[Refactor] Consolidate sequence normalization and enc-dec parsing (#33928)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
@@ -16,11 +16,8 @@ from .data import (
|
||||
TextPrompt,
|
||||
TokenInputs,
|
||||
TokensPrompt,
|
||||
build_explicit_enc_dec_prompt,
|
||||
embeds_inputs,
|
||||
to_enc_dec_tuple_list,
|
||||
token_inputs,
|
||||
zip_enc_dec_prompts,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
@@ -39,8 +36,5 @@ __all__ = [
|
||||
"EncoderDecoderInputs",
|
||||
"ProcessorInputs",
|
||||
"SingletonInputs",
|
||||
"build_explicit_enc_dec_prompt",
|
||||
"to_enc_dec_tuple_list",
|
||||
"zip_enc_dec_prompts",
|
||||
"StreamingInput",
|
||||
]
|
||||
|
||||
+146
-167
@@ -1,11 +1,10 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from collections.abc import Iterable
|
||||
from dataclasses import dataclass
|
||||
from typing import TYPE_CHECKING, Any, Generic, Literal, TypeAlias, cast
|
||||
from typing import TYPE_CHECKING, Any, Literal, TypeAlias
|
||||
|
||||
import torch
|
||||
from typing_extensions import NotRequired, TypedDict, TypeVar
|
||||
from typing_extensions import NotRequired, TypedDict
|
||||
|
||||
from vllm.sampling_params import SamplingParams
|
||||
|
||||
@@ -23,7 +22,13 @@ else:
|
||||
MultiModalUUIDDict = object
|
||||
|
||||
|
||||
class _CommonKeys(TypedDict):
|
||||
# Inputs to LLM API
|
||||
class _PromptOptions(TypedDict):
|
||||
"""
|
||||
Additional options available to all
|
||||
[`SingletonPrompt`][vllm.inputs.data.SingletonPrompt].
|
||||
"""
|
||||
|
||||
multi_modal_data: NotRequired[MultiModalDataDict | None]
|
||||
"""
|
||||
Optional multi-modal data to pass to the model,
|
||||
@@ -53,14 +58,14 @@ class _CommonKeys(TypedDict):
|
||||
"""
|
||||
|
||||
|
||||
class TextPrompt(_CommonKeys):
|
||||
class TextPrompt(_PromptOptions):
|
||||
"""Schema for a text prompt."""
|
||||
|
||||
prompt: str
|
||||
"""The input text to be tokenized before passing to the model."""
|
||||
|
||||
|
||||
class TokensPrompt(_CommonKeys):
|
||||
class TokensPrompt(_PromptOptions):
|
||||
"""Schema for a tokenized prompt."""
|
||||
|
||||
prompt_token_ids: list[int]
|
||||
@@ -73,7 +78,7 @@ class TokensPrompt(_CommonKeys):
|
||||
"""A list of token type IDs to pass to the cross encoder model."""
|
||||
|
||||
|
||||
class EmbedsPrompt(_CommonKeys):
|
||||
class EmbedsPrompt(_PromptOptions):
|
||||
"""Schema for a prompt provided via token embeddings."""
|
||||
|
||||
prompt_embeds: torch.Tensor
|
||||
@@ -83,93 +88,113 @@ class EmbedsPrompt(_CommonKeys):
|
||||
"""The prompt text corresponding to the token embeddings, if available."""
|
||||
|
||||
|
||||
class DataPrompt(_CommonKeys):
|
||||
"""Represents generic inputs handled by IO processor plugins."""
|
||||
DecoderOnlyPrompt: TypeAlias = (
|
||||
str | TextPrompt | list[int] | TokensPrompt | EmbedsPrompt
|
||||
)
|
||||
"""
|
||||
Schema of a prompt for a decoder-only model:
|
||||
|
||||
- A text prompt (string or [`TextPrompt`][vllm.inputs.data.TextPrompt])
|
||||
- A tokenized prompt (list of token IDs, or
|
||||
[`TokensPrompt`][vllm.inputs.data.TokensPrompt])
|
||||
- An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.data.EmbedsPrompt])
|
||||
|
||||
For encoder-decoder models, passing a singleton prompt is shorthand for passing
|
||||
`ExplicitEncoderDecoderPrompt(encoder_prompt=prompt, decoder_prompt=None)`.
|
||||
"""
|
||||
|
||||
|
||||
EncoderPrompt: TypeAlias = str | TextPrompt | list[int] | TokensPrompt
|
||||
"""
|
||||
Schema of a prompt for the encoder part of a encoder-decoder model:
|
||||
|
||||
- A text prompt (string or [`TextPrompt`][vllm.inputs.data.TextPrompt])
|
||||
- A tokenized prompt (list of token IDs, or
|
||||
[`TokensPrompt`][vllm.inputs.data.TokensPrompt])
|
||||
"""
|
||||
|
||||
|
||||
DecoderPrompt: TypeAlias = str | TextPrompt | list[int] | TokensPrompt
|
||||
"""
|
||||
Schema of a prompt for the decoder part of an encoder-decoder model:
|
||||
|
||||
- A text prompt (string or [`TextPrompt`][vllm.inputs.data.TextPrompt])
|
||||
- A tokenized prompt (list of token IDs, or
|
||||
[`TokensPrompt`][vllm.inputs.data.TokensPrompt])
|
||||
|
||||
Note:
|
||||
Multi-modal inputs are not supported for decoder prompts.
|
||||
"""
|
||||
|
||||
|
||||
class ExplicitEncoderDecoderPrompt(TypedDict):
|
||||
"""
|
||||
Schema for a pair of encoder and decoder singleton prompts.
|
||||
|
||||
Note:
|
||||
This schema is not valid for decoder-only models.
|
||||
"""
|
||||
|
||||
encoder_prompt: EncoderPrompt
|
||||
"""The prompt for the encoder part of the model."""
|
||||
|
||||
decoder_prompt: DecoderPrompt | None
|
||||
"""
|
||||
The prompt for the decoder part of the model.
|
||||
|
||||
Passing `None` will cause the prompt to be inferred automatically.
|
||||
"""
|
||||
|
||||
|
||||
EncoderDecoderPrompt: TypeAlias = EncoderPrompt | ExplicitEncoderDecoderPrompt
|
||||
"""
|
||||
Schema for a prompt for an encoder-decoder model.
|
||||
|
||||
You can pass a singleton encoder prompt, in which case the decoder prompt is
|
||||
considered to be `None` (i.e., infer automatically).
|
||||
"""
|
||||
|
||||
|
||||
SingletonPrompt: TypeAlias = DecoderOnlyPrompt | EncoderPrompt | DecoderPrompt
|
||||
"""
|
||||
Schema for a single prompt. This is as opposed to a data structure
|
||||
which encapsulates multiple prompts, such as
|
||||
[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt].
|
||||
"""
|
||||
|
||||
|
||||
PromptType: TypeAlias = DecoderOnlyPrompt | EncoderDecoderPrompt
|
||||
"""
|
||||
Schema for any prompt, regardless of model type.
|
||||
|
||||
This is the input format accepted by most [`LLM`][vllm.entrypoints.llm.LLM] APIs.
|
||||
"""
|
||||
|
||||
|
||||
class DataPrompt(_PromptOptions):
|
||||
"""
|
||||
Represents generic inputs that are converted to
|
||||
[`PromptType`][vllm.inputs.data.PromptType] by IO processor plugins.
|
||||
"""
|
||||
|
||||
data: Any
|
||||
"""The input data"""
|
||||
"""The input data."""
|
||||
|
||||
data_format: str
|
||||
"""The input data format"""
|
||||
"""The input data format."""
|
||||
|
||||
|
||||
SingletonPrompt: TypeAlias = str | TextPrompt | TokensPrompt | EmbedsPrompt
|
||||
"""
|
||||
Set of possible schemas for a single prompt:
|
||||
|
||||
- A text prompt ([`str`][] or [`TextPrompt`][vllm.inputs.data.TextPrompt])
|
||||
- A tokenized prompt ([`TokensPrompt`][vllm.inputs.data.TokensPrompt])
|
||||
- An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.data.EmbedsPrompt])
|
||||
|
||||
Note that "singleton" is as opposed to a data structure
|
||||
which encapsulates multiple prompts, i.e. of the sort
|
||||
which may be utilized for encoder/decoder models when
|
||||
the user desires to express both the encoder & decoder
|
||||
prompts explicitly, i.e.
|
||||
[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
|
||||
|
||||
A prompt of type [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] may be
|
||||
employed as (1) input to a decoder-only model, (2) input to
|
||||
the encoder of an encoder/decoder model, in the scenario
|
||||
where the decoder-prompt is not specified explicitly, or
|
||||
(3) as a member of a larger data structure encapsulating
|
||||
more than one prompt, i.e.
|
||||
[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
|
||||
"""
|
||||
|
||||
|
||||
_T1_co = TypeVar(
|
||||
"_T1_co", bound=SingletonPrompt, default=SingletonPrompt, covariant=True
|
||||
)
|
||||
_T2_co = TypeVar(
|
||||
"_T2_co", bound=SingletonPrompt, default=SingletonPrompt, covariant=True
|
||||
)
|
||||
|
||||
|
||||
# TODO: Make fields ReadOnly once mypy supports it
|
||||
class ExplicitEncoderDecoderPrompt(TypedDict, Generic[_T1_co, _T2_co]):
|
||||
# Outputs of processor
|
||||
class _InputOptions(TypedDict):
|
||||
"""
|
||||
Represents an encoder/decoder model input prompt,
|
||||
comprising an explicit encoder prompt and a decoder prompt.
|
||||
|
||||
The encoder and decoder prompts, respectively, may be formatted
|
||||
according to any of the
|
||||
[`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] schemas,
|
||||
and are not required to have the same schema.
|
||||
|
||||
Only the encoder prompt may have multi-modal data. mm_processor_kwargs
|
||||
should be at the top-level, and should not be set in the encoder/decoder
|
||||
prompts, since they are agnostic to the encoder/decoder.
|
||||
|
||||
Note that an
|
||||
[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
|
||||
may not be used as an input to a decoder-only model,
|
||||
and that the `encoder_prompt` and `decoder_prompt`
|
||||
fields of this data structure themselves must be
|
||||
[`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] instances.
|
||||
Additional options available to all input types.
|
||||
"""
|
||||
|
||||
encoder_prompt: _T1_co
|
||||
|
||||
decoder_prompt: _T2_co | None
|
||||
|
||||
mm_processor_kwargs: NotRequired[dict[str, Any]]
|
||||
cache_salt: NotRequired[str]
|
||||
"""Optional cache salt to be used for prefix caching."""
|
||||
|
||||
|
||||
PromptType: TypeAlias = SingletonPrompt | ExplicitEncoderDecoderPrompt[Any, Any]
|
||||
"""
|
||||
Set of possible schemas for an LLM input, including
|
||||
both decoder-only and encoder/decoder input types:
|
||||
|
||||
- A text prompt ([`str`][] or [`TextPrompt`][vllm.inputs.data.TextPrompt])
|
||||
- A tokenized prompt ([`TokensPrompt`][vllm.inputs.data.TokensPrompt])
|
||||
- An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.data.EmbedsPrompt])
|
||||
- A single data structure containing both an encoder and a decoder prompt
|
||||
([`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt])
|
||||
"""
|
||||
|
||||
|
||||
class TokenInputs(TypedDict):
|
||||
class TokenInputs(_InputOptions):
|
||||
"""Represents token-based inputs."""
|
||||
|
||||
type: Literal["token"]
|
||||
@@ -178,11 +203,6 @@ class TokenInputs(TypedDict):
|
||||
prompt_token_ids: list[int]
|
||||
"""The token IDs of the prompt."""
|
||||
|
||||
cache_salt: NotRequired[str]
|
||||
"""
|
||||
Optional cache salt to be used for prefix caching.
|
||||
"""
|
||||
|
||||
|
||||
def token_inputs(
|
||||
prompt_token_ids: list[int],
|
||||
@@ -198,7 +218,7 @@ def token_inputs(
|
||||
return inputs
|
||||
|
||||
|
||||
class EmbedsInputs(TypedDict):
|
||||
class EmbedsInputs(_InputOptions):
|
||||
"""Represents embeddings-based inputs."""
|
||||
|
||||
type: Literal["embeds"]
|
||||
@@ -207,11 +227,6 @@ class EmbedsInputs(TypedDict):
|
||||
prompt_embeds: torch.Tensor
|
||||
"""The embeddings of the prompt."""
|
||||
|
||||
cache_salt: NotRequired[str]
|
||||
"""
|
||||
Optional cache salt to be used for prefix caching.
|
||||
"""
|
||||
|
||||
|
||||
def embeds_inputs(
|
||||
prompt_embeds: torch.Tensor,
|
||||
@@ -229,96 +244,60 @@ def embeds_inputs(
|
||||
|
||||
DecoderOnlyInputs: TypeAlias = TokenInputs | EmbedsInputs | MultiModalInputs
|
||||
"""
|
||||
The inputs in [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] before they are
|
||||
passed to the model executor.
|
||||
This specifies the data required for decoder-only models.
|
||||
A processed prompt from
|
||||
[`InputPreprocessor`][vllm.inputs.preprocess.InputPreprocessor]
|
||||
which can be passed to
|
||||
[`InputProcessor`][vllm.v1.engine.input_processor.InputProcessor]
|
||||
for decoder-only models.
|
||||
"""
|
||||
|
||||
|
||||
EncoderInputs: TypeAlias = TokenInputs | MultiModalEncDecInputs
|
||||
"""
|
||||
A processed encoder prompt from
|
||||
[`InputPreprocessor`][vllm.inputs.preprocess.InputPreprocessor]
|
||||
which can be passed to
|
||||
[`InputProcessor`][vllm.v1.engine.input_processor.InputProcessor]
|
||||
for encoder-decoder models.
|
||||
"""
|
||||
|
||||
|
||||
DecoderInputs: TypeAlias = TokenInputs | MultiModalInputs
|
||||
"""
|
||||
A processed decoder prompt from
|
||||
[`InputPreprocessor`][vllm.inputs.preprocess.InputPreprocessor]
|
||||
which can be passed to
|
||||
[`InputProcessor`][vllm.v1.engine.input_processor.InputProcessor]
|
||||
for encoder-decoder models.
|
||||
"""
|
||||
|
||||
|
||||
class EncoderDecoderInputs(TypedDict):
|
||||
"""
|
||||
The inputs in [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] before they
|
||||
are passed to the model executor.
|
||||
|
||||
This specifies the required data for encoder-decoder models.
|
||||
A processed pair of encoder and decoder singleton prompts.
|
||||
[`InputPreprocessor`][vllm.inputs.preprocess.InputPreprocessor]
|
||||
which can be passed to
|
||||
[`InputProcessor`][vllm.v1.engine.input_processor.InputProcessor]
|
||||
for encoder-decoder models.
|
||||
"""
|
||||
|
||||
encoder: TokenInputs | MultiModalEncDecInputs
|
||||
encoder: EncoderInputs
|
||||
"""The inputs for the encoder portion."""
|
||||
|
||||
decoder: TokenInputs | MultiModalInputs
|
||||
decoder: DecoderInputs
|
||||
"""The inputs for the decoder portion."""
|
||||
|
||||
|
||||
SingletonInputs: TypeAlias = TokenInputs | EmbedsInputs | MultiModalInputs
|
||||
"""
|
||||
A processed [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] which can be
|
||||
passed to [`Sequence`][collections.abc.Sequence].
|
||||
"""
|
||||
|
||||
ProcessorInputs: TypeAlias = DecoderOnlyInputs | EncoderDecoderInputs
|
||||
"""
|
||||
The outputs from [`vllm.inputs.preprocess.InputPreprocessor`][].
|
||||
A processed prompt from
|
||||
[`InputPreprocessor`][vllm.inputs.preprocess.InputPreprocessor]
|
||||
which can be passed to
|
||||
[`InputProcessor`][vllm.v1.engine.input_processor.InputProcessor].
|
||||
"""
|
||||
|
||||
_T1 = TypeVar("_T1", bound=SingletonPrompt, default=SingletonPrompt)
|
||||
_T2 = TypeVar("_T2", bound=SingletonPrompt, default=SingletonPrompt)
|
||||
|
||||
|
||||
def build_explicit_enc_dec_prompt(
|
||||
encoder_prompt: _T1,
|
||||
decoder_prompt: _T2 | None,
|
||||
mm_processor_kwargs: dict[str, Any] | None = None,
|
||||
) -> ExplicitEncoderDecoderPrompt[_T1, _T2]:
|
||||
if mm_processor_kwargs is None:
|
||||
mm_processor_kwargs = {}
|
||||
return ExplicitEncoderDecoderPrompt(
|
||||
encoder_prompt=encoder_prompt,
|
||||
decoder_prompt=decoder_prompt,
|
||||
mm_processor_kwargs=mm_processor_kwargs,
|
||||
)
|
||||
|
||||
|
||||
def zip_enc_dec_prompts(
|
||||
enc_prompts: Iterable[_T1],
|
||||
dec_prompts: Iterable[_T2 | None],
|
||||
mm_processor_kwargs: Iterable[dict[str, Any]] | dict[str, Any] | None = None,
|
||||
) -> list[ExplicitEncoderDecoderPrompt[_T1, _T2]]:
|
||||
"""
|
||||
Zip encoder and decoder prompts together into a list of
|
||||
[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
|
||||
instances.
|
||||
|
||||
`mm_processor_kwargs` may also be provided; if a dict is passed, the same
|
||||
dictionary will be used for every encoder/decoder prompt. If an iterable is
|
||||
provided, it will be zipped with the encoder/decoder prompts.
|
||||
"""
|
||||
if mm_processor_kwargs is None:
|
||||
mm_processor_kwargs = cast(dict[str, Any], {})
|
||||
if isinstance(mm_processor_kwargs, dict):
|
||||
return [
|
||||
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)
|
||||
]
|
||||
return [
|
||||
build_explicit_enc_dec_prompt(encoder_prompt, decoder_prompt, mm_proc_kwargs)
|
||||
for (encoder_prompt, decoder_prompt, mm_proc_kwargs) in zip(
|
||||
enc_prompts, dec_prompts, mm_processor_kwargs
|
||||
)
|
||||
]
|
||||
|
||||
|
||||
def to_enc_dec_tuple_list(
|
||||
enc_dec_prompts: Iterable[ExplicitEncoderDecoderPrompt[_T1, _T2]],
|
||||
) -> list[tuple[_T1, _T2 | None]]:
|
||||
return [
|
||||
(enc_dec_prompt["encoder_prompt"], enc_dec_prompt["decoder_prompt"])
|
||||
for enc_dec_prompt in enc_dec_prompts
|
||||
]
|
||||
SingletonInputs: TypeAlias = DecoderOnlyInputs | MultiModalEncDecInputs
|
||||
|
||||
|
||||
@dataclass
|
||||
|
||||
+1
-109
@@ -1,88 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from typing import TYPE_CHECKING, Literal, NamedTuple, TypeAlias, TypedDict
|
||||
|
||||
from typing_extensions import TypeIs
|
||||
|
||||
from vllm.utils import length_from_prompt_token_ids_or_embeds
|
||||
|
||||
from .data import (
|
||||
EmbedsPrompt,
|
||||
ExplicitEncoderDecoderPrompt,
|
||||
ProcessorInputs,
|
||||
PromptType,
|
||||
SingletonInputs,
|
||||
SingletonPrompt,
|
||||
TextPrompt,
|
||||
TokensPrompt,
|
||||
)
|
||||
|
||||
if TYPE_CHECKING:
|
||||
import torch
|
||||
|
||||
|
||||
class ParsedStrPrompt(TypedDict):
|
||||
type: Literal["str"]
|
||||
content: str
|
||||
|
||||
|
||||
class ParsedTextPrompt(TypedDict):
|
||||
type: Literal["text"]
|
||||
content: TextPrompt
|
||||
|
||||
|
||||
class ParsedTokensPrompt(TypedDict):
|
||||
type: Literal["tokens"]
|
||||
content: TokensPrompt
|
||||
|
||||
|
||||
class ParsedEmbedsPrompt(TypedDict):
|
||||
type: Literal["embeds"]
|
||||
content: EmbedsPrompt
|
||||
|
||||
|
||||
ParsedSingletonPrompt: TypeAlias = (
|
||||
ParsedStrPrompt | ParsedTextPrompt | ParsedTokensPrompt | ParsedEmbedsPrompt
|
||||
)
|
||||
|
||||
|
||||
def parse_singleton_prompt(prompt: SingletonPrompt) -> ParsedSingletonPrompt:
|
||||
if isinstance(prompt, str):
|
||||
return ParsedStrPrompt(type="str", content=prompt)
|
||||
elif isinstance(prompt, dict):
|
||||
# Type ignores are because mypy does not correctly infer the TypedDicts
|
||||
# Pyright does succeed.
|
||||
if "prompt_embeds" in prompt:
|
||||
return ParsedEmbedsPrompt(type="embeds", content=prompt) # type: ignore[typeddict-item]
|
||||
elif "prompt_token_ids" in prompt:
|
||||
return ParsedTokensPrompt(type="tokens", content=prompt) # type: ignore[typeddict-item]
|
||||
elif "prompt" in prompt:
|
||||
return ParsedTextPrompt(type="text", content=prompt)
|
||||
raise TypeError(
|
||||
"inputs must be a string, TextPrompt, TokensPrompt, or EmbedsPrompt"
|
||||
)
|
||||
|
||||
|
||||
def is_explicit_encoder_decoder_prompt(
|
||||
prompt: PromptType,
|
||||
) -> TypeIs[ExplicitEncoderDecoderPrompt]:
|
||||
return isinstance(prompt, dict) and "encoder_prompt" in prompt
|
||||
|
||||
|
||||
def split_enc_dec_prompt(
|
||||
prompt: PromptType,
|
||||
) -> tuple[SingletonPrompt, SingletonPrompt | None]:
|
||||
if isinstance(prompt, str):
|
||||
return prompt, None
|
||||
|
||||
if "encoder_prompt" in prompt and "decoder_prompt" in prompt:
|
||||
# NOTE: This passes pyright but not mypy
|
||||
return (
|
||||
prompt["encoder_prompt"], # type: ignore[typeddict-item]
|
||||
prompt["decoder_prompt"], # type: ignore[typeddict-item]
|
||||
)
|
||||
|
||||
return prompt, None
|
||||
from .data import ProcessorInputs, SingletonInputs
|
||||
|
||||
|
||||
def split_enc_dec_inputs(
|
||||
@@ -96,30 +15,3 @@ def split_enc_dec_inputs(
|
||||
)
|
||||
|
||||
return None, inputs
|
||||
|
||||
|
||||
class PromptComponents(NamedTuple):
|
||||
text: str | None = None
|
||||
token_ids: list[int] | None = None
|
||||
embeds: "torch.Tensor | None" = None
|
||||
|
||||
|
||||
def get_prompt_components(prompt: PromptType) -> PromptComponents:
|
||||
if isinstance(prompt, str):
|
||||
return PromptComponents(text=prompt)
|
||||
|
||||
if encoder_prompt := prompt.get("encoder_prompt"):
|
||||
return get_prompt_components(encoder_prompt) # type: ignore[arg-type]
|
||||
|
||||
return PromptComponents(
|
||||
text=prompt.get("prompt"), # type: ignore[arg-type]
|
||||
token_ids=prompt.get("prompt_token_ids"), # type: ignore[arg-type]
|
||||
embeds=prompt.get("prompt_embeds"),
|
||||
)
|
||||
|
||||
|
||||
def get_prompt_len(prompt: TokensPrompt | EmbedsPrompt):
|
||||
return length_from_prompt_token_ids_or_embeds(
|
||||
prompt.get("prompt_token_ids"), # type: ignore[arg-type]
|
||||
prompt.get("prompt_embeds"), # type: ignore[arg-type]
|
||||
)
|
||||
|
||||
+66
-64
@@ -2,43 +2,51 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from collections.abc import Mapping
|
||||
from typing import Any
|
||||
from typing import Any, overload
|
||||
|
||||
from typing_extensions import assert_never
|
||||
|
||||
from vllm.config import ModelConfig, ObservabilityConfig
|
||||
from vllm.inputs.parse import split_enc_dec_prompt
|
||||
from vllm.logger import init_logger
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
|
||||
from vllm.multimodal.cache import BaseMultiModalProcessorCache
|
||||
from vllm.multimodal.inputs import (
|
||||
MultiModalDataDict,
|
||||
MultiModalEncDecInputs,
|
||||
MultiModalInputs,
|
||||
MultiModalUUIDDict,
|
||||
)
|
||||
from vllm.multimodal.processing import BaseMultiModalProcessor
|
||||
from vllm.renderers import renderer_from_config
|
||||
from vllm.renderers.inputs import (
|
||||
DecoderDictPrompt,
|
||||
DecoderOnlyDictPrompt,
|
||||
DictPrompt,
|
||||
EncoderDecoderDictPrompt,
|
||||
EncoderDictPrompt,
|
||||
SingletonDictPrompt,
|
||||
TokPrompt,
|
||||
)
|
||||
from vllm.renderers.inputs.preprocess import parse_dec_only_prompt, parse_enc_dec_prompt
|
||||
from vllm.tokenizers import TokenizerLike
|
||||
from vllm.utils.jsontree import json_iter_leaves
|
||||
from vllm.v1.metrics.stats import MultiModalCacheStats
|
||||
|
||||
from .data import (
|
||||
DecoderInputs,
|
||||
DecoderOnlyInputs,
|
||||
EmbedsInputs,
|
||||
EmbedsPrompt,
|
||||
EncoderDecoderInputs,
|
||||
EncoderInputs,
|
||||
ProcessorInputs,
|
||||
PromptType,
|
||||
SingletonInputs,
|
||||
SingletonPrompt,
|
||||
TextPrompt,
|
||||
TokenInputs,
|
||||
TokensPrompt,
|
||||
embeds_inputs,
|
||||
token_inputs,
|
||||
)
|
||||
from .parse import is_explicit_encoder_decoder_prompt, parse_singleton_prompt
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -328,9 +336,36 @@ class InputPreprocessor:
|
||||
|
||||
return inputs
|
||||
|
||||
@overload
|
||||
def _prompt_to_llm_inputs(
|
||||
self,
|
||||
prompt: SingletonPrompt,
|
||||
prompt: EncoderDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
) -> EncoderInputs: ...
|
||||
|
||||
@overload
|
||||
def _prompt_to_llm_inputs( # type: ignore[misc]
|
||||
self,
|
||||
prompt: DecoderDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
) -> DecoderInputs: ...
|
||||
|
||||
@overload
|
||||
def _prompt_to_llm_inputs( # type: ignore[misc]
|
||||
self,
|
||||
prompt: DecoderOnlyDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
) -> DecoderOnlyInputs: ...
|
||||
|
||||
def _prompt_to_llm_inputs(
|
||||
self,
|
||||
prompt: SingletonDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
@@ -346,34 +381,25 @@ class InputPreprocessor:
|
||||
|
||||
* [`SingletonInputs`][vllm.inputs.data.SingletonInputs] instance
|
||||
"""
|
||||
parsed = parse_singleton_prompt(prompt)
|
||||
if "prompt_embeds" in prompt:
|
||||
return self._process_embeds(prompt) # type: ignore[arg-type]
|
||||
|
||||
if parsed["type"] == "embeds":
|
||||
return self._process_embeds(parsed["content"])
|
||||
if parsed["type"] == "tokens":
|
||||
if "prompt_token_ids" in prompt:
|
||||
return self._process_tokens(
|
||||
parsed["content"],
|
||||
prompt, # type: ignore[arg-type]
|
||||
mm_uuids=mm_uuids,
|
||||
)
|
||||
if parsed["type"] == "text":
|
||||
|
||||
if "prompt" in prompt:
|
||||
return self._process_text(
|
||||
parsed["content"],
|
||||
tokenization_kwargs=tokenization_kwargs,
|
||||
mm_uuids=mm_uuids,
|
||||
)
|
||||
if parsed["type"] == "str":
|
||||
return self._process_text(
|
||||
TextPrompt(prompt=parsed["content"]),
|
||||
prompt, # type: ignore[arg-type]
|
||||
tokenization_kwargs=tokenization_kwargs,
|
||||
mm_uuids=mm_uuids,
|
||||
)
|
||||
|
||||
assert_never(parsed)
|
||||
assert_never(prompt) # type: ignore[arg-type]
|
||||
|
||||
def _validate_enc_inputs(
|
||||
self,
|
||||
inputs: SingletonInputs,
|
||||
) -> TokenInputs | MultiModalEncDecInputs:
|
||||
def _validate_enc_inputs(self, inputs: SingletonInputs) -> EncoderInputs:
|
||||
if inputs["type"] == "embeds":
|
||||
raise ValueError(
|
||||
"Embedding inputs are not supported for encoder-decoder models"
|
||||
@@ -387,10 +413,7 @@ class InputPreprocessor:
|
||||
|
||||
return inputs # type: ignore[return-value]
|
||||
|
||||
def _validate_dec_inputs(
|
||||
self,
|
||||
inputs: SingletonInputs,
|
||||
) -> TokenInputs | MultiModalInputs:
|
||||
def _validate_dec_inputs(self, inputs: SingletonInputs) -> DecoderInputs:
|
||||
if inputs["type"] == "embeds":
|
||||
raise ValueError(
|
||||
"Embedding inputs are not supported for encoder-decoder models"
|
||||
@@ -403,14 +426,15 @@ class InputPreprocessor:
|
||||
encoder_inputs: SingletonInputs,
|
||||
decoder_inputs: SingletonInputs | None = None,
|
||||
) -> EncoderDecoderInputs:
|
||||
if decoder_inputs is None:
|
||||
decoder_inputs = encoder_inputs
|
||||
|
||||
enc_inputs = self._validate_enc_inputs(encoder_inputs)
|
||||
dec_inputs = self._validate_dec_inputs(decoder_inputs)
|
||||
|
||||
enc_inputs_new: TokenInputs | MultiModalEncDecInputs
|
||||
dec_inputs_new: TokenInputs | MultiModalInputs
|
||||
if decoder_inputs is None:
|
||||
dec_inputs: DecoderInputs = enc_inputs # type: ignore[assignment]
|
||||
else:
|
||||
dec_inputs = self._validate_dec_inputs(decoder_inputs)
|
||||
|
||||
enc_inputs_new: EncoderInputs
|
||||
dec_inputs_new: DecoderInputs
|
||||
|
||||
if enc_inputs["type"] == "multimodal":
|
||||
enc_inputs_new = token_inputs(enc_inputs["encoder_prompt_token_ids"])
|
||||
@@ -437,7 +461,7 @@ class InputPreprocessor:
|
||||
|
||||
def _process_encoder_decoder_prompt(
|
||||
self,
|
||||
prompt: PromptType,
|
||||
prompt: EncoderDecoderDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
@@ -448,24 +472,6 @@ class InputPreprocessor:
|
||||
[`EncoderDecoderInputs`][vllm.inputs.data.EncoderDecoderInputs]
|
||||
instance.
|
||||
|
||||
There are two types of input prompts:
|
||||
singleton prompts which carry only the
|
||||
encoder prompt, and explicit encoder/decoder
|
||||
prompts which carry both the encoder and the
|
||||
decoder prompts as member variables.
|
||||
|
||||
This function handles the following scenarios:
|
||||
* Singleton encoder prompt: extract encoder prompt
|
||||
token ids & infer default decoder prompt token ids
|
||||
* Explicit encoder/decoder prompt: extract encoder
|
||||
and decoder prompt token ids
|
||||
|
||||
Note that for Explicit encoder/decoder prompts,
|
||||
each sub-prompt (encoder or decoder prompt) can
|
||||
have any possible singleton type; thus this
|
||||
method relies on helper functions to obtain
|
||||
token ids for the sub-prompts.
|
||||
|
||||
Arguments:
|
||||
|
||||
* prompt: an input prompt
|
||||
@@ -475,7 +481,8 @@ class InputPreprocessor:
|
||||
* [`EncoderDecoderInputs`][vllm.inputs.data.EncoderDecoderInputs]
|
||||
instance
|
||||
"""
|
||||
encoder_prompt, decoder_prompt = split_enc_dec_prompt(prompt)
|
||||
encoder_prompt = prompt["encoder_prompt"]
|
||||
decoder_prompt = prompt["decoder_prompt"]
|
||||
|
||||
return self._build_enc_dec_inputs(
|
||||
encoder_inputs=self._prompt_to_llm_inputs(
|
||||
@@ -495,7 +502,7 @@ class InputPreprocessor:
|
||||
|
||||
def _process_decoder_only_prompt(
|
||||
self,
|
||||
prompt: SingletonPrompt,
|
||||
prompt: DecoderOnlyDictPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
@@ -521,7 +528,7 @@ class InputPreprocessor:
|
||||
|
||||
def _preprocess(
|
||||
self,
|
||||
prompt: PromptType,
|
||||
prompt: PromptType | DictPrompt | TokPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
@@ -530,25 +537,20 @@ class InputPreprocessor:
|
||||
# Encoder-decoder model requires special mapping of
|
||||
# input prompts to encoder & decoder.
|
||||
return self._process_encoder_decoder_prompt(
|
||||
prompt,
|
||||
parse_enc_dec_prompt(prompt),
|
||||
tokenization_kwargs,
|
||||
mm_uuids=mm_uuids,
|
||||
)
|
||||
|
||||
if is_explicit_encoder_decoder_prompt(prompt):
|
||||
raise ValueError(
|
||||
"Cannot pass encoder-decoder prompt to decoder-only models"
|
||||
)
|
||||
|
||||
return self._process_decoder_only_prompt(
|
||||
prompt,
|
||||
parse_dec_only_prompt(prompt),
|
||||
tokenization_kwargs=tokenization_kwargs,
|
||||
mm_uuids=mm_uuids,
|
||||
)
|
||||
|
||||
def preprocess(
|
||||
self,
|
||||
prompt: PromptType,
|
||||
prompt: PromptType | DictPrompt | TokPrompt,
|
||||
tokenization_kwargs: dict[str, Any] | None = None,
|
||||
*,
|
||||
mm_uuids: MultiModalUUIDDict | None = None,
|
||||
|
||||
Reference in New Issue
Block a user