warp.config.launch_array_access_mode

warp.config.launch_array_access_mode : warp.config.LaunchArrayAccessMode = warp.config.LaunchArrayAccessMode.RELAXED

Kernel launch array access verification mode.

LaunchArrayAccessMode.RELAXED performs no launch array access checks and is the default. LaunchArrayAccessMode.STRICT requires every Warp array argument to be allocated on the launch device, matching Warp’s original behavior. LaunchArrayAccessMode.CHECKED raises an error before launch when Warp can determine that a cross-device Warp array argument is not accessible from the launch device. When Warp cannot verify access because an array uses an unknown custom allocator or externally wrapped allocation, checked mode emits a warning once for the launch pattern and allows the launch to proceed.

Unlike verify_cuda, this setting can be used during CUDA graph capture because checks run before each launch is recorded. For cross-GPU graph capture, enable peer or memory-pool access with Warp APIs before capture begins.

Note: Strict and checked modes impact performance.