server-basic.json

{
  // Minimal TensorSharp.Server configuration.
  // Run with:  TensorSharp.Server --config config/server-basic.json
  // Command-line options still win, so you can reuse this file across hosts and
  // override just the backend:
  //   TensorSharp.Server --config config/server-basic.json --backend ggml_cpu
  //
  // Hosts a multimodal Gemma-4 E4B (model + matching vision projector). Both
  // auto-download into ${modelRoot} on the first run (real, ungated GGUFs) and
  // are reused afterward. This is the standard -it build; point "path" at your
  // own file to host a different variant.
  "variables": {
    "modelRoot": "${TENSORSHARP_MODELS:-../models}",
    "repo": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/resolve/main"
  },

  "backend": "ggml_cuda",
  "max-tokens": 4096,
  "temperature": 0.8,
  "top-k": 40,
  "top-p": 0.95,
  "continuous-batching": true,
  "stop": ["</s>", "<|eot|>"],

  "model": {
    "path": "${modelRoot}/gemma-4-E4B-it-Q8_0.gguf",
    "urls": [ "${repo}/gemma-4-E4B-it-Q8_0.gguf" ]
  },
  "mmproj": {
    "path": "${modelRoot}/gemma-4-E4B-mmproj-F16.gguf",
    "urls": [ "${repo}/mmproj-F16.gguf" ]
  }
}