zenth815/odysseus-frontend repository overview

⁠Odysseus — ROCm / Dual-GPU Build

This image extends the official Odysseus⁠ self-hosted AI workspace with AMD ROCm support and dual AMD+NVIDIA GPU routing, built on Ubuntu 22.04 for ROCm repository compatibility.

⁠What's different from :latest

• Ubuntu 22.04 base (instead of slim Python) — required for AMD ROCm apt repos
• ROCm userspace pre-installed (rocm-dev, hipblas, rocblas, etc.)
• Optional CUDA build tools (cuda-nvcc-12-4, libcublas-dev-12-4) for dual-GPU builds
• Cookbook auto-detects GPU at first serve and compiles the right llama.cpp binary:
  • AMD only → llama-server (HIP)
  • NVIDIA only → llama-server (CUDA)
  • Both → llama-server-hip + llama-server-cuda + routing wrapper
• gfx906/MI50 specific: auto-applies -DAMDGPU_TARGETS=gfx906 -DGGML_HIP_NO_FP8=ON to avoid FP8 compile errors on ROCm 6.1.2

⁠Quick start

git clone https://github.com/Wizard815/odysseus.git -b rocmcuda
cd odysseus
cp .env.example .env

Edit .env:

COMPOSE_FILE=docker-compose.yml:docker/gpu.amd-nvidia.yml
RENDER_GID=18          # getent group render | cut -d: -f3
ROCM_VERSION=6.1.2     # 6.2.4 for RDNA2, 6.3.4 for RDNA3
ODYSSEUS_IMAGE=zenth815/odysseus-frontend:testing

docker compose build && docker compose up -d

On first model serve, the Cookbook compiles llama.cpp for your GPU — allow 10-20 min. Binaries persist in APP_DATA_DIR so subsequent starts are instant.

⁠ROCm version guide