huggingface/candle · error

seqlens_q has to be contiguous

Error message

seqlens_q has to be contiguous

What it means

The varlen path requires seqlens_q to be a contiguous tensor so it can take a single flat slice via contiguous_offsets(). Non-contiguous layouts (strided/sliced views) cannot be handed to the kernel directly.

Source

Thrown at candle-flash-attn/src/lib.rs:488

        k: &candle::CudaStorage,
        k_l: &Layout,
        v: &candle::CudaStorage,
        v_l: &Layout,
        is_bf16: bool,
    ) -> Result<(candle::CudaStorage, Shape)> {
        // https://github.com/Dao-AILab/flash-attention/blob/184b992dcb2a0890adaa19eb9b541c3e4f9d2a08/csrc/flash_attn/flash_api.cpp#L327
        let dev = q.device();
        let out_shape = q_l.shape().clone();
        let out_l = Layout::contiguous(&out_shape);

        let (seqlens_q, seqlens_q_layout) = self.seqlens_q.storage_and_layout();
        let seqlens_q = match &*seqlens_q {
            candle::Storage::Cuda(c) => c.as_cuda_slice::<u32>()?, // Should be i32!
            _ => candle::bail!("seqlens_q must be a cuda tensor"),
        };
        let seqlens_q = match seqlens_q_layout.contiguous_offsets() {
            Some((o1, o2)) => seqlens_q.slice(o1..o2),
            None => candle::bail!("seqlens_q has to be contiguous"),
        };

        let (seqlens_k, seqlens_k_layout) = self.seqlens_k.storage_and_layout();
        let seqlens_k = match &*seqlens_k {
            candle::Storage::Cuda(c) => c.as_cuda_slice::<u32>()?, // Should be i32!
            _ => candle::bail!("seqlens_k must be a cuda tensor"),
        };
        let seqlens_k = match seqlens_k_layout.contiguous_offsets() {
            Some((o1, o2)) => seqlens_k.slice(o1..o2),
            None => candle::bail!("seqlens_k has to be contiguous"),
        };

        let block_table = if let Some(block_table) = self.block_table.as_ref() {
            let (block_table_storage, block_table_layout) = block_table.storage_and_layout();
            match &*block_table_storage {
                candle::Storage::Cuda(_) => {}
                _ => candle::bail!("block_table must be a cuda tensor"),
            }

View on GitHub (pinned to d5fee525bf)

Solutions

1. Call .contiguous() on seqlens_q before passing it
2. Build seqlens fresh as a flat 1D contiguous tensor (Tensor::new/from_vec gives contiguous layout)
3. Avoid narrow/slice views for seqlens; copy the needed values into a new tensor

Example fix

// before
let seqlens_q = padded_seqlens.narrow(1, 0, batch + 1)?; // non-contiguous view
// after
let seqlens_q = padded_seqlens.narrow(1, 0, batch + 1)?.contiguous()?;

fn ensure_contiguous(t: &Tensor) -> candle::Result<Tensor> {
    if t.is_contiguous() { Ok(t.clone()) } else { t.contiguous() }
}
let seqlens_q = ensure_contiguous(&seqlens_q)?;

fn contiguous_cuda_u32(t: &Tensor) -> bool {
    t.is_contiguous() && t.device().is_cuda() && t.dtype() == candle::DType::U32
}

match flash_attn_varlen(&q, &k, &v, &seqlens_q.contiguous()?, &seqlens_k.contiguous()?, scale, max_q, max_k, None, None, None) {
    Err(e) if e.to_string().contains("has to be contiguous") => { /* log tensor layout, rebuild */ }
    r => r?,
}

When it happens

Trigger: Passing seqlens_q produced by narrow/slice/pad operations that leave a non-contiguous layout, or .transpose/.index_select views without .contiguous().

Common situations: Deriving cu_seqlens from a padded batch tensor by slicing rows/columns, reusing cached seqlens tensors that were reshaped with non-unit strides.

Related errors

• seqlens_k has to be contiguous
• seqlens_q has to be contiguous
• seqlens_k has to be contiguous
• seqlens_q must be a cuda tensor
• seqlens_k must be a cuda tensor

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/6a87c67eb7f7ec44. Report an issue: GitHub.