What causes "compile with '--features flash-attn'" in huggingface/candle?

Recompile with the feature: `cargo run --release --features candle-transformers/flash-attn` (CUDA required).

compile with '--features flash-attn'

Error message

compile with '--features flash-attn'

What it means

granite.rs defines a `flash_attn` function whose fallback build (without the `flash-attn` cargo feature) panics with `unimplemented!`. When `CausalSelfAttention` forward selects the flash path because `config.use_flash_attn` is true, this stub is hit and the program aborts.

Source

Thrown at candle-transformers/src/models/granite.rs:212

    span: tracing::Span,
    span_rot: tracing::Span,
    max_position_embeddings: usize,
}

#[cfg(feature = "flash-attn")]
fn flash_attn(
    q: &Tensor,
    k: &Tensor,
    v: &Tensor,
    softmax_scale: f32,
    causal: bool,
) -> Result<Tensor> {
    candle_flash_attn::flash_attn(q, k, v, softmax_scale, causal)
}

#[cfg(not(feature = "flash-attn"))]
fn flash_attn(_: &Tensor, _: &Tensor, _: &Tensor, _: f32, _: bool) -> Result<Tensor> {
    unimplemented!("compile with '--features flash-attn'")
}

impl CausalSelfAttention {
    fn apply_rotary_emb(&self, x: &Tensor, index_pos: usize, cache: &Cache) -> Result<Tensor> {
        let _enter = self.span_rot.enter();
        let (_b_sz, _, seq_len, _hidden_size) = x.dims4()?;
        let cos = cache.cos.narrow(0, index_pos, seq_len)?;
        let sin = cache.sin.narrow(0, index_pos, seq_len)?;
        candle_nn::rotary_emb::rope(x, &cos, &sin)
    }

    fn forward(
        &self,
        x: &Tensor,
        index_pos: usize,
        block_idx: usize,
        cache: &mut Cache,
    ) -> Result<Tensor> {

View on GitHub (pinned to d5fee525bf)

Solutions

Recompile with the feature: `cargo run --release --features candle-transformers/flash-attn` (CUDA required).
Set `config.use_flash_attn = false` before creating the model.
If you cannot enable CUDA, keep the default attention implementation and accept slower inference.

Example fix

// before
cargo build --release
// after
cargo build --release --features candle-transformers/flash-attn

Defensive patterns

Strategy: validation

Validate before calling

#[cfg(not(feature = "flash-attn"))]
if config.use_flash_attn {
    return Err(anyhow::anyhow!("granite: rebuild with --features candle-transformers/flash-attn or set use_flash_attn=false"));
}

Type guard

fn flash_ok(cfg: &granite::Config) -> bool {
    !cfg.use_flash_attn || cfg!(feature = "flash-attn")
}

Try / catch

let config = if flash_ok(config) { config } else { config.with_flash_attn(false) };

Prevention

Validate config flags against enabled cargo features at startup
Keep CUDA-specific features behind explicit feature checks
Search checkpoint config.json for flash-attention keys when porting models
Run a one-token forward pass as a startup self-check

When it happens

Trigger: Running the Granite model with `use_flash_attn: true` in Config while compiled without `--features flash-attn`.

Common situations: Checkpoint config enables flash attention; default builds without the feature; attempting to use the feature on non-CUDA systems where candle-flash-attn cannot run.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/bcfc3dc43ab8863c. Report an issue: GitHub.