PERSONAFLEX-WIN.md

PersonaPlex Setup Guide (Windows / PowerShell)

Summary of what we did

• Confirmed there is a hosted demo and a local Web UI.
• Installed Python dependencies from moshi/requirements.txt.
• Diagnosed build failures on Python 3.13:
  • safetensors tried to compile from source and failed because Rust/Cargo wasn’t on PATH.
  • sentencepiece tried to compile from source and failed because the project’s CMake files are too old for CMake 4.
• Recommended switching to Python 3.11 (or 3.10) to use prebuilt wheels and avoid source builds.
• Launched the server and hit a CUDA error because PyTorch was CPU‑only.
• Attempted CUDA builds for RTX 5090:
  • cu126 stable PyTorch failed with “no kernel image” (no sm_120 support).
  • Nightly cu128 PyTorch installed successfully and detected the RTX 5090.
• Noted that moshi pins torch<2.5, so nightly PyTorch is outside the declared range (expected warning).
• Hit a warmup crash due to missing Triton when torch.compile runs on Windows.
• Worked around it by disabling torch.compile via NO_TORCH_COMPILE=1.

Web demo and local Web UI

• Hosted demo: https://research.nvidia.com/labs/adlr/personaplex/
• Local Web UI: https://localhost:8998 (after starting the server)

Known Windows pitfalls we resolved

1. Python 3.13 is too new for several wheels

   • safetensors and sentencepiece may build from source and fail.
   • Fix: install Python 3.11 or 3.10 and use a fresh venv.

2. CMake 4 incompatibility

   • sentencepiece fails with “CMake < 3.5 removed” on CMake 4.
   • Fix: avoid source builds by using Python 3.11/3.10 wheels.

3. CUDA not available in PyTorch

   • Error: “Torch not compiled with CUDA enabled.”
   • Fix: install a CUDA‑enabled PyTorch build or run CPU mode.

Recommended setup (Python 3.11)

1. Create and activate venv:
   • py -3.11 -m venv .venv
   • ./.venv/Scripts/Activate.ps1
2. Install dependencies:
   • python -m pip install --upgrade pip
   • pip install -r moshi/requirements.txt
   • pip install moshi/.
3. Start server (CPU):
   • python -m moshi.server --device cpu --ssl $sslDir.FullName

GPU setup (RTX 5090)

Stable path (may fail on RTX 5090)

1. Update NVIDIA drivers to latest.
2. Install CUDA‑enabled PyTorch:
   • pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
   • If cu124 fails, try: https://download.pytorch.org/whl/cu126
3. If you see “no kernel image is available” or warnings about sm_120, the build does not support RTX 5090.

Experimental path (nightly, works with RTX 5090)

1. Uninstall existing PyTorch in the venv:
   • python -m pip uninstall -y torch torchvision torchaudio
2. Install nightly CUDA 12.8 build:
   • python -m pip install --upgrade --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
3. Verify CUDA:
   • python -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'no cuda')"
4. Start server:
   • python -m moshi.server --ssl $sslDir.FullName

Note: Nightly PyTorch is outside moshi’s declared range (torch<2.5). If you get runtime issues, fall back to CPU mode or wait for a compatible release.

Windows nightly + Triton note

If you see an error like “Cannot find a working triton installation,” it’s coming from torch.compile during warmup. On Windows, Triton support is inconsistent. The simplest fix is to disable compile:

• $env:NO_TORCH_COMPILE="1"
• python -m moshi.server --ssl $sslDir.FullName

Windows: Install older Python using Python Install Manager

If py -3.11 fails because Python 3.11 isn’t installed:

1. Uninstall any existing Python Install Manager / Python Manager

   • Old Windows Python install managers can conflict with the newest installer.
   • Remove them first from “Apps & features” or “Programs and Features.”

2. Install the latest Python Install Manager from python.org

   • Download Python 3.11 x64 from https://www.python.org/downloads/windows/
   • Ensure “Install launcher for all users” is checked.

3. Use the install manager to install the older version

   • Run: py -3.11 -m pip --version to verify it resolves.
   • Then create the venv: py -3.11 -m venv .venv

4. Activate and proceed with installs

   • ./.venv/Scripts/Activate.ps1
   • pip install -r moshi/requirements.txt
   • pip install moshi/.

Troubleshooting quick map

• Build errors for safetensors or sentencepiece → use Python 3.11/3.10.
• CUDA error on start → install CUDA PyTorch or run CPU mode.
• TritonMissing during warmup → set NO_TORCH_COMPILE=1 and retry.
• Web UI not reachable → confirm server is running and open https://localhost:8998.