Tencent Hunyuan's HPC-Ops operator library has contributed two production-hardened kernels to vLLM's main branch as first-class backends. The Attention backend replaces fixed split-KV scheduling with a per-step load-balanced dynamic scheduler plus a fused RoPE+QK-Norm+KV-write prologue, achieving up to 2.95× speedup over static scheduling on mixed-length decode batches and 2.25× on average over FlashInfer and FlashAttention. The MoE backend fuses the entire MoE pipeline (routing, Gate-Up GEMM, activation/quantization, Down GEMM, top-k reduction) into a compact FP8 execution path using occupancy-first design and PDL-chained stages, delivering 1.59× speedup at TP8/EP1 and 1.21× at TP1/EP8 over Triton and CUTLASS. End-to-end on the Hy3 295B MoE model across 8× H20 GPUs, the combined backends cut TTFT by ~24% and TPOT by ~17% versus vLLM defaults. Both backends integrate via standard vLLM backend interfaces with no source modifications required, and are optimized for NVIDIA Hopper architecture (especially H20).
