Tencent Hunyuan AI Infra Open-Sources New Project: HPC-Ops Inference Core Operators Fully Upgraded
AI BriefingThe Tencent Hunyuan AI Infra team has open-sourced an upgraded HPC-Ops inference operator library, introducing five core operators. Attention uses dynamic runtime workload scheduling, delivering up to 2.95x acceleration on long texts and a 17% increase in end-to-end QPM; Router GEMM uses a dual-BF16 combination to achieve FP32 precision, with up to 3.22x speedup over CuBLAS FP32; FusedMoE improves performance by 1.2x–1.6x compared with vLLM and SGLang; Fused AllReduce+Norm delivers up to 1.68x speedup over mainstream solutions; and Sampler fuses decoding and sampling into 2 CUDA Kernels, achieving 4.0x–7.5x speedup over vLLM. All capabilities come from production practice and are fully open-sourced.
Source: WeChat Official Account: Tencent Hunyuan · mp.weixin.qq.com