🐛 [Bug] Running refit_engine_example.py produces incorrect results.

[953250587]

Bug Description

run refit_engine_example.py get “AssertionError: Refit Result is not correct. Refit failed”

To Reproduce

python refit_engine_example.py

Environment

• Torch-TensorRT Version : 2.6.0
• PyTorch Version :2.6.0+cu124
• OS: Linux
• Python version: 3.10.16
• CUDA version: 12.4
• Tensorrt version: 10.8.0.43

Additional context

There are some warning.
It seems that the following code caused the above issue:
TRTInterpreter.check_weight_equal(sd[sd_weight_name], np_map[engine_weight_name]). This code in TensorRT/py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py in line 537.
The numerical difference between sd[sd_weight_name] and np_map[engine_weight_name] is significant. So “weight_name_map” will not record the corresponding weights. However, I don't know why this issue occurs. Could it be due to an incorrect version of TensorRT?

[narendasan]

@cehongwang can you take a look

[cehongwang]

This is a version issue. Can you try with the latest Torch-TRT version with TRT 10.8?

[953250587]

This is a version issue. Can you try with the latest Torch-TRT version with TRT 10.8?

I tried switching TensorRT to 10.7.0.post1 first and then back to 10.8.0.43, and surprisingly, the issue was resolved. I have no idea why. Anyway, thank you for your help!

[cehongwang]

I encountered the same thing before. You are welcome! Let me know if you have any other questions.