#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
#include "op_plugin/utils/custom_functions/aclops/inner_compute.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
c10::SmallVector<int64_t, SIZE> get_rstd_shape(const at::Tensor &self, const at::Tensor &gamma)
{
c10::SmallVector<int64_t, SIZE> ret;
auto rstd_dim = self.dim() - gamma.dim();
for (int64_t i = 0; i < self.dim(); i++) {
if (i < rstd_dim) {
ret.emplace_back(self.size(i));
} else {
ret.emplace_back(1);
}
}
return ret;
}
}
std::tuple<at::Tensor, at::Tensor> npu_rms_norm(const at::Tensor &self, const at::Tensor &gamma, double epsilon)
{
TORCH_CHECK(self.dim() >= gamma.dim(), "The gamma shape should not be bigger than self shape." + OPS_ERROR(ErrCode::PARAM));
at::Tensor y = npu_preparation::apply_tensor(self.sizes(), self.options(), self);
auto rstd_shape = get_rstd_shape(self, gamma);
at::Tensor rstd = npu_preparation::apply_tensor(rstd_shape, self.options().dtype(at::kFloat), self);
at_npu::native::OpCommand cmd;
cmd.Name("RmsNorm")
.Input(self, "x")
.Input(gamma, "gamma")
.Output(y, "y")
.Output(rstd, "rstd")
.Attr("epsilon", static_cast<float>(epsilon))
.Run();
return std::make_tuple(y, rstd);
}
}