What causes "conv_transpose3d: unsupported dtype {:?}" in tracel-ai/burn?

Cast the latent/input to float before conv_transpose3d: x.cast(DType::F32).

conv_transpose3d: unsupported dtype {:?}

Error message

conv_transpose3d: unsupported dtype {:?}

What it means

conv_transpose3d in the burn-flex backend dispatches on input dtype to conv_transpose3d_f32/f64/f16/bf16 kernels; any other dtype triggers the panic. As with all conv/conv-transpose ops, only float tensors are valid, and the dynamic-dtype backend detects violations only at runtime.

Source

Thrown at crates/burn-flex/src/ops/module.rs:304

            DType::F64 => conv_transpose::conv_transpose2d_f64(x, weight, bias, &options),
            DType::F16 => conv_transpose::conv_transpose2d_f16(x, weight, bias, &options),
            DType::BF16 => conv_transpose::conv_transpose2d_bf16(x, weight, bias, &options),
            dtype => panic!("conv_transpose2d: unsupported dtype {:?}", dtype),
        }
    }

    fn conv_transpose3d(
        x: FloatTensor<Flex>,
        weight: FloatTensor<Flex>,
        bias: Option<FloatTensor<Flex>>,
        options: ConvTransposeOptions<3>,
    ) -> FloatTensor<Flex> {
        match x.dtype() {
            DType::F32 => conv_transpose::conv_transpose3d_f32(x, weight, bias, &options),
            DType::F64 => conv_transpose::conv_transpose3d_f64(x, weight, bias, &options),
            DType::F16 => conv_transpose::conv_transpose3d_f16(x, weight, bias, &options),
            DType::BF16 => conv_transpose::conv_transpose3d_bf16(x, weight, bias, &options),
            dtype => panic!("conv_transpose3d: unsupported dtype {:?}", dtype),
        }
    }

    fn avg_pool2d(
        x: FloatTensor<Flex>,
        kernel_size: [usize; 2],
        stride: [usize; 2],
        padding: [usize; 2],
        count_include_pad: bool,
        ceil_mode: bool,
    ) -> FloatTensor<Flex> {
        match x.dtype() {
            DType::F32 => pool::avg_pool2d_f32(
                x,
                kernel_size,
                stride,
                padding,
                count_include_pad,

View on GitHub (pinned to d16f7ba2ed)

Solutions

Cast the latent/input to float before conv_transpose3d: x.cast(DType::F32).
If latents are persisted as integers, dequantize/cast immediately after loading, before the decoder stack.
Check .dtype() on input and weights just before the call to trace the origin of the mismatch.
Extend the dtype match in crates/burn-flex/src/ops/module.rs conv_transpose3d if a new dtype is required.

Example fix

// before
let vol = conv_transpose3d(stored_latent_i8, weight, bias, options);
// panic: conv_transpose3d: unsupported dtype I8

// after
let x = stored_latent_i8.cast(burn::tensor::DType::F32);
let vol = conv_transpose3d(x, weight, bias, options);

Defensive patterns

Strategy: validation

Validate before calling

if !matches!(x.dtype(), DType::F32 | DType::F64 | DType::F16 | DType::BF16) {
    x = x.cast(DType::F32);
}
let vol = conv_transpose3d(x, weight, bias, options);

Type guard

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

Try / catch

let vol = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| conv_transpose3d(x.clone(), w.clone(), b.clone(), opts.clone())))
    .unwrap_or_else(|_| conv_transpose3d(x.cast(DType::F32), w, b, opts));

Prevention

Store latents as float (or store scale/zero-point and dequantize on load).
Cast immediately after loading int-quantized 3D latents, before the decoder.
Keep a round-trip save/load test that asserts latent dtype is preserved.
Print dtype at each decoder stage when debugging new 3D models.

When it happens

Trigger: Calling conv_transpose3d with a tensor of dtype I8/I16/I32/I64/U8/Bool/etc.; decoding quantized 3D latents (volumetric GAN/autoencoder outputs) stored as integers.

Common situations: 3D generative models (medical volume synthesis, video voxel decoders) consuming int-quantized latents; data pipelines saving latents as int8 for storage and forgetting to convert on load.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/b445f87e330aac25. Report an issue: GitHub.