opendatalab/MinerU · error · ValueError

OCR_INFERENCE_PRECISION must be one of: auto, fp32, fp16

Error message

OCR_INFERENCE_PRECISION must be one of: auto, fp32, fp16

What it means

Raised by BaseOCRV20._resolve_inference_dtype when the OCR_INFERENCE_PRECISION constant (typically sourced from an environment variable) lowercases to something other than auto, fp32, or fp16. The value controls whether the OCR network runs float32 on CPU / forced-fp32, or float16 elsewhere.

Source

Thrown at mineru/model/utils/pytorchocr/base_ocr_v20.py:32

class BaseOCRV20:
    def __init__(self, config, **kwargs):
        self.config = config
        self.build_net(**kwargs)
        self.ocr_inference_dtype = torch.float32
        self.net.eval()


    def build_net(self, **kwargs):
        self.net = BaseModel(self.config, **kwargs)

    def _resolve_inference_dtype(self, device):
        """根据常量和设备类型解析 OCR 网络推理使用的浮点精度。"""
        precision = OCR_INFERENCE_PRECISION.lower()
        device_name = str(device).lower()
        is_cpu = device_name.startswith("cpu")

        if precision not in {"auto", "fp32", "fp16"}:
            raise ValueError(
                "OCR_INFERENCE_PRECISION must be one of: auto, fp32, fp16"
            )
        if precision == "fp32" or is_cpu:
            return torch.float32
        return torch.float16

    def _apply_inference_precision(self, device):
        """将 OCR 网络移动到目标设备，并在非 CPU 半精度场景下切到 fp16。"""
        self.net.to(device)
        self.ocr_inference_dtype = self._resolve_inference_dtype(device)
        if self.ocr_inference_dtype == torch.float16:
            self.net.to(dtype=torch.float16)

    def _to_inference_dtype(self, tensor):
        """将浮点输入 tensor 转为 OCR 推理精度，整型/布尔辅助输入保持原 dtype。"""
        if torch.is_tensor(tensor) and torch.is_floating_point(tensor):
            inference_dtype = getattr(self, "ocr_inference_dtype", torch.float32)
            return tensor.to(dtype=inference_dtype)

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Set OCR_INFERENCE_PRECISION to one of: auto (fp16 on GPU, fp32 on CPU), fp32, or fp16.
2. Unset the variable to fall back to the built-in default (auto).
3. Check for typos like 'float16' vs 'fp16' in your environment/deployment config.

Example fix

# before
export OCR_INFERENCE_PRECISION=bf16

# after
export OCR_INFERENCE_PRECISION=fp16   # or fp32 / auto

import os
precision = os.getenv("OCR_INFERENCE_PRECISION", "auto").lower()
assert precision in {"auto", "fp32", "fp16"}, f"bad OCR_INFERENCE_PRECISION: {precision!r}"

When it happens

Trigger: Setting OCR_INFERENCE_PRECISION to values like 'bf16', 'fp8', 'float16', or 'FP16 ' (with trailing space handled, but e.g. 'fp_16' is not) before running mineru's pytorchocr pipeline.

Common situations: Users trying bfloat16 on newer GPUs; typo in .env or CI variables; deployment manifests that set the variable to an unsupported precision name.

Related errors

• Language {lang} not supported. Allowed values: {allowed_valu
• Language {lang} not supported
• not support limit type, image
• Unsupported lmdeploy device type: {device_type}
• Unsupported lmdeploy backend: {lm_backend}

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/211cc1c82e7760e2. Report an issue: GitHub.