[Frontend] Consolidate tokenizer init code (#26276)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-10-06 09:34:52 +00:00
committed by GitHub
parent 77c95f72f7
commit 391612e78b
8 changed files with 46 additions and 70 deletions
+6 -14
View File
@@ -28,7 +28,7 @@ from vllm.sampling_params import SamplingParams
from vllm.tasks import SupportedTask
from vllm.tracing import init_tracer
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
from vllm.transformers_utils.tokenizer import AnyTokenizer, init_tokenizer_from_configs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.usage.usage_lib import UsageContext
from vllm.utils import Device, as_list, cancel_task_threadsafe, cdiv, deprecate_kwargs
from vllm.v1.engine import EngineCoreRequest
@@ -104,20 +104,8 @@ class AsyncLLM(EngineClient):
"logger list; enabling logging without default stat loggers"
)
if self.model_config.skip_tokenizer_init:
self.tokenizer = None
else:
# Tokenizer (+ ensure liveness if running in another process).
self.tokenizer = init_tokenizer_from_configs(
model_config=vllm_config.model_config
)
# Processor (converts Inputs --> EngineCoreRequests).
self.processor = Processor(
vllm_config=vllm_config,
tokenizer=self.tokenizer,
mm_registry=mm_registry,
)
self.processor = Processor(vllm_config, mm_registry=mm_registry)
# OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
self.output_processor = OutputProcessor(
@@ -257,6 +245,10 @@ class AsyncLLM(EngineClient):
cancel_task_threadsafe(getattr(self, "output_handler", None))
@property
def tokenizer(self) -> Optional[AnyTokenizer]:
return self.processor.tokenizer
async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
return await self.engine_core.get_supported_tasks_async()
+10 -12
View File
@@ -23,7 +23,7 @@ from vllm.pooling_params import PoolingParams
from vllm.sampling_params import SamplingParams
from vllm.tasks import SupportedTask
from vllm.tracing import init_tracer
from vllm.transformers_utils.tokenizer import AnyTokenizer, init_tokenizer_from_configs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.usage.usage_lib import UsageContext
from vllm.utils import Device
from vllm.v1.engine import EngineCoreRequest
@@ -95,18 +95,8 @@ class LLMEngine:
self.dp_group = None
self.should_execute_dummy_batch = False
if self.model_config.skip_tokenizer_init:
self.tokenizer = None
else:
# Tokenizer (+ ensure liveness if running in another process).
self.tokenizer = init_tokenizer_from_configs(
model_config=vllm_config.model_config
)
# Processor (convert Inputs --> EngineCoreRequests)
self.processor = Processor(
vllm_config=vllm_config, tokenizer=self.tokenizer, mm_registry=mm_registry
)
self.processor = Processor(vllm_config, mm_registry=mm_registry)
# OutputProcessor (convert EngineCoreOutputs --> RequestOutput).
self.output_processor = OutputProcessor(
@@ -214,6 +204,14 @@ class LLMEngine:
def validate_outputs(cls, outputs, output_type):
return outputs
@property
def tokenizer(self) -> Optional[AnyTokenizer]:
return self.processor.tokenizer
@tokenizer.setter
def tokenizer(self, tokenizer: Optional[AnyTokenizer]) -> None:
self.processor.tokenizer = tokenizer
def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
return self.engine_core.get_supported_tasks()
+11 -8
View File
@@ -37,15 +37,13 @@ class Processor:
def __init__(
self,
vllm_config: VllmConfig,
tokenizer: AnyTokenizer,
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
):
) -> None:
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.lora_config = vllm_config.lora_config
self.structured_outputs_config = vllm_config.structured_outputs_config
self.tokenizer = tokenizer
self.generation_config_fields = self.model_config.try_get_generation_config()
@@ -54,11 +52,18 @@ class Processor:
self.input_preprocessor = InputPreprocessor(
self.model_config,
self.tokenizer,
mm_registry,
mm_processor_cache=self.mm_processor_cache,
)
@property
def tokenizer(self) -> Optional[AnyTokenizer]:
return self.input_preprocessor.tokenizer
@tokenizer.setter
def tokenizer(self, tokenizer: Optional[AnyTokenizer]) -> None:
self.input_preprocessor.tokenizer = tokenizer
def _validate_logprobs(
self,
params: SamplingParams,
@@ -511,10 +516,8 @@ class Processor:
else:
raise ValueError(f"The {prompt_type} prompt cannot be empty")
if self.model_config.skip_tokenizer_init:
tokenizer = None
else:
tokenizer = self.tokenizer
tokenizer = self.tokenizer
if tokenizer is not None:
max_input_id = max(prompt_ids or [], default=0)
# NOTE: tokenizer.max_token_id is the tokenizer’s vocab size while