modelship-cuda ghcr.io/modelship-ai/modelship:latest-cuda https://ghcr.io/modelship-ai/modelship bridge sh false https://github.com/modelship-ai/modelship/issues https://github.com/modelship-ai/modelship Self-hosted, OpenAI-compatible inference for the agentic era. Modelship runs your whole AI stack -- reasoning LLMs with universal tool calling, embeddings, speech-to-text, text-to-speech, and image generation -- as multiple models sharing your GPUs behind a single OpenAI-compatible gateway. Built on Ray Serve with pluggable backends (vLLM, llama.cpp, Diffusers, plugins). This is the GPU variant, and requires the NVIDIA Driver plugin + NVIDIA Container Toolkit set up on this Unraid box. For a CPU-only host, use the separate "modelship-cpu" template instead. IMPORTANT: create your models.yaml at the path below BEFORE starting the container -- it will not start without one. See https://github.com/modelship-ai/modelship/blob/main/docs/model-configuration.md and https://github.com/modelship-ai/modelship/tree/main/config/examples for the format and ready-made examples. IMPORTANT: the default --shm-size=2g in Extra Parameters is sized only for a single small quick-start model. Ray sizes its object store to roughly 30% of the RAM it sees as available AT STARTUP -- if --shm-size is smaller than that, Ray crashes immediately. Raise --shm-size (roughly 30% of the RAM you want to give modelship) for bigger or multi-model / GPU stacks. AI: http://[IP]:[PORT:8000]/ https://raw.githubusercontent.com/modelship-ai/modelship-unraid/main/icon.png --shm-size=2g --runtime=nvidia -e MSHIP_UID=99 -e MSHIP_GID=100 GPU variant -- needs the NVIDIA Driver plugin and NVIDIA Container Toolkit. For CPU-only hosts, use the "modelship-cpu" template instead. 8000 8079 /mnt/user/appdata/modelship/models.yaml /mnt/user/appdata/modelship/cache INFO true modelship all all