Simplify (and fix) passing of guided decoding backend options (#17008)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-04-29 19:02:23 +00:00
committed by GitHub
parent 2fa2a50bf9
commit a6977dbd15
17 changed files with 309 additions and 217 deletions
+26 -10
View File
@@ -18,9 +18,9 @@ from vllm import version
from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
ConfigFormat, ConfigType, DecodingConfig, Device,
DeviceConfig, DistributedExecutorBackend,
GuidedDecodingBackendV1, HfOverrides,
KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
ModelConfig, ModelImpl, MultiModalConfig,
GuidedDecodingBackend, GuidedDecodingBackendV1,
HfOverrides, KVTransferConfig, LoadConfig, LoadFormat,
LoRAConfig, ModelConfig, ModelImpl, MultiModalConfig,
ObservabilityConfig, ParallelConfig, PoolerConfig,
PrefixCachingHashAlgo, PromptAdapterConfig,
SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
@@ -317,7 +317,12 @@ class EngineArgs:
bool] = SchedulerConfig.enable_chunked_prefill
disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input
guided_decoding_backend: str = DecodingConfig.guided_decoding_backend
guided_decoding_backend: GuidedDecodingBackend = DecodingConfig.backend
guided_decoding_disable_fallback: bool = DecodingConfig.disable_fallback
guided_decoding_disable_any_whitespace: bool = \
DecodingConfig.disable_any_whitespace
guided_decoding_disable_additional_properties: bool = \
DecodingConfig.disable_additional_properties
logits_processor_pattern: Optional[str] = None
speculative_config: Optional[Dict[str, Any]] = None
@@ -498,9 +503,17 @@ class EngineArgs:
title="DecodingConfig",
description=DecodingConfig.__doc__,
)
guided_decoding_group.add_argument("--guided-decoding-backend",
**guided_decoding_kwargs["backend"])
guided_decoding_group.add_argument(
'--guided-decoding-backend',
**guided_decoding_kwargs["guided_decoding_backend"])
"--guided-decoding-disable-fallback",
**guided_decoding_kwargs["disable_fallback"])
guided_decoding_group.add_argument(
"--guided-decoding-disable-any-whitespace",
**guided_decoding_kwargs["disable_any_whitespace"])
guided_decoding_group.add_argument(
"--guided-decoding-disable-additional-properties",
**guided_decoding_kwargs["disable_additional_properties"])
guided_decoding_group.add_argument(
"--reasoning-parser",
# This choices is a special case because it's not static
@@ -1244,7 +1257,11 @@ class EngineArgs:
if self.enable_prompt_adapter else None
decoding_config = DecodingConfig(
guided_decoding_backend=self.guided_decoding_backend,
backend=self.guided_decoding_backend,
disable_fallback=self.guided_decoding_disable_fallback,
disable_any_whitespace=self.guided_decoding_disable_any_whitespace,
disable_additional_properties=\
self.guided_decoding_disable_additional_properties,
reasoning_backend=self.reasoning_parser
if self.enable_reasoning else None,
)
@@ -1335,9 +1352,8 @@ class EngineArgs:
recommend_to_remove=True)
return False
# remove backend options when doing this check
if self.guided_decoding_backend.split(':')[0] \
not in get_args(GuidedDecodingBackendV1):
if self.guided_decoding_backend not in get_args(
GuidedDecodingBackendV1):
_raise_or_fallback(
feature_name=
f"--guided-decoding-backend={self.guided_decoding_backend}",
+1 -1
View File
@@ -2091,7 +2091,7 @@ class LLMEngine:
tokenizer = self.get_tokenizer(lora_request=lora_request)
guided_decoding.backend = guided_decoding.backend or \
self.decoding_config.guided_decoding_backend
self.decoding_config.backend
if self.decoding_config.reasoning_backend is not None:
logger.debug("Building with reasoning backend %s",
+2 -2
View File
@@ -615,9 +615,9 @@ class MQLLMEngineClient(EngineClient):
build_guided_decoding_logits_processor_async(
sampling_params=params,
tokenizer=await self.get_tokenizer(lora_request),
default_guided_backend=(self.decoding_config.guided_decoding_backend
default_guided_backend=(self.decoding_config.backend
if self.decoding_config
else DecodingConfig.guided_decoding_backend),
else DecodingConfig.backend),
model_config=self.model_config,
reasoning_backend=self.decoding_config.reasoning_backend,
)