tracel-ai/burn · error

deform_conv2d_backward: unsupported dtype {:?}

Error message

deform_conv2d_backward: unsupported dtype {:?}

What it means

deform_conv2d_backward mirrors the forward deform_conv2d dispatch: it matches the input dtype, casts tensors to f32 to run the f32 backward kernels, then casts gradients back with cast_from_f32. Unrecognized dtypes hit the catch-all panic. It fires during backward passes, so it typically surfaces inside .backward()/training loops rather than at the original call site.

Source

Thrown at crates/burn-flex/src/ops/module.rs:243

                    cast_to_f32(weight, to),
                    mask.map(|m| cast_to_f32(m, to)),
                    bias.map(|b| cast_to_f32(b, to)),
                    cast_to_f32(output_grad, to),
                    options.stride,
                    options.padding,
                    options.dilation,
                    options.weight_groups,
                    options.offset_groups,
                );
                (
                    cast_from_f32(xg, from),
                    cast_from_f32(og, from),
                    cast_from_f32(wg, from),
                    mg.map(|m| cast_from_f32(m, from)),
                    bg.map(|b| cast_from_f32(b, from)),
                )
            }
            dtype => panic!("deform_conv2d_backward: unsupported dtype {:?}", dtype),
        };
        DeformConv2dBackward::new(x_grad, offset_grad, weight_grad, mask_grad, bias_grad)
    }

    fn conv3d(
        x: FloatTensor<Flex>,
        weight: FloatTensor<Flex>,
        bias: Option<FloatTensor<Flex>>,
        options: ConvOptions<3>,
    ) -> FloatTensor<Flex> {
        match x.dtype() {
            DType::F32 => conv::conv3d_f32(x, weight, bias, &options),
            DType::F64 => conv::conv3d_f64(x, weight, bias, &options),
            DType::F16 => conv::conv3d_f16(x, weight, bias, &options),
            DType::BF16 => conv::conv3d_bf16(x, weight, bias, &options),
            dtype => panic!("conv3d: unsupported dtype {:?}", dtype),
        }
    }

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Ensure the forward deform_conv2d inputs are float (F32/F16/BF16) so the saved tensors used in backward keep a supported dtype.
2. Audit any cast/quantize ops between the forward call and backward() that could convert saved tensors to integers.
3. Reproduce by checking dtype of x_grad inputs in a minimal test: run forward, print .dtype(), then backward.
4. If a non-float dtype must be supported, extend the match arms in crates/burn-flex/src/ops/module.rs deform_conv2d_backward.

Example fix

// before
let grads = loss.backward(); // x saved as I8 by earlier quantize op
// panic: deform_conv2d_backward: unsupported dtype I8

// after
let x = quantized_x.dequantize().cast(burn::tensor::DType::F32);
let out = deform_conv2d(x, offset, mask, weight, bias, options);
let grads = loss.backward();

// before backward, ensure forward inputs were float
assert!(matches!(x_saved.dtype(), DType::F32 | DType::F64 | DType::F16 | DType::BF16),
        "saved input for deform_conv2d_backward must be float, got {:?}", x_saved.dtype());

fn backward_safe(t: &Tensor<Flex>) -> bool {
    matches!(t.dtype(), DType::F32 | DType::F64 | DType::F16 | DType::BF16)
}

let grads = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| loss.backward()))
    .unwrap_or_else(|_| { // recast saved tensors to f32 and rerun
        let x = x_saved.cast(DType::F32);
        deform_conv2d(x, offset, mask, w, b, opts); /* rebuild graph */ loss.backward()
    });

When it happens

Trigger: Running backward on a deform_conv2d node where the saved input/weight tensors carry a dtype other than F32/F64/F16/BF16 (e.g. an int dtype from a quantized or mis-cast forward); dtype of x changed between forward and backward.

Common situations: Training Deformable DETR / DCN models with a mixed-precision or quantization setup that downcasts activations to integers; a custom autograd graph where casts between forward and backward changed the recorded dtype.

Related errors

• conv1d: unsupported dtype {:?}
• conv2d: unsupported dtype {:?}
• deform_conv2d: unsupported dtype {:?}
• conv3d: unsupported dtype {:?}
• adaptive_avg_pool2d_backward: unsupported dtype {:?}

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/d59d5538758c7e08. Report an issue: GitHub.