# From here is where you define the config for llama-swap.
healthCheckTimeout: 3600 # Set it to one hour so model downloads don't stop halfway through.
# 262144
models:
  GLM47:
    aliases: 
      - "glm-coder"
    cmd: >
          llama-server 
          --port ${PORT} 
          -m /models/GLM-4.7-Flash-MXFP4_MOE.gguf  
          --fit-ctx 230000 
          --temp 0.7 --top-p 1.0 --min-p 0.01

  Qwen3.6-35B-A3B:
    aliases: 
      - "qwen-omni"
    cmd:  > 
          llama-server
          --port ${PORT}
          -m /models/Qwen3.6-35B-A3B-MXFP4_MOE.gguf
          --mmproj /models/Qwen-mmproj-F16.gguf
          --fit-ctx 230000
          --fit-target 2048
          --temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.0 --min-p 0.00 --no-mmap


  Qwen3.6-Opus:
    aliases: 
      - "qwen-opus"
    cmd:  > 
          llama-server
          --port ${PORT}
          --fit-ctx 262144
          -m /models/Qwen3.6-Opus.gguf
          --fit-target 2048
          --temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.0 --min-p 0.00
          --spec-type mtp --spec-draft-n-max 3 -np 1 --no-mmap

  kokoro-tts:
    proxy: http://${MODEL_ID}:8880
    name: "kokoro TTS"
    useModelName: "tts-1"
    checkEndpoint: /health
    cmd: |
      docker run --rm --name ${MODEL_ID} --network nerd-network
      --gpus 'device=0'
      --env 'API_LOG_LEVEL=INFO'
      ghcr.io/remsky/kokoro-fastapi-gpu:latest
    cmdStop: docker stop ${MODEL_ID}
          

  Qwen-Image:
    proxy: http://127.0.0.1:${PORT}
    checkEndpoint: /
    aliases: 
      - "qwen-image"
    cmd:  > 
          sd-server
          --listen-port ${PORT} --fa --offload-to-cpu
          --diffusion-model /models/sd/unet/qwen-image-2512-Q4_K_M.gguf
          --llm /models/sd/text_encoders/Qwen2.5-VL-7B-Instruct-UD-Q4_K_XL.gguf
          --llm_vision /models/sd/text_encoders/Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf
          --vae /models/sd/vae/qwen_image_vae.safetensors