Module cudnn 

CudnnActor — Phase 2 cuDNN slice. Wraps a [cudarc::cudnn::Cudnn] handle and exposes the v9 frontend graph API plus legacy ConvForward / Activation / Softmax shims for back-compat.

Module layout

kernel/cudnn/
├── mod.rs           — CudnnActor, CudnnMsg, CudnnInner, props
├── graph.rs         — v9 frontend graph spec (TensorSpec, OpSpec,
│                     OperationGraphSpec) + plan cache
├── conv.rs          — ConvFwdRequest<T>, ConvBwdDataRequest<T>,
│                     ConvBwdFilterRequest<T>
├── norm.rs          — BatchNormRequest<T>, LayerNormRequest<T>,
│                     InstanceNormRequest<T>, GroupNormRequest<T>,
│                     NormBwdRequest<T>
├── pool.rs          — PoolFwdRequest<T>, PoolBwdRequest<T>
├── attention.rs     — MultiHeadAttnFwdRequest<T>,
│                     MultiHeadAttnBwdRequest<T>
├── rnn.rs           — RnnFwdRequest<T>, RnnBwdRequest<T>
└── activation.rs    — ActivationFwdRequest<T>, SoftmaxFwdRequest<T>,
                      DropoutFwdRequest<T>, LrnFwdRequest<T>

Op coverage

Family	Forward	Backward	Notes
Conv	✓	✓ data + filter	1D/2D/3D, NCHW + NHWC, groups, dilation
Pool	✓	✓	max, avg, avg-exclude-padding
BatchNorm	✓	✓	training, inference, persistent
LayerNorm	✓	✓	training, inference
InstanceNorm	✓	✓
GroupNorm	✓	✓
Activation	✓	(fused with conv epilogue)	relu, sigmoid, tanh, gelu, gelu_approx, swish, elu, softplus, identity
Softmax	✓	(planned)	instance + channel mode
Dropout	✓	(planned)
LRN	✓	(planned)
Attention	✓	✓	causal, sliding-window, GQA/MQA, dropout
RNN/LSTM/GRU	✓	✓	uni + bi, multi-layer, dropout

Dtype matrix

Every request type is generic over T: crate::dtype::CudnnSupported. Implementations cover f32, f64, i8, plus half::f16 and half::bf16 under the f16 feature.

Re-exports

pub use activation::ActivationFwdRequest;

pub use activation::ActivationKind;

pub use activation::DropoutFwdRequest;

pub use activation::LrnFwdRequest;

pub use activation::LrnParams;

pub use activation::SoftmaxFwdRequest;

pub use activation::SoftmaxMode;

pub use attention::AttentionMask;

pub use attention::AttentionParams;

pub use attention::MultiHeadAttnBwdRequest;

pub use attention::MultiHeadAttnFwdRequest;

pub use conv::ConvBwdDataRequest;

pub use conv::ConvBwdFilterRequest;

pub use conv::ConvDescParams;

pub use conv::ConvFwdRequest;

pub use conv::EpilogueKind;

pub use graph::cache_key;

pub use graph::CachedPlan;

pub use graph::DtypeTag;

pub use graph::NormMode;

pub use graph::NormPhase;

pub use graph::OpSpec;

pub use graph::OperationGraphSpec;

pub use graph::PlanCache;

pub use graph::PlanCacheKey;

pub use graph::PointwiseMode;

pub use graph::PoolKind;

pub use graph::ReduceOp;

pub use graph::TensorLayout;

pub use graph::TensorSpec;

pub use graph::DEFAULT_PLAN_CACHE_SIZE;

pub use norm::BatchNormRequest;

pub use norm::GroupNormRequest;

pub use norm::InstanceNormRequest;

pub use norm::LayerNormRequest;

pub use norm::NormBwdRequest;

pub use pool::PoolBwdRequest;

pub use pool::PoolFwdRequest;

pub use pool::PoolMode;

pub use pool::PoolParams;

pub use rnn::RnnBwdRequest;

pub use rnn::RnnDirection;

pub use rnn::RnnFwdRequest;

pub use rnn::RnnMode;

pub use rnn::RnnParams;

Modules

activation

Activation, dropout, and LRN requests for the cuDNN actor.

attention

Multi-head attention (cudnnFusedAttnFwd/cudnnFusedAttnBwd) request types.

conv

Convolution requests for the cuDNN actor (Phase 2 frontend graph API).

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.

norm

Normalisation requests for the cuDNN actor: BatchNorm (training + inference + persistent), LayerNorm, InstanceNorm, GroupNorm, RMSNorm.

pool

Pooling requests (max / avg, fwd + bwd) for the cuDNN actor.

rnn

RNN / LSTM / GRU forward + backward training requests.

Structs

ActivationRequest

Legacy F2 Activation request.

ConvForwardRequest

Legacy F2 ConvForward request (NCHW, f32 only).

ConvParams

Convolution parameters (cuDNN 2D conv subset).

CudnnActor

SoftmaxRequest

Legacy F2 Softmax request.

Enums

CudnnMsg

Mailbox message for CudnnActor.