undefined symbol:_ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_ib

Issue: undefined symbol:_ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_ib

Environment

• OS: Ubuntu 22.04.5 LTS (glibc 2.35)
• CUDA: 12.8 (Driver 595.45.04)
• GPUs: 8x NVIDIA B300 SXM6 AC (275GB each)
• Python: 3.11.15
• Model: DeepSeek-V3.2 (671B parameters)

Problem Description

When running the generate script in a Docker environment, I encounter an error loading the libtilert.so library due to an undefined symbol related to PyTorch's CUDA checks.

Steps to Reproduce

pip install tilert==0.1.3
python -c "import tilert"  # Fails with symbol errors

Tested PyTorch Versions

Complete Error Logs

PyTorch 2.6.0 + tilert 0.1.3

OSError: /usr/local/lib/python3.11/dist-packages/tilert/libtilert.so: 
undefined symbol: _ZN3c108ListType3getERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEENS_4Type24SingletonOrSharedTypePtrIS9_EE

PyTorch 2.10.0 + tilert 0.1.3

OSError: /usr/local/lib/python3.11/dist-packages/tilert/libtilert.so: 
undefined symbol: _ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_ib

PyTorch 2.9.1 + tilert 0.1.3

ImportError: /usr/local/lib/python3.11/dist-packages/torch/lib/libtorch_cuda.so: 
undefined symbol: ncclCommWindowDeregister

Attempted Solutions

1. Manual Installation in Ubuntu 22.04

• Created custom container with Ubuntu 22.04 (glibc 2.35)
• Tested 7+ PyTorch versions, all failed with different symbol errors
• tilert 0.1.3 imports successfully only with specific PyTorch versions, but model conversion fails

Start with the pip install and python -c "import tilert" reproduction in the Docker environment, then compare the listed PyTorch versions and their complete symbol errors. Review the tilert==0.1.3 import and model-conversion behavior; done means the library imports and model conversion works with a supported PyTorch and CUDA combination.