vLLM's compiled extension fails to import against a source-built PyTorch with a different CXX11 ABI setting

import vllm raises ImportError: /workspace/vllm-abo/vllm/_C.abi3.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb. The .so file exists and vllm is installed; nothing about the version pins looks wrong.

The environment had PyTorch and torchvision compiled from source with _GLIBCXX_USE_CXX11_ABI=1, while vLLM's _C.abi3.so was linked against a torch built with the other ABI setting. The mangled symbol encodes the std::string signature, so the two builds are binary-incompatible even at matching version numbers (torch 2.5.1, torchvision 0.20.1, vllm 0.6.5.dev398+cu124 in the report).

Install the official PyTorch wheels rather than source builds, or rebuild vLLM against exactly the torch you have and check torch._C._GLIBCXX_USE_CXX11_ABI before building.

A dependency bump or a from-source rebuild of any torch-adjacent package can introduce this without changing any visible version number. Related to gotcha-vllm-v1-memory-footprint-growth.