#!/usr/bin/env bash
# Canonical Qwen3.8-27B launcher for OpenAI-compatible coding clients.
#
# Qwen3.8 is multimodal. The default requires both the text GGUF and its
# source-matched projector; startup fails closed if either side is missing or
# incompatible. Set QWEN38_VISION=off only for an intentional text-only run.
# Exact target-verified speculation is considered by default on eligible
# requests. `auto` first measures ordinary decode, tries request-history lookup
# before an available MTP head, and independently drops either proposer for the
# generation after two consecutive measured four-round windows are not better
# than ordinary decode. Unsupported request semantics remain on ordinary target
# decode.
set -euo pipefail

script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"

export MODEL="${MODEL:-/opt/hf2q/models/qwen3.8/Qwen3.8-27B-Q4_K_M.gguf}"
export MMPROJ="${MMPROJ:-/opt/hf2q/models/qwen3.8/Qwen3.8-27B-mmproj-F16.gguf}"
export VISION_MODE="${QWEN38_VISION:-required}"
export HF2Q_QWEN_SPECULATION="${QWEN38_SPECULATION:-auto}"
# Qwen3.8 K=3 verifies four target positions at once. The native qL4
# decision/cache gate and matched ABBA receipt qualify the weight-amortized
# K-quant width-four route here. Process-wide mlx-native defaults remain
# unchanged outside this family launcher.
export HF2Q_DECODE_MVN="${HF2Q_DECODE_MVN:-0}"
export HF2Q_DECODE_MV_EXT="${HF2Q_DECODE_MV_EXT:-1}"

case "$HF2Q_QWEN_SPECULATION" in
    off|auto) ;;
    *)
        echo "QWEN38_SPECULATION must be off or auto (got: $HF2Q_QWEN_SPECULATION)" >&2
        exit 3
        ;;
esac
case "$HF2Q_DECODE_MVN:$HF2Q_DECODE_MV_EXT" in
    0:1|0:0|1:0|1:1) ;;
    *)
        echo "HF2Q_DECODE_MVN and HF2Q_DECODE_MV_EXT must each be 0 or 1" >&2
        exit 3
        ;;
esac

echo "Qwen3.8 exact speculation policy: $HF2Q_QWEN_SPECULATION" >&2
echo "Qwen3.8 K-quant width routing: mvN=$HF2Q_DECODE_MVN mv_ext=$HF2Q_DECODE_MV_EXT" >&2

exec "$script_dir/serve_qwen36_opencode.sh"