#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/OpAdapter.h"
#include "op_plugin/utils/op_api_common.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
c10::SmallVector<int64_t, SIZE> quantize_reshape_size(
const at::Tensor& self,
int64_t axis)
{
c10::SmallVector<int64_t, SIZE> out_size;
for (int64_t i = 0; i < self.dim(); i++) {
if (i != axis) {
out_size.emplace_back(1);
} else {
out_size.emplace_back(self.sizes()[i]);
}
}
return out_size;
}
at::Tensor& npu_quantize_out_nocheck(
at::Tensor& result,
const at::Tensor& self,
const at::Tensor& scales,
const at::Tensor& zero_points,
int64_t axis,
int64_t dtype)
{
auto reshape_size = quantize_reshape_size(self, axis);
at::Tensor scales_reshape = scales.reshape(reshape_size);
at::Tensor zp_reshape = zero_points.defined() ? zero_points.reshape(reshape_size) : (at::Tensor());
string dtype_str = "torch.qint8";
if (dtype == static_cast<int64_t>(at::ScalarType::QUInt8)) {
dtype_str = "torch.quint8";
} else if (dtype == static_cast<int64_t>(at::ScalarType::QInt32)) {
dtype_str = "torch.qint32";
}
at_npu::native::OpCommand cmd;
cmd.Name("Quantize")
.Input(self)
.Input(scales_reshape);
if (zero_points.defined()) {
cmd.Input(zp_reshape);
} else {
cmd.Input();
}
cmd.Output(result)
.Attr("axis", axis)
.Attr("dtype", dtype_str)
.Run();
return result;
}
at::Tensor& npu_ascend_quant_v2(
at::Tensor& result,
const at::Tensor& self,
const at::Tensor& scales,
const at::Tensor& zero_points,
at::ScalarType dtype)
{
at_npu::native::OpCommand cmd;
cmd.Name("AscendQuantV2")
.Input(self)
.Input(scales)
.Input(zero_points)
.Output(result)
.Attr("sqrt_mode", false)
.Attr("round_mode", "round")
.Attr("dst_type", dtype)
.Run();
return result;
}
}
at::Tensor npu_quantize(
const at::Tensor& self,
const at::Tensor& scales,
const c10::optional<at::Tensor>& zero_points_opt,
int64_t dtype,
int64_t axis,
bool div_mode)
{
const at::Tensor& zero_points = c10::value_or_else(zero_points_opt, [] { return at::Tensor(); });
if (!div_mode) {
return op_api::npu_quantize(self, scales, zero_points_opt, dtype, axis, div_mode);
}
axis = op_plugin::utils::make_warp_dim(axis, self.dim());
TORCH_CHECK(scales.dim() == 1, "Scales' dim should be equal to 1." + OPS_ERROR(ErrCode::PARAM));
if (zero_points.defined()) {
TORCH_CHECK(
zero_points.dim() == 1,
"Zero points' dim should be equal to 1." + OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(
scales.sizes()[0] == zero_points.sizes()[0],
"Scales' size should be equal to zero points' size." + OPS_ERROR(ErrCode::PARAM));
}
TORCH_CHECK(
axis >= 0 && axis <= self.sizes().size() - 1,
"Axis should be in range [-rank(x), rank(x) - 1]." + OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(
scales.sizes()[0] == self.sizes()[axis],
"Length of scales must equal to the specified dimension." + OPS_ERROR(ErrCode::PARAM));
auto output_dtype = at::kInt;
if (dtype == static_cast<int64_t>(at::ScalarType::QInt8)) {
output_dtype = at::kChar;
} else if (dtype == static_cast<int64_t>(at::ScalarType::QUInt8)) {
output_dtype = at::kByte;
} else if (dtype == static_cast<int64_t>(at::ScalarType::QInt32)) {
output_dtype = at::kInt;
}
at::Tensor result = npu_preparation::apply_tensor(self, self.options().dtype(output_dtype));
npu_quantize_out_nocheck(result, self, scales, zero_points, axis, dtype);
return result;
}
}