tracel-ai/burn · error

an autodiff float primitive must have an enabled autodiff co

Error message

an autodiff float primitive must have an enabled autodiff context

What it means

In q_matmul_fq_arms, when the lhs float tensor is stored as an Autodiff primitive (DispatchTensorKind::Autodiff), the code requires the dispatch autodiff context to be DispatchAutodiffContext::Enabled so it can unwrap the gradation strategy. If the context is Disabled while the tensor is an autodiff primitive, the invariant is broken and burn panics: an autodiff-wrapped tensor cannot be used without its gradient-tracking context.

Source

Thrown at crates/burn-dispatch/src/ops/qtensor.rs:125

                                    TensorPrimitive::QFloat(rhs.quantized()),
                                );
                                wrap_q_matmul_autodiff!($Backend, output, strategy)
                            })
                        }
                        #[cfg(not(feature = "autodiff"))]
                        DispatchAutodiffContext::Enabled(_) => {
                            panic!("autodiff context requires the `autodiff` feature")
                        }
                    }
                }
            )*
            #[cfg(feature = "autodiff")]
            (DispatchTensorKind::Autodiff(lhs), rhs) => match (*lhs, rhs) {
                $(
                    #[cfg($cfg)]
                    (DispatchTensorKind::$Backend(lhs), DispatchTensorKind::$Backend(rhs)) => {
                        let DispatchAutodiffContext::Enabled(strategy) = $autodiff else {
                            panic!("an autodiff float primitive must have an enabled autodiff context")
                        };
                        with_autodiff_backend!($Backend, strategy, |B| {
                            let output = B::q_matmul(
                                TensorPrimitive::Float(lhs.autodiff()),
                                TensorPrimitive::QFloat(rhs.quantized()),
                            );
                            wrap_q_matmul_autodiff!($Backend, output, strategy)
                        })
                    }
                )*
                #[allow(unreachable_patterns)]
                _ => panic!("q_matmul inputs are on different backends"),
            },
            #[allow(unreachable_patterns)]
            _ => panic!("q_matmul inputs are on different backends"),
        }
    }};
}

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Enable the autodiff context (gradation) around the matmul so the context matches the tensor primitive
2. Use the tensor's inner float value (detach / .inner()) when working outside a gradation context
3. Ensure tensors produced by an Autodiff backend are only consumed within that same context
4. Rebuild the tensor through the same backend/context chain that created it

Example fix

// before
let out = dispatch::q_matmul(autodiff_lhs, q_rhs, DispatchAutodiffContext::Disabled);
// after
let out = dispatch::q_matmul(autodiff_lhs, q_rhs, DispatchAutodiffContext::Enabled(strategy));

fn autodiff_tensor_in_enabled_ctx(t: &DispatchTensor, ctx: &DispatchAutodiffContext) -> bool {
    !matches!(t.kind, DispatchTensorKind::Autodiff(_))
        || matches!(ctx, DispatchAutodiffContext::Enabled(_))
}

fn is_autodiff(t: &DispatchTensor) -> bool {
    matches!(t.kind, DispatchTensorKind::Autodiff(_))
}

When it happens

Trigger: Calling q_matmul where lhs.kind is DispatchTensorKind::Autodiff but the autodiff argument is DispatchAutodiffContext::Disabled — e.g. an autodiff tensor from a gradation-enabled scope used inside a context that was constructed with autodiff disabled.

Common situations: Using a tensor returned from an Autodiff backend scope inside a plain (non-grad) dispatch scope; constructing a DispatchTensor manually with mismatched autodiff context; caching autodiff tensors across contexts in a module or server that toggles gradation.

Related errors

• Autodiff float tensor is on the wrong backend (expected {bac
• Expected autodiff-wrapped float tensor for backend {backend}
• Autodiff should not wrap an autodiff device.
• an autodiff float primitive must have an enabled autodiff co
• autodiff context requires the `autodiff` feature

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/0a6380d39dfcde43. Report an issue: GitHub.