Struct TensorWasmMemoryCreator 

pub struct TensorWasmMemoryCreator { /* private fields */ }

A MemoryCreator that hands out TensorWasmLinearMemory instances.

Wrap in Arc and pass to wasmtime::Config::with_host_memory.

Implementations

impl TensorWasmMemoryCreator

pub fn new(device_id: DeviceId) -> Self

Construct without an underlying pool. New memories allocate fresh UnifiedBuffers on every new_memory call. Uses the default MemoryCreatorConfig (per-instance cap = HARD_MAX_LINEAR_MEMORY_BYTES).

pub fn with_config(device_id: DeviceId, config: MemoryCreatorConfig) -> Self

Construct without an underlying pool, with a custom MemoryCreatorConfig (e.g. a tighter per-instance cap for multi-tenant deployments). The configured cap is clamped to HARD_MAX_LINEAR_MEMORY_BYTES — see MemoryCreatorConfig::effective_max_linear_memory_bytes.

pub fn with_tenant_context( device_id: DeviceId, tenant_ctx: Arc<TenantContext>, ) -> Self

Construct a tenant-aware creator that accounts every fresh UnifiedBuffer against the tenant’s GPU memory cap.

Roadmap feature #8 builder: holds an Arc<TenantContext> and routes the no-pool / pool-fallback allocation paths through UnifiedBuffer::new_on_with_tenant_context. On a cap violation the underlying error is tensor_wasm_core::error::TensorWasmError::GpuMemoryExhausted; it surfaces here as a String because Wasmtime’s MemoryCreator::new_memory returns Result<_, String>.

Pool-carved memories are intentionally unmetered — see the note on MemoryCreatorState::tenant_ctx for the rationale and the v0.4 follow-up.

pub fn with_pool_and_tenant_context( device_id: DeviceId, pool: Arc<UnifiedMemoryPool>, tenant_ctx: Arc<TenantContext>, ) -> Self

Tenant-aware variant of Self::with_pool.

Behaves like Self::with_pool for the pool-carve hot path (no extra counter traffic) and like Self::with_tenant_context for the fallback fresh-allocation path.

pub fn with_pool(device_id: DeviceId, pool: Arc<UnifiedMemoryPool>) -> Self

Construct with a pre-allocated pool. new_memory carves out of this slab; on exhaustion it falls back to a fresh UnifiedBuffer so a short slab does not turn into a fatal error.

Lifetime contract. Pool-backed linear memories form an all-or-nothing batch: the caller must drop every Wasm instance handed out by this creator before dropping the last Arc<UnifiedMemoryPool> reference and calling UnifiedMemoryPool::reset. See new_memory for the unsafe rationale.

pub fn with_pool_and_config( device_id: DeviceId, pool: Arc<UnifiedMemoryPool>, config: MemoryCreatorConfig, ) -> Self

Construct with both a pre-allocated pool and a custom MemoryCreatorConfig.

pub fn device_id(&self) -> DeviceId

The device this creator targets.

pub fn pool(&self) -> Option<&Arc<UnifiedMemoryPool>>

The underlying pool, if one was provided at construction.

pub fn config(&self) -> &MemoryCreatorConfig

The tunable configuration handed to this creator.

Trait Implementations

impl Clone for TensorWasmMemoryCreator

fn clone(&self) -> TensorWasmMemoryCreator

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Default for TensorWasmMemoryCreator

fn default() -> Self

Returns the “default value” for a type.

impl MemoryCreator for TensorWasmMemoryCreator

fn new_memory( &self, _ty: MemoryType, minimum: usize, maximum: Option<usize>, reserved_size_in_bytes: Option<usize>, guard_size_in_bytes: usize, ) -> Result<Box<dyn LinearMemory>, String>

Create a new LinearMemory object from the specified parameters.