Installation fails on Linux with modern pip/setuptools — 4 bugs with fixes

[ArielRi]

Describe the bug

./start.sh fails on Linux with modern pip/setuptools before the webui ever starts. There are 4 distinct bugs in the installer and requirements-voice-gpu.txt that affect any clean Ubuntu 24.04 install.

Bug 1 — ModuleNotFoundError: No module named 'pkg_resources'
setuptools >= 71 removed pkg_resources. The installer does not pin setuptools, so fresh conda envs and pip's isolated build envs in /tmp both get a modern version that breaks openai-whisper's build.

Bug 2 — nvidia-cudnn-cu12 version conflict
requirements-voice-gpu.txt pins nvidia-cudnn-cu12==8.9.7.29, but torch==2.5.1+cu124 requires ==9.1.0.70. pip cannot resolve this and aborts.

Bug 3 — cmake not listed as system prerequisite
pyopenjtalk requires cmake to compile. It is not mentioned in the README or install docs, causing a hard FileNotFoundError.

Bug 4 — conda env created with wrong permissions
After the first failed run, installer_files/env has permissions that block pip from reinstalling packages, even as the same user. Requires manual sudo chown to recover.

---

To Reproduce

1. Clone the repo on Ubuntu 24.04 LTS (clean install, NVIDIA GPU)
2. Run ./start.sh
3. Select G (NVIDIA GPU)
4. Observe failure during pip install -r requirements-voice-gpu.txt:
   • First run: ModuleNotFoundError: No module named 'pkg_resources'
   • After attempting to fix setuptools manually: nvidia-cudnn-cu12 version conflict
   • After fixing cudnn: FileNotFoundError: cmake not found (pyopenjtalk)
   • After first failed run: Permission denied on reinstall attempts

---

Expected behavior

./start.sh should complete installation and launch the webui without requiring manual intervention.

---

Screenshots

N/A — all errors are terminal output (reproduced exactly as shown above).

---

Additional context

• OS: Ubuntu 24.04 LTS
• Python: 3.10.20 (conda managed by installer)
• GPU: NVIDIA (cu124)
• pip: 25.0 / 26.x
• voice-pro: latest (ABUS Launcher v3.0)

Workaround that works:

source installer_files/conda/etc/profile.d/conda.sh
conda activate installer_files/env

# Fix permissions (Bug 4)
sudo chown -R $USER:$USER installer_files/env

# Fix pkg_resources (Bug 1)
pip install "setuptools==69.5.1" --force-reinstall --no-user
pip install openai-whisper==20240930 --no-build-isolation --no-user

# Fix cmake (Bug 3)
sudo apt-get install -y cmake build-essential

# Fix cudnn conflict (Bug 2)
grep -v "nvidia-cudnn-cu12" requirements-voice-gpu.txt > /tmp/requirements-fixed.txt
pip install -r /tmp/requirements-fixed.txt --no-build-isolation --no-user

conda deactivate
./start.sh  # succeeds

Suggested fixes:

• Pin setuptools<71 in the installer bootstrap, or install openai-whisper with --no-build-isolation
• Remove nvidia-cudnn-cu12==8.9.7.29 from requirements-voice-gpu.txt (torch pulls the correct version)
• Add cmake and build-essential to Linux prerequisites in the README
• Add chmod -R u+w installer_files/env after env creation in start.sh