Struct ModelContext 

pub struct ModelContext { /* private fields */ }

Explicitly shared allocation, compilation, and scheduling domain for models.

Implementations

impl ModelContext

pub fn new(options: RuntimeOptions) -> Result<Self>

Construct a context; hardware and compiler initialization remain lazy.

pub fn runtime(&self) -> &Runtime

Coordinated runtime shared by every model and tensor in this context.

pub fn compiler(&self) -> Result<Compiler>

Lazily obtain the compiler for the selected device’s actual target.

pub fn allocate(&self, desc: TensorDesc) -> Result<DeviceTensor>

Allocate initialized device storage. Empty tensors allocate nothing.

pub fn allocate_with( &self, desc: TensorDesc, placement: MemoryPlacement, ) -> Result<DeviceTensor>

Allocate a tensor with explicit storage placement. Host-visible IO can be submitted and read back directly without transfer staging or copies.

pub fn tensor( &self, desc: TensorDesc, view: BufferView, producer: Completion, ) -> Result<DeviceTensor>

Wrap existing tracked storage without copying. The producer must belong to this runtime or represent already completed initialization.

pub fn upload(&self, desc: TensorDesc, bytes: &[u8]) -> Result<DeviceTensor>

Copy host bytes into owned staging and enqueue upload. The caller may release or mutate its source as soon as this method returns.

pub fn copy_into( &self, destination: &mut DeviceTensor, byte_offset: usize, source: &DeviceTensor, ) -> Result<()>

Copy a contiguous tensor into a byte-aligned subrange of another tensor. Both tensors must be contiguous and have the same dtype. This updates the destination’s producer dependency, preserving failures from earlier writes. Slot owners survive until the copy drains, even if the source is dropped.

pub fn concatenate(&self, inputs: &[DeviceTensor]) -> Result<DeviceTensor>

Concatenate contiguous tensors along their first axis on the device. Inputs must have identical trailing dimensions, dtype and layout. Empty inputs are allowed; scalars and an empty input list are not.

pub fn download(&self, tensor: &DeviceTensor) -> Result<TensorReadback>

Queue readback after a tensor’s producer, without a host wait.

pub fn validate(&self, tensor: &DeviceTensor) -> Result<()>

Reject a tensor from another allocation domain before any work is queued.

Trait Implementations

impl Clone for ModelContext

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.