PaddlePaddle/PaddleOCR · error · TypeError

Unsupported data type: {pd_dtype}

Error message

Unsupported data type: {pd_dtype}

What it means

_pd_dtype_to_np_dtype maps Paddle Inference tensor data types (FLOAT32/INT64/INT32/UINT8/INT8/FLOAT64) to numpy dtypes for building TRT input arrays. Any other Paddle DataType — e.g. FLOAT16, BFLOAT16, BOOL — reaches the else branch and raises TypeError. In practice this means the model contains an input whose precision has no numpy equivalent usable for engine calibration.

Source

Thrown at tools/infer/utility.py:534

    pp_model_path = pp_model_file.split(".")[0]
    convert(pp_model_path, trt_config)


def _pd_dtype_to_np_dtype(pd_dtype):
    if pd_dtype == inference.DataType.FLOAT64:
        return np.float64
    elif pd_dtype == inference.DataType.FLOAT32:
        return np.float32
    elif pd_dtype == inference.DataType.INT64:
        return np.int64
    elif pd_dtype == inference.DataType.INT32:
        return np.int32
    elif pd_dtype == inference.DataType.UINT8:
        return np.uint8
    elif pd_dtype == inference.DataType.INT8:
        return np.int8
    else:
        raise TypeError(f"Unsupported data type: {pd_dtype}")


def load_config(file_path):
    _, ext = os.path.splitext(file_path)
    if ext not in [".yml", ".yaml"]:
        raise ValueError(f"only support yaml files for now, got {file_path}")
    with open(file_path, "rb") as file:
        config = yaml.load(file, Loader=yaml.SafeLoader)
    return config


def get_output_tensors(args, mode, predictor):
    output_names = predictor.get_output_names()
    output_tensors = []
    if mode == "rec" and args.rec_algorithm in ["CRNN", "SVTR_LCNet", "SVTR_HGNet"]:
        output_name = "softmax_0.tmp_0"
        if output_name in output_names:
            return [predictor.get_output_handle(output_name)]

View on GitHub (pinned to 2661c7c0ef)

Solutions

1. Re-export the inference model in FP32 (remove --fp16/half precision at export time) before TRT conversion
2. Extend _pd_dtype_to_np_dtype with the needed mapping (e.g. FLOAT16 -> np.float16) if the data is representable
3. Skip TensorRT and run plain GPU Paddle Inference for this model

Example fix

# tools/infer/utility.py — before
    else:
        raise TypeError(f"Unsupported data type: {pd_dtype}")

# after
    elif pd_dtype == inference.DataType.FLOAT16:
        return np.float16
    else:
        raise TypeError(f"Unsupported data type: {pd_dtype}")

SUPPORTED = {inference.DataType.FLOAT64, inference.DataType.FLOAT32,
             inference.DataType.INT64, inference.DataType.INT32,
             inference.DataType.UINT8, inference.DataType.INT8}
for name in predictor.get_input_names():
    t = predictor.get_input_handle(name).type()
    assert t in SUPPORTED, f'input {name} has dtype {t}, not usable for TRT build — export FP32'

def dtype_supported(handle) -> bool:
    return handle.type() in {
        inference.DataType.FLOAT64, inference.DataType.FLOAT32,
        inference.DataType.INT64, inference.DataType.INT32,
        inference.DataType.UINT8, inference.DataType.INT8,
    }

try:
    _pd_dtype_to_np_dtype(handle.type())
except TypeError:
    raise SystemExit(f'input {name} dtype {handle.type()} unsupported for TRT — re-export model in FP32')

When it happens

Trigger: Loading a model with a FLOAT16/BF16/BOOL input while building TensorRT engines with dynamic shapes; TRT cache regeneration on a model exported at half precision.

Common situations: FP16-exported inference models used with --use_tensorrt; new Paddle versions exposing additional DataType enum values; mixed-precision exports of newer architectures.

Related errors

• Configuration Error: 'trt_dynamic_shapes' must be defined in
• Invalid input name {repr(name)} found in `dynamic_shapes`
• Input name {repr(name)} not found in `dynamic_shapes`
• Invalid input name {repr(name)} found in `dynamic_shape_inpu

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/a88cca9fb9693e82. Report an issue: GitHub.