PyTorch 用 Helion 为 vLLM 构建高性能可移植线性后端
PyTorch 将 Helion 集成进 vLLM 的线性后端，用单个 Helion GEMM 实现覆盖 Standard GEMM、Split-K 和 Swap-AB 等变体，并通过按 shape 自动调优选择最佳配置。在 NVIDIA Hopper GPU 上，该后端结合按 shape 调优与混合调度，在评测模型中超过 vLLM 默认的 CUTLASS 和 DeepGEMM 后端，部分工作负载吞吐提升超过 10%。
otuning to end users. This favors performance and maintainability over out-of-the-box usability, but Helion’s automated tuning framework makes the additional optimization step practical for performance-sensitive deployments. Acknowledgments This work was supported by many contributors across the OCTO and vLLM teams at Red Hat, as well as the Helion team at Meta. In particular, we would like to thank our colleagues: Richard Zou and Jongsok Choi for their feedback and support throughout this work.
匹配字段正文
