channings
channings
PaddleSlim
@JunrQ I also find this difference
遇到同样的问题
请教一下，速度多少token/s，我在部署32k的int4，awq和gptq都不到1t/s..很困惑
@jiangjiajun 目前prior_box的转换依赖输入的shape来生成，所以需要输入的shape是固定的，考虑未来支持动态输入
@znsoftm - paddle2onnx don't support quanted model. - if your model is contruct by paddlepaddle>2.0.0rc1，lstm is supported. Old version of paddlepaddle is not supported. - ctc_align is not support now....
@znsoftm can you share your paddle model(saved) to us, ctc_align is a operator with [LodTensor](https://www.paddlepaddle.org.cn/documentation/docs/en/beginners_guide/basic_concept/lod_tensor_en.html) Inputs, which may support by onnx when the inputs of ctc_align in your model is...
@PhilCuriosity maskrcnn包含控制流op，目前paddle2onnx尚不支持。
@yxyphoebe 可以参考这个教程 https://aistudio.baidu.com/aistudio/projectdetail/1479970
请PaddleOCR的release/2.0分支现在支持导出为动态shape的模型： https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.0/tools/export_model.py 这个是导出inference模型的教程： https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.0/doc/doc_ch/inference.md