Error when trying to convert LLM models to MNN with llmexport.py script

[mikewalshchicago]

Following the instructions, I downloaded a model from huggingface (git lfs is installed, and I verified the full files are present in the local directory) and execute the following command:

python3 llmexport.py --path ~/qwen/Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v1/ --dst_path ~/qwen/final --export mnn --mnnconvert ~/MNN/build/MNNConvert --quant_bit 4

The compilation fails however, always with the following error:

<class 'torch.fx.experimental.symbolic_shapes.GuardOnDataDependentSymNode'>: Could not extract specialized integer from data-dependent expression u0 (unhinted: u0). (Size-like symbols: none)

I have all of the requirements from the requirements.txt file installed, and I have tried 3 different models with the same result. Is a specific version of pytorch required to get past this error?

[rap20171201]

I got the same question , which version of MNN you used? I use MNN3.2.3, but i see requirment installed pymnn3.2.5. Does it caused by?

[rap20171201]

I think this is due to the use of incompatible torch and onnx versions. When I modified torch to 2.8.0 and onnx to 1.19.0, the error was resolved