vLLM Integrates Tencent Hunyuan Production Kernels, Boosting Hopper Inference
vllm_project · x · 2026-07-06
The vLLM project announced that serving the Hunyuan3 (Hy3) model on NVIDIA Hopper is now significantly optimized. The HPC-Ops attention and MoE backend, originally used by Tencent's Hunyuan AI Infra team in their production environment, has become a first-class citizen backend in the vLLM main branch.
The solution includes a step-wise load-balancing decode scheduler and a fully fused FP8 MoE pipeline. It achieves up to a 2.95x speedup on mixed-length decoding compared to static split-KV scheduling, and reduces Hy3's TTFT by about 24% and TPOT by roughly 17%. It can be used without forking or modifying the source code.
More from Infra
- Anthropic's S-1 Reveals Over $600B in Infrastructure Commitments Across Google, Amazon, Microsoft — cedric_chee · 2026-10-26
- Anthropic S-1 reveals 48 cents of compute cost per dollar of Q2 revenue — nmasc_ · 2026-10-26
- Persistent AI joins UR network to route private AI inference with encrypted, IP-hidden traffic — 0xSammy · 2026-10-26
- Anthropic's disclosed compute commitments: $114.5B Google, $161.2B Broadcom top the list — BenBajarin · 2026-10-26
- SpaceX strikes $40B deal with Apollo Global to lease AI chips and data center hardware — Polymarket · 2026-10-26
- SpaceX AI revenue up 830%, Claude designs 318K antivenom proteins in 64 hours — PeterDiamandis · 2026-10-26