Support MetaX and Vastai tokenizer compatibility images

vllm_wrapper.sh

@@ -9,17 +9,39 @@ fi
 
 model_dir="$2"
 shift 2
-prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir")"
+args=("$@")
+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" "$@"
+    exec "$vllm_real" serve "$model_dir" "${args[@]}"
 fi
 
-args=("$@")
+effective_model="$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]}" == "$model_dir" ]]; then
+    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} && "${args[i+1]}" == "$tokenizer_dir" ]]; then
         args[i+1]="$prepared"
-    elif [[ "${args[i]}" == "--tokenizer=$model_dir" ]]; then
+    elif [[ "${args[i]}" == "--tokenizer=$tokenizer_dir" && -n "$tokenizer_dir" ]]; then
         args[i]="--tokenizer=$prepared"
     fi
 done
-exec "$vllm_real" serve "$prepared" "${args[@]}"
+exec "$vllm_real" serve "$effective_model" "${args[@]}"