❓ Unable to freeze tensor of type Int64/Float64 into constant layer, try to compile model with truncate_long_and_double enabled?

[panahikhas]

❓ RuntimeError: [Error thrown at core/conversion/converters/converter_util.cpp:251] Unable to freeze tensor of type Int64/Float64 into constant layer, try to compile model with truncate_long_and_double enabled:

1. A pre-trained Torch model like Resnet18 was loaded
2. The model was quantized using pytorch_quantization.quant_modules.initialize()
3. The quantized model was calibrated
4. The model was fine-tuned (QAT)
5. I tried to convert the fine-tuned model to TensorRT using
   trt_mod = torch_tensorrt.compile(qat_model, inputs=[torch_tensorrt.Input([32, 3, 32, 32])], enabled_precisions={torch.int8})
   but I encountered the error below:

File "/home/i2027/anaconda3/envs/p/lib/python3.10/site-packages/torch_tensorrt/_compile.py", line 133, in compile
return torch_tensorrt.ts.compile(
File "/home/i2027/anaconda3/envs/p/lib/python3.10/site-packages/torch_tensorrt/ts/_compiler.py", line 139, in compile
compiled_cpp_mod = _C.compile_graph(module._c, _parse_compile_spec(spec))
RuntimeError: [Error thrown at core/conversion/converters/converter_util.cpp:251] Unable to freeze tensor of type Int64/Float64 into constant layer, try to compile model with truncate_long_and_double enabled

I have checked the model parameters and all of them were of type float32. I don't know why TorchTensorRT complains about Int64/Float64! Please note that I have managed to convert a simple CNN to TensorRT using the method described above successfully. However, I failed to convert an existing torchvision model using the steps above. I will be grateful for any hint.

Environment

• PyTorch Version: 2.0.1+cu118

• CPU Architecture: x86

• OS: Ubuntu 20.04

• How you installed PyTorch: pip

• Python version: 3.10.11

• CUDA version: 12.1

• GPU models and configuration: GeForce GTX 1080 - 12 GB

• All pckages versions:

• • torch==2.0.1+cu118
• • torch_tensorrt==1.4.0
• • torchvision==0.15.2+cu118
• • pytorch_quantization==2.1.2
• • torchvision==0.15.2+cu118

Additional context

The code is available at https://github.com/panahikhas/TensorRT-QAT/blob/main/torch-tensorrt-QAT.py to reproduce the results.

[narendasan]

Model parameters and all of them were of type float32.

truncate_long_and_double does not apply just to model weights, it also deals with things such as tensor indexing which in PyTorch/TorchScript happens in int64 but when we convert to tensorrt this needs to be in int32.

What happens if you add truncate_long_and_double=True to your compile settings?

[panahikhas]

Thank you very much for your response.
If I add the argument truncate_long_and_double=True the accuracy become 10% from 80%.
I don't know the reason but the problem was from PyCharm. I was working with it and it had errors. But when I used the terminal or VSCode there was no problem and no warning/error about int64.