[Question]: Data consistency across multiple devices in P2P scenarios

[LIZAIHAN]

Question

I noticed that L1 bypass configurations (__ldcg) are not used in the code (e.g., nvshmemi_memcpy_threadgroup()). How does nvshmem ensure data consistency across multiple devices in P2P scenarios?

[zhang662817]

https://github.com/NVIDIA/nvshmem/blob/devel/src/include/non_abi/device/common/nvshmemi_common_device.cuh#L278

Does this code have a cache consistency risk?