Trident: Unifying Guarded Dispatch and Host Execution for PyTorch Triton Workloads

Jinjie Liu, Xiaoyan Liu, Shuhan Zhang +5

User-written Triton kernels enable high-performance GPU computation within PyTorch, but their end-to-end latency can remain dominated by host-side orchestration, especially when de…