docling-project/docling · error · ValueError

Unsupported numpy dtype for gRPC inline (non-binary) encodin

Error message

Unsupported numpy dtype for gRPC inline (non-binary) encoding: {tensor.dtype!s}. Supported non-binary dtypes: bool, uint8/uint16/uint32/uint64, int8/int16/int32/int64, float32/float64, BYTES.

What it means

Raised as ValueError by _encode_contents when building a non-binary (inline) KServe gRPC tensor from a numpy array whose dtype is not in the supported set (bool, u8-u64, i8-i64, f32/f64, object/BYTES). Notably float16 and bfloat16 are rejected.

Source

Thrown at docling/models/inference_engines/common/kserve_v2_grpc.py:131

    flat = tensor.flatten()
    if tensor.dtype == np.float32:
        contents.fp32_contents.extend(flat.tolist())
    elif tensor.dtype == np.float64:
        contents.fp64_contents.extend(flat.tolist())
    elif tensor.dtype in (np.int8, np.int16, np.int32):
        contents.int_contents.extend(flat.astype(np.int32).tolist())
    elif tensor.dtype == np.int64:
        contents.int64_contents.extend(flat.tolist())
    elif tensor.dtype in (np.uint8, np.uint16, np.uint32):
        contents.uint_contents.extend(flat.astype(np.uint32).tolist())
    elif tensor.dtype == np.uint64:
        contents.uint64_contents.extend(flat.tolist())
    elif tensor.dtype == np.bool_:
        contents.bool_contents.extend(flat.tolist())
    elif tensor.dtype == object:
        contents.bytes_contents.extend(encode_bytes_element(value) for value in flat)
    else:
        raise ValueError(
            f"Unsupported numpy dtype for gRPC inline (non-binary) encoding: {tensor.dtype!s}. "
            "Supported non-binary dtypes: bool, uint8/uint16/uint32/uint64, "
            "int8/int16/int32/int64, float32/float64, BYTES."
        )


def _decode_contents(
    contents: Any, np_dtype: np.dtype[Any], shape: tuple[int, ...]
) -> np.ndarray:
    """Decode an InferTensorContents message to a numpy array (non-binary path)."""
    canonical_dtype = np.dtype(np_dtype)

    if canonical_dtype == np.dtype(np.float32):
        data = list(contents.fp32_contents)
    elif canonical_dtype == np.dtype(np.float64):
        data = list(contents.fp64_contents)
    elif canonical_dtype in (
        np.dtype(np.int8),

View on GitHub (pinned to 61d76f1ff3)

Solutions

Cast to a supported dtype before the request: arr.astype(np.float32).
For half-precision pipelines, cast at the boundary right before sending.
If binary transport is enabled (use_binary_data=True), raw bytes are sent instead and the dtype restriction applies to the declared datatype string — but casting to fp32 is still the simplest fix.

Example fix

# before
payload = pixels.astype(np.float16)  # -> ValueError on encode

# after
payload = pixels.astype(np.float32)  # supported inline dtype

Defensive patterns

Strategy: validation

Validate before calling

import numpy as np
SUPPORTED = {np.bool_, np.uint8, np.uint16, np.uint32, np.uint64,
             np.int8, np.int16, np.int32, np.int64, np.float32, np.float64, np.dtype(object)}
if tensor.dtype not in SUPPORTED:
    tensor = tensor.astype(np.float32)

Type guard

import numpy as np

def is_supported_inline_dtype(arr: np.ndarray) -> bool:
    return arr.dtype in (np.bool_, np.uint8, np.uint16, np.uint32, np.uint64,
                         np.int8, np.int16, np.int32, np.int64,
                         np.float32, np.float64) or arr.dtype == object

Prevention

Cast tensors to float32/uint8 at the client boundary, especially from fp16 pipelines.
Standardize one preprocessing dtype for all remote calls.
Document the supported dtype set next to your request builder.

When it happens

Trigger: Sending a np.float16 image tensor, a float128, a datetime64, or a structured dtype through the inline tensor-contents path of the gRPC client.

Common situations: Half-precision preprocessing pipelines (fp16 common on GPUs); mixed-precision exports; object arrays containing non-bytes elements hitting the BYTES branch expectations.

Related errors

AI-assisted analysis of docling-project/docling@61d76f1ff3 (2026-08-14). Data as JSON: /api/errors/a206ea5a348a948e. Report an issue: GitHub.