Exporting onnx -> trt with fp16 problems

[Bobbelty]

I downloaded the pretrained DEIM-D-FINE-X with Object365. Running onnx export worked well and turning the onnx to .engine also worked which I did with --fp16 as mentioned in the readme. For some, running inference on this gives 0 results. Running the onnx on the same image gives results.

I also tried exporting onnx to trt without --fp16. Running inference on this model works. Should this be the case that it doesn't run properly as fp16 or is the readme not updated to reflect some changes?

Thanks!

[LeBron-Jian]

Use the official code to directly convert onnxruntime to fp16; I also modified some code; if I don't modify it, an error will be reported; I wonder if you have also modified it

The error message is as follows：

    return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
           ~~~~~~^^^^^^^^^^^^^
RuntimeError: expected m1 and m2 to have the same dtype, but got: float != struct c10::Half