Let max_num_batched_tokens use human_readable_int for large numbers (#18968)

Signed-off-by: mgoin <[email protected]>

Author: mgoin

vllm/engine/arg_utils.py

@@ -224,7 +224,7 @@ def parse_dataclass(val: str, cls=dataclass_cls) -> Any:
         elif contains_type(type_hints, int):
             kwargs[name]["type"] = int
             # Special case for large integers
-            if name in {"max_model_len"}:
+            if name in {"max_model_len", "max_num_batched_tokens"}:
                 kwargs[name]["type"] = human_readable_int
         elif contains_type(type_hints, float):
             kwargs[name]["type"] = float