Module tensor 

Zero-copy device tensor views (§5.4) and their DLPack alignment (§5.3).

These are thin, non-owning views over device-resident memory used at the kernel boundary. They intentionally use raw device pointers; the owning session/EP guarantees the backing memory outlives the view.

DLPack correspondence (§5.3)

A TensorView carries exactly the fields a DLPack DLTensor needs to be reconstructed, so import/export is a field-wise mapping rather than a copy:

DLTensor	TensorView
data	TensorView::data
byte_offset	TensorView::byte_offset
shape / ndim	TensorView::shape
strides (elements)	TensorView::strides
device	TensorView::device
dtype	TensorView::dtype

DLPack keeps byte_offset separate from data (the base pointer is the allocation start; byte_offset selects the element origin) and permits negative strides. Both are representable here: strides are i64 and the offset is applied lazily by TensorView::data_ptr. Use TensorView::validate to check the invariants of an imported view before handing it to a kernel.

Structs

DevicePtr

An opaque immutable device pointer (a host pointer for CPU tensors).

DevicePtrMut

An opaque mutable device pointer.

ExternalMmapRegion

One validated read-only range in an ONNX external-data mmap.

TensorMut

Mutable, non-owning view of a tensor on any device.

TensorView

Immutable, non-owning view of a tensor on any device.

Enums

TensorBacking

Provenance relevant to kernels that can consume lazy external weights.