Nvshmem4py nvshmem.core.tensor does not support dtype torch.uint64, which is wired

Accelerated Computing GPU-Accelerated Libraries
1

nvshmem signal is of dtype uint64, but nvshmem4py does not support torch.uint64, which is wiered.

Signaling Operations — NVSHMEM 3.3.0 documentation

image
image1560×812 111 KB

nvshmem4py torch extension does not support torch.uint64

image
image1948×1450 241 KB

2

Hi, thank you for the report.

Torch integers above uint8_t only have limited support so we left them out of this release. nvshmem.core.put_signal() takes an object of type cuda.core.Buffer allocatable with nvshmem.core.buffer() and of size 8 bytes. We also support CuPy arrays of type uint64 if that suits your needs.

We can investigate torch uint64 for the next release.

3

that helps