Failed to convert PaddlePaddle model: (Unimplemented) the 0th elementwise MUST be ir::FloatAttribute

[fish-kong]

bug描述 Describe the Bug

PC环境：
paddlepddle-gpu=3.3.0
paddleocr=3.3.2
paddlex=3.3.11
paddle2onnx=2.1.0
Debian GNU/Linux 12,RTX5090,cuda13.0,cudnn9.13

对PP-OCRv5_mobile_det进行微调，微调后的模型能够正常导出inference模型，在导出onnx模型的时候报错
命令如下：

paddle2onnx --model_dir output/ppocr_mc20251216/ \
--model_filename inference.json \
--params_filename inference.pdiparams \
--save_file ./ppocr_mobile_ocr20260207/ppocr_mobile_ocr20260207.onnx \
--enable_onnx_checker True 

报错如下：

[Paddle2ONNX] Start parsing the Paddle model file...
[Paddle2ONNX] Use opset_version = 11 for ONNX export.
2026-02-08 11:46:26 [ERROR]     Failed to convert PaddlePaddle model: (Unimplemented) the 0th elementwise MUST be ir::FloatAttribute
  [Hint: Expected array_list[0].isa<::pir::FloatAttribute>() == true, but received array_list[0].isa<::pir::FloatAttribute>():0 != true:1.] (at /github/workspace/paddle2onnx/parser/pir_parser.cc:814)

报错之后，我用paddle2onnx对去年12月微调并导出的inference模型，也是det模型，能够成功导出onnx。当时的环境是4090显卡，cuda是12.6，但是paddlepaddle_gpu不是最新版，所以怀疑是paddlepaddle的问题，而非paddle2onnx的问题

其他补充信息 Additional Supplementary Information

1.尝试更换 opset_version 从11到18，都是这个错误。
2.20260208更换为paddlepaddle_gpu更换为3.2.1解决上述问题，能够正常导出onnx模型，并推理成功

[risemeup1]

需要paddle2onnx来适配paddle最新版本，但是暂无人力适配了，后续paddle2onnx由外部开发者自己适配，可以找我来合入PR

[tyoungroy]

The code lines printing error log is
https://github.com/PaddlePaddle/Paddle2ONNX/blob/9ca238f8e6bda3b143ff805a84c2917ad6913eb1/paddle2onnx/parser/pir_parser.cc#L808-L818

it may be temporally fixed by adding a code checking FloatAttribute when array_list.size() > 0 or casting.
But, not sure it works correctly.
@risemeup1 would you leave simple comment that?

[tyoungroy]

This is temporally working for me.

-        if (array_list.size() > 0) {
+       if (array_list.size() > 0 && array_list[0].isa<::pir::FloatAttribute>()) {

I'm not sure of the converted onnx works correctly but when i check it by netron the onnx model looks fine.
I hope someone checks and share results.

[tyoungroy]

This is temporally working for me.

-        if (array_list.size() > 0) {
+       if (array_list.size() > 0 && array_list[0].isa<::pir::FloatAttribute>()) {

I'm not sure of the converted onnx works correctly but when i check it by netron the onnx model looks fine. I hope someone checks and share results.

It causes another error. In my case, when the GetOpAttr is called the actual value is ::pir::DoubleAttribute() so that the garbage data is given.
when i check as
if (array_list.size() > 0 && (array_list[0].isa<::pir::FloatAttribute>() || array_list[0].isa<::pir::DoubleAttribute>())) { PADDLE_ENFORCE_EQ( (array_list[0].isa<::pir::FloatAttribute>() || array_list[0].isa<::pir::DoubleAttribute>()), true, ::common::errors::Unimplemented("the 0th elementwise MUST be " "ir::FloatAttribute")); for (size_t i = 0; i < array_list.size(); ++i) { if (array_list[0].isa<::pir::FloatAttribute>()) { res->push_back(array_list[i].dyn_cast<::pir::FloatAttribute>().data()); } else { //::pir::DoubleAttribute>() res->push_back(static_cast<float>(array_list[i].dyn_cast<::pir::DoubleAttribute>().data())); } } }
the model is converted well and the converted onnx works as well.