DeepSpeed

7e284fb5 - fix: replace tensor.is_cuda with a cuda-build guard in the cache update

fix: replace tensor.is_cuda with a cuda-build guard in the cache update The torchcuda hook flags .is_cuda; on_accelerator() has different semantics on the CPU accelerator (CPU tensors are 'on accelerator'), which un-guarded the capture check on CPU-only builds. Gate on torch.version.cuda instead. Signed-off-by: Guokai Ma <guokai.ma@intel.com>