Struct Tensor

#[repr(C)]
pub struct Tensor { /* private fields */ }

ABI stable owned Shape for ffi

Implementations

impl Tensor

pub fn data_ptr(&self) -> *const c_void

Get the data pointer of the Tensor

Returns

• *mut core::ffi::c_void - The data pointer of the Tensor

pub fn data_ptr_mut(&mut self) -> *mut c_void

Get the data pointer of the Tensor

Returns

• *mut core::ffi::c_void - The data pointer of the Tensor

pub fn is_contiguous(&self) -> bool

Check if the Tensor is contiguous

Returns

• bool - True if the Tensor is contiguous, false otherwise

pub fn data_as_slice<T: AsDLDataType>(&self) -> Result<&[T]>

pub fn data_as_slice_mut<T: AsDLDataType>(&self) -> Result<&mut [T]>

Returns the tensor data as a mutable slice.

This method takes &self rather than &mut self by design: like std::fs::File::write, the metadata of a Tensor (shape, dtype, device) is governed by Rust’s ownership rules, but writing to the underlying data buffer (CPU memory or a GPU pointer) is a side-effect outside Rust’s aliasing model. Most C/CUDA kernel APIs accept a non-mut Tensor and mutate its data content, so requiring &mut self here would force artificial mutability annotations throughout the deep-learning stack with no real safety benefit.

Safety contract (caller responsibility)

If the Tensor has been cloned (via ObjectArc), the caller must ensure no other clone is concurrently reading the data.

pub fn shape(&self) -> &[i64]

pub fn ndim(&self) -> usize

pub fn numel(&self) -> usize

pub fn strides(&self) -> &[i64]

pub fn dtype(&self) -> DLDataType

pub fn device(&self) -> DLDevice

impl Tensor

pub fn from_nd_alloc<TNDAlloc>( alloc: TNDAlloc, shape: &[i64], dtype: DLDataType, device: DLDevice, ) -> Self

where TNDAlloc: NDAllocator,

Arguments

• alloc - The NDAllocator
• shape - The shape of the Tensor
• dtype - The data type of the Tensor
• device - The device of the Tensor

Returns

• Tensor - The created Tensor

pub fn from_env_alloc( shape: &[i64], dtype: DLDataType, device: DLDevice, ) -> Result<Self>

Create a Tensor with the environment allocator of this thread, as C++ Tensor::FromEnvAlloc(TVMFFIEnvTensorAlloc, ...) does.

Kernel libraries allocate intermediate tensors this way, so that they come from the caller’s allocator, such as the one a framework sets with TVMFFIEnvSetDLPackManagedTensorAllocator.

Arguments

• shape - The shape of the Tensor
• dtype - The data type of the Tensor
• device - The device of the Tensor

Returns

• Result<Tensor> - The created Tensor, or the error the allocator raised, which is a RuntimeError when no allocator is set

pub fn from_slice<T: AsDLDataType>(slice: &[T], shape: &[i64]) -> Result<Self>

Create a Tensor from a slice

Arguments

• slice - The slice to create the Tensor from
• shape - The shape of the Tensor

Returns

• Tensor - The created Tensor

Trait Implementations

impl AnyCompatible for Tensor

unsafe fn is_undefined(field: *const Self) -> bool

Whether field, which holds a value of this type or all-zero bytes, holds all-zero bytes that are no value of this type, as a null object reference, a string of type index None or a zero device type: a field that a creator left undefined for reflection to set, as C++ UnsafeInit leaves a reference undefined (null). Unlike a C++ undefined reference, it is no value; a nullable reference is Option<T>, whose None is a defined value.

fn type_str() -> String

the type string of the type

unsafe fn copy_to_any_view(src: &Self, data: &mut TVMFFIAny)

the value to copy to TVMFFIAny

unsafe fn check_any_strict(data: &TVMFFIAny) -> bool

unsafe fn copy_from_any_view_after_check(data: &TVMFFIAny) -> Self

Copy value from TVMFFIAny after checking caller must ensure that the value is compatible with the type

unsafe fn move_to_any(src: Self, data: &mut TVMFFIAny)

consume the value to move to Any

unsafe fn move_from_any_after_check(data: &mut TVMFFIAny) -> Self

the value to move from TVMFFIAny NOTE: pay very careful attention to avoid memory leak!

unsafe fn try_cast_from_any_view(data: &TVMFFIAny) -> Result<Self, ()>

try to cast the value from AnyView

fn get_mismatch_type_info(data: &TVMFFIAny) -> String

Get the type key of a type when TryCastFromAnyView fails.

fn type_schema() -> String

The type schema of the type: the JSON object C++ TypeTraits<T>::TypeSchema writes, which the type schemas of functions and fields embed.

fn to_any(&self) -> Any

Borrow self into an owned Any, increfing object-backed values.

impl<'a> ArgIntoRef for &'a Tensor

type Target = Tensor

fn to_ref(&self) -> &Self::Target

impl ArgIntoRef for Tensor

type Target = Tensor

fn to_ref(&self) -> &Self::Target

impl Clone for Tensor

fn clone(&self) -> Tensor

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl From<&Tensor> for Tensor

fn from(value: &Tensor) -> Self

Converts to this type from the input type.

impl From<&Tensor> for TensorView

fn from(tensor: &Tensor) -> Self

Converts to this type from the input type.

impl<'a> IntoArgHolder for &'a Tensor

type Target = &'a Tensor

fn into_arg_holder(self) -> Self::Target

impl IntoArgHolder for Tensor

type Target = Tensor

fn into_arg_holder(self) -> Self::Target

impl ObjectRefCore for Tensor

type ContainerType = <ObjectArc<TensorObj> as Deref>::Target

fn data(this: &Self) -> &ObjectArc<Self::ContainerType>

fn into_data(this: Self) -> ObjectArc<Self::ContainerType>

unsafe fn from_data(data: ObjectArc<Self::ContainerType>) -> Self

Construct a reference view from an owning container handle.

fn same_as<Other: ObjectRefCore>(&self, other: &Other) -> bool

Return whether two object references point to the same allocation.

fn downcast_ref<T: ObjectCore>(&self) -> Option<&T>

Borrow the underlying object as T when its runtime type is T or a subtype.

impl TryFrom<Any> for Tensor

type Error = Error

The type returned in the event of a conversion error.

fn try_from(value: Any) -> Result<Self, Self::Error>

Performs the conversion.

impl<'a> TryFrom<AnyView<'a>> for Tensor

type Error = Error

The type returned in the event of a conversion error.

fn try_from(value: AnyView<'a>) -> Result<Self, Self::Error>

Performs the conversion.