Struct ModelDefinition 

pub struct ModelDefinition { /* private fields */ }

Immutable model code and weights in a shared context. Shape-specific plans allocate private scratch and graph storage for each bounded inference slot.

Implementations

impl ModelDefinition

pub fn context(&self) -> &ModelContext

Shared scheduling and allocation domain.

pub fn launch_config( &self, id: KernelId, workload_bits: &[u64], ) -> Result<LaunchConfig>

Evaluate a loaded kernel’s compiler launch program during graph preparation.

pub fn weight_bytes(&self) -> usize

Physical immutable weight bytes, counted once regardless of plan or slot count.

pub unsafe fn prepare( &self, commands: &[Command], inputs: &[(Region, TensorDesc)], outputs: &[(Region, TensorDesc)], capacity: usize, ) -> Result<PreparedModel>

Prepare a shape with shared weights and fixed-address private slots. Input/output descriptors select leading bytes of their corresponding regions. Commands must restrict accesses to the selected shape.

Safety

Dispatch bindings and scalar arguments must exactly describe the trusted compiled code, including accessible extents and every read or write. Writable argument aliasing must be safe for that kernel’s execution.

pub unsafe fn fragment( &self, commands: &[Command], inputs: &[(Region, TensorDesc)], outputs: &[(Region, TensorDesc)], ) -> Result<ModelFragment>

Validate a shape’s commands once, without allocating inference slots. The fragment owns shared weights and code independently of this definition.

Safety

Dispatch contracts, accessible extents and writable aliases must satisfy the same requirements as Self::prepare.