Yu Luo

Haihao Shen, Hengyu Meng, Bo Dong et al. · mit

In recent years, Transformer-based language models have become the standard approach for natural language processing tasks. However, stringent throughput and latency requirements in industrial applications are limiting their adoption. To mitigate the gap, model compression techniques such as structured pruning are being used to improve inference efficiency. However, most existing neural network inference runtimes lack adequate support for structured sparsity. In this paper, we propose an efficient sparse deep learning inference software stack for Transformer-based language models where the weights are pruned with constant block size. Our sparse software accelerator leverages Intel Deep Learning Boost to maximize the performance of sparse matrix - dense matrix multiplication (commonly abbreviated as SpMM) on CPUs. Our SpMM kernel outperforms the existing sparse libraries (oneMKL, TVM, and LIBXSMM) by an order of magnitude on a wide range of GEMM shapes under 5 representative sparsity ratios (70%, 75%, 80%, 85%, 90%). Moreover, our SpMM kernel shows up to 5x speedup over dense GEMM kernel of oneDNN, a well-optimized dense library widely used in industry. We apply our sparse accelerator on widely-used Transformer-based language models including Bert-Mini, DistilBERT, Bert-Base, and BERT-Large. Our sparse inference software shows up to 1.5x speedup over Neural Magic's Deepsparse under same configurations on Xeon on Amazon Web Services under proxy production latency constraints. We also compare our solution with two framework-based inference solutions, ONNX Runtime and PyTorch, and demonstrate up to 37x speedup over ONNX Runtime and 345x over PyTorch on Xeon under the latency constraints. All the source code is publicly available on Github: https://github.com/intel/intel-extension-for-transformers.

Haihao Shen, Hanwen Chang, Bo Dong et al.

Large language models (LLMs) have demonstrated remarkable performance and tremendous potential across a wide range of tasks. However, deploying these models has been challenging due to the astronomical amount of model parameters, which requires a demand for large memory capacity and high memory bandwidth. In this paper, we propose an effective approach that can make the deployment of LLMs more efficiently. We support an automatic INT4 weight-only quantization flow and design a special LLM runtime with highly-optimized kernels to accelerate the LLM inference on CPUs. We demonstrate the general applicability of our approach on popular LLMs including Llama2, Llama, GPT-NeoX, and showcase the extreme inference efficiency on CPUs. The code is publicly available at: https://github.com/intel/intel-extension-for-transformers.

Xiangshuo Qiao, Xianxin Li, Xiaozhe Qu et al.

Vision-Language Models pre-trained on large-scale image-text datasets have shown superior performance in downstream tasks such as image retrieval. Most of the images for pre-training are presented in the form of open domain common-sense visual elements. Differently, video covers in short video search scenarios are presented as user-originated contents that provide important visual summaries of videos. In addition, a portion of the video covers come with manually designed cover texts that provide semantic complements. In order to fill in the gaps in short video cover data, we establish the first large-scale cover-text benchmark for Chinese short video search scenarios. Specifically, we release two large-scale datasets CBVS-5M/10M to provide short video covers, and the manual fine-labeling dataset CBVS-20K to provide real user queries, which serves as an image-text benchmark test in the Chinese short video search field. To integrate the semantics of cover text in the case of modality missing, we propose UniCLIP where cover texts play a guiding role during training, however are not relied upon by inference. Extensive evaluation on CBVS-20K demonstrates the excellent performance of our proposal. UniCLIP has been deployed to Tencent's online video search systems with hundreds of millions of visits and achieved significant gains. The dataset and code are available at https://github.com/QQBrowserVideoSearch/CBVS-UniCLIP.

Hanting Chen, Yasheng Wang, Kai Han et al.

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.

Yue Cai, Yu Luo, Jie Ling et al.

Magnetic Resonance Imaging (MRI) is a widely used imaging technique, however it has the limitation of long scanning time. Though previous model-based and learning-based MRI reconstruction methods have shown promising performance, most of them have not fully utilized the edge prior of MR images, and there is still much room for improvement. In this paper, we build a joint edge optimization model that not only incorporates individual regularizers specific to both the MR image and the edges, but also enforces a co-regularizer to effectively establish a stronger correlation between them. Specifically, the edge information is defined through a non-edge probability map to guide the image reconstruction during the optimization process. Meanwhile, the regularizers pertaining to images and edges are incorporated into a deep unfolding network to automatically learn their respective inherent a-priori information.Numerical experiments, consisting of multi-coil and single-coil MRI data with different sampling schemes at a variety of sampling factors, demonstrate that the proposed method outperforms other compared methods.