What causes "sliding window is not supported" in huggingface/candle?

Set use_sliding_window to false (or remove it / set it to null) in the SmolLM3Config before constructing the model

sliding window is not supported

Error message

sliding window is not supported

What it means

SmolLM3 attention layers reject model configurations that enable sliding-window attention. The candle implementation of SmolLM3 only implements global (full) attention, so when cfg.use_sliding_window is true (or Some(true)) it aborts construction of the attention layer via candle::bail!. This is an explicit unsupported-feature guard, not an unexpected failure.

Source

Thrown at candle-transformers/src/models/smol/smollm3.rs:174

    head_dim: usize,
    hidden_size: usize,
    // utils
    rotary_emb: Option<Arc<SmolLM3RotaryEmbedding>>,
    kv_cache: KvCache,
    // NoPE flag
    skip_rope: bool,
}

impl SmolLM3Attention {
    pub(crate) fn new(
        cfg: &Config,
        layer_idx: usize,
        rotary_emb: Option<Arc<SmolLM3RotaryEmbedding>>,
        vb: VarBuilder,
    ) -> Result<Self> {
        let use_sliding_window = cfg.use_sliding_window.unwrap_or(false);
        if use_sliding_window {
            candle::bail!("sliding window is not supported")
        }

        let head_dim = cfg.head_dim();
        let num_heads = cfg.num_attention_heads;
        let num_kv_heads = cfg.num_key_value_heads;
        let num_kv_groups = num_heads / num_kv_heads;

        let attention_bias = cfg.attention_bias.unwrap_or(false);

        let q_proj = linear_b(
            cfg.hidden_size,
            num_heads * head_dim,
            attention_bias,
            vb.pp("q_proj"),
        )?;

        let k_proj = linear_b(
            cfg.hidden_size,

View on GitHub (pinned to d5fee525bf)

Solutions

Set use_sliding_window to false (or remove it / set it to null) in the SmolLM3Config before constructing the model
Check whether a newer candle version has added sliding-window support for SmolLM3 and upgrade
Use a model checkpoint that does not rely on sliding-window attention

Example fix

// before
let cfg = SmolLM3Config { use_sliding_window: Some(true), .. };
// after
let cfg = SmolLM3Config { use_sliding_window: Some(false), .. };

Defensive patterns

Strategy: validation

Validate before calling

if cfg.use_sliding_window.unwrap_or(false) {
    return Err(anyhow::anyhow!("SmolLM3 sliding window unsupported; disable use_sliding_window"));
}

Try / catch

match SmolLM3Model::new(&cfg, vb) {
    Ok(m) => m,
    Err(e) if e.to_string().contains("sliding window is not supported") => {
        let mut cfg = cfg.clone(); cfg.use_sliding_window = Some(false);
        SmolLM3Model::new(&cfg, vb)?
    }
    Err(e) => return Err(e.into()),
}

Prevention

Always normalize use_sliding_window to false for SmolLM3 in candle
Sanitize HF config.json fields before mapping them into SmolLM3Config
Pin a candle version and check its SmolLM3 feature support before loading new checkpoints

When it happens

Trigger: Calling SmolLM3Attention::new (during model load) with a SmolLM3Config whose use_sliding_window field is Some(true), e.g. loading a checkpoint/config.json that specifies sliding-window attention.

Common situations: Loading a SmolLM3 variant released with sliding-window attention enabled in its HF config; hand-editing config.json to enable use_sliding_window; using a config ported from transformers where the flag defaults on for certain layers.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/2f9cbbea5866fc72. Report an issue: GitHub.