Tencent Hunyuan AI Infra open-sources upgraded HPC-Ops inference core operators
AI summary
Tencent Hunyuan AI Infra team has open-sourced an upgraded HPC-Ops inference operator library, introducing five core operators. Attention uses runtime dynamic load scheduling, accelerating long-text workloads by up to 2.95x and raising end-to-end QPM by 17%; Router GEMM achieves FP32 precision through a dual BF16 combination, up to 3.22x faster than CuBLAS FP32; FusedMoE delivers 1.2x–1.6x performance gains over vLLM and SGLang; Fused AllReduce+Norm is up to 1.68x faster than mainstream solutions; and Sampler fuses decoding sampling into 2 CUDA kernels, 4.0x–7.5x faster than vLLM. All capabilities come from production practice and are fully open-source.
Key Takeaways
- Tencent Hunyuan AI Infra team has open-sourced an upgraded HPC-Ops inference operator…
- Attention uses runtime dynamic load scheduling,accelerating long-text workloads by up to 2.
- 95x and raising end-to-end QPM by 17%
- Router GEMM achieves FP32 precision through a dual BF16 combination,up to 3.
- 22x faster than CuBLAS FP32
Quick Facts
- Published
- Updated
- Source
- 公众号:腾讯混元
- Category
- Products & Apps
- AI score
- 78
腾讯混元把推理全链路的瓶颈都加速了一遍,Sampler 算子比 vLLM 快 4-7 倍,Attention 动态调度根治长尾延迟,这套开源算子库可以直接用,做推理部署的可以抄作业。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Tencent Hunyuan AI Infra open-sources upgraded HPC-Ops inference core operators. June 11, 2026. https://new.nat6.net/en/news/cmq99cd9g09skslldniw7u2ub/