Which version of rocm does Deepspeed support?

I don't know if my analysis is correct
I encountered this issue during compilation

https://github.com/deepspeedai/DeepSpeed/issues/7565#issuecomment-3305379549

When I checked the source code, I found that
deepspeed/ops/csrc/deepspeed4science/evoformer_attn/gemm_kernel_utils.h:234

CUTLASS_DEVICE int32_t warp_uniform(int32_t value)
{
    return (int32_t)__shfl_sync(0xffffffff, (unsigned)value, 0);
}

/opt/rocm-6.4.1/include/hip/amd_detail/amd_warp_sync_functions.h:230

template <typename MaskT, typename T>
__device__ inline
T __shfl_sync(MaskT mask, T var, int srcLane,
              int width = warpSize) {
  static_assert(
      __hip_internal::is_integral<MaskT>::value && sizeof(MaskT) == 8,
      "The mask must be a 64-bit integer. "
      "Implicitly promoting a smaller integer is almost always an error.");
  __hip_adjust_mask_for_wave32(mask);
  __hip_check_mask(mask);
  return __shfl(var, srcLane, width);
}

That is to say, in this version, it seems that rocm compilation is bound to encounter errors because the type sizeof (MaskT)==8 is always incorrect

But it seems that no one has reported this, so I'm not quite sure now
Could it be that the linker encountered an error during compilation? Found the abnormal header file?