#!/bin/bash
# Launch llama-server for the Genoria Qwen3.6-27B Fable-Fusion-711 demo.
# Every knob is a runtime env var so it can be tuned from Space settings
# without a rebuild. Defaults follow DavidAU's model card ("thinking mode"):
#   temperature=1.0 top_p=0.95 top_k=20 min_p=0.0 repetition_penalty=1.0
set -euo pipefail

THREADS="${THREADS:-$(nproc)}"
KV_QUANT="${KV_QUANT:-q8_0}"

ARGS=(
  --model "$MODEL_PATH"
  --alias "${MODEL_ALIAS:-genoria/qwen3.6-27b-fable-fusion-711}"
  --host 0.0.0.0
  --port "${PORT:-7860}"
  --ctx-size "${CTX_SIZE:-8192}"
  --n-predict "${MAX_TOKENS:-8192}"
  --threads "$THREADS"
  --n-gpu-layers "${NGL:-999}"
  --cache-type-k "$KV_QUANT"
  --cache-type-v "$KV_QUANT"
  --temp "${TEMP:-1.0}"
  --top-p "${TOP_P:-0.95}"
  --top-k "${TOP_K:-20}"
  --min-p "${MIN_P:-0.0}"
  --repeat-penalty "${REPEAT_PENALTY:-1.0}"
  --presence-penalty "${PRESENCE_PENALTY:-0.0}"
  --jinja
  --no-warmup
)

# Vision (image-text-to-text) via the repo's mmproj; disable with ENABLE_VISION=0
if [[ "${ENABLE_VISION:-1}" == "1" && -f "${MMPROJ_PATH:-}" ]]; then
  ARGS+=(--mmproj "$MMPROJ_PATH")
fi

# Escape hatch for any extra llama-server flags
if [[ -n "${EXTRA_ARGS:-}" ]]; then
  # shellcheck disable=SC2206
  ARGS+=(${EXTRA_ARGS})
fi

echo "Starting llama-server: model=$MODEL_PATH threads=$THREADS ctx=${CTX_SIZE:-8192} kv=$KV_QUANT"
exec /app/llama-server "${ARGS[@]}"