[konstde00]

The nightly python workflow fails on all five Python versions with ModuleNotFoundError: No module named 'deepspeed.moe.v2opt' when ds_report imports DeepSpeed (for example https://github.com/Supercomputing-System-AI-Lab/X-MoE/actions/runs/40053119642).

deepspeed/moe/v2opt/ has no __init__.py, so find_packages (setup.py:320) skips it and a regular pip install . leaves out the 13 v2opt modules, even though deepspeed/moe/sharded_moe.py:37 imports .v2opt.gating. An editable install (pip install -e ., as in the README) still works because Python finds the directory as a namespace package, which is why this doesn't show up in a development checkout.

This adds the same empty __init__.py that the liheng/GroupedGEMM, ELMoE, X-MoE-4D, zixian/pp_dev and zixian/shared_exp_impl branches already have, so merging those branches later won't conflict.

Tests

• Non-editable pip install .[dev,autotuning] on Python 3.12: before, ds_report fails with the error above. After, all 13 v2opt modules are installed, and with triton present ds_report and import deepspeed.moe.v2opt.gating work.
• python workflow on my fork: master fails on all five versions with the same error (https://github.com/konstde00/X-MoE/actions/runs/40109822074). This branch passes on 3.11 and 3.12, and 3.8 to 3.10 stop at the triton import described below (https://github.com/konstde00/X-MoE/actions/runs/40109820064).

On 3.8, 3.9 and 3.10 the job now stops at No module named 'triton', because the v2opt kernels import triton and the gh-builder images for those versions ship torch 1.12.1 without it. cpu-torch-latest installs CPU torch, so it stops at the same import (and would then hit its --torch_ver="2.5" check). I didn't touch these here, since whether to drop old Python versions from CI or guard the triton import is your call.