What causes "Requires `autodiff` feature" in tracel-ai/burn?

Enable the feature: add `features = ["autodiff"]` to the burn-dispatch/burn dependency in Cargo.toml and rebuild.

Requires `autodiff` feature

Error message

Requires `autodiff` feature

What it means

The Dispatch backend (burn-dispatch) is compiled without the `autodiff` feature, so AutodiffBackend::backward is a stub that always panics with unimplemented!. The trait impl exists to satisfy the type system, but gradients are genuinely unavailable in this build.

Source

Thrown at crates/burn-dispatch/src/backend.rs:1027

            true,
        );
        let output = Dispatch::float_mul(x.clone(), x.clone());
        let gradients = <Dispatch as AutodiffBackend>::backward(output);
        let gradient = <Dispatch as AutodiffBackend>::grad(&x, &gradients).unwrap();
        assert_eq!(gradient.autodiff, DispatchAutodiffContext::Disabled);
        assert!(!matches!(gradient.kind, DispatchTensorKind::Autodiff(_)));
    }
}

// NOTE: placeholder for autodiff module requirements
#[cfg(not(feature = "autodiff"))]
impl AutodiffBackend for Dispatch {
    type InnerBackend = Dispatch;

    type Gradients = bool;

    fn backward(_tensor: DispatchTensor) -> Self::Gradients {
        unimplemented!("Requires `autodiff` feature")
    }

    fn grad(_tensor: &DispatchTensor, _grads: &Self::Gradients) -> Option<DispatchTensor> {
        unimplemented!("Requires `autodiff` feature")
    }

    fn grad_remove(
        _tensor: &DispatchTensor,
        _grads: &mut Self::Gradients,
    ) -> Option<DispatchTensor> {
        unimplemented!("Requires `autodiff` feature")
    }

    fn grad_replace(_tensor: &DispatchTensor, _grads: &mut Self::Gradients, _grad: DispatchTensor) {
        unimplemented!("Requires `autodiff` feature")
    }

    fn inner(_tensor: DispatchTensor) -> DispatchTensor {

View on GitHub (pinned to d16f7ba2ed)

Solutions

Enable the feature: add `features = ["autodiff"]` to the burn-dispatch/burn dependency in Cargo.toml and rebuild.
Use a backend that implements autodiff (e.g. Autodiff<Backend>) directly for training workloads.
Restrict backward/grad calls to code paths compiled under cfg(feature = "autodiff").
If only inference is needed, remove the training path so backward is never invoked.

Example fix

// before (Cargo.toml)
burn-dispatch = "0.x"
// after
burn-dispatch = { version = "0.x", features = ["autodiff"] }

Defensive patterns

Strategy: validation

Validate before calling

#[cfg(not(feature = "autodiff"))]
compile_error!("training code requires the `autodiff` feature on burn-dispatch");

Prevention

Enable `features = ["autodiff"]` whenever training code is compiled
Never call backward in inference-only builds
Add a compile_error!/cfg guard in binaries that need gradients
Separate train and infer binaries with distinct feature sets

When it happens

Trigger: Calling t.backward() (or any training-loop code invoking AutodiffBackend::backward) on a DispatchTensor when the burn-dispatch crate was built without the autodiff feature.

Common situations: Running inference-oriented builds (default features trimmed) then executing training code; Cargo.toml missing `features = ["autodiff"]` on the burn/burn-dispatch dependency; release profiles that disable default features.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/3457c0db97d01ed3. Report an issue: GitHub.