macros:
  llama: |
    llama-server --port ${PORT}
  mlx: |
    mlx_lm.server --port ${PORT}

  # - small `n` are not recommended
  # - MoEs require long drafts
  # - dense models: can reduce `--draft-min` and `--draft-max`
  next_edit: |
    --spec-type ngram-mod
    --spec-ngram-mod-n-match 24
    --spec-ngram-mod-n-min 24
    --spec-ngram-mod-n-max 32

models:
  sweep-next-edit-1.5b:
    name: Sweep Next-Edit 1.5B
    unlisted: true
    aliases:
      - sweep-next-edit
    cmd: |
      ${llama} --hf-repo sweepai/sweep-next-edit-1.5B
      ${next_edit}
  sweep-next-edit-v2-7b:
    name: Sweep Next-Edit v2 7B
    unlisted: true
    aliases:
      - sweep-next-edit-v2
    cmd: |
      ${llama} --hf-repo henrik3/sweep-next-edit-v2-7B-GGUF
      ${next_edit}
  gpt-oss-20b:
    name: GPT OSS (20b)
    cmd: |
      ${llama} --hf-repo unsloth/gpt-oss-20b-GGUF
  gpt-oss-120b:
    name: GPT OSS (120b)
    cmd: |
      ${llama} --hf-repo unsloth/gpt-oss-120b-GGUF
  qwen3.6-27b:
    name: Qwen 3.6 27B
    cmd: |
      ${llama} --hf-repo unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0
      --spec-type draft-mtp
      --flash-attn on
      --ctx-size 65536
  gemma-4-coder-fable5-composer2.5:
    name: Gemma 4 Coder (Fable 5, Composer 2.5)
    useModelName: yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF
    cmd: |
      ${llama} --hf-repo yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q8_0

groups:
  edit_prediction:
    persistent: true # not unloaded by other groups
    swap: true # only run one model from this group
    exclusive: false
    members:
      - sweep-next-edit-1.5b
      - sweep-next-edit-v2-7b

hooks:
  on_startup:
    preload:
      - sweep-next-edit-v2-7b

peers:
  prefect:
    proxy: http://prefect:11343
    models:
      - sweep-next-edit
      - sweep-next-edit-v2
      - zeta20
      - zeta21