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 12, 13:30 UTC
[Bugfix] Fix logic for choosing default prefix caching setting (#29393)
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Thomas Parnell committed
Nov 25, 2025 at 15:05 UTC
516c3f784723f0f25d65305ef7051ae016e18d42
2 files changed
+7
-2
tests/engine/test_arg_utils.py
+2
-1
@@ -277,8 +277,9 @@ def test_prefix_cache_default():
277
parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
278
args = parser.parse_args([])
279
280
+ # should be None by default (depends on model).
281
engine_args = EngineArgs.from_cli_args(args=args)
281
- assert engine_args.enable_prefix_caching, "prefix caching should default to on."
282
+ assert engine_args.enable_prefix_caching is None
283
284
# with flag to turn it on.
285
args = parser.parse_args(["--enable-prefix-caching"])
vllm/engine/arg_utils.py
+5
-1
@@ -880,7 +880,11 @@ class EngineArgs:
880
"--num-gpu-blocks-override", **cache_kwargs["num_gpu_blocks_override"]
881
)
882
cache_group.add_argument(
883
- "--enable-prefix-caching", **cache_kwargs["enable_prefix_caching"]
883
+ "--enable-prefix-caching",
884
+ **{
885
+ **cache_kwargs["enable_prefix_caching"],
886
+ "default": None,
887
+ },
888
)
889
cache_group.add_argument(
890
"--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"]