Open source of AI Infra core technology: theoretical ingestion up 30%

HPC-Ops, an official open source production-grade high-performance LLLM core algorithm, based on CUDA and Cue's zero construction, the hybrid model reasoning 30%, the DeepSeek model 17%; in single-calculations, Attention increased by a maximum of 2.22 times the single-calculator performance, and GroupGEM increased by a maximum of 1.88 times the DeepGEM and FusedMoE increased by a maximum of 1.49 times the value of TensorRT-LLLM; it optimized the mainstream reasoning in the country and addressed the high cost and non-matching of target hardware in the existing mainstream account。