Simplify TokenizerGroup (#16790)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-04-24 04:43:56 -07:00
committed by GitHub
parent 14288d1332
commit 0a05ed57e6
24 changed files with 80 additions and 752 deletions
+8 -18
View File
@@ -7,9 +7,8 @@ import json
import re
import threading
from dataclasses import MISSING, dataclass, fields
from typing import (TYPE_CHECKING, Any, Callable, Dict, List, Literal,
Optional, Tuple, Type, TypeVar, Union, cast, get_args,
get_origin)
from typing import (Any, Callable, Dict, List, Literal, Optional, Tuple, Type,
TypeVar, Union, cast, get_args, get_origin)
import torch
from typing_extensions import TypeIs
@@ -23,7 +22,7 @@ from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
ModelConfig, ModelImpl, MultiModalConfig,
ObservabilityConfig, ParallelConfig, PoolerConfig,
PoolType, PrefixCachingHashAlgo, PromptAdapterConfig,
PrefixCachingHashAlgo, PromptAdapterConfig,
SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
TaskOption, TokenizerPoolConfig, VllmConfig,
get_attr_docs, get_field)
@@ -39,9 +38,6 @@ from vllm.utils import FlexibleArgumentParser, GiB_bytes, is_in_ray_actor
# yapf: enable
if TYPE_CHECKING:
from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
logger = init_logger(__name__)
ALLOWED_DETAILED_TRACE_MODULES = ["model", "worker", "all"]
@@ -185,13 +181,12 @@ class EngineArgs:
enforce_eager: Optional[bool] = None
max_seq_len_to_capture: int = 8192
disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce
# The following three fields are deprecated and will be removed in a future
# release. Setting them will have no effect. Please remove them from your
# configurations.
tokenizer_pool_size: int = TokenizerPoolConfig.pool_size
# Note: Specifying a tokenizer pool by passing a class
# is intended for expert use only. The API may change without
# notice.
tokenizer_pool_type: Union[PoolType, Type["BaseTokenizerGroup"]] = \
TokenizerPoolConfig.pool_type
tokenizer_pool_extra_config: dict[str, Any] = \
tokenizer_pool_type: str = TokenizerPoolConfig.pool_type
tokenizer_pool_extra_config: dict = \
get_field(TokenizerPoolConfig, "extra_config")
limit_mm_per_prompt: dict[str, int] = \
get_field(MultiModalConfig, "limit_per_prompt")
@@ -1187,11 +1182,6 @@ class EngineArgs:
enable_expert_parallel=self.enable_expert_parallel,
max_parallel_loading_workers=self.max_parallel_loading_workers,
disable_custom_all_reduce=self.disable_custom_all_reduce,
tokenizer_pool_config=TokenizerPoolConfig.create_config(
self.tokenizer_pool_size,
self.tokenizer_pool_type,
self.tokenizer_pool_extra_config,
),
ray_workers_use_nsight=self.ray_workers_use_nsight,
placement_group=placement_group,
distributed_executor_backend=self.distributed_executor_backend,
-2
View File
@@ -526,8 +526,6 @@ class _AsyncLLMEngine(LLMEngine):
)
async def check_health_async(self) -> None:
if self.tokenizer:
self.tokenizer.check_health()
self.model_executor.check_health()
+6 -24
View File
@@ -55,7 +55,7 @@ from vllm.tracing import (SpanAttributes, SpanKind, extract_trace_context,
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.transformers_utils.tokenizer_group import (
BaseTokenizerGroup, init_tokenizer_from_configs)
TokenizerGroup, init_tokenizer_from_configs)
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
usage_message)
from vllm.utils import (Counter, Device, deprecate_kwargs,
@@ -66,7 +66,6 @@ from vllm.worker.model_runner_base import InputProcessingError
logger = init_logger(__name__)
_LOCAL_LOGGING_INTERVAL_SEC = 5
_G = TypeVar("_G", bound=BaseTokenizerGroup, default=BaseTokenizerGroup)
_O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
_R = TypeVar("_R", default=Any)
@@ -205,7 +204,7 @@ class LLMEngine:
return outputs_
tokenizer: Optional[BaseTokenizerGroup]
tokenizer: Optional[TokenizerGroup]
def __init__(
self,
@@ -321,11 +320,6 @@ class LLMEngine:
self.parallel_config.disable_custom_all_reduce,
})
if self.tokenizer:
# Ping the tokenizer to ensure liveness if it runs in a
# different process.
self.tokenizer.ping()
self.cached_scheduler_outputs = [
SchedulerOutputState()
for _ in range(self.parallel_config.pipeline_parallel_size)
@@ -537,21 +531,12 @@ class LLMEngine:
if model_executor := getattr(self, "model_executor", None):
model_executor.shutdown()
def get_tokenizer_group(
self,
group_type: Type[_G] = BaseTokenizerGroup,
) -> _G:
tokenizer_group = self.tokenizer
if tokenizer_group is None:
def get_tokenizer_group(self) -> TokenizerGroup:
if self.tokenizer is None:
raise ValueError("Unable to get tokenizer because "
"skip_tokenizer_init is True")
if not isinstance(tokenizer_group, group_type):
raise TypeError("Invalid type of tokenizer group. "
f"Expected type: {group_type}, but "
f"found type: {type(tokenizer_group)}")
return tokenizer_group
return self.tokenizer
def get_tokenizer(
self,
@@ -559,11 +544,10 @@ class LLMEngine:
) -> AnyTokenizer:
return self.get_tokenizer_group().get_lora_tokenizer(lora_request)
def _init_tokenizer(self) -> BaseTokenizerGroup:
def _init_tokenizer(self) -> TokenizerGroup:
return init_tokenizer_from_configs(
model_config=self.model_config,
scheduler_config=self.scheduler_config,
parallel_config=self.parallel_config,
lora_config=self.lora_config)
def _verify_args(self) -> None:
@@ -1952,8 +1936,6 @@ class LLMEngine:
return self.model_executor.is_sleeping
def check_health(self) -> None:
if self.tokenizer:
self.tokenizer.check_health()
self.model_executor.check_health()
def is_tracing_enabled(self) -> bool:
-1
View File
@@ -101,7 +101,6 @@ class MQLLMEngineClient(EngineClient):
self.tokenizer = init_tokenizer_from_configs(
model_config=self.model_config,
scheduler_config=engine_config.scheduler_config,
parallel_config=engine_config.parallel_config,
lora_config=engine_config.lora_config)
self.input_preprocessor = InputPreprocessor(self.model_config,
self.tokenizer)