#!/bin/bash
set -e

echo "=== Ternary-Bonsai startup ==="

# FIX: Remove old incorrect usage if env still points to drafter
if [ -f "/models/Ternary-Bonsai-27B-dspark-Q4_1.gguf" ]; then
  echo "Found drafter file (will keep for optional speculative decoding, but NOT as main model)"
  ls -lh /models/
fi

# Download CORRECT main model if missing - 7.17GB
if [ ! -f "/models/Ternary-Bonsai-27B-Q2_0.gguf" ]; then
  echo "Downloading main model Ternary-Bonsai-27B-Q2_0.gguf (7.2GB)... this takes a few minutes on first run"
  python3 -c 'from huggingface_hub import hf_hub_download; hf_hub_download(repo_id="prism-ml/Ternary-Bonsai-27B-gguf", filename="Ternary-Bonsai-27B-Q2_0.gguf", local_dir="/models")'
else
  echo "Found main model: /models/Ternary-Bonsai-27B-Q2_0.gguf"
  ls -lh /models/Ternary-Bonsai-27B-Q2_0.gguf
fi

echo "Starting llama-server on 0.0.0.0:7860 with CORRECT model..."

# For HF Spaces CPU: no -ngl, no draft-model (draft only helps on CUDA and slows CPU)
# If you have CUDA, add: --draft-model /models/Ternary-Bonsai-27B-dspark-Q4_1.gguf -ngl 99
exec llama-server \
  -m /models/Ternary-Bonsai-27B-Q2_0.gguf \
  --host 0.0.0.0 \
  --port 7860 \
  -c 8192 \
  -n 4096 \
  -t 8 \
  --cache-type-k f16 \
  --cache-type-v f16