Struct ModelFragment 

pub struct ModelFragment { /* private fields */ }

Validated, shape-specific model commands with shared immutable weights. Record into a caller-owned graph to connect stages without copies, or build a standalone bounded inference pool with Self::prepare.

Implementations

impl ModelFragment

pub unsafe fn reuse_private_scratch(self) -> Self

Reuse private temporaries across fragments recorded into the same graph. Input/output allocations and immutable weights are never recycled. Each inference slot still owns a separate workspace. Graph hazards order reads before a later fragment overwrites that storage, potentially serializing otherwise independent branches in exchange for lower peak memory.

Safety

Every private scratch byte read by this fragment must first be written by this fragment on every execution, including padding and conditional paths. Neither initial zeroes nor another fragment’s values may be used. All scratch accesses must already be declared by the kernel contracts.

pub fn prepare(&self, capacity: usize) -> Result<PreparedModel>

Allocate independent fixed-address slots, exposing their actual model IO.

pub fn record( &self, graph: &mut Graph, inputs: &[DeviceTensor], ) -> Result<Vec<DeviceTensor>>

Append model commands directly to an unprepared graph. Inputs are bound without a device copy; outputs and scratch are allocated for this graph. Adjacent stages can consume the returned tensors before graph preparation. No pool, host staging, submission or intermediate synchronization is added.

Inputs must describe the complete accessed regions of their allocations. Partial aliases which would require a copy are rejected. Initialization producers are drained here; this is a preparation-time operation.