#!/bin/sh
set -eu

MODEL_SPEC="${MODEL_SPEC:-unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL}"
HOST="${HOST:-0.0.0.0}"
PORT="${PORT:-7860}"
CTX_SIZE="${CTX_SIZE:-8192}"
THREADS="${THREADS:-2}"
PARALLEL="${PARALLEL:-1}"
REASONING_MODE="${REASONING_MODE:-off}"
CACHE_TYPE_K="${CACHE_TYPE_K:-q8_0}"
CACHE_TYPE_V="${CACHE_TYPE_V:-q8_0}"
FLASH_ATTN="${FLASH_ATTN:-on}"
SPEC_TYPE="${SPEC_TYPE:-draft-mtp}"

# NOTE: the b8840 CPU build does not support draft-mtp (needs a newer
# llama.cpp build than this image ships). ngram-cache is a self-speculation
# mode that needs no external draft model: it speculates from n-grams already
# seen in context/output, verified by the target model, so it's zero-cost on
# accuracy. Set SPEC_TYPE=none in Space secrets to disable entirely.
SPEC_FLAGS="--spec-type $SPEC_TYPE"

exec /app/llama-server \
  -hf "$MODEL_SPEC" \
  --host "$HOST" \
  --port "$PORT" \
  --ctx-size "$CTX_SIZE" \
  --threads "$THREADS" \
  --threads-batch "$THREADS" \
  --parallel "$PARALLEL" \
  --cache-type-k "$CACHE_TYPE_K" \
  --cache-type-v "$CACHE_TYPE_V" \
  --reasoning "$REASONING_MODE" \
  -fa "$FLASH_ATTN" \
  --spec-type draft-mtp \
  --spec-draft-n-max 4