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 15, 11:00 UTC
[V1] Improve VLLM_ALLOW_INSECURE_SERIALIZATION logging (#17860)
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Russell Bryant committed
May 8, 2025 at 12:57 UTC
226a4272cf6d184f20c6638f29a409347c6c474b
1 file changed
+10
-7
vllm/v1/serial_utils.py
+10
-7
@@ -40,6 +40,11 @@ MMF_CLASS_TO_FACTORY: dict[type[BaseMultiModalField], str] = {
40
bytestr = Union[bytes, bytearray, memoryview, zmq.Frame]
41
42
43
+def _log_insecure_serialization_warning():
44
+ logger.warning_once("Allowing insecure serialization using pickle due to "
45
+ "VLLM_ALLOW_INSECURE_SERIALIZATION=1")
46
+
47
+
48
class MsgpackEncoder:
49
"""Encoder with custom torch tensor and numpy array serialization.
50
@@ -60,9 +65,7 @@ class MsgpackEncoder:
65
self.aux_buffers: Optional[list[bytestr]] = None
66
self.size_threshold = size_threshold
67
if envs.VLLM_ALLOW_INSECURE_SERIALIZATION:
63
- logger.warning(
64
- "Allowing insecure serialization using pickle due to "
65
- "VLLM_ALLOW_INSECURE_SERIALIZATION=1")
68
+ _log_insecure_serialization_warning()
69
70
def encode(self, obj: Any) -> Sequence[bytestr]:
71
try:
@@ -119,7 +122,9 @@ class MsgpackEncoder:
122
for item in itemlist]
123
124
if not envs.VLLM_ALLOW_INSECURE_SERIALIZATION:
122
- raise TypeError(f"Object of type {type(obj)} is not serializable")
125
+ raise TypeError(f"Object of type {type(obj)} is not serializable"
126
+ "Set VLLM_ALLOW_INSECURE_SERIALIZATION=1 to allow "
127
+ "fallback to pickle-based serialization.")
128
129
if isinstance(obj, FunctionType):
130
# `pickle` is generally faster than cloudpickle, but can have
@@ -202,9 +207,7 @@ class MsgpackDecoder:
207
dec_hook=self.dec_hook)
208
self.aux_buffers: Sequence[bytestr] = ()
209
if envs.VLLM_ALLOW_INSECURE_SERIALIZATION:
205
- logger.warning(
206
- "Allowing insecure deserialization using pickle due to "
207
- "VLLM_ALLOW_INSECURE_SERIALIZATION=1")
210
+ _log_insecure_serialization_warning()
211
212
def decode(self, bufs: Union[bytestr, Sequence[bytestr]]) -> Any:
213
if isinstance(bufs, (bytes, bytearray, memoryview, zmq.Frame)):