Module graph 

cuDNN v9 frontend graph builder — Rust-level “spec” objects that describe a backend descriptor DAG (tensors → ops → operation graph → engine config → execution plan) plus a plan cache keyed on op-shape signatures.

The spec layer is fully host-buildable. The actual cudnnBackendCreateDescriptor / cudnnBackendFinalize calls live in [Self::build_into] which only fires when a real cuDNN handle is plumbed in. Unit tests round-trip the spec without touching FFI.

What we build

TensorSpec*  ──► OpSpec  ──► OperationGraphSpec
                                  │
                                  ▼
                      EngineHeurSpec ──► EnginecfgSpec
                                             │
                                             ▼
                                      ExecutionPlanSpec
                                             │
                                             ▼
                                      VariantPackSpec

Plan-cache key is op-kind + dtype + tensor-spec digest, so two requests with identical shapes hit the same cached plan.

Structs

CachedPlan

Cached execution-plan handle. On a host without cuDNN this is just a marker that “this signature was prepared” — the actual BackendDescriptor lives only on a real GPU build.

OperationGraphSpec

Top-level operation graph.

PlanCache

LRU plan cache. The cuDNN actor wraps this in a Mutex and shares one instance across all op kinds — entries are tagged by op_kind in the key.

PlanCacheKey

Plan-cache key: op-kind + dtype + signature digest. Compact (24 bytes) so the LRU lookup is cheap.

TensorSpec

One tensor in a backend graph: unique id, dtype, dims, strides, alignment.

Enums

DtypeTag

cuDNN scalar dtype tag, decoupled from a T: CudaDtype parameter so spec-level objects are dyn-friendly.

NormMode

Normalisation kind.

NormPhase

Normalisation training phase.

OpSpec

One op in a backend graph. Each OpSpec references TensorSpecs by their uid; the actual TensorSpec values live on the parent OperationGraphSpec.

PointwiseMode

Pointwise op mode.

PoolKind

Pooling kind.

ReduceOp

Reduction op tag.

TensorLayout

Tensor layout: NCHW, NHWC, or fully arbitrary nd-strided.

Constants

DEFAULT_PLAN_CACHE_SIZE

Default LRU capacity for the plan cache (matches the existing cuDNN ConvForward cache + cuBLASLt heuristic cache).

Functions

cache_key

Build a PlanCacheKey from an op-kind tag + dtype + an OperationGraphSpec. The signature is the full graph digest.