#!/usr/bin/env bash
# Launch baidu/Unlimited-OCR via llama-server (OpenAI-compatible, multimodal).
# Foreground — Ctrl+C to stop. All paths overridable via env.
set -euo pipefail

export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-1}"

BIN="${LLAMA_SERVER_BIN:-/usr/local/bin/llama-server}"
MODEL="${OCR_MODEL:-$HOME/models/Unlimited-OCR-Q4_K_M.gguf}"
MMPROJ="${OCR_MMPROJ:-$HOME/models/mmproj-Unlimited-OCR-F16.gguf}"
HOST="${OCR_HOST:-127.0.0.1}"
PORT="${OCR_PORT:-11436}"
CTX="${OCR_CTX:-32768}"
TEMP="${OCR_TEMP:-0}"
REPEAT_PENALTY="${OCR_REPEAT_PENALTY:-1.2}"

[[ -x $BIN    ]] || { echo "MISSING (or not executable): $BIN" >&2; exit 1; }
[[ -f $MODEL  ]] || { echo "MISSING: $MODEL" >&2; exit 1; }
[[ -f $MMPROJ ]] || { echo "MISSING: $MMPROJ" >&2; exit 1; }

exec "$BIN" \
  --model "$MODEL" \
  --mmproj "$MMPROJ" \
  --alias Unlimited-OCR \
  --host "$HOST" \
  --port "$PORT" \
  --temp "$TEMP" \
  --repeat-penalty "$REPEAT_PENALTY" \
  --ctx-size "$CTX" \
  --n-gpu-layers 99 \
  --flash-attn on \
  --cont-batching \
  --parallel 1 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --metrics