enginex-qwen25-compat/vllm_wrapper.sh

#!/usr/bin/env bash
set -euo pipefail

vllm_real="$(dirname "$(readlink -f "$0")")/vllm_real"

if [[ "${1:-}" != "serve" || -z "${2:-}" ]]; then
    exec "$vllm_real" "$@"
fi

model_dir="$2"
shift 2
args=("$@")
tokenizer_dir=""
for ((i=0; i<${#args[@]}; i++)); do
    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} ]]; then
        tokenizer_dir="${args[i+1]}"
        break
    elif [[ "${args[i]}" == --tokenizer=* ]]; then
        tokenizer_dir="${args[i]#--tokenizer=}"
        break
    fi
done

if [[ -n "$tokenizer_dir" && -d "$tokenizer_dir" ]]; then
    prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir" "$tokenizer_dir")"
else
    prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir")"
    tokenizer_dir=""
fi
if [[ -z "$prepared" ]]; then
    exec "$vllm_real" serve "$model_dir" "${args[@]}"
fi

effective_model="$prepared"
if [[ -n "$tokenizer_dir" && "$(readlink -f "$tokenizer_dir")" != "$(readlink -f "$model_dir")" ]]; then
    # A vendor entrypoint's --tokenizer points at a copied tokenizer tree.
    # Keep model weights at the original path; redirect only that tokenizer.
    effective_model="$model_dir"
fi
for ((i=0; i<${#args[@]}; i++)); do
    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} && "${args[i+1]}" == "$tokenizer_dir" ]]; then
        args[i+1]="$prepared"
    elif [[ "${args[i]}" == "--tokenizer=$tokenizer_dir" && -n "$tokenizer_dir" ]]; then
        args[i]="--tokenizer=$prepared"
    fi
done
exec "$vllm_real" serve "$effective_model" "${args[@]}"