What causes "SpectralNorm is not supported yet." in huggingface/candle?

Use a Mimi checkpoint whose conv layers use weight norm (the standard Mimi weights) instead of spectral norm

SpectralNorm is not supported yet.

Error message

SpectralNorm is not supported yet.

What it means

The Mimi conv1d layer supports no normalization, WeightNorm, and TimeGroupNorm, but SpectralNorm is not implemented in candle. When the conv config requests SpectralNorm, new bails immediately. This is an explicit unsupported-feature guard during model construction.

Source

Thrown at candle-transformers/src/models/mimi/conv.rs:80

        k_size: usize,
        causal: bool,
        norm: Option<Norm>,
        bias: bool,
        cfg: candle_nn::Conv1dConfig,
        vb: VarBuilder,
    ) -> Result<Self> {
        let conv = match norm {
            None | Some(Norm::TimeGroupNorm) => {
                if bias {
                    candle_nn::conv1d(in_c, out_c, k_size, cfg, vb.pp("conv"))?
                } else {
                    candle_nn::conv1d_no_bias(in_c, out_c, k_size, cfg, vb.pp("conv"))?
                }
            }
            Some(Norm::WeightNorm) => {
                conv1d_weight_norm(in_c, out_c, k_size, bias, cfg, vb.pp("conv"))?
            }
            Some(Norm::SpectralNorm) => candle::bail!("SpectralNorm is not supported yet."),
        };
        let norm = match norm {
            None | Some(Norm::WeightNorm) | Some(Norm::SpectralNorm) => None,
            Some(Norm::TimeGroupNorm) => {
                if causal {
                    candle::bail!("GroupNorm doesn't support causal evaluation.")
                }
                let norm = candle_nn::group_norm(1, out_c, 1e-5, vb.pp("norm"))?;
                Some(norm)
            }
        };
        Ok(Self {
            conv,
            norm,
            span: tracing::span!(tracing::Level::TRACE, "norm-conv1d"),
        })
    }
}

View on GitHub (pinned to d5fee525bf)

Solutions

Use a Mimi checkpoint whose conv layers use weight norm (the standard Mimi weights) instead of spectral norm
Change the norm field in the conv config to Norm::WeightNorm or None if your source model allows it
Implement spectral normalization in candle (power iteration on the weight) before enabling this path

Example fix

// before
let cfg = Conv1dConfig { norm: Some(Norm::SpectralNorm), .. };
// after
let cfg = Conv1dConfig { norm: Some(Norm::WeightNorm), .. };

Defensive patterns

Strategy: validation

Validate before calling

if cfg.norm == Some(Norm::SpectralNorm) {
    return Err("candle mimi does not support SpectralNorm; use WeightNorm or None");
}

Try / catch

let layer = Conv1d::new(in_c, out_c, k, cfg, causal, vb)
    .map_err(|e| format!("unsupported norm: {e}"))?;

Prevention

Verify checkpoint norm types before loading (spectral norm is unimplemented)
Prefer the standard Mimi weights, which use weight norm
Normalize configs to supported variants when converting from PyTorch

When it happens

Trigger: Constructing a Mimi conv1d layer (Conv1d::new) from a config whose norm field is Some(Norm::SpectralNorm), typically loaded from a Mimi checkpoint/ops config that specifies spectral normalization.

Common situations: Loading a Mimi/EnCodec-style codec checkpoint that used spectral norm in some layers; converting models that rely on spectral norm; reusing configs from PyTorch implementations where spectral norm is available.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/5ed9d5b15477ec2d. Report an issue: GitHub.