xai-org/x-algorithm · error · PyTypeError

rows must contain only 1-dim uint8 numpy arrays

Error message

rows must contain only 1-dim uint8 numpy arrays

What it means

PyTypeError raised by embedding_gather in the xai-recsys-engine Rust extension when one of the rows passed in is not a 1-dimensional numpy array of dtype uint8. Each element of rows is extracted to PyReadwriteArray1<'py, u8>; anything else (list, wrong dtype, 2-D array) fails extraction and triggers this error.

Source

Thrown at phoenix/crates/serving/xai-recsys-engine/src/emb_table.rs:414

    Ok(())
}

#[pyfunction]
pub fn embedding_gather<'py>(
    py: Python<'py>,
    table: PyReadonlyArray1<'py, u8>,
    row_indexes: PyReadonlyArray1<'py, u32>,
    rows: Bound<'py, PyTuple>,
    row_size: usize,
    num_threads: usize,
) -> PyResult<()> {
    let table_slice = table.as_slice()?;
    let row_index_slice = row_indexes.as_slice()?;

    let mut arrs: Vec<PyReadwriteArray1<'py, u8>> = Vec::with_capacity(rows.len());
    let err = || PyTypeError::new_err("rows must contain only 1-dim uint8 numpy arrays");
    for item in rows.iter() {
        arrs.push(item.extract().map_err(|_| err())?);
    }
    let mut row_slices: Vec<&mut [u8]> = Vec::with_capacity(arrs.len());
    for item in arrs.iter_mut() {
        row_slices.push(item.as_slice_mut()?);
    }

    py.detach(|| {
        embedding_gather_into(
            table_slice,
            row_index_slice,
            &mut row_slices,
            row_size,
            num_threads,
        )
    })
    .map_err(|e| PyValueError::new_err(e.0))
}

View on GitHub (pinned to 24c60942c5)

Solutions

1. Ensure every element of rows is numpy 1-D with dtype uint8: np.asarray(row, dtype=np.uint8)
2. Check for accidental 2-D arrays (shape (1, d)) — squeeze or index them to 1-D
3. If you have float embeddings, quantize/cast them explicitly before calling

Example fix

# before
rows = [emb.numpy() for emb in embs]  # float32
# after
rows = [np.asarray(emb, dtype=np.uint8) for emb in embs]

import numpy as np
rows = [np.asarray(r, dtype=np.uint8) for r in rows]

def is_valid_rows(rows) -> bool:
    return all(isinstance(r, np.ndarray) and r.dtype == np.uint8 and r.ndim == 1 for r in rows)

try: embedding_gather(rows, ...) except TypeError as e: raise ValueError('rows must be 1-D uint8 arrays') from e

When it happens

Trigger: Calling emb_table.embedding_gather(rows, ...) where rows contains a Python list, a numpy array with dtype != uint8 (e.g. float32/int64), or a multi-dimensional array instead of np.array(..., dtype=np.uint8) 1-D vectors.

Common situations: Feeding embeddings produced by a float model directly; converting from torch tensors with .numpy() without casting to uint8; accidentally passing row indexes or nested lists.

Related errors

• Unsupported emb_table dtype: {embeddings.dtype}
• {key}: got {arr.shape}/{arr.dtype}, manifest says {meta['sha
• Only bfloat16 is supported for keys.
• {name} must have dtype torch.int32
• Type mismatch: {self.q_dtype} != {self.k_dtype}

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/0cc2ba200c351e0a. Report an issue: GitHub.