A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 22:00 UTC
[Misc] Add deprecation warning for beam search (#6402)
Woosuk Kwon committed
Jul 13, 2024 at 11:52 UTC
eeceadaecc80cc51c4e9ddae0cb99a33d379452d
2 files changed
+17
vllm/envs.py
+5
@@ -41,6 +41,7 @@ if TYPE_CHECKING:
41
NVCC_THREADS: Optional[str] = None
42
VLLM_USE_PRECOMPILED: bool = False
43
VLLM_INSTALL_PUNICA_KERNELS: bool = False
44
+ VLLM_NO_DEPRECATION_WARNING: bool = False
45
CMAKE_BUILD_TYPE: Optional[str] = None
46
VERBOSE: bool = False
47
@@ -251,6 +252,10 @@ environment_variables: Dict[str, Callable[[], Any]] = {
252
lambda: os.getenv("VLLM_XLA_CACHE_PATH", "~/.vllm/xla_cache/"),
253
"VLLM_FUSED_MOE_CHUNK_SIZE":
254
lambda: int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "65536")),
255
+
256
+ # If set, vllm will skip the deprecation warnings.
257
+ "VLLM_NO_DEPRECATION_WARNING":
258
+ lambda: bool(int(os.getenv("VLLM_NO_DEPRECATION_WARNING", "0"))),
259
}
260
261
# end-env-vars-definition
vllm/sampling_params.py
+12
@@ -8,6 +8,11 @@ import torch
8
from pydantic import Field
9
from typing_extensions import Annotated
10
11
+import vllm.envs as envs
12
+from vllm.logger import init_logger
13
+
14
+logger = init_logger(__name__)
15
+
16
_SAMPLING_EPS = 1e-5
17
18
@@ -184,6 +189,13 @@ class SamplingParams:
189
190
self._verify_args()
191
if self.use_beam_search:
192
+ if not envs.VLLM_NO_DEPRECATION_WARNING:
193
+ logger.warning(
194
+ "[IMPORTANT] We plan to discontinue the support for beam "
195
+ "search in the next major release. Please refer to "
196
+ "https://github.com/vllm-project/vllm/issues/6226 for "
197
+ "more information. Set VLLM_NO_DEPRECATION_WARNING=1 to "
198
+ "suppress this warning.")
199
self._verify_beam_search()
200
else:
201
self._verify_non_beam_search()