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 8, 06:15 UTC
[KV Connector] Add temporary, off-by-default `VLLM_DISABLE_REQUEST_ID_RANDOMIZATION` workaround (#34415)
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Seiji Eicher committed
Feb 14, 2026 at 23:26 UTC
79c7e092350e4ae82d679ea4b2cdaaa4b580944b
2 files changed
+15
-1
vllm/envs.py
+6
@@ -169,6 +169,7 @@ if TYPE_CHECKING:
169
VLLM_XGRAMMAR_CACHE_MB: int = 0
170
VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
171
VLLM_ALLOW_INSECURE_SERIALIZATION: bool = False
172
+ VLLM_DISABLE_REQUEST_ID_RANDOMIZATION: bool = False
173
VLLM_NIXL_SIDE_CHANNEL_HOST: str = "localhost"
174
VLLM_NIXL_SIDE_CHANNEL_PORT: int = 5600
175
VLLM_MOONCAKE_BOOTSTRAP_PORT: int = 8998
@@ -1236,6 +1237,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
1237
"VLLM_ALLOW_INSECURE_SERIALIZATION": lambda: bool(
1238
int(os.getenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "0"))
1239
),
1240
+ # Temporary: skip adding random suffix to internal request IDs. May be
1241
+ # needed for KV connectors that match request IDs across instances.
1242
+ "VLLM_DISABLE_REQUEST_ID_RANDOMIZATION": lambda: bool(
1243
+ int(os.getenv("VLLM_DISABLE_REQUEST_ID_RANDOMIZATION", "0"))
1244
+ ),
1245
# IP address used for NIXL handshake between remote agents.
1246
"VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.getenv(
1247
"VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
vllm/v1/engine/input_processor.py
+9
-1
@@ -5,6 +5,7 @@ import time
5
from collections.abc import Mapping
6
from typing import Any, Literal, cast
7
8
+import vllm.envs as envs
9
from vllm.config import VllmConfig
10
from vllm.inputs.data import (
11
ProcessorInputs,
@@ -296,7 +297,14 @@ class InputProcessor:
297
" passed to vLLM; use the request_id field."
298
)
299
request.external_req_id = request.request_id
299
- request.request_id = f"{request.external_req_id}-{random_uuid():.8}"
300
+ if envs.VLLM_DISABLE_REQUEST_ID_RANDOMIZATION:
301
+ logger.warning_once(
302
+ "VLLM_DISABLE_REQUEST_ID_RANDOMIZATION is set and will be "
303
+ "removed in a future release. Duplicate externally-provided "
304
+ "request IDs may cause failures and/or subtle correctness errors."
305
+ )
306
+ else:
307
+ request.request_id = f"{request.external_req_id}-{random_uuid():.8}"
308
309
def process_inputs(
310
self,