huggingface/candle · error

prediction_type not implemented yet: sample

Error message

prediction_type not implemented yet: sample

What it means

EulerAncestralDiscreteScheduler::step computes the original sample from the model output, but only Epsilon and VPrediction prediction types are implemented; PredictionType::Sample is explicitly rejected with this bail. It is a declared unsupported-path guard rather than a computational failure.

Source

Thrown at candle-transformers/src/models/stable_diffusion/euler_ancestral_discrete.rs:191

    /// Performs a backward step during inference.
    fn step(&mut self, model_output: &Tensor, timestep: usize, sample: &Tensor) -> Result<Tensor> {
        let step_index = self
            .timesteps
            .iter()
            .position(|&p| p == timestep)
            .ok_or_else(|| Error::Msg("timestep out of this schedulers bounds".to_string()))?;

        let sigma_from = &self.sigmas[step_index];
        let sigma_to = &self.sigmas[step_index + 1];

        // 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
        let pred_original_sample = match self.config.prediction_type {
            PredictionType::Epsilon => (sample - (model_output * *sigma_from))?,
            PredictionType::VPrediction => {
                ((model_output * (-sigma_from / (sigma_from.powi(2) + 1.0).sqrt()))?
                    + (sample / (sigma_from.powi(2) + 1.0))?)?
            }
            PredictionType::Sample => bail!("prediction_type not implemented yet: sample"),
        };

        let sigma_up = (sigma_to.powi(2) * (sigma_from.powi(2) - sigma_to.powi(2))
            / sigma_from.powi(2))
        .sqrt();
        let sigma_down = (sigma_to.powi(2) - sigma_up.powi(2)).sqrt();

        // 2. convert to a ODE derivative
        let derivative = ((sample - pred_original_sample)? / *sigma_from)?;
        let dt = sigma_down - *sigma_from;
        let prev_sample = (sample + derivative * dt)?;

        let noise = prev_sample.randn_like(0.0, 1.0)?;

        prev_sample + noise * sigma_up
    }

    fn add_noise(&self, original: &Tensor, noise: Tensor, timestep: usize) -> Result<Tensor> {

View on GitHub (pinned to d5fee525bf)

Solutions

  1. Set prediction_type to PredictionType::Epsilon (most SD checkpoints) or PredictionType::VPrediction in the scheduler config
  2. Use a different scheduler implementation in candle that supports Sample prediction if your model requires it
  3. Check the model's config for prediction_type and align it with what candle supports

Example fix

// before
let cfg = EulerAncestralDiscreteSchedulerConfig { prediction_type: PredictionType::Sample, .. };
// after
let cfg = EulerAncestralDiscreteSchedulerConfig { prediction_type: PredictionType::Epsilon, .. };
Defensive patterns

Strategy: validation

Validate before calling

match config.prediction_type {
    PredictionType::Epsilon | PredictionType::VPrediction => {},
    other => return Err(anyhow::anyhow!("prediction_type {:?} unsupported by euler_ancestral_discrete", other)),
}

Type guard

fn supports_sample_pred(p: &PredictionType) -> bool {
    matches!(p, PredictionType::Epsilon | PredictionType::VPrediction)
}

Try / catch

match scheduler.step(&model_out, t, &mut latents) {
    Err(e) if e.to_string().contains("prediction_type not implemented") => {
        anyhow::bail!("rebuild scheduler with Epsilon or VPrediction")
    }
    r => r?,
}

Prevention

When it happens

Trigger: Using a scheduler built with EulerAncestralDiscreteSchedulerConfig { prediction_type: PredictionType::Sample, .. } and calling step(), or loading a Stable Diffusion checkpoint/config whose prediction_type is 'sample'.

Common situations: Copied a config from another pipeline (e.g. x-prediction/sample-style models); defaulted prediction_type incorrectly when constructing the scheduler config; model card specifies 'sample' prediction which candle does not support for this scheduler.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/80a7367c25454929. Report an issue: GitHub.