Crate tensor_wasm_mem 

CUDA Unified Memory allocator and Wasmtime MemoryCreator integration.

Two byte-buffer types:

• unified::UnifiedBuffer — CUDA Unified Memory when the unified-memory feature is enabled; a heap buffer otherwise.
• pinned_host::GuardedHostBuffer — guarded host memory for the explicit-fallback path used by cargo build --no-default-features. On all hosts the buffer is bracketed by PROT_NONE / PAGE_NOACCESS guard pages catching OOB reads/writes at the OS level. The managed memory path (UnifiedBuffer on CUDA hosts) still cannot use guard pages — that limitation is unchanged. The previous name PinnedHostBuffer is preserved as a deprecated alias.

Pools (pool::UnifiedMemoryPool) amortise the cost of CUDA allocations by carving sub-slices from a single slab. Hints (advise) forward to cudaMemAdvise on CUDA hosts and are no-ops otherwise.

Modules

advise

cudaMemAdvise hint helpers.

isolation

Memory isolation policy for TensorWasm Wasm instances.

pinned_host

GuardedHostBuffer — host buffer used when the unified-memory feature is disabled.

pool

UnifiedMemoryPool — a bump allocator that amortises the cost of CUDA Unified Memory allocations.

unified

UnifiedBuffer — a region of memory that is addressable from both CPU and GPU when a CUDA backing feature is enabled.

wasm_memory

Wasmtime MemoryCreator backed by UnifiedBuffer.