Crate tenferro_ad 

Automatic differentiation APIs for tenferro.

This crate is the explicit opt-in boundary for traced and eager automatic differentiation. Primal graph construction and execution live in tenferro-runtime; tensor storage lives in tenferro-tensor, and CPU execution lives in tenferro-cpu.

Use EagerRuntime and EagerTensor for PyTorch-style immediate execution where tracked variables accumulate gradients after backward(). Use TracedTensorAdExt or AdContext for JAX-style graph transforms such as grad, vjp, and jvp on tenferro_runtime::TracedTensor values. AdContext is the explicit place to add extension AD rule sets for operation-family crates such as tenferro-linalg.

User-facing guides live at https://tensor4all.org/tenferro-rs/guides/autodiff.html and https://tensor4all.org/tenferro-rs/guides/choosing-an-api.html.

Examples

use tenferro_ad::AdContext;
use tenferro_runtime::TracedTensor;

let ad = AdContext::builder().build().unwrap();
let x = TracedTensor::from_vec_col_major(vec![], vec![3.0_f64]).unwrap();
let loss = (&x * &x).unwrap();
let dx = ad.grad(&loss, &x).unwrap();
assert_eq!(dx.rank, 0);

Re-exports

pub use traced::TracedTensorAdExt;

Modules

error

extension

Eager AD support for out-of-tree extension primitives.

traced

Structs

AdContext

Explicit automatic-differentiation context.

AdContextBuilder

Builder for AdContext.

ContextId

Opaque identifier for an eager AD runtime, used in Error::ContextMismatch.

DotGeneralConfig

DotGeneral dimension configuration.

EagerRuntime

Shared eager execution context for tensors on a backend.

EagerRuntimeCacheStats

Stats for caches owned by an EagerRuntime.

EagerTensor

Eager tensor with reverse-mode autodiff over concrete tensor values.

GatherConfig

StableHLO gather dimension configuration.

PadConfig

StableHLO pad configuration.

ScatterConfig

StableHLO scatter dimension configuration.

SliceConfig

Slice configuration.

Enums

CompareDir

Comparison direction.

DType

Runtime scalar dtype tag.

EagerBackend

Error

Errors produced by einsum, eval, and other tenferro operations.

Tensor

Dynamic tensor enum over the supported scalar types.

Type Aliases

Result

Result type alias for tenferro operations.