[Bug]: vLLM segmentation fault due to incompatible protobuf version on s390x

[dmholtz]

Describe the bug

protobuf is pinned to version 6.33.6 in the lockfile, introduced by the installation of torch-spyre[cpsat] extra in #671. Beforehand, it was pinned to version 7.35.1.

Since #671, vLLM crashes with a segmentation fault if protobuf is not manually re-installed to a 7.x version.

[root@spyre]# vllm
INFO 09-23 13:38:52 [__init__.py:52] Available plugins for group vllm.platform_plugins:
INFO 09-23 13:38:52 [__init__.py:54] - spyre_inference -> spyre_inference:register
INFO 09-23 13:38:52 [__init__.py:57] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.
INFO 09-23 13:38:52 [__init__.py:254] Platform plugin spyre_inference is activated
INFO 09-23 13:38:59 [importing.py:88] Triton not installed or not compatible; certain GPU-related functions will not be available.
!!!!!!! Segfault encountered !!!!!!!
  File "<unknown>", line 0, in upb_strtable_lookup2

Segmentation fault (core dumped)

The segmentation fault can be avoided by uv sync --locked && uv pip install protobuf==7.35.1

How to reproduce

uv sync --locked
vllm

Additional context

Recently published container builds are not affected by this, suggesting they work around this by either not syncing the locked dependencies or overriding the protobuf version externally.

Checklist

• I have searched for similar issues.