Split reshape of ND block-quantized tensor is not yet suppor

Error message

Split reshape of ND block-quantized tensor is not yet supported.

What it means

A Split reshape ([32,4] -> [32,2,2]) of an ND block-quantized tensor would split data inside existing quantization blocks, invalidating the scale layout. This case is explicitly not implemented and panics.

Source

Thrown at crates/burn-cubecl/src/ops/base.rs:392

        match analysis_values {
            ReshapeAnalysis::IsContiguous => {
                if let Some(block_size) = scheme.block_size()
                    && block_size.len() > 1
                    && !is_unsqueeze
                {
                    // General reshape (e.g. [32, 4] -> [16, 8]): only valid if
                    // reshaped dimension is aligned with the block boundaries.
                    unimplemented!("Reshape of ND block-quantized tensor is not yet supported.");
                }
            }
            ReshapeAnalysis::Broadcasted => {} // only preprends unit dims
            ReshapeAnalysis::Split => {
                if let Some(block_size) = scheme.block_size()
                    && block_size.len() > 1
                {
                    // Split reshape (e.g. [32, 4] -> [32, 2, 2]): only valid if
                    // reshaped dimension is aligned with the block boundaries.
                    unimplemented!(
                        "Split reshape of ND block-quantized tensor is not yet supported."
                    );
                }
            }
            other => unreachable!("Reshape analysis {other:?} should not update strides."),
        }
    }

    let shape_last = *shape.last().unwrap();

    // The per-tensor scale is a scalar in its own region, so only the block grid moves.
    let shape_scales = match scheme.block_size() {
        None => scales.meta.shape().clone(), // always [1], invariant under reshape
        Some(block_size) if block_size.len() == 1 && shape_last < (block_size[0] as usize) => {
            // If the new last dimension is smaller than the block size,
            // it means a single block now spans across multiple rows.
            if scales.meta.shape().num_elements() > 1 {
                unimplemented!("Reshape would split a block across multiple rows.");

View on GitHub (pinned to d16f7ba2ed)

Solutions

Dequantize before splitting, then re-quantize the new shape
Avoid splitting dimensions covered by ND quantization blocks
Use 1D block sizes so split reshapes (aligned) are permitted
Upgrade burn and retest

Example fix

// before
let heads = x.reshape([b, h, s, d]); // split of ND block-quant -> panic
// after
let heads = x.dequantize().reshape([b, h, s, d]);

Defensive patterns

Strategy: validation

Validate before calling

fn split_reshape_safe(scheme: &QuantScheme, analysis: ReshapeAnalysis) -> bool {
    !(analysis == ReshapeAnalysis::Split && scheme.block_size().map_or(false, |b| b.len() > 1))
}

Type guard

fn supports_split(scheme: &QuantScheme) -> bool {
    scheme.block_size().map_or(true, |b| b.len() <= 1)
}

Try / catch

// Guard before split reshape:
if supports_split(&scheme) { x.reshape(split_shape) } else { x.dequantize().reshape(split_shape) }

Prevention

Do not split dims covered by ND quant blocks
Dequantize before adding dimensions (heads, groups)
Use 1D blocks where split reshapes are needed

When it happens

Trigger: Calling q_reshape where the reshape analysis classifies as Split and the quantization scheme has block_size with more than one dimension (ND blocks).

Common situations: Splitting batched/feature dims of ND-block-quantized activations; reshaping to add heads/groups dimension in attention code on quantized tensors.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/92345fd30c92b359. Report an issue: GitHub.