huggingface/candle · error

top_p must be between 0 and 1, got {}

Error message

top_p must be between 0 and 1, got {}

What it means

Raised when the optional top_p (nucleus sampling) value in the Voxtral GenerationConfig lies outside the closed interval [0.0, 1.0]. top_p is a cumulative-probability cutoff, so any value outside 0..=1 is invalid; the library checks this at candle_transformers/src/models/voxtral/model.rs:893 and bails before sampling starts.

Source

Thrown at candle-transformers/src/models/voxtral/model.rs:893

        input_ids: &Tensor,
        input_features: Option<&Tensor>,
        config: VoxtralGenerationConfig,
    ) -> Result<Vec<u32>> {
        // Validate inputs
        if config.max_new_tokens == 0 {
            return input_ids.i(0)?.to_vec1::<u32>(); // Get first batch
        }

        if config.temperature < 0.0 {
            candle::bail!(
                "Temperature must be non-negative, got {}",
                config.temperature
            );
        }

        if let Some(p) = config.top_p {
            if !(0.0..=1.0).contains(&p) {
                candle::bail!("top_p must be between 0 and 1, got {}", p);
            }
        }

        let mut final_cache = if let Some(cache) = config.cache {
            cache
        } else {
            // Get the dtype from the language model by creating a small embedding
            let dummy_token = Tensor::new(&[1u32], &config.device)?;
            let dummy_embed = self.language_model.embed(&dummy_token)?;
            let model_dtype = dummy_embed.dtype();
            VoxtralCache::new(true, model_dtype, &self.text_config, &config.device)?
        };
        let mut tokens = input_ids.i(0)?.to_vec1::<u32>()?; // Get first batch
        let initial_len = tokens.len();

        for idx in 0..config.max_new_tokens {
            let (input, index_pos) = if idx == 0 {
                (input_ids.clone(), 0)

View on GitHub (pinned to d5fee525bf)

Solutions

1. Set top_p to a value in 0.0..=1.0 (e.g. Some(0.9)); use None to disable nucleus sampling
2. Clamp before calling: config.top_p = config.top_p.map(|p| p.clamp(0.0, 1.0))
3. Fix deserialized config files so top_p is expressed as a fraction, not a percentage
4. If you meant to disable top-p filtering, set the field to None rather than 0 or a negative sentinel

Example fix

// before
let config = GenerationConfig { top_p: Some(50.0), ..Default::default() };
// after
let config = GenerationConfig { top_p: Some(0.9), ..Default::default() };

fn ensure_valid_top_p(p: Option<f64>) -> Result<Option<f64>, String> {
    match p {
        Some(v) if !(0.0..=1.0).contains(&v) => {
            Err(format!("top_p must be between 0 and 1, got {}", v))
        }
        other => Ok(other),
    }
}

fn is_valid_top_p(p: f64) -> bool { (0.0..=1.0).contains(&p) && p.is_finite() }

let top_p = ensure_valid_top_p(config.top_p)
    .map_err(|e| eprintln!("invalid sampling config: {e}"))
    .ok();

When it happens

Trigger: Calling Voxtral generation with Some(p) for config.top_p where p < 0.0 or p > 1.0 (e.g. top_p: Some(1.5) or Some(-0.1)). top_p = None skips the check entirely.

Common situations: Confusing top_p with top_k (an integer count) and passing a value like 50; scaling bugs that multiply top_p by a factor; hand-edited config files with percentages (e.g. 90 instead of 0.9); sentinel values like -1 meaning 'disabled' instead of using None.

Related errors

• Temperature must be non-negative, got {}
• text_embeddings cannot be empty
• {} is a dummy type and cannot be constructed
• {} is a dummy type and cannot be converted
• {} is a dummy type and cannot be converted to scalar

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/1b0779135e8f25d7. Report an issue: GitHub.