Unable to launch single kernel for bge-rerank-base model

[lixiaoqi-LXQ]

Issue Description

Hi, I am trying to use MonoNN with bge-rerank-base model, but it doesn’t seem to take effect, my profiling result shows below.

I have tested MonoNN with the bert_base model following the provided example, and it works fine.

Steps to Reproduce

1. set up the environment for MonoNN
2. download the bge-rerank-base model
3. convert the onnx model to tensroflw .pb format:
   1. pip install onnx-tf==1.10.0
   2. onnx-tf convert -i BAAI/bge-reranker-base/onnx/model.onnx -o bge-reranker-base-tf
4. obtain the frozen graph following the instructions
5. the minimal code snippet can be found here.

Environment

• Driver Version: 555.42.06
• CUDA Version: 11.6
• OS: Ubuntu 20.04.6

---

Thanks in advance for your help!

[lixiaoqi-LXQ]

I’ve added a screenshot of the output from running the code snippet:

I’m not sure if the final line is the key issue. From my observations, it seems like MonoNN is not entering the key optimization phase and exits early. Could this be the root cause?