lanczos3 interpolation backward is not supported by PyTorch/

Error message

lanczos3 interpolation backward is not supported by PyTorch/tch backend

What it means

The burn-tch backend does not implement the backward pass for Lanczos3 interpolation; LibTorch exposes no lanczos upsampling backward, so the backend panics. Forward Lanczos3 is likewise unsupported in tch (module.rs:433). It is an unimplemented-feature panic, hit only during gradient computation.

Source

Thrown at crates/burn-tch/src/ops/module.rs:481

            }
            InterpolateMode::Bilinear => tch::Tensor::upsample_bilinear2d_backward(
                &grad.tensor,
                output_size,
                input_size,
                align_corners,
                None,
                None,
            ),
            InterpolateMode::Bicubic => tch::Tensor::upsample_bicubic2d_backward(
                &grad.tensor,
                output_size,
                input_size,
                align_corners,
                None,
                None,
            ),
            InterpolateMode::Lanczos3 => {
                panic!("lanczos3 interpolation backward is not supported by PyTorch/tch backend")
            }
        };

        TchTensor::new(tensor)
    }

    fn attention(
        query: TchTensor,
        key: TchTensor,
        value: TchTensor,
        mask: Option<TchTensor>,
        attn_bias: Option<TchTensor>,
        options: AttentionModuleOptions,
    ) -> TchTensor {
        if attn_bias.is_some() {
            return attention_fallback::<Self>(query, key, value, mask, attn_bias, options);
        }

View on GitHub (pinned to d16f7ba2ed)

Solutions

Change the interpolate mode to Bilinear or Bicubic, which have supported tch backward kernels
Use InterpolateMode::Nearest if exactness of lanczos filtering is not required
Run the training/backward graph on a backend that implements Lanczos3 (e.g. burn-cube/wgpu) and keep tch for inference only
Validate the interpolation mode against the selected backend at startup and fail early with a clear message

Example fix

// before
let options = InterpolateOptions::new(InterpolateMode::Lanczos3);
// after
let options = InterpolateOptions::new(InterpolateMode::Bicubic);

Defensive patterns

Strategy: validation

Validate before calling

fn ensure_interpolate_backward_supported(options: &InterpolateOptions) {
    match options.mode {
        InterpolateMode::Lanczos3 | InterpolateMode::NearestExact => {
            panic!("mode {:?} has no backward in burn-tch; use Nearest/Bilinear/Bicubic", options.mode)
        }
        _ => {}
    }
}

Try / catch

let result = std::panic::catch_unwind(|| loss.backward());
if result.is_err() {
    eprintln!("unsupported interpolation backward; rebuild graph with Bicubic");
}

Prevention

Restrict training configs to modes with tch backward support: Nearest, Bilinear, Bicubic
Keep Lanczos3 configs separate per backend (inference-only on tch)
Test backward passes of resize layers whenever you change backend
Validate InterpolateOptions against the backend before constructing the training loop

When it happens

Trigger: Backward pass of an interpolate/upsample op configured with InterpolateOptions { mode: InterpolateMode::Lanczos3 } under the tch (LibTorch) backend — i.e. training or grad computation on a resize layer set to lanczos3.

Common situations: Reusing a configuration or model definition written for a backend that supports Lanczos3 (e.g. burn-cube/wgpu) and then training it on burn-tch; switching backends for GPU/CPU training without revisiting interpolation settings.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/ef4c64fe358cb9fb. Report an issue: GitHub.