klaster/apps/llama/configs/config.yaml

# yaml-language-server: $schema=https://raw.githubusercontent.com/mostlygeek/llama-swap/refs/heads/main/config-schema.json
healthCheckTimeout: 600
logToStdout: "both" # proxy and upstream

macros:
  base_args: "--no-warmup --port ${PORT}"
  common_args: "--fit-target 1536 --fit-ctx 32768 --no-warmup --port ${PORT}"
  gemma_sampling: "--prio 2 --temp 1.0 --repeat-penalty 1.0 --min-p 0.00 --top-k 64 --top-p 0.95"
  qwen35_sampling: "--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00"
  qwen35_35b_args: "--temp 1.0 --min-p 0.00 --top-p 0.95 --top-k 20"
  qwen35_35b_heretic_mmproj: "--mmproj-url https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/main/mmproj-F16.gguf --mmproj /root/.cache/llama.cpp/unsloth_Qwen3.5-35B-A3B-GGUF_mmproj-F16.gguf"
  thinking_on: "--chat-template-kwargs '{\"enable_thinking\": true}'"
  thinking_off: "--chat-template-kwargs '{\"enable_thinking\": false}'"

hooks:
  on_startup:
    preload:
      - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"

groups:
  always:
    persistent: true
    exclusive: false
    swap: false
    members:
      - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"

models:
  "gemma3-12b":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/gemma-3-12b-it-GGUF:Q4_K_M
        ${gemma_sampling}
        ${common_args}

  "gemma3-12b-novision":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/gemma-3-12b-it-GGUF:Q4_K_M
        ${gemma_sampling}
        --no-mmproj
        ${common_args}

  "gemma3-4b":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/gemma-3-4b-it-GGUF:Q4_K_M
        ${gemma_sampling}
        ${common_args}

  "gemma3-4b-novision":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/gemma-3-4b-it-GGUF:Q4_K_M
        ${gemma_sampling}
        --no-mmproj
        ${common_args}

  "Qwen3-Coder-Next-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3-Coder-Next-GGUF:Q4_K_M
        --ctx-size 65536
        --predict 8192
        --temp 1.0
        --min-p 0.01
        --top-p 0.95
        --top-k 40
        --repeat-penalty 1.0
        ${common_args}

  "Qwen3.5-35B-A3B-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
        ${qwen35_35b_args}
        ${common_args}

  "Qwen3.5-35B-A3B-GGUF-nothink:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
        ${qwen35_35b_args}
        ${common_args}
        ${thinking_off}

  # The "heretic" version does not provide the mmproj
  # so providing url to the one from the non-heretic version.
  "Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf mradermacher/Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M
        ${qwen35_35b_heretic_mmproj}
        ${qwen35_35b_args}
        ${common_args}

  "Qwen3.5-35B-A3B-heretic-GGUF-nothink:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf mradermacher/Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M
        ${qwen35_35b_heretic_mmproj}
        ${qwen35_35b_args}
        ${common_args}
        ${thinking_off}

  "Qwen3.5-0.8B-GGUF:Q4_K_XL":
    ttl: 0
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-0.8B-GGUF:Q4_K_XL
        ${qwen35_sampling}
        ${base_args}
        ${thinking_on}

  "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL":
    ttl: 0
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-0.8B-GGUF:Q4_K_XL
        --ctx-size 4096
        ${qwen35_sampling}
        ${base_args}
        ${thinking_off}

  "Qwen3.5-2B-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-2B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_on}

  "Qwen3.5-2B-GGUF-nothink:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-2B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_off}

  "Qwen3.5-4B-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_on}

  "Qwen3.5-4B-GGUF-nothink:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_off}

  "Qwen3.5-9B-GGUF:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_on}

  "Qwen3.5-9B-GGUF-nothink:Q4_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_off}

  "Qwen3.5-9B-GGUF:Q3_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-9B-GGUF:Q3_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_on}

  "Qwen3.5-9B-GGUF-nothink:Q3_K_M":
    ttl: 600
    cmd: |
      /app/llama-server
        -hf unsloth/Qwen3.5-9B-GGUF:Q3_K_M
        ${qwen35_sampling}
        ${common_args}
        ${thinking_off}