[Core] Rename input data types (#8688)
This commit is contained in:
+29
-8
@@ -1,7 +1,8 @@
|
||||
from .data import (EncoderDecoderLLMInputs, ExplicitEncoderDecoderPrompt,
|
||||
LLMInputs, PromptType, SingletonPrompt, TextPrompt,
|
||||
TokensPrompt, build_explicit_enc_dec_prompt,
|
||||
to_enc_dec_tuple_list, zip_enc_dec_prompts)
|
||||
from .data import (DecoderOnlyInputs, EncoderDecoderInputs,
|
||||
ExplicitEncoderDecoderPrompt, PromptType, SingletonInputs,
|
||||
SingletonPrompt, TextPrompt, TokenInputs, TokensPrompt,
|
||||
build_explicit_enc_dec_prompt, to_enc_dec_tuple_list,
|
||||
token_inputs, zip_enc_dec_prompts)
|
||||
from .registry import InputContext, InputRegistry
|
||||
|
||||
INPUT_REGISTRY = InputRegistry()
|
||||
@@ -19,8 +20,11 @@ __all__ = [
|
||||
"PromptType",
|
||||
"SingletonPrompt",
|
||||
"ExplicitEncoderDecoderPrompt",
|
||||
"LLMInputs",
|
||||
"EncoderDecoderLLMInputs",
|
||||
"TokenInputs",
|
||||
"token_inputs",
|
||||
"SingletonInputs",
|
||||
"DecoderOnlyInputs",
|
||||
"EncoderDecoderInputs",
|
||||
"build_explicit_enc_dec_prompt",
|
||||
"to_enc_dec_tuple_list",
|
||||
"zip_enc_dec_prompts",
|
||||
@@ -31,9 +35,9 @@ __all__ = [
|
||||
|
||||
|
||||
def __getattr__(name: str):
|
||||
if name == "PromptInput":
|
||||
import warnings
|
||||
import warnings
|
||||
|
||||
if name == "PromptInput":
|
||||
msg = ("PromptInput has been renamed to PromptType. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
@@ -41,4 +45,21 @@ def __getattr__(name: str):
|
||||
|
||||
return PromptType
|
||||
|
||||
if name == "LLMInputs":
|
||||
msg = ("LLMInputs has been renamed to DecoderOnlyInputs. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
warnings.warn(DeprecationWarning(msg), stacklevel=2)
|
||||
|
||||
return DecoderOnlyInputs
|
||||
|
||||
if name == "EncoderDecoderLLMInputs":
|
||||
msg = (
|
||||
"EncoderDecoderLLMInputs has been renamed to EncoderDecoderInputs. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
warnings.warn(DeprecationWarning(msg), stacklevel=2)
|
||||
|
||||
return EncoderDecoderInputs
|
||||
|
||||
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
|
||||
|
||||
+62
-16
@@ -1,5 +1,5 @@
|
||||
from typing import (TYPE_CHECKING, Any, Dict, Generic, Iterable, List,
|
||||
Optional, Tuple, Union)
|
||||
Optional, Tuple, Union, cast)
|
||||
|
||||
from typing_extensions import NotRequired, TypedDict, TypeVar
|
||||
|
||||
@@ -51,7 +51,7 @@ class TokensPrompt(TypedDict):
|
||||
|
||||
SingletonPrompt = Union[str, TextPrompt, TokensPrompt]
|
||||
"""
|
||||
Set of possible schemas for a single LLM input:
|
||||
Set of possible schemas for a single prompt:
|
||||
|
||||
- A text prompt (:class:`str` or :class:`TextPrompt`)
|
||||
- A tokenized prompt (:class:`TokensPrompt`)
|
||||
@@ -120,13 +120,8 @@ both decoder-only and encoder/decoder input types:
|
||||
"""
|
||||
|
||||
|
||||
class LLMInputs(TypedDict):
|
||||
"""
|
||||
The inputs in :class:`~vllm.LLMEngine` before they are
|
||||
passed to the model executor.
|
||||
|
||||
This specifies the data required for decoder-only models.
|
||||
"""
|
||||
class TokenInputs(TypedDict):
|
||||
"""Represents token-based inputs."""
|
||||
prompt_token_ids: List[int]
|
||||
"""The token IDs of the prompt."""
|
||||
|
||||
@@ -150,7 +145,40 @@ class LLMInputs(TypedDict):
|
||||
"""
|
||||
|
||||
|
||||
class EncoderDecoderLLMInputs(LLMInputs):
|
||||
def token_inputs(
|
||||
prompt_token_ids: List[int],
|
||||
prompt: Optional[str] = None,
|
||||
multi_modal_data: Optional["MultiModalDataDict"] = None,
|
||||
mm_processor_kwargs: Optional[Dict[str, Any]] = None,
|
||||
) -> TokenInputs:
|
||||
"""Construct :class:`TokenInputs` from optional values."""
|
||||
inputs = TokenInputs(prompt_token_ids=prompt_token_ids)
|
||||
|
||||
if prompt is not None:
|
||||
inputs["prompt"] = prompt
|
||||
if multi_modal_data is not None:
|
||||
inputs["multi_modal_data"] = multi_modal_data
|
||||
if mm_processor_kwargs is not None:
|
||||
inputs["mm_processor_kwargs"] = mm_processor_kwargs
|
||||
|
||||
return inputs
|
||||
|
||||
|
||||
SingletonInputs = TokenInputs
|
||||
"""
|
||||
A processed :class:`SingletonPrompt` which can be passed to
|
||||
:class:`vllm.sequence.Sequence`.
|
||||
"""
|
||||
|
||||
DecoderOnlyInputs = TokenInputs
|
||||
"""
|
||||
The inputs in :class:`~vllm.LLMEngine` before they are
|
||||
passed to the model executor.
|
||||
This specifies the data required for decoder-only models.
|
||||
"""
|
||||
|
||||
|
||||
class EncoderDecoderInputs(TokenInputs):
|
||||
"""
|
||||
The inputs in :class:`~vllm.LLMEngine` before they are
|
||||
passed to the model executor.
|
||||
@@ -204,11 +232,12 @@ def zip_enc_dec_prompts(
|
||||
be zipped with the encoder/decoder prompts.
|
||||
"""
|
||||
if mm_processor_kwargs is None:
|
||||
mm_processor_kwargs = {}
|
||||
if isinstance(mm_processor_kwargs, Dict):
|
||||
mm_processor_kwargs = cast(Dict[str, Any], {})
|
||||
if isinstance(mm_processor_kwargs, dict):
|
||||
return [
|
||||
build_explicit_enc_dec_prompt(encoder_prompt, decoder_prompt,
|
||||
mm_processor_kwargs)
|
||||
build_explicit_enc_dec_prompt(
|
||||
encoder_prompt, decoder_prompt,
|
||||
cast(Dict[str, Any], mm_processor_kwargs))
|
||||
for (encoder_prompt,
|
||||
decoder_prompt) in zip(enc_prompts, dec_prompts)
|
||||
]
|
||||
@@ -229,9 +258,9 @@ def to_enc_dec_tuple_list(
|
||||
|
||||
|
||||
def __getattr__(name: str):
|
||||
if name == "PromptInput":
|
||||
import warnings
|
||||
import warnings
|
||||
|
||||
if name == "PromptInput":
|
||||
msg = ("PromptInput has been renamed to PromptType. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
@@ -239,4 +268,21 @@ def __getattr__(name: str):
|
||||
|
||||
return PromptType
|
||||
|
||||
if name == "LLMInputs":
|
||||
msg = ("LLMInputs has been renamed to DecoderOnlyInputs. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
warnings.warn(DeprecationWarning(msg), stacklevel=2)
|
||||
|
||||
return DecoderOnlyInputs
|
||||
|
||||
if name == "EncoderDecoderLLMInputs":
|
||||
msg = (
|
||||
"EncoderDecoderLLMInputs has been renamed to EncoderDecoderInputs. "
|
||||
"The original name will be removed in an upcoming version.")
|
||||
|
||||
warnings.warn(DeprecationWarning(msg), stacklevel=2)
|
||||
|
||||
return EncoderDecoderInputs
|
||||
|
||||
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
|
||||
|
||||
@@ -4,9 +4,9 @@ from typing_extensions import TypeIs
|
||||
|
||||
from vllm.utils import is_list_of
|
||||
|
||||
from .data import (EncoderDecoderLLMInputs, ExplicitEncoderDecoderPrompt,
|
||||
LLMInputs, PromptType, SingletonPrompt, TextPrompt,
|
||||
TokensPrompt)
|
||||
from .data import (DecoderOnlyInputs, EncoderDecoderInputs,
|
||||
ExplicitEncoderDecoderPrompt, PromptType, SingletonPrompt,
|
||||
TextPrompt, TokensPrompt)
|
||||
|
||||
|
||||
class ParsedText(TypedDict):
|
||||
@@ -100,7 +100,7 @@ def is_explicit_encoder_decoder_prompt(
|
||||
return isinstance(prompt, dict) and "encoder_prompt" in prompt
|
||||
|
||||
|
||||
def is_valid_encoder_decoder_llm_inputs(
|
||||
inputs: Union[LLMInputs, EncoderDecoderLLMInputs],
|
||||
) -> TypeIs[EncoderDecoderLLMInputs]:
|
||||
def is_encoder_decoder_inputs(
|
||||
inputs: Union[DecoderOnlyInputs, EncoderDecoderInputs],
|
||||
) -> TypeIs[EncoderDecoderInputs]:
|
||||
return "encoder_prompt_token_ids" in inputs
|
||||
|
||||
+18
-18
@@ -10,7 +10,7 @@ from vllm.prompt_adapter.request import PromptAdapterRequest
|
||||
from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
|
||||
from vllm.utils import print_warning_once
|
||||
|
||||
from .data import (EncoderDecoderLLMInputs, LLMInputs, PromptType,
|
||||
from .data import (DecoderOnlyInputs, EncoderDecoderInputs, PromptType,
|
||||
SingletonPrompt)
|
||||
from .parse import is_explicit_encoder_decoder_prompt, parse_singleton_prompt
|
||||
|
||||
@@ -306,7 +306,7 @@ class InputPreprocessor:
|
||||
encoder_comps: PromptComponents,
|
||||
decoder_comps: DecoderPromptComponents,
|
||||
mm_processor_kwargs: Dict[str, Any],
|
||||
) -> EncoderDecoderLLMInputs:
|
||||
) -> EncoderDecoderInputs:
|
||||
encoder_prompt, encoder_prompt_ids, encoder_mm_data, _ = encoder_comps
|
||||
decoder_prompt, decoder_prompt_ids, decoder_mm_data, _ = decoder_comps
|
||||
|
||||
@@ -324,7 +324,7 @@ class InputPreprocessor:
|
||||
decoder_prompt_ids,
|
||||
force_bos=(encoder_mm_data is None and decoder_mm_data is None)))
|
||||
|
||||
return EncoderDecoderLLMInputs(
|
||||
return EncoderDecoderInputs(
|
||||
prompt_token_ids=decoder_prompt_ids,
|
||||
prompt=decoder_prompt,
|
||||
multi_modal_data=decoder_mm_data,
|
||||
@@ -338,11 +338,11 @@ class InputPreprocessor:
|
||||
self,
|
||||
prompt: PromptType,
|
||||
request_id: str,
|
||||
) -> EncoderDecoderLLMInputs:
|
||||
) -> EncoderDecoderInputs:
|
||||
'''
|
||||
For encoder/decoder models only:
|
||||
Process an input prompt into an
|
||||
:class:`EncoderDecoderLLMInputs` instance.
|
||||
:class:`EncoderDecoderInputs` instance.
|
||||
|
||||
There are two types of input prompts:
|
||||
singleton prompts which carry only the
|
||||
@@ -369,7 +369,7 @@ class InputPreprocessor:
|
||||
|
||||
Returns:
|
||||
|
||||
* :class:`EncoderDecoderLLMInputs` instance
|
||||
* :class:`EncoderDecoderInputs` instance
|
||||
'''
|
||||
|
||||
encoder_comps: PromptComponents
|
||||
@@ -411,7 +411,7 @@ class InputPreprocessor:
|
||||
self,
|
||||
prompt: PromptType,
|
||||
request_id: str,
|
||||
) -> EncoderDecoderLLMInputs:
|
||||
) -> EncoderDecoderInputs:
|
||||
"""Async version of :meth:`_process_encoder_decoder_prompt`."""
|
||||
encoder_comps: PromptComponents
|
||||
decoder_comps: DecoderPromptComponents
|
||||
@@ -455,17 +455,17 @@ class InputPreprocessor:
|
||||
self,
|
||||
prompt_comps: PromptComponents,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest],
|
||||
) -> LLMInputs:
|
||||
) -> DecoderOnlyInputs:
|
||||
(prompt, prompt_token_ids, multi_modal_data,
|
||||
mm_processor_kwargs) = prompt_comps
|
||||
|
||||
prompt_token_ids = self._apply_prompt_adapter(
|
||||
prompt_token_ids, prompt_adapter_request=prompt_adapter_request)
|
||||
|
||||
return LLMInputs(prompt_token_ids=prompt_token_ids,
|
||||
prompt=prompt,
|
||||
multi_modal_data=multi_modal_data,
|
||||
mm_processor_kwargs=mm_processor_kwargs)
|
||||
return DecoderOnlyInputs(prompt_token_ids=prompt_token_ids,
|
||||
prompt=prompt,
|
||||
multi_modal_data=multi_modal_data,
|
||||
mm_processor_kwargs=mm_processor_kwargs)
|
||||
|
||||
def _process_decoder_only_prompt(
|
||||
self,
|
||||
@@ -473,10 +473,10 @@ class InputPreprocessor:
|
||||
request_id: str,
|
||||
lora_request: Optional[LoRARequest] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
) -> LLMInputs:
|
||||
) -> DecoderOnlyInputs:
|
||||
'''
|
||||
For decoder-only models:
|
||||
Process an input prompt into an :class:`LLMInputs` instance.
|
||||
Process an input prompt into an :class:`DecoderOnlyInputs` instance.
|
||||
|
||||
Arguments:
|
||||
|
||||
@@ -487,7 +487,7 @@ class InputPreprocessor:
|
||||
|
||||
Returns:
|
||||
|
||||
* :class:`LLMInputs` instance
|
||||
* :class:`DecoderOnlyInputs` instance
|
||||
'''
|
||||
|
||||
prompt_comps = self._extract_prompt_components(
|
||||
@@ -507,7 +507,7 @@ class InputPreprocessor:
|
||||
request_id: str,
|
||||
lora_request: Optional[LoRARequest] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
) -> LLMInputs:
|
||||
) -> DecoderOnlyInputs:
|
||||
"""Async version of :meth:`_process_decoder_only_prompt`."""
|
||||
prompt_comps = await self._extract_prompt_components_async(
|
||||
prompt,
|
||||
@@ -526,7 +526,7 @@ class InputPreprocessor:
|
||||
request_id: str,
|
||||
lora_request: Optional[LoRARequest] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
) -> Union[LLMInputs, EncoderDecoderLLMInputs]:
|
||||
) -> Union[DecoderOnlyInputs, EncoderDecoderInputs]:
|
||||
"""Preprocess the input prompt."""
|
||||
if self.is_encoder_decoder_model():
|
||||
# Encoder-decoder model requires special mapping of
|
||||
@@ -554,7 +554,7 @@ class InputPreprocessor:
|
||||
request_id: str,
|
||||
lora_request: Optional[LoRARequest] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
|
||||
) -> Union[LLMInputs, EncoderDecoderLLMInputs]:
|
||||
) -> Union[DecoderOnlyInputs, EncoderDecoderInputs]:
|
||||
"""Async version of :meth:`preprocess`."""
|
||||
if self.is_encoder_decoder_model():
|
||||
# Encoder-decoder model requires special mapping of
|
||||
|
||||
@@ -12,7 +12,7 @@ from vllm.logger import init_logger
|
||||
from vllm.utils import (get_allowed_kwarg_only_overrides, print_warning_once,
|
||||
resolve_mm_processor_kwargs)
|
||||
|
||||
from .data import LLMInputs
|
||||
from .data import DecoderOnlyInputs
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.config import ModelConfig
|
||||
@@ -100,7 +100,7 @@ class _MultiModalCounts(UserDict):
|
||||
raise KeyError(msg) from exc
|
||||
|
||||
|
||||
InputProcessor = Callable[[InputContext, LLMInputs], LLMInputs]
|
||||
InputProcessor = Callable[[InputContext, DecoderOnlyInputs], DecoderOnlyInputs]
|
||||
"""Preprocess the inputs to the model."""
|
||||
|
||||
|
||||
@@ -134,7 +134,7 @@ class InputRegistry:
|
||||
# Avoid circular import
|
||||
from vllm.sequence import SequenceData
|
||||
|
||||
dummy_seq_data = SequenceData.from_token_counts((0, seq_len))
|
||||
dummy_seq_data = SequenceData.from_prompt_token_counts((0, seq_len))
|
||||
dummy_multi_modal_data = None
|
||||
|
||||
return dummy_seq_data, dummy_multi_modal_data
|
||||
@@ -245,8 +245,11 @@ class InputRegistry:
|
||||
|
||||
return seq_data, mm_data
|
||||
|
||||
def _default_input_processor(self, ctx: InputContext,
|
||||
inputs: LLMInputs) -> LLMInputs:
|
||||
def _default_input_processor(
|
||||
self,
|
||||
ctx: InputContext,
|
||||
inputs: DecoderOnlyInputs,
|
||||
) -> DecoderOnlyInputs:
|
||||
"""The default input processor is a no-op."""
|
||||
return inputs
|
||||
|
||||
@@ -279,7 +282,7 @@ class InputRegistry:
|
||||
.get(model_cls, self._default_input_processor)
|
||||
|
||||
def process_input(self, model_config: "ModelConfig",
|
||||
inputs: LLMInputs) -> LLMInputs:
|
||||
inputs: DecoderOnlyInputs) -> DecoderOnlyInputs:
|
||||
"""
|
||||
Apply an input processor to an instance of model inputs.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user