pytorch/torch/csrc/utils/tensor_new.cpp

#include <torch/csrc/python_headers.h>
#include <torch/csrc/utils/tensor_new.h>

#include <pybind11/pybind11.h>
#include <torch/csrc/DynamicTypes.h>
#include <torch/csrc/Exceptions.h>
#include <torch/csrc/Size.h>
#include <torch/csrc/autograd/variable.h>
#include <torch/csrc/utils/cuda_lazy_init.h>
#include <torch/csrc/utils/numpy_stub.h>
#include <torch/csrc/utils/python_arg_parser.h>
#include <torch/csrc/utils/python_numbers.h>
#include <torch/csrc/utils/python_scalars.h>
#include <torch/csrc/utils/python_strings.h>
#include <torch/csrc/utils/tensor_numpy.h>
#include <torch/csrc/autograd/generated/variable_factories.h>

#include <ATen/ATen.h>
#include <ATen/DLConvertor.h>
#include <ATen/dlpack.h>
#include <ATen/InitialTensorOptions.h>
#include <ATen/NamedTensorUtils.h>
#include <ATen/TracerMode.h>
#include <c10/core/Backend.h>
#include <c10/core/DispatchKeySet.h>
#include <c10/core/Layout.h>
#include <c10/util/Exception.h>
#include <c10/util/irange.h>
#include <c10/util/Optional.h>

#include <stdexcept>
#include <vector>

using at::Backend;
using at::Device;
using at::IntArrayRef;
using at::kCPU;
using at::kCUDA;
using at::kLong;
using at::kInt;
using at::Scalar;
using at::ScalarType;
using at::Storage;
using at::Tensor;
using at::TensorOptions;
using at::Type;
using c10::optional;

namespace torch { namespace utils {
namespace {
const int MAX_DIMS = 128;

TensorOptions build_options(c10::TensorOptions options, at::ScalarType scalar_type, const c10::optional<Device>& device=c10::nullopt) {
  options = options.dtype(scalar_type);
  if (device.has_value()) {
    return options.device(device);
  }
  return options;
}

void maybe_initialize_cuda(const Device device) {
  if (device.is_cuda()) {
    torch::utils::cuda_lazy_init();
  }
}

// NB: It appears there is some consistency invariant between options and device, where
// if device is non-empty, its type must be consistent with the device type in
// options.
// TODO: Refactor this so we just pass everything in via options

Tensor dispatch_ones(c10::TensorOptions options, at::ScalarType scalar_type, const optional<Device>& device, IntArrayRef sizes) {
  maybe_initialize_cuda(options.device());
  pybind11::gil_scoped_release no_gil;
  return torch::ones(sizes, build_options(options, scalar_type, device));
}

Tensor new_with_sizes(c10::TensorOptions options, at::ScalarType scalar_type, const optional<Device>& device, IntArrayRef sizes) {
  maybe_initialize_cuda(options.device());
  pybind11::gil_scoped_release no_gil;
  return torch::empty(sizes, build_options(options, scalar_type, device));
}

Tensor new_with_storage(c10::TensorOptions options, at::ScalarType scalar_type, Storage storage) {
  auto tensor = at::empty({}, build_options(options, scalar_type));
  tensor.set_(std::move(storage));
  return tensor;
}

Tensor new_with_tensor(c10::TensorOptions options, at::ScalarType scalar_type, const Tensor& other) {
  options = options.dtype(scalar_type);
  TORCH_CHECK_TYPE(other.options().type_equal(options), "expected ",
                   options, " (got ", other.options(), ")");
  return other.alias();
}

std::vector<int64_t> compute_sizes(PyObject* seq, ScalarType scalar_type) {
  bool is_storage = isStorage(seq);
  std::vector<int64_t> sizes;
  THPObjectPtr handle;
  while (PySequence_Check(seq)) {
    auto length = PySequence_Length(seq);
    if (length < 0) throw python_error();
    if (is_storage) {
      length /= elementSize(scalar_type);
    }
    sizes.push_back(length);
    if (sizes.size() > MAX_DIMS) {
      throw ValueError("too many dimensions '%s'", Py_TYPE(seq)->tp_name);
    }
    if (length == 0) break;
    handle = THPObjectPtr(PySequence_GetItem(seq, 0));
    if (!handle) {
      throw ValueError("could not determine the shape of object type '%s'", Py_TYPE(seq)->tp_name);
    }
    seq = handle.get();
  }

  return sizes;
}

ScalarType infer_scalar_type(PyObject *obj) {
#ifdef USE_NUMPY
  if (is_numpy_available()) {
    if (PyArray_Check(obj)) {
      return numpy_dtype_to_aten(PyArray_TYPE((PyArrayObject*)obj));
    }
    if (PyArray_CheckScalar(obj)) {
      THPObjectPtr arr(PyArray_FromScalar(obj, nullptr));
      return numpy_dtype_to_aten(PyArray_TYPE((PyArrayObject*) arr.get()));
    }
  }
#endif
  if (PyFloat_Check(obj)) {
    // this is always guaranteed to be a floating-point type, and makes it more
    // convenient to write e.g. torch.tensor(0.) than torch.tensor(0., dtype=torch.Tensor.dtype).
    return torch::tensors::get_default_scalar_type();
  }
  if (THPUtils_checkLong(obj)) {
    return ScalarType::Long;
  }
  if (PyBool_Check(obj)) {
    return ScalarType::Bool;
  }
  if (PyComplex_Check(obj)) {
    switch (torch::tensors::get_default_scalar_type()) {
      case ScalarType::Float: return ScalarType::ComplexFloat;
      case ScalarType::Double: return ScalarType::ComplexDouble;
      default: TORCH_CHECK(false, "invalid default scalar type for complex");
    }
  }
  if (THPVariable_Check(obj)) {
    const auto& var = THPVariable_Unpack(obj);
    return var.scalar_type();
  }
  if (THPUtils_checkString(obj)) {
    throw TypeError("new(): invalid data type '%s'", Py_TYPE(obj)->tp_name);
  }
  if (PySequence_Check(obj)) {
    c10::optional<ScalarType> scalarType;
    auto length = PySequence_Length(obj);
    if (length < 0) throw python_error();
    // match NumPy semantics, except use default tensor type instead of double.
    if (length == 0) return torch::tensors::get_default_scalar_type();
    for (const auto i : c10::irange(length)) {
      THPObjectPtr handle(PySequence_GetItem(obj, i));
      if (!handle) throw python_error();
      auto cur_item = handle.get();
      if (cur_item == obj) throw TypeError("new(): self-referential lists are incompatible");
      ScalarType item_scalarType = infer_scalar_type(cur_item);
      scalarType = (scalarType) ?
          at::promoteTypes(*scalarType, item_scalarType) : item_scalarType;
      if (scalarType == ScalarType::ComplexDouble) {
        // this won't change (unless we hit undefined, but that will fail later).
        return *scalarType;
      }
    }
    return *scalarType;
  }
  AT_ERROR("Could not infer dtype of ", Py_TYPE(obj)->tp_name);
}

void recursive_store(char* data, IntArrayRef sizes, IntArrayRef strides, int64_t dim,
                            ScalarType scalarType, int elementSize, PyObject* obj) {
  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(data != nullptr);

  int64_t ndim = sizes.size();
  if (dim == ndim) {
    torch::utils::store_scalar(data, scalarType, obj);
    return;
  }

  auto n = sizes[dim];
  auto seq = THPObjectPtr(PySequence_Fast(obj, "not a sequence"));
  if (!seq) throw python_error();
  // NOLINTNEXTLINE(bugprone-branch-clone)
  auto seq_size = PySequence_Fast_GET_SIZE(seq.get());
  if (seq_size != n) {
    throw ValueError("expected sequence of length %lld at dim %lld (got %lld)",
      (long long)n, (long long)dim, (long long)seq_size);
  }

  PyObject** items = PySequence_Fast_ITEMS(seq.get());
  for(const auto i : c10::irange(n)) {
#ifdef USE_NUMPY
    if (is_numpy_available() && PyArray_Check(items[i])) {
      TORCH_WARN_ONCE(
        "Creating a tensor from a list of numpy.ndarrays is extremely slow. "
        "Please consider converting the list to a single numpy.ndarray with "
        "numpy.array() before converting to a tensor.");
    }
#endif
    recursive_store(data, sizes, strides, dim + 1, scalarType, elementSize, items[i]);
    data += strides[dim] * elementSize;
  }
}

Tensor internal_new_from_data(
    c10::TensorOptions options,
    at::ScalarType scalar_type,
    c10::optional<Device> device_opt,
    PyObject* data,
    bool copy_variables,
    bool copy_numpy,
    bool type_inference,
    bool pin_memory = false) {
  if (THPUtils_checkString(data)) {
    throw TypeError("new(): invalid data type '%s'", Py_TYPE(data)->tp_name);
  }

  if (THPVariable_Check(data)) {
    TORCH_CHECK(!pin_memory, "Can't pin tensor constructed from a variable");
    // TODO: use MaybeOwned
    auto var = THPVariable_Unpack(data);
    if (copy_variables) {
      var = var.detach();
    }
    // infer the scalar type and device type; it's not expected to infer the layout since these constructors
    // are defined per-layout-type (e.g. tensor vs sparse_coo_tensor).
    const auto& inferred_scalar_type = type_inference ? var.scalar_type() : scalar_type;
    auto device = device_opt.has_value() ? *device_opt : var.device();
    pybind11::gil_scoped_release no_gil;
    maybe_initialize_cuda(device);
    return var.to(device, inferred_scalar_type, /*non_blocking=*/false, /*copy=*/copy_variables);
  }

#ifdef USE_NUMPY
  if (PyObject_HasAttrString(data, "__cuda_array_interface__")) {
    TORCH_CHECK(!pin_memory, "Can't pin tensor constructed from __cuda_array_interface__");
    auto tensor = tensor_from_cuda_array_interface(data);
    const auto& inferred_scalar_type = type_inference ? tensor.scalar_type() : scalar_type;
    auto device = device_opt.has_value() ? *device_opt : options.device();
    pybind11::gil_scoped_release no_gil;
    maybe_initialize_cuda(device);
    return tensor.to(device, inferred_scalar_type, /*non_blocking=*/false, /*copy=*/copy_numpy);
  }

  if (is_numpy_available() && PyArray_Check(data)) {
    TORCH_CHECK(!pin_memory, "Can't pin tensor constructed from numpy");
    auto tensor = tensor_from_numpy(data, /*warn_if_not_writeable=*/!copy_numpy);
    const auto& inferred_scalar_type = type_inference ? tensor.scalar_type() : scalar_type;
    auto device = device_opt.has_value() ? *device_opt : options.device();
    pybind11::gil_scoped_release no_gil;
    maybe_initialize_cuda(device);
    return tensor.to(device, inferred_scalar_type, /*non_blocking=*/false, /*copy=*/copy_numpy);
  }
#endif

  auto device = device_opt.has_value() ? *device_opt : options.device();

  auto sizes = compute_sizes(data, scalar_type);

  ScalarType inferred_scalar_type = type_inference ? infer_scalar_type(data) : scalar_type;
  // This exists to prevent us from tracing the call to empty().  The actual
  // autograd code doesn't really matter, because requires_grad is always false
  // here.
  Tensor tensor;
  {
    at::AutoDispatchBelowADInplaceOrView guard;  // TODO: remove
    at::tracer::impl::NoTracerDispatchMode tracer_guard;
    c10::impl::ExcludeDispatchKeyGuard pythonmode_guard(c10::DispatchKey::Python);
    c10::impl::ExcludeDispatchKeyGuard pythonmode_snapshot_guard(c10::DispatchKey::PythonTLSSnapshot);
    // functorch uses FuncTorchDynamicLayerBackMode as a mode key to wrap all
    // tensors returned from operators in special TensorWrapper tensor extension
    // The problem with this is that TensorWrapper does not have storage so
    // accessing the data_ptr (for recursive_store) internal asserts.
    // As a quick hack, the guard here prevents functorch from wrapping the empty
    // tensor in a TensorWrapper and instead when `tensor.to` is called later,
    // the tensor gets wrapped. A more long-term solution is to think about
    // what the extensibility mechanism for this function (internal_new_from_data)
    // looks like for mode-based dispatch keys and C++ tensor extensions.
    c10::impl::ExcludeDispatchKeyGuard functorch_guard(c10::DispatchKey::FuncTorchDynamicLayerBackMode);

    if (isStorage(data)) {
      ScalarType storage_scalar_type;
      bool is_typed_storage = false;
      Storage storage = createStorageGetType(data, storage_scalar_type, is_typed_storage);
      TORCH_CHECK(!is_typed_storage || storage_scalar_type == scalar_type,
          "Expected a Storage of type ", scalar_type,
          " or an _UntypedStorage, but got ", storage_scalar_type);
      tensor = at::empty(sizes, at::initialTensorOptions().dtype(is_typed_storage ? storage_scalar_type : inferred_scalar_type).pinned_memory(pin_memory).device(storage.device()));
      tensor.set_(storage);

    } else {
      TensorOptions opts = at::initialTensorOptions().dtype(inferred_scalar_type);

      // If the device is Meta, take the shortcut. We don't want to allocate an
      // empty CPU tensor which would break our contract for meta tensors.
      if (device == at::kMeta) {
        return at::empty(sizes, opts.device(device));
      }
      tensor = at::empty(sizes, opts.pinned_memory(pin_memory));
      if (c10::multiply_integers(tensor.sizes()) != 0) {
        recursive_store(
            (char*)tensor.data_ptr(), tensor.sizes(), tensor.strides(), 0,
            inferred_scalar_type, tensor.dtype().itemsize(), data);
      }
    }
  }
  pybind11::gil_scoped_release no_gil;
  maybe_initialize_cuda(device);
  // However, it is VERY important that we trace the to() call here (even
  // though the reason this is important is a hack).  Without *some* factory
  // function call that is traced at construction time, we will consider
  // a tensor constant as originating from "outside" the trace, and if you
  // try to return it directly we will fail with the error saying no
  // "no observable data dependence".  In an ideal world, we wouldn't trace
  // a to() call but I need to think harder about what exactly we should trace
  // in this case.
  return tensor.to(device, inferred_scalar_type, /*non_blocking=*/false, /*copy=*/false);
}

Tensor new_from_data_copy(
    c10::TensorOptions options,
    at::ScalarType scalar_type,
    c10::optional<Device> device,
    PyObject* data) {
  return internal_new_from_data(options, scalar_type, device, data,
                                /*copy_variables=*/true, /*copy_numpy=*/true,
                                /*type_inference=*/false);
}

Tensor legacy_new_from_sequence(
    c10::TensorOptions options,
    at::ScalarType scalar_type,
    c10::optional<Device> device,
    PyObject* data) {
  if (!PySequence_Check(data)) {
    throw TypeError("new(): data must be a sequence (got %s)", Py_TYPE(data)->tp_name);
  }
  return internal_new_from_data(options, scalar_type, device, data,
                                /*copy_variables=*/false, /*copy_numpy=*/false,
                                /*type_inference=*/false);
}

// "base" here refers to the Tensor type on which the function was invoked, e.g.:
// in x.new(y), 'x' is the base.
// TODO: Rewrite this using dispatchKeyToTensorOptions
void check_base_legacy_new(c10::DispatchKey dispatch_key, at::Layout expected_layout) {
  if (expected_layout == c10::kStrided) {
    constexpr c10::DispatchKeySet expected_key_set({
        c10::DispatchKey::CPU,
        c10::DispatchKey::CUDA,
        c10::DispatchKey::HIP,
        c10::DispatchKey::XLA,
        c10::DispatchKey::Lazy,
        c10::DispatchKey::XPU,
        c10::DispatchKey::HPU,
    });
    TORCH_CHECK(expected_key_set.has(dispatch_key),
        "new(): expected key in ",
        expected_key_set,
        " but got: ",
        dispatch_key);
  } else if(expected_layout == c10::kSparse) {
    // NOTE: no sparse XLA or Lazy
    constexpr c10::DispatchKeySet expected_key_set({
        c10::DispatchKey::SparseCPU,
        c10::DispatchKey::SparseCUDA,
        c10::DispatchKey::SparseHIP,
        c10::DispatchKey::SparseXPU,
    });
    TORCH_CHECK(expected_key_set.has(dispatch_key),
        "new(): expected key in ",
        expected_key_set,
        " but got: ",
        dispatch_key);
  } else {
    TORCH_INTERNAL_ASSERT(false, "unexpected layout");
  }
}

// TODO: Make this accept options instead of dispatch key
void check_legacy_ctor_device(c10::DispatchKey dispatch_key, c10::optional<Device> device) {
  if (device.has_value()) {
    TORCH_CHECK(dispatchKeyToDeviceType(dispatch_key) == device.value().type(),
             "legacy constructor expects device type: ", dispatchKeyToDeviceType(dispatch_key),
             " but device type: ", device.value().type(), " was passed");
  }
}

enum class CtorOrNew {
  CTOR,
  NEW,
};

Tensor legacy_sparse_tensor_generic_ctor_new(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs, CtorOrNew ctor_or_new) {
  auto options = dispatchKeyToTensorOptions(dispatch_key);
  static PythonArgParser parser({
    "new(*, Device? device=None)",
    "new(*, int64_t cdata)|hidden",
    "new(Tensor indices, Tensor values, *, Device? device=None)",
    "new(Tensor indices, Tensor values, IntArrayRef size, *, Device? device=None)",
    "new(IntArrayRef size, *, Device? device=None)",
  });
  if (ctor_or_new == CtorOrNew::NEW) check_base_legacy_new(dispatch_key, c10::kSparse);
  ParsedArgs<4> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    auto deviceOptional = r.deviceOptional(0);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    return at::empty({0}, build_options(options, scalar_type, deviceOptional));
  } else if (r.idx == 1) {
    auto cdata = reinterpret_cast<void*>(r.toInt64(0));
    return at::unsafeTensorFromTH(cdata, true);
  } else if (r.idx == 2) {
    // Note: this signature doesn't have a dtype, even though it has a device; it probably shouldn't
    // have a device (we should infer it).
    auto deviceOptional = r.deviceOptional(2);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    at::OptionalDeviceGuard device_guard(deviceOptional);
    return at::sparse_coo_tensor(r.tensor(0), r.tensor(1));
  } else if (r.idx == 3) {
    // Note: this signature doesn't have a dtype, even though it has a device; it probably shouldn't
    // have a device (we should infer it).
    auto deviceOptional = r.deviceOptional(3);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    at::OptionalDeviceGuard device_guard(deviceOptional);
    return at::sparse_coo_tensor(r.tensor(0), r.tensor(1), r.intlist(2));
  } else if (r.idx == 4) {
    PyObject* arg = r.pyobject(0);
    auto deviceOptional = r.deviceOptional(1);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    if (!THPSize_Check(arg) && PyTuple_GET_SIZE(args) >= 1 && arg == PyTuple_GET_ITEM(args, 0)) {
      // new(sequence) binds to this signature but should be treated differently
      // unless the sequences is a torch.Size
      if (ctor_or_new == CtorOrNew::CTOR) {
        throw TypeError("torch.SparseTensor(sequence) only accepts sizes.  Please use torch.sparse_coo_tensor() " \
                        "or construct a strided tensor and convert it to sparse via to_sparse.");
      } else {
        throw TypeError("SparseTensor.new(sequence) only accepts sizes.  Please use torch.sparse_coo_tensor() " \
                        "or construct a strided tensor and convert it to sparse via to_sparse.");
      }
    }
    return new_with_sizes(options, scalar_type, r.deviceOptional(1), r.intlist(0));
  }
  throw std::runtime_error("new(): invalid arguments");
}

Tensor legacy_sparse_tensor_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  return legacy_sparse_tensor_generic_ctor_new(dispatch_key, scalar_type, args, kwargs, CtorOrNew::CTOR);
}

Tensor legacy_sparse_tensor_new(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  return legacy_sparse_tensor_generic_ctor_new(dispatch_key, scalar_type, args, kwargs, CtorOrNew::NEW);
}

// NB: device_idx here is NOT a DeviceIndex, but index into PythonArgs
c10::TensorOptions typeIdWithDefault(PythonArgs& r, int64_t device_idx, c10::DispatchKey dispatch_key) {
  auto options = dispatchKeyToTensorOptions(dispatch_key);
  if (!r.isNone(device_idx)) {
    // TODO: This line doesn't seem to be exercised at all in tests
    options = options.device(r.device(device_idx).type());
  }
  return options;
}

} // namespace

Tensor legacy_tensor_generic_ctor_new(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs, CtorOrNew ctor_or_new) {
  auto options = dispatchKeyToTensorOptions(dispatch_key);
  static PythonArgParser parser({
    "new(*, Device? device=None)",
    "new(Storage storage)",
    "new(*, int64_t cdata)|hidden",
    "new(Tensor other)",
    "new(Tensor other, *, Device? device=None)|hidden",  // prevent Tensor matching with IntArrayRef, PyObject*
    "new(IntArrayRef size, *, Device? device=None)",
    "new(PyObject* data, *, Device? device=None)",
  });

  if (isSparse(dispatchKeyToBackend(dispatch_key))) {
    return legacy_sparse_tensor_generic_ctor_new(dispatch_key, scalar_type, args, kwargs, ctor_or_new);
  }

  if (ctor_or_new == CtorOrNew::NEW) check_base_legacy_new(dispatch_key, c10::kStrided);

  ParsedArgs<2> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    auto deviceOptional = r.deviceOptional(0);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    at::OptionalDeviceGuard device_guard(deviceOptional);
    return at::empty({0}, build_options(options, scalar_type));
  } else if (r.idx == 1) {
    at::ScalarType storage_scalar_type;
    bool is_typed_storage = false;
    at::Storage storage = r.storage(0, storage_scalar_type, is_typed_storage);
    if (storage_scalar_type != at::ScalarType::Undefined && is_typed_storage) {
      TORCH_CHECK(
        storage_scalar_type == scalar_type,
        "Expected a Storage of type ", scalar_type,
        " or an _UntypedStorage, but got type ", storage_scalar_type,
        " for argument 1 'storage'");
    }
    return new_with_storage(options, scalar_type, storage);
  } else if (r.idx == 2) {
    auto cdata = reinterpret_cast<void*>(r.toInt64(0));
    return at::unsafeTensorFromTH(cdata, true);
  } else if (r.idx == 3) {
    return new_with_tensor(options, scalar_type, r.tensor(0));
  } else if (r.idx == 4) {
    if (ctor_or_new == CtorOrNew::CTOR) {
      TORCH_CHECK(false, "Legacy tensor constructor of the form torch.Tensor(tensor, device=device) " \
                  "is not supported.  Use torch.tensor(...) or torch.as_tensor(...) instead.");
    } else {
      TORCH_CHECK(false, "Legacy tensor new of the form tensor.new(tensor, device=device) " \
                  "is not supported.  Use torch.as_tensor(...) instead.");
    }
  } else if (r.idx == 5) {
    PyObject* arg = r.pyobject(0);
    auto deviceOptional = r.deviceOptional(1);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    if (!THPSize_Check(arg) && PyTuple_GET_SIZE(args) >= 1 && arg == PyTuple_GET_ITEM(args, 0)) {
      // new(sequence) binds to this signature but should be treated differently
      // unless the sequences is a torch.Size
      return legacy_new_from_sequence(options, scalar_type, deviceOptional, r.pyobject(0));
    }
    return new_with_sizes(options, scalar_type, r.deviceOptional(1), r.intlist(0));
  } else if (r.idx == 6) {
    auto deviceOptional = r.deviceOptional(1);
    check_legacy_ctor_device(dispatch_key, deviceOptional);
    return legacy_new_from_sequence(options, scalar_type, deviceOptional, r.pyobject(0));
  }
  throw std::runtime_error("new(): invalid arguments");
}

Tensor legacy_tensor_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  return legacy_tensor_generic_ctor_new(dispatch_key, scalar_type, args, kwargs, CtorOrNew::CTOR);
}

Tensor legacy_tensor_new(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  return legacy_tensor_generic_ctor_new(dispatch_key, scalar_type, args, kwargs, CtorOrNew::NEW);
}

Tensor indexing_tensor_from_data(
    c10::TensorOptions options,
    at::ScalarType scalar_type,
    c10::optional<Device> device,
    PyObject* data) {
  // Specific to tensor indexing, converts an indexing list to an
  // indexing tensor (type Byte or Long)
  ScalarType inferred_scalar_type = infer_scalar_type(data);
  if (inferred_scalar_type == ScalarType::Byte || inferred_scalar_type == ScalarType::Bool) {
    return internal_new_from_data(options, inferred_scalar_type, device, data,
                                  /*copy_variables=*/false, /*copy_numpy=*/false,
                                  /*type_inference=*/false);
  } else {
    return internal_new_from_data(options, scalar_type, device, data,
                                  /*copy_variables=*/false, /*copy_numpy=*/false,
                                  /*type_inference=*/false);
  }
}

Tensor sparse_csr_tensor_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  TORCH_INTERNAL_ASSERT(!isSparseCsr(dispatchKeyToBackend(dispatch_key)));
  TORCH_INTERNAL_ASSERT(!isSparse(dispatchKeyToBackend(dispatch_key)));
  static PythonArgParser parser({
      "sparse_csr_tensor(PyObject* crow_indices, PyObject* col_indices, PyObject* values, IntArrayRef size, *, ScalarType dtype=None, Layout? layout=None, Device? device=None, bool pin_memory=False, bool requires_grad=False)",
      "sparse_csr_tensor(PyObject* crow_indices, PyObject* col_indices, PyObject* values, *, ScalarType dtype=None, Layout? layout=None, Device? device=None, bool pin_memory=False, bool requires_grad=False)",
  });
  const int NUM_ARGS = 9, CROW_INDICES_ARG = 0, COL_INDICES_ARG = 1, VALUES_ARG = 2;
  ParsedArgs<NUM_ARGS> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  auto safe_get_attr_string = [](PyObject *o, const char *attr_name) -> PyObject* {
    // Clear error indicator if attribute does not exists.
    // Otherwise subsequent Python C API calls might return bogus values.
    // See https://github.com/pytorch/pytorch/issues/58520 for more details
    auto rc = PyObject_GetAttrString(o, attr_name);
    if (!rc) {
      if (!PyErr_ExceptionMatches(PyExc_AttributeError)) {
        throw python_error();
      }
      // Warning: a wrong attribute error may be suppressed here
      PyErr_Clear();
    }
    return rc;
  };
  THPObjectPtr crow_indices_dtype_attr(safe_get_attr_string(r.pyobject(CROW_INDICES_ARG), "dtype"));
  THPObjectPtr col_indices_dtype_attr(safe_get_attr_string(r.pyobject(COL_INDICES_ARG), "dtype"));
  at::ScalarType crow_indices_scalar_type = crow_indices_dtype_attr ? reinterpret_cast<THPDtype*>(
    crow_indices_dtype_attr.get())->scalar_type : kInt;
  at::ScalarType col_indices_scalar_type = col_indices_dtype_attr ? reinterpret_cast<THPDtype*>(
    col_indices_dtype_attr.get())->scalar_type : kInt;

  if (r.idx == 0) {
    const int SIZE_ARRAY_ARG = 3, TYPE_INFERENCE_ARG = 4, DEVICE_TYPE_ARG = 6, REQ_GRAD_ARG = 8;
    bool type_inference = r.isNone(TYPE_INFERENCE_ARG);
    const auto inferred_options = typeIdWithDefault(r, DEVICE_TYPE_ARG, dispatch_key);
    const auto inferred_scalar_type = r.scalartypeWithDefault(TYPE_INFERENCE_ARG, scalar_type);
    at::OptionalDeviceGuard device_guard(r.deviceOptional(DEVICE_TYPE_ARG));

    Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG),
                                           r.pyobject(VALUES_ARG), /*copy_variables=*/false, /*copy_numpy=*/true,
                                           /*type_inference=*/type_inference);
    Tensor crow_indices =  internal_new_from_data(values.options(),
      crow_indices_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG), r.pyobject(CROW_INDICES_ARG),
      /*copy_variables=*/false, /*copy_numpy=*/true,
      /*type_inference=*/true);
    Tensor col_indices = internal_new_from_data(values.options(),
      col_indices_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG), r.pyobject(COL_INDICES_ARG),
      /*copy_variables=*/false, /*copy_numpy=*/true,
      /*type_inference=*/true);

    return at::sparse_csr_tensor(crow_indices, col_indices, values, r.intlist(SIZE_ARRAY_ARG),
                                 values.options().layout(at::kSparseCsr)).set_requires_grad(r.toBool(REQ_GRAD_ARG));
  } else if (r.idx == 1) {
    const int TYPE_INFERENCE_ARG = 3, DEVICE_TYPE_ARG = 5, REQ_GRAD_ARG = 7;
    bool type_inference = r.isNone(TYPE_INFERENCE_ARG);
    const auto inferred_options = typeIdWithDefault(r, DEVICE_TYPE_ARG, dispatch_key);
    const auto inferred_scalar_type = r.scalartypeWithDefault(TYPE_INFERENCE_ARG, scalar_type);
    at::OptionalDeviceGuard device_guard(r.deviceOptional(DEVICE_TYPE_ARG));

    Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG),
                                           r.pyobject(VALUES_ARG), /*copy_variables=*/false, /*copy_numpy=*/true,
                                           /*type_inference=*/type_inference);
    Tensor crow_indices = internal_new_from_data(values.options(),
      crow_indices_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG),
      r.pyobject(CROW_INDICES_ARG), /*copy_variables=*/false, /*copy_numpy=*/true,
      /*type_inference=*/true);
    Tensor col_indices = internal_new_from_data(values.options(), col_indices_scalar_type, r.deviceOptional(DEVICE_TYPE_ARG),
      r.pyobject(COL_INDICES_ARG), /*copy_variables=*/false, /*copy_numpy=*/true,
      /*type_inference=*/true);
    return at::sparse_csr_tensor(crow_indices, col_indices, values,
                                 values.options().layout(at::kSparseCsr)).set_requires_grad(r.toBool(REQ_GRAD_ARG));
  }
  throw std::runtime_error("sparse_csr_tensor(): invalid arguments");
}

Tensor _sparse_csr_tensor_unsafe_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  TORCH_INTERNAL_ASSERT(!isSparseCsr(dispatchKeyToBackend(dispatch_key)));
  TORCH_INTERNAL_ASSERT(!isSparse(dispatchKeyToBackend(dispatch_key)));
  enum {
    ARG_CROW_INDICES = 0,
    ARG_COL_INDICES,
    ARG_VALUES,
    ARG_SIZE,
    ARG_TYPE,
    ARG_DEVICE,
    ARG_REQUIRES_GRAD,
    ARGS_COUNT
  };
  static PythonArgParser parser({
    "_sparse_csr_tensor_unsafe(PyObject* crow_indices, PyObject* col_indices, PyObject* values, IntArrayRef size, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
  });

  ParsedArgs<ARGS_COUNT> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  bool type_inference = r.isNone(ARG_TYPE);
  const auto inferred_options = typeIdWithDefault(r, ARG_DEVICE, dispatch_key);
  const auto inferred_scalar_type = r.scalartypeWithDefault(ARG_TYPE, scalar_type);
  at::OptionalDeviceGuard device_guard(r.deviceOptional(ARG_DEVICE));
  Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(ARG_DEVICE), r.pyobject(ARG_VALUES),
                                         /*copy_variables=*/false, /*copy_numpy=*/true,
                                         /*type_inference=*/type_inference);

  Tensor crow_indices = internal_new_from_data(values.options(), kInt, r.deviceOptional(ARG_DEVICE), r.pyobject(ARG_CROW_INDICES),
                                          /*copy_variables=*/false, /*copy_numpy=*/true,
                                          /*type_inference=*/true);

  Tensor col_indices = internal_new_from_data(values.options(), kInt, r.deviceOptional(ARG_DEVICE), r.pyobject(ARG_COL_INDICES),
                                          /*copy_variables=*/false, /*copy_numpy=*/true,
                                          /*type_inference=*/true);

  return at::_sparse_csr_tensor_unsafe(crow_indices, col_indices, values, r.intlist(ARG_SIZE), values.options().layout(at::kSparseCsr)).set_requires_grad(r.toBool(ARG_REQUIRES_GRAD));
}

// Note [Ensuring sparse values and indices match devices]
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// In all places where we construct indices, we read out options from values
// (rather than use inferred_options).  Why?  This handles the case when
// values is a CUDA tensor, but indices is a non-Tensor value (and the device
// argument is not set).  Example:
//
//  torch.sparse_coo_tensor(([0, 1],), self.empty(2, 0).cuda(), (4, 0))
//
// Sparse tensors require both indices and values to live on the same device.
// If values lives on CUDA, we can infer where the indices should live, and
// should accept even ordinary index sequences (and just make sure we write them
// into the correct device).  values is the ONLY way we know that the index
// tensor should go to CUDA, so we have to get the information in somehow.
//
// This code is kind of jank.  For one, the dtype in options is silently ignored
// by internal_new_from_data.  Also, in classic janky code style, it used to
// not work quite right: if values lives on "cuda:1", before all we said was
// "this needs to be CUDA" and indices would be allocated on the wrong tensor.
// Options is more right and gets this correct.

Tensor sparse_coo_tensor_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  TORCH_INTERNAL_ASSERT(!isSparse(dispatchKeyToBackend(dispatch_key)));
  TORCH_INTERNAL_ASSERT(!isSparseCsr(dispatchKeyToBackend(dispatch_key)));
  static PythonArgParser parser({
    "sparse_coo_tensor(PyObject* indices, PyObject* values, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
    "sparse_coo_tensor(PyObject* indices, PyObject* values, IntArrayRef size, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
    "sparse_coo_tensor(IntArrayRef size, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
  });

  ParsedArgs<6> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    bool type_inference = r.isNone(2);
    const auto inferred_options = typeIdWithDefault(r, 3, dispatch_key);
    const auto inferred_scalar_type = r.scalartypeWithDefault(2, scalar_type);
    at::OptionalDeviceGuard device_guard(r.deviceOptional(3));
    // if no dtype provided, infer type based on value type.
    Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(3), r.pyobject(1),
                                           /*copy_variables=*/false, /*copy_numpy=*/true,
                                           /*type_inference=*/type_inference);
    // See Note [Ensuring sparse values and indices match devices]
    Tensor indices = internal_new_from_data(values.options(), kLong, r.deviceOptional(3), r.pyobject(0),
                                            /*copy_variables=*/false, /*copy_numpy=*/true,
                                            /*type_inference=*/false);
    return at::sparse_coo_tensor(indices, values, values.options().layout(at::kSparse)).set_requires_grad(r.toBool(4));
  } else if (r.idx == 1) {
    bool type_inference = r.isNone(3);
    const auto inferred_options = typeIdWithDefault(r, 4, dispatch_key);
    const auto inferred_scalar_type = r.scalartypeWithDefault(3, scalar_type);
    at::OptionalDeviceGuard device_guard(r.deviceOptional(4));
    Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(4), r.pyobject(1),
                                           /*copy_variables=*/false, /*copy_numpy=*/true,
                                           /*type_inference=*/type_inference);
    // See Note [Ensuring sparse values and indices match devices]
    Tensor indices = internal_new_from_data(values.options(), kLong, r.deviceOptional(4), r.pyobject(0),
                                            /*copy_variables=*/false, /*copy_numpy=*/true,
                                            /*type_inference=*/false);
    return at::sparse_coo_tensor(indices, values, r.intlist(2), values.options().layout(at::kSparse)).set_requires_grad(r.toBool(5));
  } else if (r.idx == 2) {
    const auto inferred_options = typeIdWithDefault(r, 2, dispatch_key);
    const auto inferred_scalar_type = r.scalartypeWithDefault(1, scalar_type);
    at::OptionalDeviceGuard device_guard(r.deviceOptional(2));
    return at::sparse_coo_tensor(r.intlist(0), inferred_options.dtype(inferred_scalar_type).layout(at::kSparse)).set_requires_grad(r.toBool(3));
  }
  throw std::runtime_error("sparse_coo_tensor(): invalid arguments");
}

Tensor _sparse_coo_tensor_unsafe_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  TORCH_INTERNAL_ASSERT(!isSparse(dispatchKeyToBackend(dispatch_key)));
  TORCH_INTERNAL_ASSERT(!isSparseCsr(dispatchKeyToBackend(dispatch_key)));
  enum {
    ARG_INDICES = 0,
    ARG_VALUES,
    ARG_SIZE,
    ARG_TYPE,
    ARG_DEVICE,
    ARG_REQUIRES_GRAD,
    ARGS_COUNT
  };
  static PythonArgParser parser({
    "_sparse_coo_tensor_unsafe(PyObject* indices, PyObject* values, IntArrayRef size, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
  });

  ParsedArgs<ARGS_COUNT> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  bool type_inference = r.isNone(ARG_TYPE);
  const auto inferred_options = typeIdWithDefault(r, ARG_DEVICE, dispatch_key);
  const auto inferred_scalar_type = r.scalartypeWithDefault(ARG_TYPE, scalar_type);
  at::OptionalDeviceGuard device_guard(r.deviceOptional(ARG_DEVICE));
  Tensor values = internal_new_from_data(inferred_options, inferred_scalar_type, r.deviceOptional(ARG_DEVICE), r.pyobject(ARG_VALUES),
                                         /*copy_variables=*/false, /*copy_numpy=*/true,
                                         /*type_inference=*/type_inference);
  // See Note [Ensuring sparse values and indices match devices]
  Tensor indices = internal_new_from_data(values.options(), kLong, r.deviceOptional(ARG_DEVICE), r.pyobject(ARG_INDICES),
                                          /*copy_variables=*/false, /*copy_numpy=*/true,
                                          /*type_inference=*/false);
  return at::_sparse_coo_tensor_unsafe(indices, values, r.intlist(ARG_SIZE), values.options().layout(at::kSparse)).set_requires_grad(r.toBool(ARG_REQUIRES_GRAD));
}

void _validate_sparse_coo_tensor_args(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  auto options = dispatchKeyToTensorOptions(dispatch_key);
  static PythonArgParser parser({
    "_validate_sparse_coo_tensor(PyObject* indices, PyObject* values, IntArrayRef size)",
  });

  ParsedArgs<3> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  Tensor values = internal_new_from_data(
      options, scalar_type, c10::nullopt, r.pyobject(1),
      /*copy_variables=*/false, /*copy_numpy=*/true, /*type_inference=*/true);
  // See Note [Ensuring sparse values and indices match devices]
  Tensor indices = internal_new_from_data(
      values.options(), kLong, c10::nullopt, r.pyobject(0),
      /*copy_variables=*/false, /*copy_numpy=*/true, /*type_inference=*/false);
  at::native::_validate_sparse_coo_tensor_args(indices, values, r.intlist(2));
}


void _validate_sparse_csr_tensor_args(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  auto options = dispatchKeyToTensorOptions(dispatch_key);
  static PythonArgParser parser({
    "_validate_sparse_csr_tensor(PyObject* crow_indices, PyObject* col_indices, PyObject* values, IntArrayRef size)",
  });

  ParsedArgs<4> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  Tensor values = internal_new_from_data(
      options, scalar_type, c10::nullopt, r.pyobject(2),
      /*copy_variables=*/false, /*copy_numpy=*/true, /*type_inference=*/true);
  // See Note [Ensuring sparse values and indices match devices]
  Tensor crow_indices = internal_new_from_data(
      values.options(), kInt, c10::nullopt, r.pyobject(0),
      /*copy_variables=*/false, /*copy_numpy=*/true, /*type_inference=*/true);
  Tensor col_indices = internal_new_from_data(
      values.options(), kInt, c10::nullopt, r.pyobject(1),
      /*copy_variables=*/false, /*copy_numpy=*/true, /*type_inference=*/true);

  at::native::_validate_sparse_csr_tensor_args(crow_indices, col_indices, values, r.intlist(3));
}

Tensor tensor_ctor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  static PythonArgParser parser({
    "tensor(PyObject* data, *, ScalarType dtype=None, Device? device=None, bool pin_memory=False, bool requires_grad=False, DimnameList? names=None)",
  });

  constexpr int ctor_num_args = 6;
  ParsedArgs<ctor_num_args> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    PyObject* data = r.pyobject(0);
    if (THPVariable_Check(data)) {
      auto ret = PyErr_WarnEx(PyExc_UserWarning,
        "To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() "
        "or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).", 1);
      if (ret != 0) throw python_error();
    }

    bool type_inference = r.isNone(1);
    bool pin_memory = r.toBool(3);
    bool args_requires_grad = r.toBool(4);
    auto new_tensor = internal_new_from_data(
               typeIdWithDefault(r, 2, dispatch_key),
               r.scalartypeWithDefault(1, scalar_type),
               r.deviceOptional(2),
               data,
               /*copy_variables=*/true,
               /*copy_numpy=*/true,
               /*type_inference=*/type_inference,
               pin_memory);
    auto names = r.toDimnameListOptional(5);
    if (names) {
      at::namedinference::propagate_names(new_tensor, *names, /*validate_names=*/true);
    }
    new_tensor.detach_(); // ensure new_tensor a leaf node
    new_tensor.set_requires_grad(args_requires_grad);
    return new_tensor;
  }
  throw std::runtime_error("tensor(): invalid arguments");
}

Tensor as_tensor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  // TODO: add requires_grad once we decide on semantics for sharing data.
  static PythonArgParser parser({
    "as_tensor(PyObject* data, *, ScalarType dtype=None, Device? device=None)",
  });

  ParsedArgs<3> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    bool type_inference = r.isNone(1);
    return internal_new_from_data(
        typeIdWithDefault(r, 2, dispatch_key),
        r.scalartypeWithDefault(1, scalar_type),
        r.deviceOptional(2),
        r.pyobject(0),
        /*copy_variables=*/false,
        /*copy_numpy=*/false,
        /*type_inference=*/type_inference);
  }
  throw std::runtime_error("tensor(): invalid arguments");
}

Tensor new_tensor(c10::DispatchKey dispatch_key, at::ScalarType scalar_type, PyObject* args, PyObject* kwargs) {
  static PythonArgParser parser({
    "new_tensor(PyObject* data, *, ScalarType dtype=None, Device? device=None, bool requires_grad=False)",
  });

  ParsedArgs<4> parsed_args;
  auto r = parser.parse(args, kwargs, parsed_args);
  if (r.idx == 0) {
    PyObject* data = r.pyobject(0);
    if (THPVariable_Check(data)) {
      auto ret = PyErr_WarnEx(PyExc_UserWarning,
        "To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() "
        "or sourceTensor.clone().detach().requires_grad_(True), rather than tensor.new_tensor(sourceTensor).", 1);
      if (ret != 0) throw python_error();
    }

    bool args_requires_grad = r.toBool(3);
    auto new_tensor = new_from_data_copy(
               typeIdWithDefault(r, 2, dispatch_key),
               r.scalartypeWithDefault(1, scalar_type),
               r.deviceOptional(2),
               data);
    new_tensor.detach_(); // ensure new_tensor a leaf node
    new_tensor.set_requires_grad(args_requires_grad);
    return new_tensor;
  }
  throw std::runtime_error("new_tensor(): invalid arguments");
}

Tensor tensor_frombuffer(PyObject* buffer, ScalarType dtype, int64_t count, int64_t offset, bool requires_grad) {
  auto elsize = at::elementSize(dtype);
  size_t actual_count = 0;

  Py_buffer view;
  if (PyObject_GetBuffer(buffer, &view, PyBUF_WRITABLE) < 0) {
    TORCH_CHECK(
        PyObject_GetBuffer(buffer, &view, PyBUF_SIMPLE) >= 0,
        "could not retrieve buffer from object");
    TORCH_WARN_ONCE(
        "The given buffer is not writable, and PyTorch does "
        "not support non-writable tensors. This means you can write to the "
        "underlying (supposedly non-writable) buffer using the tensor. "
        "You may want to copy the buffer to protect its data or make it writable "
        "before converting it to a tensor. This type of warning will be "
        "suppressed for the rest of this program.");
    PyErr_Clear();
  }

  Py_INCREF(view.obj);
  THPObjectPtr obj(view.obj);

  auto len = view.len;
  auto buf = view.buf;
  PyBuffer_Release(&view);

  TORCH_CHECK_VALUE(
      len > 0 && count != 0,
      "both buffer length (", len, ") and count (", count, ") must not be 0");
  TORCH_CHECK_VALUE(
      offset >= 0 && offset < len,
      "offset (", offset, " bytes) must be non-negative and no greater than "
      "buffer length (", len, " bytes) minus 1");
  TORCH_CHECK_VALUE(
      count > 0 || (len - offset) % elsize == 0,
      "buffer length (", len - offset, " bytes) after offset (", offset, " bytes) "
      "must be a multiple of element size (", elsize, ")");

  if (count < 0) {
    actual_count = (len - offset) / elsize;
  } else {
    actual_count = static_cast<size_t>(count);
  }

  TORCH_CHECK_VALUE(
      static_cast<size_t>(offset) + actual_count * elsize <= len,
      "requested buffer length (", actual_count, " * ", elsize, " bytes) "
      "after offset (", offset, " bytes) must not be greater than actual "
      "buffer length (", len, " bytes)");

  auto offset_buf = static_cast<char*>(buf) + offset;
  auto options = TensorOptions()
      .dtype(dtype)
      .device(c10::kCPU);

  auto tensor = at::for_blob(offset_buf, static_cast<int64_t>(actual_count))
                    .options(options)
                    .deleter([obj = obj.release()](void*) {
                      pybind11::gil_scoped_acquire gil;
                      Py_DECREF(obj);
                    })
                    .make_tensor();
  tensor.set_requires_grad(requires_grad);
  return tensor;
}

Tensor tensor_fromDLPack(PyObject *data) {
  DLManagedTensor * dlMTensor = (DLManagedTensor *)PyCapsule_GetPointer(data, "dltensor");
  TORCH_CHECK(dlMTensor,
    "from_dlpack received an invalid capsule. "
    "Note that DLTensor capsules can be consumed only once, "
    "so you might have already constructed a tensor from it once.");

  // atensor steals the ownership of the underlying storage. It also passes a
  // destructor function that will be called when the underlying storage goes
  // out of scope. When the destructor is called, the dlMTensor is destructed too.
  auto atensor = at::fromDLPack(dlMTensor);

  // Make sure this capsule will never be used again.
  PyCapsule_SetName(data, "used_dltensor");

  // It is possible that the call to at::fromDLPack is the very first
  // call to create a Tensor in PyTorch. If so, then _lazy_init has
  // not been called, and the attempt to call createPyObject will fail
  // because cuda ATen types have not been registered in Python yet.
  // so if we have a cuda tensor, then we need to make sure
  // we have called _lazy_init here
  if(atensor.is_cuda()) {
    py::module::import("torch.cuda").attr("init")();
  }
  return atensor;
}

Tensor asarray(
    PyObject* obj,
    c10::optional<ScalarType> dtype,
    c10::optional<Device> device,
    c10::optional<bool> copy,
    bool requires_grad) {
  Tensor tensor;

  bool force_copy = copy.value_or(false);
  bool force_alias = !copy.value_or(true);
  bool should_warn_numpy_not_writable = false;

  auto dtype_unwrapped =
      dtype.value_or(torch::tensors::get_default_scalar_type());

  // Check whether 'obj' is a 'Tensor'
  if (THPVariable_Check(obj)) {
    tensor = THPVariable_Unpack(obj);
  }

#ifdef USE_NUMPY
  // Check whether 'obj' is a NumPy Array
  if (is_numpy_available() && PyArray_Check(obj)) {
    tensor = tensor_from_numpy(obj, /*warn_if_not_writeable=*/false);
    should_warn_numpy_not_writable = !PyArray_ISWRITEABLE((PyArrayObject*) obj);
  }
#endif

  // Check whether 'obj' is a 'DLPack' capsule
  if (!tensor.defined() && PyCapsule_IsValid(obj, "dltensor") != 0) {
    tensor = tensor_fromDLPack(obj);
  }

  // Check whether 'obj' implements the buffer protocol
  if (!tensor.defined() && PyObject_CheckBuffer(obj) != 0) {
    tensor = tensor_frombuffer(obj, dtype_unwrapped, -1, 0, requires_grad);
  }

  if (tensor.defined()) {
    // Given an aliasable tensor, should we copy it?
    bool wrong_device = device.has_value() && device.value() != tensor.device();
    bool wrong_dtype =
        dtype.has_value() && dtype.value() != tensor.scalar_type();
    bool needs_copying = !copy.has_value() && (wrong_device || wrong_dtype);

    // Given a defined tensor, we copy it if either we have to (copy=True) or
    // if we need to (copy=None) because of mismatched device or dtype.
    if (force_copy || needs_copying) {
      if (wrong_device || wrong_dtype) {
        tensor = tensor.to(
            device.value_or(tensor.device()),
            dtype.value_or(tensor.scalar_type()));
      } else {
        tensor = tensor.clone();
      }
    } else {
      // If we are not copying, we have to check whther we have the tensor
      // in the right device, with the right dtype.
      TORCH_CHECK_VALUE(
          !wrong_device,
          "can't alias tensor from device '", tensor.device(),
          "' to '", device.value(), "'.");
      TORCH_CHECK_VALUE(
          !wrong_dtype,
          "can't alias tensor with dtype '", tensor.scalar_type(),
          "' into dtype '", dtype.value(), "'.");
      // If tensor is a NumPy Array view, we warn the user about non-writeable
      // arrays if this is the case.
      if (should_warn_numpy_not_writable) {
        warn_numpy_not_writeable();
      }
    }

    // Setting 'requires_grad' when the tensor is not a leaf does not work.
    // Whenever that happens, we have to use 'detach'.
    if (!tensor.is_leaf() && !requires_grad) {
      tensor = tensor.detach();
    } else {
      tensor.set_requires_grad(requires_grad);
    }
  } else {
    // Undefined tensor means it does not implement neither DLPack nor
    // the buffer protocol. Last case is a sequence, in which case we must
    // copy (copy can't be false).
    TORCH_CHECK_VALUE(
        !force_alias, "can't alias arbitrary sequence into a tensor.");

    // Make tensor from sequence, inferring its type, and then convert
    // it to the desired type.
    // Type inference is activated only if the dtype has not been specified.
    // Otherwise, we force the unwrapped dtype.
    tensor = internal_new_from_data(
        TensorOptions(), dtype_unwrapped, device, obj,
        /* copy_variables = */ false, /* copy_numpy = */ false, /* type_inference = */ !dtype.has_value());
    tensor.set_requires_grad(requires_grad);
  }

  return tensor;
}

}} // namespace torch::utils