pip-installing vLLM/torch silently downgrades NCCL below the 2.30.4 floor; import dies at undefined symbol and check_nccl_so() cannot catch it

[dmvevents]

Problem

import deep_ep fails with a raw dynamic-linker error after installing vLLM (or anything that pins
torch==2.11.0) into an environment that satisfied the documented NCCL floor:

ImportError: /opt/DeepEP/deep_ep/_C.cpython-312-x86_64-linux-gnu.so: undefined symbol: ncclGetLsaDevicePointer

The chain (each step verified on H200/EFA, deep_ep 2.0.0, torch 2.11.0+cu130, vllm 0.26.0):

1. The environment starts at nvidia-nccl-cu13==2.30.4 (the README floor: "NCCL 2.30.4 and above").
2. pip install vllm==0.26.0 pins torch==2.11.0, and torch 2.11.0 pins nvidia-nccl-cu13==2.28.9
   (exact ==) — pip silently downgrades NCCL as a side effect.
3. NCCL 2.28.x lacks the GIN/LSA device-API symbols _C.so was built against, so the next
   import deep_ep dies at the linker with no version hint.
4. pip install --no-deps --force-reinstall nvidia-nccl-cu13==2.30.4 fully heals it — deep_ep + vllm
   0.26.0 + torch 2.11.0 then import and run together cleanly.

Why check_nccl_so() cannot catch this case

The guard in deep_ep/__init__.py compares the loaded libnccl.so (from /proc/self/maps) against
the lib at find_nccl_root() — i.e. the NCCL currently installed. A wheel downgrade moves both sides
together: torch loads 2.28.9 and find_nccl_root() now also resolves to 2.28.9. Loaded == reference, the
guard passes, and the failure surfaces one line later at import deep_ep._C as an opaque linker error.
The invariant that actually broke is "loaded == what _C.so was built against", which the build does
not record anywhere the guard can read.

Suggested fixes (either or both)

1. Declare the floor in wheel metadata — install_requires=["nvidia-nccl-cu13>=2.30.4"] (per CUDA
   variant). pip would then refuse the torch-driven downgrade loudly instead of performing it silently.
   Today setup.py locates NCCL at build time via find_pkgs.find_nccl_root() but ships no constraint.
2. Record the built-against NCCL version into the wheel (e.g. write it into the generated envs.py
   next to the existing persistent envs) and extend check_nccl_so() to compare the loaded lib against
   that, so this failure mode produces the guard's existing clear message instead of
   undefined symbol: ....

Environment

• deep_ep 2.0.0 (wheel built against nvidia-nccl-cu13 2.30.4)
• torch 2.11.0+cu130, vllm 0.26.0, python 3.12
• 2× p5en.48xlarge (H200), EFA — though the failure is host-side and hardware-independent

Happy to provide the full install transcript.