[Frontend] Remove AsyncMicrobatchTokenizer. (#45759)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
+12
-16
@@ -38,10 +38,7 @@ from vllm.multimodal.processing import BaseMultiModalProcessor
|
||||
from vllm.multimodal.processing import ProcessorInputs as MMProcessorInputs
|
||||
from vllm.multimodal.registry import MultiModalTimingRegistry
|
||||
from vllm.tokenizers import TokenizerLike
|
||||
from vllm.utils.async_utils import (
|
||||
AsyncMicrobatchTokenizer,
|
||||
make_async,
|
||||
)
|
||||
from vllm.utils.async_utils import make_async
|
||||
from vllm.utils.counter import AtomicCounter
|
||||
from vllm.utils.torch_utils import set_default_torch_num_threads
|
||||
from vllm.v1.metrics.stats import MultiModalCacheStats
|
||||
@@ -92,8 +89,9 @@ class BaseRenderer(ABC, Generic[_T]):
|
||||
# to keep the asyncio event loop responsive under concurrent load.
|
||||
self._mm_executor: Executor = self._executor
|
||||
|
||||
# Lazy initialization since offline LLM doesn't use async
|
||||
self._async_tokenizer: AsyncMicrobatchTokenizer | None = None
|
||||
# Offloading tokenizer encode & decode to thread pool.
|
||||
self._async_tokenizer_encode = make_async(self._encode, executor=self._executor)
|
||||
self._async_tokenizer_decode = make_async(self._decode, executor=self._executor)
|
||||
|
||||
self.mm_processor: BaseMultiModalProcessor | None = None
|
||||
self._readonly_mm_processor: BaseMultiModalProcessor | None = None
|
||||
@@ -146,13 +144,11 @@ class BaseRenderer(ABC, Generic[_T]):
|
||||
|
||||
return tokenizer
|
||||
|
||||
def get_async_tokenizer(self) -> AsyncMicrobatchTokenizer:
|
||||
if self._async_tokenizer is None:
|
||||
self._async_tokenizer = AsyncMicrobatchTokenizer(
|
||||
self.get_tokenizer(), executor=self._executor
|
||||
)
|
||||
def _decode(self, *args, **kwargs):
|
||||
return self.get_tokenizer().decode(*args, **kwargs)
|
||||
|
||||
return self._async_tokenizer
|
||||
def _encode(self, *args, **kwargs):
|
||||
return self.get_tokenizer().encode(*args, **kwargs)
|
||||
|
||||
def get_mm_processor(self) -> "BaseMultiModalProcessor":
|
||||
if self.mm_processor is None:
|
||||
@@ -436,8 +432,7 @@ class BaseRenderer(ABC, Generic[_T]):
|
||||
prompt: TextPrompt,
|
||||
params: TokenizeParams,
|
||||
) -> TokensPrompt:
|
||||
tokenizer = self.get_async_tokenizer()
|
||||
prompt_token_ids = await tokenizer.encode(
|
||||
prompt_token_ids = await self._async_tokenizer_encode(
|
||||
prompt["prompt"],
|
||||
**params.get_encode_kwargs(),
|
||||
)
|
||||
@@ -451,8 +446,9 @@ class BaseRenderer(ABC, Generic[_T]):
|
||||
return prompt
|
||||
|
||||
async def _detokenize_prompt_async(self, prompt: TokensPrompt) -> TokensPrompt:
|
||||
tokenizer = self.get_async_tokenizer()
|
||||
prompt["prompt"] = await tokenizer.decode(prompt["prompt_token_ids"])
|
||||
prompt["prompt"] = await self._async_tokenizer_decode(
|
||||
prompt["prompt_token_ids"]
|
||||
)
|
||||
|
||||
return prompt
|
||||
|
||||
|
||||
Reference in New Issue
Block a user