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 11, 10:00 UTC
[Bugfix] Fail instead of ignoring when CompilationConfig gets invalid args (#30708)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Dec 15, 2025 at 15:18 UTC
a450c64a30ab6d450b23587611a726af965618b1
2 files changed
+4
-12
tests/benchmarks/test_param_sweep.py
-8
@@ -23,14 +23,6 @@ class TestParameterSweepItem:
23
{"compilation_config.use_inductor_graph_partition": True},
24
"--compilation-config.use_inductor_graph_partition=true",
25
),
26
- (
27
- {"compilation_config.use_inductor": False},
28
- "--compilation-config.use_inductor=false",
29
- ),
30
- (
31
- {"compilation_config.use_inductor": True},
32
- "--compilation-config.use_inductor=true",
33
- ),
26
],
27
)
28
def test_nested_boolean_params(self, input_dict, expected):
vllm/config/compilation.py
+4
-4
@@ -8,7 +8,7 @@ from dataclasses import field
8
from pathlib import Path
9
from typing import TYPE_CHECKING, Any, ClassVar, Literal
10
11
-from pydantic import Field, TypeAdapter, field_validator
11
+from pydantic import ConfigDict, Field, TypeAdapter, field_validator
12
from pydantic.dataclasses import dataclass
13
14
import vllm.envs as envs
@@ -96,7 +96,7 @@ class CUDAGraphMode(enum.Enum):
96
97
98
@config
99
-@dataclass
99
+@dataclass(config=ConfigDict(extra="forbid"))
100
class PassConfig:
101
"""Configuration for custom Inductor passes.
102
@@ -251,7 +251,7 @@ class DynamicShapesType(str, enum.Enum):
251
252
253
@config
254
-@dataclass
254
+@dataclass(config=ConfigDict(extra="forbid"))
255
class DynamicShapesConfig:
256
"""Configuration to control/debug torch compile dynamic shapes."""
257
@@ -290,7 +290,7 @@ class DynamicShapesConfig:
290
291
292
@config
293
-@dataclass
293
+@dataclass(config=ConfigDict(extra="forbid"))
294
class CompilationConfig:
295
"""Configuration for compilation.
296