Tencent Hunyuan’s HPC-Ops operator library contributed two high-performance kernels to vLLM targeting Hopper GPUs. The attention backend uses dynamic load-balanced scheduling with fused RoPE/QK-Norm operations, reporting up to 2.95x speedup over a static split-KV schedule on mixed-length decode workloads, while the MoE backend provides a fully fused FP8 pipeline that removes intermediate materialization, running 1.59x faster than the best baseline at TP8/EP1. End-to-end tests on Hy3 models across eight H20 GPUs show roughly 24% lower time-to-first-token and 17% better time-per-output-token versus vLLM’s default backends.
