torch 🔗

def as_bytelike(self) -> bytes | bytearray | memoryview:
    """Returns the tensor data buffer as a byte-like object.

    Returns:
        The data of the tensor as a byte-like object.
    """
    if self._c_data is not None:
        if isinstance(self._c_data, CBytes):
            return self._c_data.value
        return self._c_data.as_memoryview()
    return CBytes.from_c_obj(
        self._c_ptr_or_raise.data, self._c_ptr_or_raise.size
    ).value

def as_memoryview(self) -> memoryview:
    """Returns the tensor data buffer as memoryview.

    Returns:
        The data of the tensor as memoryview.
    """
    data = self.as_bytelike()
    return data if isinstance(data, memoryview) else memoryview(data)

def as_numpy(self) -> "np.ndarray":
    """Returns the tensor data buffer as a NumPy array.

    Returns:
        The data of the tensor as a NumPy array.
    """
    if not HAS_NUMPY:
        msg = "NumPy is not available"
        raise NotImplementedError(msg)

    if isinstance(self._c_data, CNumpyArray):
        return self._c_data.value

    dtype = TensorTypeMapper.type_to_numpy(self.data_type)
    return CNumpyArray.from_c_obj(
        self._c_ptr_or_raise.data,
        self._c_ptr_or_raise.size,
        self.shape,
        dtype,
    ).value

def as_torch(self) -> "torch.Tensor":
    """Returns the tensor as a PyTorch tensor.

    Returns:
        The tensor as a PyTorch tensor.

    Raises:
        NotImplementedError: If PyTorch is not installed.
    """
    if not HAS_TORCH:
        msg = "PyTorch is not available"
        raise NotImplementedError(msg)

    if isinstance(self._data, torch.Tensor):
        return self._data

    dtype = TensorTypeMapper.type_to_torch(self.data_type)
    buf = ffi.buffer(self._c_ptr_or_raise.data, self._c_ptr_or_raise.size)

    return torch.frombuffer(buf, dtype=dtype).reshape(self.shape)

@classmethod
def empty(cls) -> "Tensor":
    """Initializes a new empty `Tensor` object.

    The object has a NULL pointer in place of the C struct.

    Returns:
        A new `Tensor` object
    """
    inst = cls.__new__(cls)
    inst._dims = None
    inst._data = None
    inst._data_type = None
    inst._c_data = None
    inst._c_obj_ptr = ffi.new("struct vaccel_torch_tensor **")
    inst._c_obj = inst._c_obj_ptr[0]
    inst._c_obj_ptr[0] = ffi.NULL
    inst._c_size = ffi.sizeof("struct vaccel_torch_tensor *")
    return inst

@classmethod
def from_buffer(
    cls,
    dims: list[int],
    data_type: TensorType,
    data: bytes | bytearray | memoryview,
) -> "Tensor":
    """Initializes a new `Tensor` object from byte-like data.

    Args:
        dims: The dims to be passed to the C struct.
        data_type: The data_type to be passed to the C struct.
        data: The data to be passed to the C struct.

    Returns:
        A new `Tensor` object
    """
    inst = cls.__new__(cls)
    inst._dims = dims
    inst._data = data
    inst._data_type = data_type
    inst._c_data = CBytes(inst._data)
    inst._c_obj_ptr = ffi.NULL
    inst._c_obj_data = inst._c_data._c_ptr
    super().__init__(inst)
    return inst

@classmethod
def from_c_obj(cls, c_obj: ffi.CData) -> "Tensor":
    """Initializes a new `Tensor` object from an existing C struct.

    Args:
        c_obj: A pointer to a `struct vaccel_torch_tensor` C object.

    Returns:
        A new `Tensor` object
    """
    type_str = ffi.getctype(ffi.typeof(c_obj))
    if type_str != "struct vaccel_torch_tensor *":
        msg = f"Expected 'struct vaccel_torch_tensor *', got '{type_str}'"
        raise TypeError(msg)

    inst = cls.__new__(cls)
    inst._dims = None
    inst._data = None
    inst._data_type = None
    inst._c_data = None
    inst._c_obj_ptr = ffi.NULL
    inst._c_obj_data = ffi.NULL
    inst._c_obj = c_obj
    inst._c_size = ffi.sizeof(inst._c_obj)
    return inst

@classmethod
def from_numpy(cls, data: "np.ndarray") -> "Tensor":
    """Initializes a new `Tensor` object from a NumPy array.

    Args:
        data: The NumPy array containing the tensor data.

    Returns:
        A new `Tensor` object

    Raises:
        NotImplementedError: If NumPy is not installed.
    """
    if not HAS_NUMPY:
        msg = "NumPy is not available"
        raise NotImplementedError(msg)

    inst = cls.__new__(cls)
    inst._dims = list(data.shape)
    inst._data = data
    inst._data_type = TensorTypeMapper.type_from_numpy(inst._data.dtype)
    inst._c_data = CNumpyArray(inst._data)
    inst._c_obj_ptr = ffi.NULL
    inst._c_obj_data = inst._c_data._c_ptr
    super().__init__(inst)
    return inst

@classmethod
def from_torch(cls, data: "torch.Tensor") -> "Tensor":
    """Initializes a new `Tensor` object from a PyTorch tensor.

    Args:
        data: The PyTorch tensor containing the tensor data.

    Returns:
        A new `Tensor` object

    Raises:
        NotImplementedError: If PyTorch is not installed.
    """
    if not HAS_TORCH:
        msg = "PyTorch is not available"
        raise NotImplementedError(msg)

    inst = cls.__new__(cls)
    inst._dims = list(data.shape)
    inst._data = data.contiguous().cpu()
    inst._data_type = TensorTypeMapper.type_from_torch(inst._data.dtype)
    inst._c_obj_ptr = ffi.NULL
    data_ptr_int = inst._data.untyped_storage().data_ptr()
    data_ptr = ffi.cast("void *", data_ptr_int)
    data_size = inst._data.untyped_storage().nbytes()
    inst._c_data = CBytes.from_c_obj(data_ptr, data_size)
    inst._c_obj_data = inst._c_data._c_ptr
    super().__init__(inst)
    return inst

def to_bytes(self) -> bytes:
    """Returns the tensor data buffer as bytes.

    Returns:
        The data of the tensor as bytes.
    """
    data = self.as_bytelike()
    return data if isinstance(data, bytes) else bytes(data)

def torch_jitload_forward(
    self,
    resource: Resource,
    in_tensors: list[Tensor],
    nr_out_tensors: int = 1,
    run_options: Buffer | None = None,
) -> list[Tensor]:
    """Performs the Torch jitload forward operation.

    Wraps the `vaccel_torch_jitload_forward()` C operation.

    Args:
        resource: A resource with the model to run.
        in_tensors: The input tensors for the inference.
        nr_out_tensors: The number of output tensors. Defaults to 1.
        run_options: The inference options.

    Returns:
        The output tensors

    Raises:
        FFIError: If the C operation fails.
    """
    run_options_ptr = (
        ffi.NULL if run_options is None else run_options._c_ptr
    )
    c_in_tensors = CList.from_ptrs(in_tensors)
    c_out_tensors = CList.from_ptrs([Tensor.empty()] * nr_out_tensors)

    ret = lib.vaccel_torch_jitload_forward(
        self._c_ptr_or_raise,
        resource._c_ptr,
        run_options_ptr,
        c_in_tensors._c_ptr,
        len(c_in_tensors),
        c_out_tensors._c_ptr,
        len(c_out_tensors),
    )
    if ret != 0:
        raise FFIError(ret, "Torch jitload forward operation failed")

    return [Tensor.from_c_obj(t) for t in c_out_tensors.value]