xai-org/x-algorithm · error · ValueError

Unsupported emb_table dtype: {embeddings.dtype}

Error message

Unsupported emb_table dtype: {embeddings.dtype}

What it means

_maybe_init_rows lazily fills uninitialized embedding rows by generating random float32 values and bit-casting them into the table's storage dtype: uint16 stores the high 16 bits (bfloat16-style truncation via values.view(np.uint32) >> 16) and uint32 stores the full bits. Any other dtype has no defined bit-cast path, so it raises rather than writing garbage. Reached from lookup_h2d_embeddings.

Source

Thrown at phoenix/xrex/inference/h2d.py:412

    )
    if embeddings.ndim != 2:
        raise ValueError(f"Expected 2D emb_table, got shape={embeddings.shape}")
    row_size = embeddings.shape[1]
    scale = 1.0 / math.sqrt(row_size)
    unique_rows = np.unique(row_indexes)
    for row in unique_rows:
        idx = int(row)
        first = int(embeddings[idx, 0])
        if first != 0:
            continue
        rng = np.random.default_rng(idx)
        values = rng.uniform(-scale, scale, size=row_size).astype(np.float32)
        if embeddings.dtype == np.uint16:
            embeddings[idx] = (values.view(np.uint32) >> 16).astype(np.uint16)
        elif embeddings.dtype == np.uint32:
            embeddings[idx] = values.view(np.uint32)
        else:
            raise ValueError(f"Unsupported emb_table dtype: {embeddings.dtype}")

View on GitHub (pinned to 24c60942c5)

Solutions

1. Re-create the table with the dtype from create_h2d_state ('bfloat16'->np.uint16, 'float32'->np.uint32)
2. Convert an existing table: table.astype(np.uint32) for float32 values or proper bfloat16 bit conversion for uint16
3. If a new storage dtype is genuinely needed, add an explicit branch in _maybe_init_rows with a correct bit-cast

Example fix

# before
table = np.zeros((vocab, dim), dtype=np.float32)
rows = lookup_h2d_embeddings(idx, table, ...)  # lazy init -> error
# after
table = np.zeros((vocab, dim), dtype=np.uint32)
rows = lookup_h2d_embeddings(idx, table, ...)

import numpy as np
assert table.dtype in (np.uint16, np.uint32), f"bad table dtype {table.dtype}"

def is_supported_table_dtype(a: np.ndarray) -> bool:
    return a.dtype in (np.dtype(np.uint16), np.dtype(np.uint32))

When it happens

Trigger: Calling lookup_h2d_embeddings with an emb_table whose dtype is float32, float16, int32, or anything other than np.uint16/np.uint32 while some row indexes are uninitialized. Note a table created via create_h2d_state will always be uint16/uint32, so this only happens with externally supplied tables.

Common situations: Passing a native float32 numpy table built for debugging instead of the uint view; reusing a dtype from a different pipeline; a checkpoint conversion step that emits float32 arrays directly.

Related errors

• Expected 2D emb_table, got shape={embeddings.shape}
• rows must contain only 1-dim uint8 numpy arrays
• type checking expression %s failed: invalid argument type: e
• {key}: got {arr.shape}/{arr.dtype}, manifest says {meta['sha
• Only bfloat16 is supported for keys.

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/04953e1b4350b614. Report an issue: GitHub.