run-server.sh

#!/usr/bin/env bash
set -euo pipefail

SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
LLAMA_SERVER=${LLAMA_SERVER:-llama-server}

if ! command -v "$LLAMA_SERVER" >/dev/null 2>&1; then
  echo "llama-server was not found. Install llama.cpp or set LLAMA_SERVER=/path/to/llama-server." >&2
  exit 1
fi

if [[ -n "${MIMO_MODEL:-}" ]]; then
  MODEL=$MIMO_MODEL
else
  shopt -s nullglob
  CANDIDATES=("$SCRIPT_DIR"/MiMo-V2.5-coder-Q2-00001-of-*.gguf)
  shopt -u nullglob

  if [[ ${#CANDIDATES[@]} -eq 0 ]]; then
    echo "No first GGUF shard found next to run-server.sh." >&2
    exit 1
  fi

  MODEL=${CANDIDATES[0]}
fi

ARGS=(
  --model "$MODEL"
  --host "${MIMO_HOST:-127.0.0.1}"
  --port "${MIMO_PORT:-8080}"
  --ctx-size "${MIMO_CTX:-100000}"
  --parallel "${MIMO_PARALLEL:-1}"
  --batch-size "${MIMO_BATCH:-512}"
  --ubatch-size "${MIMO_UBATCH:-128}"
  --threads "${MIMO_THREADS:-12}"
  --threads-batch "${MIMO_THREADS_BATCH:-18}"
  --prio "${MIMO_PRIO:-0}"
  --poll "${MIMO_POLL:-80}"
  --flash-attn on
  --jinja
  --fit "${MIMO_FIT:-on}"
  --fit-target "${MIMO_FIT_TARGET:-4096}"
  --fit-ctx "${MIMO_FIT_CTX:-100000}"
  --gpu-layers "${MIMO_GPU_LAYERS:-auto}"
  --cache-type-k "${MIMO_CACHE_K:-f16}"
  --cache-type-v "${MIMO_CACHE_V:-f16}"
  --reasoning "${MIMO_REASONING:-off}"
)

if [[ "${MIMO_CPU_MOE:-0}" == "1" ]]; then
  ARGS+=(--cpu-moe)
fi

if [[ -n "${MIMO_DEVICE:-}" ]]; then
  ARGS+=(--device "$MIMO_DEVICE")
fi

if [[ -n "${MIMO_TOOLS:-}" ]]; then
  ARGS+=(--tools "$MIMO_TOOLS")
fi

exec "$LLAMA_SERVER" "${ARGS[@]}" "$@"