tracel-ai/burn · error

ctc_loss_backward: 2 * max_target_len + 1 = {} exceeds the k

Error message

ctc_loss_backward: 2 * max_target_len + 1 = {} exceeds the kernel's shared-memory alpha capacity ({}). Reduce target length or raise SHARED_ALPHA_CAPACITY.

What it means

RouterChannel::change_client_backend panics when the tensor description's dtype is neither float, int, nor bool. The router dispatches dtype-specific bridge calls (change_backend_float/int/bool) and has no branch for other dtypes (e.g. QFloat).

Source

Thrown at crates/burn-cubecl/src/kernel/ctc.rs:610

) -> (CubeTensor, CubeTensor, CubeTensor) {
    // Manual stride indexing below requires a contiguous physical layout;
    // fusion-produced tensors may arrive with layouts that break that
    // assumption. No-op when already contiguous.
    let log_probs = into_contiguous(log_probs);
    let targets = into_contiguous(targets);
    let input_lengths = into_contiguous(input_lengths);
    let target_lengths = into_contiguous(target_lengths);

    let log_probs_shape = log_probs.shape();
    let [max_input_length, batch_size, _c] = log_probs_shape.dims::<3>();
    let target_shape = targets.shape();
    let max_target_len = target_shape.dims::<2>()[1];
    let max_l_prime = 2 * max_target_len + 1;

    assert!(
        max_l_prime as u32 <= SHARED_ALPHA_CAPACITY,
        "ctc_loss_backward: 2 * max_target_len + 1 = {} exceeds the kernel's shared-memory \
         alpha capacity ({}). Reduce target length or raise SHARED_ALPHA_CAPACITY.",
        max_l_prime,
        SHARED_ALPHA_CAPACITY,
    );

    let hw_max = log_probs.client.properties().hardware.max_cube_dim.0;
    let cube_dim_x = (max_l_prime as u32).min(hw_max).min(256);

    let client = log_probs.client.clone();
    let device = log_probs.device.clone();
    let f_dtype = log_probs.dtype;
    let i_dtype = targets.dtype;

    // Pre-fill alpha/beta with -inf so positions the kernel doesn't touch
    // (s >= 2U+1, or t outside the valid range for an individual batch
    // element) are not read as stale zeros by the gradient composition.
    let shape_abt = Shape::new([max_input_length, batch_size, max_l_prime]);
    let neg_inf = InputScalar::new(f32::NEG_INFINITY, dtype_to_storage_type(f_dtype));
    let alpha_out = crate::ops::numeric::full_device_dtype(

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Do not migrate quantized tensors with change_client_backend; dequantize first, move, then re-quantize on the target backend.
2. Verify tensor dtype is Float/Int/Bool before attempting backend migration.
3. If this stems from a newly added DType variant, update burn-router to add the corresponding change_backend_* bridge branch.

Example fix

// before
handle.change_client_backend::<B2>(device); // panics for QFloat
// after
assert!(!matches!(desc.dtype, DType::QFloat(_)));
let float_t = dequantize(q_tensor);
let moved = float_t.change_client_backend::<B2>(device);

fn can_migrate(dtype: burn_tensor::DType) -> bool {
    matches!(dtype, burn_tensor::DType::F32 | burn_tensor::DType::F64
        | burn_tensor::DType::I32 | burn_tensor::DType::I64
        | burn_tensor::DType::U32 | burn_tensor::DType::U64
        | burn_tensor::DType::Bool)
}
// call: assert!(can_migrate(desc.dtype)) before change_client_backend

fn is_quantized(dtype: &burn_tensor::DType) -> bool {
    matches!(dtype, burn_tensor::DType::QFloat(_))
}

std::panic::catch_unwind(std::panic::AssertUnwindSafe(||
    handle.change_client_backend::<B2>(&device)
)).map_err(|_| anyhow::anyhow!("dtype not supported for backend migration"))

When it happens

Trigger: Calling change_client_backend (client backend migration, e.g. moving a tensor between backends/devices via register/change_client) on a tensor whose DType is not Float/Int/Bool — practically a quantized (QFloat) tensor.

Common situations: Moving a model containing quantized tensors between backends, or a new DType variant added upstream without updating the router's dtype dispatch chain.

Related errors

• Invalid broadcast shapes: Next grad shape {:?}, Previous gra
• Can't differentiate embedding backward.
• Can't differentiate linear_x_backward.
• Can't differentiate linear_weight_backward.
• Can't differentiate linear_bias_backward.

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/98e8605c71c1dd9c. Report an issue: GitHub.