+ extremely_dangerous

Signed-off-by: wang.yuqi <noooop@126.com>

Author: noooop

vllm/config.py

@@ -3631,13 +3631,20 @@ def _get_and_verify_max_len(
                 f"{model_max_length} in model's config.json). This may lead "
                 "to incorrect model outputs or CUDA errors.")
             if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
-                logger.warning(
-                    "%s Make sure the value is correct and within the "
-                    "model context size.", msg)
+                logger.extremely_dangerous(
+                    "Using VLLM_ALLOW_LONG_MAX_MODEL_LEN may lead"
+                    " to the following exceptions:\n"
+                    "- If the model uses rope position encoding, "
+                    "positions exceeding derived_max_model_len "
+                    "lead to nan.\n"
+                    "- If the model uses absolute position encoding, "
+                    "positions exceeding derived_max_model_len "
+                    "will cause a CUDA array out-of-bounds error.\n"
+                    "For testing purposes only, do not implement "
+                    "any functionality relying "
+                    "on VLLM_ALLOW_LONG_MAX_MODEL_LEN.")
             else:
-                raise ValueError(
-                    f"{msg} To allow overriding this maximum, set "
-                    "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
+                raise ValueError(msg)
     return int(max_model_len)
 
 

vllm/logger.py

@@ -71,6 +71,12 @@ def _print_warning_once(logger: Logger, msg: str, *args: Hashable) -> None:
     logger.warning(msg, *args, stacklevel=2)
 
 
+def _extremely_dangerous(logger: Logger, msg: str, *args: Hashable) -> None:
+    template = (f"\n===== EXTREMELY DANGEROUS =====\n"
+                f"{msg}\n===============================\n")
+    logger.warning(template, *args, stacklevel=2)
+
+
 class _VllmLogger(Logger):
     """
     Note:
@@ -101,6 +107,9 @@ def warning_once(self, msg: str, *args: Hashable) -> None:
         """
         _print_warning_once(self, msg, *args)
 
+    def extremely_dangerous(self, param, *args):
+        _extremely_dangerous(param, *args)
+
 
 def _configure_vllm_root_logger() -> None:
     logging_config = dict[str, Any]()
@@ -148,6 +157,7 @@ def init_logger(name: str) -> _VllmLogger:
         "debug_once": _print_debug_once,
         "info_once": _print_info_once,
         "warning_once": _print_warning_once,
+        "extremely_dangerous": _extremely_dangerous
     }
 
     for method_name, method in methods_to_patch.items():