tracel-ai/burn · error

only float tensors may use an autodiff primitive

Error message

only float tensors may use an autodiff primitive

What it means

Generated by extract_extension (autodiff path): within an `Autodiff` dispatch variant, the inner `BackendTensor` must itself be the `Autodiff` variant (i.e. a float tensor under gradient tracking). The match panics when the inner tensor is Int, Bool, Quantized, or otherwise not an autodiff float — only float tensors can carry autodiff state.

Source

Thrown at crates/burn-backend-extension/src/routing.rs:772

    if autodiff {
        let target = ir::with_backend(ty, quote!(#backend_alias));
        let context = quote!(#dispatch_root::DispatchAutodiffContext);
        quote! {
            let #name = <#target as #extension_trait<#backend_alias>>::map_from_dispatch(#name, |__tensor| {
                let __input_context = __tensor.autodiff;
                match __tensor.kind {
                    #dispatch_kind::Autodiff(inner) => {
                        let #context::Enabled(_) = __input_context else {
                            panic!("an autodiff float primitive must have an enabled autodiff context")
                        };
                        let tensor = match *inner {
                            #dispatch_kind::#backend(tensor) => tensor,
                            #[allow(unreachable_patterns)]
                            _ => #mismatch,
                        };
                        match tensor {
                            #backend_tensor::Autodiff(tensor) => #backend_tensor::Float(tensor),
                            _ => panic!("only float tensors may use an autodiff primitive"),
                        }
                    }
                    #dispatch_kind::#backend(tensor) => match tensor {
                        #backend_tensor::Float(tensor) => {
                            let #context::Disabled = __input_context else {
                                panic!("an enabled float tensor must use an autodiff primitive")
                            };
                            #backend_tensor::Float(
                                <#backend_alias as #autodiff_trait>::from_inner(tensor)
                            )
                        }
                        #backend_tensor::Int(tensor) => #backend_tensor::Int(tensor),
                        #backend_tensor::Bool(tensor) => #backend_tensor::Bool(tensor),
                        #backend_tensor::Quantized(tensor) => #backend_tensor::Quantized(tensor),
                        #backend_tensor::Autodiff(_) => {
                            panic!("autodiff float input reached concrete dispatch")
                        }
                    },

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Pass non-float tensors as their concrete kind (not wrapped in Autodiff)
2. Verify upstream code did not change the tensor kind; keep autodiff wrapping only for float tensors
3. Adjust the routed signature so the input is declared with its actual kind

Example fix

// before
DispatchTensorKind::Autodiff(Box::new(DispatchTensorKind::Candle(BackendTensor::Int(t)))) // panics
// after
DispatchTensorKind::Candle(BackendTensor::Int(t)) // pass Int directly

if let DispatchTensorKind::Autodiff(inner) = &t.kind { assert!(matches!(**inner_inner_is_autodiff_float(inner)), "only float tensors may be autodiff-wrapped"); }

fn is_autodiff_float(t: &DispatchTensor) -> bool { matches!(&t.kind, DispatchTensorKind::Autodiff(inner) if matches!(inner.as_ref(), DispatchTensorKind::Target(BackendTensor::Autodiff(_)))) }

When it happens

Trigger: Wrapping a non-float tensor (Int indices, Bool mask, Quantized weights) in the Autodiff dispatch kind and passing it to an autodiff-routed extension input.

Common situations: Generic helper that wraps every input in `Autodiff(...)` regardless of dtype; tensor kind changed to Int/Bool upstream while routing metadata still says autodiff float; quantized models routed through autodiff primitives.

Related errors

• autodiff float input reached concrete dispatch
• Expected autodiff-wrapped float tensor for backend {backend}
• autodiff context requires the `autodiff` feature
• Operation not marked for autodiff.
• an autodiff float primitive must have an enabled autodiff co

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/f403679ffeaee712. Report an issue: GitHub.