[V0 Deprecation] Remove Prompt Adapters (#20588)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-07-23 16:36:48 -07:00
committed by GitHub
parent 78c13e30e1
commit 82ec66f514
60 changed files with 126 additions and 1727 deletions
+11 -38
View File
@@ -30,9 +30,9 @@ from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
LogprobsMode, LoRAConfig, ModelConfig, ModelDType,
ModelImpl, MultiModalConfig, ObservabilityConfig,
ParallelConfig, PoolerConfig, PrefixCachingHashAlgo,
PromptAdapterConfig, SchedulerConfig, SchedulerPolicy,
SpeculativeConfig, TaskOption, TokenizerMode,
VllmConfig, get_attr_docs, get_field)
SchedulerConfig, SchedulerPolicy, SpeculativeConfig,
TaskOption, TokenizerMode, VllmConfig, get_attr_docs,
get_field)
from vllm.logger import init_logger
from vllm.platforms import CpuArchEnum, current_platform
from vllm.plugins import load_general_plugins
@@ -358,11 +358,6 @@ class EngineArgs:
max_cpu_loras: Optional[int] = LoRAConfig.max_cpu_loras
lora_dtype: Optional[Union[str, torch.dtype]] = LoRAConfig.lora_dtype
lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size
# PromptAdapter fields
enable_prompt_adapter: bool = False
max_prompt_adapters: int = PromptAdapterConfig.max_prompt_adapters
max_prompt_adapter_token: int = \
PromptAdapterConfig.max_prompt_adapter_token
num_scheduler_steps: int = SchedulerConfig.num_scheduler_steps
multi_step_stream_outputs: bool = SchedulerConfig.multi_step_stream_outputs
@@ -437,6 +432,8 @@ class EngineArgs:
ParallelConfig.enable_multimodal_encoder_data_parallel
async_scheduling: bool = SchedulerConfig.async_scheduling
# DEPRECATED
enable_prompt_adapter: bool = False
def __post_init__(self):
# support `EngineArgs(compilation_config={...})`
@@ -729,23 +726,6 @@ class EngineArgs:
lora_group.add_argument("--default-mm-loras",
**lora_kwargs["default_mm_loras"])
# PromptAdapter related configs
prompt_adapter_kwargs = get_kwargs(PromptAdapterConfig)
prompt_adapter_group = parser.add_argument_group(
title="PromptAdapterConfig",
description=PromptAdapterConfig.__doc__,
)
prompt_adapter_group.add_argument(
"--enable-prompt-adapter",
action=argparse.BooleanOptionalAction,
help="If True, enable handling of PromptAdapters.")
prompt_adapter_group.add_argument(
"--max-prompt-adapters",
**prompt_adapter_kwargs["max_prompt_adapters"])
prompt_adapter_group.add_argument(
"--max-prompt-adapter-token",
**prompt_adapter_kwargs["max_prompt_adapter_token"])
# Speculative arguments
speculative_group = parser.add_argument_group(
title="SpeculativeConfig",
@@ -850,6 +830,12 @@ class EngineArgs:
parser.add_argument('--disable-log-stats',
action='store_true',
help='Disable logging statistics.')
parser.add_argument('--enable-prompt-adapter',
action='store_true',
deprecated=True,
help='[DEPRECATED] Prompt adapter has been '
'removed. Setting this flag to True or False'
' has no effect on vLLM behavior.')
return parser
@@ -1234,11 +1220,6 @@ class EngineArgs:
load_config = self.create_load_config()
prompt_adapter_config = PromptAdapterConfig(
max_prompt_adapters=self.max_prompt_adapters,
max_prompt_adapter_token=self.max_prompt_adapter_token) \
if self.enable_prompt_adapter else None
decoding_config = DecodingConfig(
backend=self.guided_decoding_backend,
disable_fallback=self.guided_decoding_disable_fallback,
@@ -1266,7 +1247,6 @@ class EngineArgs:
load_config=load_config,
decoding_config=decoding_config,
observability_config=observability_config,
prompt_adapter_config=prompt_adapter_config,
compilation_config=self.compilation_config,
kv_transfer_config=self.kv_transfer_config,
kv_events_config=self.kv_events_config,
@@ -1342,12 +1322,6 @@ class EngineArgs:
recommend_to_remove=False)
return False
# No Prompt Adapter so far.
if self.enable_prompt_adapter:
_raise_or_fallback(feature_name="--enable-prompt-adapter",
recommend_to_remove=False)
return False
# No text embedding inputs so far.
if self.enable_prompt_embeds:
_raise_or_fallback(feature_name="--enable-prompt-embeds",
@@ -1469,7 +1443,6 @@ class EngineArgs:
if (is_gpu and not use_sliding_window and not use_spec_decode
and not self.enable_lora
and not self.enable_prompt_adapter
and model_config.runner_type != "pooling"):
self.enable_chunked_prefill = True
logger.warning(
-10
View File
@@ -29,7 +29,6 @@ from vllm.model_executor.guided_decoding import (
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.outputs import PoolingRequestOutput, RequestOutput
from vllm.pooling_params import PoolingParams
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sampling_params import SamplingParams
from vllm.sequence import ExecuteModelRequest
from vllm.transformers_utils.tokenizer import AnyTokenizer
@@ -435,7 +434,6 @@ class _AsyncLLMEngine(LLMEngine):
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
data_parallel_rank: Optional[int] = None,
tokenization_kwargs: Optional[dict[str, Any]] = None,
@@ -468,7 +466,6 @@ class _AsyncLLMEngine(LLMEngine):
processed_inputs = await self.input_preprocessor.preprocess_async(
prompt,
lora_request=lora_request,
prompt_adapter_request=prompt_adapter_request,
tokenization_kwargs=tokenization_kwargs,
)
@@ -491,7 +488,6 @@ class _AsyncLLMEngine(LLMEngine):
params=params,
arrival_time=arrival_time,
lora_request=lora_request,
prompt_adapter_request=prompt_adapter_request,
trace_headers=trace_headers,
priority=priority,
)
@@ -861,7 +857,6 @@ class AsyncLLMEngine(EngineClient):
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
data_parallel_rank: Optional[int] = None,
tokenization_kwargs: Optional[dict[str, Any]] = None,
@@ -889,7 +884,6 @@ class AsyncLLMEngine(EngineClient):
arrival_time=arrival_time or time.time(),
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
priority=priority,
data_parallel_rank=data_parallel_rank,
tokenization_kwargs=tokenization_kwargs,
@@ -904,7 +898,6 @@ class AsyncLLMEngine(EngineClient):
request_id: str,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
data_parallel_rank: Optional[int] = None,
) -> AsyncGenerator[RequestOutput, None]:
@@ -922,8 +915,6 @@ class AsyncLLMEngine(EngineClient):
request_id: The unique id of the request.
lora_request: LoRA request to use for generation, if any.
trace_headers: OpenTelemetry trace headers.
prompt_adapter_request: Prompt Adapter request to use
for generation, if any.
priority: The priority of the request.
Only applicable with priority scheduling.
data_parallel_rank: The (global) data parallel rank that must
@@ -983,7 +974,6 @@ class AsyncLLMEngine(EngineClient):
sampling_params,
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
priority=priority,
data_parallel_rank=data_parallel_rank,
):
+18 -50
View File
@@ -44,7 +44,6 @@ from vllm.multimodal.processing import EncDecMultiModalProcessor
from vllm.outputs import (PoolingRequestOutput, RequestOutput,
RequestOutputFactory)
from vllm.pooling_params import PoolingParams
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sampling_params import RequestOutputKind, SamplingParams
from vllm.sequence import (ExecuteModelRequest, ParallelSampleSequenceGroup,
PoolingSequenceGroupOutput, Sequence, SequenceGroup,
@@ -223,7 +222,6 @@ class LLMEngine:
self.load_config = vllm_config.load_config
self.decoding_config = vllm_config.decoding_config or DecodingConfig( # noqa
)
self.prompt_adapter_config = vllm_config.prompt_adapter_config # noqa
self.observability_config = vllm_config.observability_config or ObservabilityConfig( # noqa
)
@@ -294,8 +292,6 @@ class LLMEngine:
# Feature flags
"enable_lora":
bool(self.lora_config),
"enable_prompt_adapter":
bool(self.prompt_adapter_config),
"enable_prefix_caching":
self.cache_config.enable_prefix_caching,
"enforce_eager":
@@ -542,9 +538,6 @@ class LLMEngine:
self.lora_config.verify_with_model_config(self.model_config)
self.lora_config.verify_with_scheduler_config(
self.scheduler_config)
if self.prompt_adapter_config:
self.prompt_adapter_config.verify_with_model_config(
self.model_config)
def _add_processed_request(
self,
@@ -553,7 +546,6 @@ class LLMEngine:
params: Union[SamplingParams, PoolingParams],
arrival_time: float,
lora_request: Optional[LoRARequest],
prompt_adapter_request: Optional[PromptAdapterRequest],
trace_headers: Optional[Mapping[str, str]] = None,
priority: int = 0,
) -> Optional[SequenceGroup]:
@@ -569,7 +561,6 @@ class LLMEngine:
arrival_time=arrival_time,
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
priority=priority,
)
return None
@@ -583,11 +574,10 @@ class LLMEngine:
encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
seq = Sequence(seq_id, decoder_inputs, block_size, eos_token_id,
lora_request, prompt_adapter_request)
lora_request)
encoder_seq = (None if encoder_inputs is None else Sequence(
seq_id, encoder_inputs, block_size, eos_token_id, lora_request,
prompt_adapter_request))
seq_id, encoder_inputs, block_size, eos_token_id, lora_request))
# Create a SequenceGroup based on SamplingParams or PoolingParams
if isinstance(params, SamplingParams):
@@ -598,7 +588,6 @@ class LLMEngine:
arrival_time=arrival_time,
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
encoder_seq=encoder_seq,
priority=priority)
elif isinstance(params, PoolingParams):
@@ -608,7 +597,6 @@ class LLMEngine:
params,
arrival_time=arrival_time,
lora_request=lora_request,
prompt_adapter_request=prompt_adapter_request,
encoder_seq=encoder_seq,
priority=priority)
else:
@@ -637,7 +625,6 @@ class LLMEngine:
lora_request: Optional[LoRARequest] = None,
tokenization_kwargs: Optional[dict[str, Any]] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
) -> None:
"""Add a request to the engine's request pool.
@@ -658,7 +645,6 @@ class LLMEngine:
the current monotonic time.
lora_request: The LoRA request to add.
trace_headers: OpenTelemetry trace headers.
prompt_adapter_request: The prompt adapter request to add.
priority: The priority of the request.
Only applicable with priority scheduling.
@@ -719,7 +705,6 @@ class LLMEngine:
prompt,
tokenization_kwargs=tokenization_kwargs,
lora_request=lora_request,
prompt_adapter_request=prompt_adapter_request,
)
self._add_processed_request(
@@ -728,7 +713,6 @@ class LLMEngine:
params=params,
arrival_time=arrival_time,
lora_request=lora_request,
prompt_adapter_request=prompt_adapter_request,
trace_headers=trace_headers,
priority=priority,
)
@@ -741,7 +725,6 @@ class LLMEngine:
arrival_time: float,
lora_request: Optional[LoRARequest],
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
encoder_seq: Optional[Sequence] = None,
priority: int = 0,
) -> SequenceGroup:
@@ -769,17 +752,15 @@ class LLMEngine:
if self.vllm_config.speculative_config is not None:
draft_size = \
self.vllm_config.speculative_config.num_speculative_tokens + 1
seq_group = SequenceGroup(
request_id=request_id,
seqs=[seq],
arrival_time=arrival_time,
sampling_params=sampling_params,
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
encoder_seq=encoder_seq,
priority=priority,
draft_size=draft_size)
seq_group = SequenceGroup(request_id=request_id,
seqs=[seq],
arrival_time=arrival_time,
sampling_params=sampling_params,
lora_request=lora_request,
trace_headers=trace_headers,
encoder_seq=encoder_seq,
priority=priority,
draft_size=draft_size)
return seq_group
@@ -790,7 +771,6 @@ class LLMEngine:
pooling_params: PoolingParams,
arrival_time: float,
lora_request: Optional[LoRARequest],
prompt_adapter_request: Optional[PromptAdapterRequest],
encoder_seq: Optional[Sequence] = None,
priority: int = 0,
) -> SequenceGroup:
@@ -798,15 +778,13 @@ class LLMEngine:
# Defensive copy of PoolingParams, which are used by the pooler
pooling_params = pooling_params.clone()
# Create the sequence group.
seq_group = SequenceGroup(
request_id=request_id,
seqs=[seq],
arrival_time=arrival_time,
lora_request=lora_request,
pooling_params=pooling_params,
prompt_adapter_request=prompt_adapter_request,
encoder_seq=encoder_seq,
priority=priority)
seq_group = SequenceGroup(request_id=request_id,
seqs=[seq],
arrival_time=arrival_time,
lora_request=lora_request,
pooling_params=pooling_params,
encoder_seq=encoder_seq,
priority=priority)
return seq_group
def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
@@ -1834,16 +1812,6 @@ class LLMEngine:
def pin_lora(self, lora_id: int) -> bool:
return self.model_executor.pin_lora(lora_id)
def add_prompt_adapter(
self, prompt_adapter_request: PromptAdapterRequest) -> bool:
return self.model_executor.add_prompt_adapter(prompt_adapter_request)
def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool:
return self.model_executor.remove_prompt_adapter(prompt_adapter_id)
def list_prompt_adapters(self) -> List[int]:
return self.model_executor.list_prompt_adapters()
def start_profile(self) -> None:
self.model_executor.start_profile()
-4
View File
@@ -10,7 +10,6 @@ from vllm import PoolingParams
from vllm.inputs import PromptType
from vllm.lora.request import LoRARequest
from vllm.outputs import RequestOutput
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sampling_params import SamplingParams
from vllm.utils import Device
@@ -33,7 +32,6 @@ class RPCProcessRequest:
request_id: str
lora_request: Optional[LoRARequest] = None
trace_headers: Optional[Mapping[str, str]] = None
prompt_adapter_request: Optional[PromptAdapterRequest] = None
priority: int = 0
def __init__(
@@ -43,7 +41,6 @@ class RPCProcessRequest:
request_id: str,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
) -> None:
super().__init__()
@@ -53,7 +50,6 @@ class RPCProcessRequest:
self.request_id = request_id
self.lora_request = lora_request
self.trace_headers = trace_headers
self.prompt_adapter_request = prompt_adapter_request
self.priority = priority
+1 -8
View File
@@ -45,7 +45,6 @@ from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.outputs import PoolingRequestOutput, RequestOutput
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sampling_params import SamplingParams
from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
from vllm.utils import Device
@@ -448,7 +447,6 @@ class MQLLMEngineClient(EngineClient):
request_id: str,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
) -> AsyncGenerator[RequestOutput, None]:
"""Generate outputs for a request.
@@ -465,8 +463,6 @@ class MQLLMEngineClient(EngineClient):
request_id: The unique id of the request.
lora_request: LoRA request to use for generation, if any.
trace_headers: OpenTelemetry trace headers.
prompt_adapter_request: Prompt Adapter request to use
for generation, if any.
priority: Priority of the request (lower means earlier handling).
Any priority other than 0 will lead to an error if the
scheduling policy is not "priority".
@@ -474,8 +470,7 @@ class MQLLMEngineClient(EngineClient):
return cast(
AsyncGenerator[RequestOutput, None],
self._process_request(prompt, sampling_params, request_id,
lora_request, trace_headers,
prompt_adapter_request, priority))
lora_request, trace_headers, priority))
def encode(
self,
@@ -521,7 +516,6 @@ class MQLLMEngineClient(EngineClient):
request_id: str,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
) -> Union[AsyncGenerator[RequestOutput, None], AsyncGenerator[
PoolingRequestOutput, None]]:
@@ -575,7 +569,6 @@ class MQLLMEngineClient(EngineClient):
request_id=request_id,
lora_request=lora_request,
trace_headers=trace_headers,
prompt_adapter_request=prompt_adapter_request,
priority=priority,
))
+6 -8
View File
@@ -304,14 +304,12 @@ class MQLLMEngine:
self._send_outputs(rpc_err)
try:
self.engine.add_request(
request_id=request_id,
prompt=request.prompt,
params=request.params,
lora_request=request.lora_request,
trace_headers=request.trace_headers,
prompt_adapter_request=request.prompt_adapter_request,
priority=request.priority)
self.engine.add_request(request_id=request_id,
prompt=request.prompt,
params=request.params,
lora_request=request.lora_request,
trace_headers=request.trace_headers,
priority=request.priority)
if self.log_requests:
logger.info("Added request %s.", request.request_id)
-2
View File
@@ -16,7 +16,6 @@ from vllm.lora.request import LoRARequest
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.outputs import CompletionOutput, PoolingRequestOutput, RequestOutput
from vllm.pooling_params import PoolingParams
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.utils import Device, collect_from_async_generator, random_uuid
@@ -55,7 +54,6 @@ class EngineClient(ABC):
request_id: str,
lora_request: Optional[LoRARequest] = None,
trace_headers: Optional[Mapping[str, str]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
priority: int = 0,
) -> AsyncGenerator[RequestOutput, None]:
"""Generate outputs for a request."""