As AIDC increasingly adopt accelerators from various brands, issues such as incompatibility among heterogeneous computing resources, insufficient adaptation between model frameworks and underlying chips often arise. These issues lead to challenges in compute scheduling and hinder the effective scaling of computational resources.

Insufficient computing power scheduling in tasks like large-scale data processing and model training leads to uneven allocation of resources, wasting computing power and lowering utilization rates. Inadequate scheduling for different AI tasks affects overall AIDC service performance.

Optimization for AI tasks like data loading, training, fine-tuning, and inference of large models is often incomplete or only partially accelerated. This results in slow storage access, low computing and memory utilization rates, and inefficient communication, impacting the completion of AI tasks.

Data scientists often lack skills to utilize intelligent computing hardware, while enterprise IT personnel lack cluster management capabilities for large model training. There's a need for a flexible, user-friendly cluster environment and toolchain for fine-tuning large AI models.