast structured pruning method named SlimLLM for large language models. For channel and attention head pruning, we evaluate the importance based on the entire channel or head, rather than merely aggregating the importance of individual elements within a sub-module. This approach enables a more holistic consideration of the interdependence among elements within the sub-module. In addition, we design a simple linear regression strategy for the output matrix to quickly recover performance. We also propose layer-based importance ratio to determine the pruning ratio for each layer. Based on the LLaMA benchmark results, our SlimLLM outperforms other methods and achieves state-of-the-art performance.

Pangu Light: Weight Re-Initialization for Pruning and Accelerating LLMs

Large Language Models (LLMs) deliver state-of-the-art capabilities across numerous tasks, but their immense size and inference costs pose significant computational challenges for practical deployment. While structured pruning offers a promising avenue for model compression, existing methods often struggle with the detrimental effects of aggressive, simultaneous width and depth reductions, leading to substantial performance degradation. This paper argues that a critical, often overlooked, aspect in making such aggressive joint pruning viable is the strategic re-initialization and adjustment of remaining weights to improve the model post-pruning training accuracies. We introduce Pangu Light, a framework for LLM acceleration centered around structured pruning coupled with novel weight re-initialization techniques designed to address this ``missing piece''. Our framework systematically targets multiple axes, including model width, depth, attention heads, and RMSNorm, with its effectiveness rooted in novel re-initialization methods like Cross-Layer Attention Pruning (CLAP) and Stabilized LayerNorm Pruning (SLNP) that mitigate performance drops by providing the network a better training starting point. Further enhancing efficiency, Pangu Light incorporates specialized optimizations such as absorbing Post-RMSNorm computations and tailors its strategies to Ascend NPU characteristics. The Pangu Light models consistently exhibit a superior accuracy-efficiency trade-off, outperforming prominent baseline pruning methods like Nemotron and established LLMs like Qwen3 series. For instance, on Ascend NPUs, Pangu Light-32B's 81.6 average score and 2585 tokens/s throughput exceed Qwen3-32B's 80.9 average score and 2225 tokens/s.

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.

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.

SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs

Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.

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.

Post-Training Quantization for Diffusion Transformer via Hierarchical Timestep Grouping

Diffusion Transformer (DiT) has now become the preferred choice for building image generation models due to its great generation capability. Unlike previous convolution-based UNet models, DiT is purely composed of a stack of transformer blocks, which renders DiT excellent in scalability like large language models. However, the growing model size and multi-step sampling paradigm bring about considerable pressure on deployment and inference. In this work, we propose a post-training quantization framework tailored for Diffusion Transforms to tackle these challenges. We firstly locate that the quantization difficulty of DiT mainly originates from the time-dependent channel-specific outliers. We propose a timestep-aware shift-and-scale strategy to smooth the activation distribution to reduce the quantization error. Secondly, based on the observation that activations of adjacent timesteps have similar distributions, we utilize a hierarchical clustering scheme to divide the denoising timesteps into multiple groups. We further design a re-parameterization scheme which absorbs the quantization parameters into nearby module to avoid redundant computations. Comprehensive experiments demonstrate that out PTQ method successfully quantize the Diffusion Transformer into 8-bit weight and 8-bit activation (W8A8) with state-of-the-art FiD score. And our method can further quantize DiT model into 4-bit weight and 8-bit activation (W4A8) without sacrificing generation quality.