[misc] soft drop beam search (#8763)

Author: youkaichao

vllm/envs.py

@@ -62,6 +62,7 @@
     VLLM_TORCH_PROFILER_DIR: Optional[str] = None
     VLLM_USE_TRITON_AWQ: bool = False
     VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
+    VLLM_ALLOW_DEPRECATED_BEAM_SEARCH: bool = False
 
 
 def get_default_cache_root():
@@ -195,6 +196,10 @@ def get_default_config_root():
     lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in
              ("true", "1")),
 
+    # If set, allowing the use of deprecated beam search implementation
+    "VLLM_ALLOW_DEPRECATED_BEAM_SEARCH":
+    lambda: os.environ.get("VLLM_ALLOW_DEPRECATED_BEAM_SEARCH", "0") == "1",
+
     # Internal flag to enable Dynamo graph capture
     "VLLM_TEST_DYNAMO_GRAPH_CAPTURE":
     lambda: int(os.environ.get("VLLM_TEST_DYNAMO_GRAPH_CAPTURE", "0")),

vllm/sampling_params.py

@@ -8,6 +8,7 @@
 import torch
 from typing_extensions import Annotated
 
+import vllm.envs as envs
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)
@@ -260,6 +261,10 @@ def __post_init__(self) -> None:
 
         self._verify_args()
         if self.use_beam_search:
+            if not envs.VLLM_ALLOW_DEPRECATED_BEAM_SEARCH:
+                raise ValueError(
+                    "Using beam search as a sampling parameter is deprecated, and will be removed in the future release. Please use the `vllm.LLM.use_beam_search` method for dedicated beam search instead, or set the environment variable `VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1` to suppress this error. For more details, see https://github.com/vllm-project/vllm/issues/8306 ."  # noqa
+                )
             self._verify_beam_search()
         else:
             self._verify_non_beam_search()