Qingcheng Jizhi Chitu

0 reviews

Chitu is a high-performance inference engine designed for large-scale AI models, developed by Qingcheng.AI in collaboration with Tsinghua University. It enables enterprises to deploy large models efficiently, significantly reducing computational costs and enhancing inference speed. Chitu supports a gradual transition from small-scale experiments to large-scale deployments, offering flexible and efficient solutions that lower technical barriers and initial investments. This facilitates a smoother path for businesses toward intelligent transformation and broader application of large models in various scenarios. Key Features and Functionality: - Hardware Independence: Chitu breaks the dependency on specific hardware by supporting FP8 and FP4 precision inference on various platforms, including Ascend, Muxi, and non-H-series NVIDIA GPUs. - Enhanced Performance: Compared to foreign engines deploying the full version of DeepSeek, Chitu reduces GPU usage by 50% and accelerates inference speed by 3.15 times. The Ascend FP4 acceleration library decreases computational requirements by 75%, doubling single-card performance. - Versatile Deployment Options: Chitu is compatible with diverse hardware ecosystems, supporting pure GPU, pure CPU, and CPU+GPU hybrid deployments, as well as single GPU and large-scale cluster deployments. - Broad Model Compatibility: It supports mainstream models like Qwen and DeepSeek, accommodating various parameter sizes from 0.6B to 1T. Chitu also offers standard interfaces compatible with OpenAI HTTP and ComfyUI for image generation. Primary Value and User Solutions: Chitu addresses the critical challenge of deploying large AI models by providing a cost-effective, high-performance inference engine that is hardware-agnostic. By supporting low-precision data types on multiple domestic chips, it enhances inference efficiency and reduces computational costs. This empowers enterprises to transition smoothly from experimental phases to large-scale implementations, accelerating the adoption of AI technologies across various business applications.

Qingcheng Jizhi Bagualu

0 reviews

BaGuaLu is an advanced large-scale model training acceleration system designed to optimize the pre-training of AI models on GPU clusters. By implementing comprehensive system optimizations, BaGuaLu enhances the performance of training tasks on domestic A512 GPU clusters by an average of 30%. It also introduces a novel parallel scheme that delivers an additional 10% performance boost and improves distributed communication efficiency by 50%. Furthermore, BaGuaLu extends its capabilities to domestic supercomputers, scaling up to 100,000 servers to facilitate the accelerated pre-training of models with trillions of parameters. Key Features and Functionality: - Comprehensive System Optimization: Achieves a 30% average performance improvement in training tasks on domestic A512 GPU clusters. - Enhanced Parallel Processing: Introduces a new parallel scheme that provides an additional 10% performance enhancement. - Improved Distributed Communication: Boosts distributed communication efficiency by 50%, facilitating faster data exchange during training. - Scalability: Capable of expanding to 100,000 servers, enabling the accelerated pre-training of models with trillions of parameters. Primary Value and Problem Solved: BaGuaLu addresses the challenges associated with training large-scale AI models, which require substantial computational and memory resources. By optimizing system performance and scalability, BaGuaLu significantly reduces training times and resource consumption, making it feasible to develop and deploy models with trillions of parameters. This advancement empowers researchers and organizations to push the boundaries of AI capabilities, leading to more accurate and efficient models across various applications.