#include "op_plugin/OpApiInterface.h"
#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "torch_npu/csrc/framework/utils/UtilForOpAdapter.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
const static int64_t ROTATE_HALF = 0;
const static int64_t ROTATE_INTERLEAVED = 1;
static bool isRotaryMulBackwardMixDtypeSupport(
const at::Tensor& grad,
const at::Tensor& self,
const at::Tensor& r1,
const at::Tensor& r2)
{
return self.dtype() == r1.dtype() && self.dtype() == r2.dtype() && self.dtype() == grad.dtype() ? false : true;
}
static at::Tensor npu_dtype_cast_impl_op_api(const at::Tensor& self, at::ScalarType dtype)
{
return self.dtype() == dtype ? self : self.to(dtype);
}
std::tuple<at::Tensor, at::Tensor, at::Tensor> npu_rotary_mul_backward(
const at::Tensor& grad,
const at::Tensor& self,
const at::Tensor& r1,
const at::Tensor& r2,
c10::string_view rotary_mode)
{
static bool notNeedCheck = c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend950;
TORCH_CHECK((notNeedCheck || (rotary_mode == "half" || rotary_mode == "interleave")),
"The rotary_mode of npu_rotary_mul_backward should be half or interleave, but got ", rotary_mode,
OPS_ERROR(ErrCode::PARAM));
DO_COMPATIBILITY(aclnnRotaryPositionEmbeddingGrad, acl_op::npu_rotary_mul_backward(grad, self, r1, r2, rotary_mode));
int64_t mode = op_plugin::utils::get_rotary_mode(rotary_mode);
if (c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend910B1) {
return acl_op::npu_rotary_mul_backward(grad, self, r1, r2, rotary_mode);
}
at::Tensor dx = npu_preparation::apply_tensor_without_format(grad.sizes(), self.options());
at::Tensor dcos = npu_preparation::apply_tensor_without_format(r1.sizes(), self.options());
at::Tensor dsin = npu_preparation::apply_tensor_without_format(r2.sizes(), self.options());
bool isMixDataType = isRotaryMulBackwardMixDtypeSupport(grad, self, r1, r2);
at::Tensor x = (r1.requires_grad() || r2.requires_grad()) ? self : at::Tensor();
if (isMixDataType) {
at::Tensor gradCast = npu_dtype_cast_impl_op_api(grad, self.scalar_type());
at::Tensor cosCast = npu_dtype_cast_impl_op_api(r1, self.scalar_type());
at::Tensor sinCast = npu_dtype_cast_impl_op_api(r2, self.scalar_type());
EXEC_NPU_CMD(aclnnRotaryPositionEmbeddingGrad, gradCast, cosCast, sinCast, x, mode, dx, dcos, dsin);
} else {
EXEC_NPU_CMD(aclnnRotaryPositionEmbeddingGrad, grad, r1, r2, x, mode, dx, dcos, dsin);
}
return std::tuple<at::Tensor, at::Tensor, at::Tensor>(dx, dcos, dsin);
}
}