[Frontend][3/n] Improve pooling entrypoints | scoring. (#28631)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
@@ -0,0 +1,112 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from http import HTTPStatus
|
||||
|
||||
from fastapi import APIRouter, Depends, Request
|
||||
|
||||
from vllm.entrypoints.openai.engine.protocol import ErrorResponse
|
||||
from vllm.entrypoints.openai.utils import validate_json_request
|
||||
from vllm.entrypoints.utils import load_aware_call, with_cancellation
|
||||
from vllm.logger import init_logger
|
||||
|
||||
from .protocol import RerankRequest, ScoreRequest
|
||||
from .serving import ServingScores
|
||||
|
||||
router = APIRouter()
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
def score(request: Request) -> ServingScores | None:
|
||||
return request.app.state.serving_scores
|
||||
|
||||
|
||||
def rerank(request: Request) -> ServingScores | None:
|
||||
return request.app.state.serving_scores
|
||||
|
||||
|
||||
@router.post(
|
||||
"/score",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
@with_cancellation
|
||||
@load_aware_call
|
||||
async def create_score(request: ScoreRequest, raw_request: Request):
|
||||
handler = score(raw_request)
|
||||
if handler is None:
|
||||
raise NotImplementedError("The model does not support Score API")
|
||||
|
||||
return await handler(request, raw_request)
|
||||
|
||||
|
||||
@router.post(
|
||||
"/v1/score",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
@with_cancellation
|
||||
@load_aware_call
|
||||
async def create_score_v1(request: ScoreRequest, raw_request: Request):
|
||||
logger.warning(
|
||||
"To indicate that Score API is not part of standard OpenAI API, we "
|
||||
"have moved it to `/score`. Please update your client accordingly."
|
||||
)
|
||||
|
||||
return await create_score(request, raw_request)
|
||||
|
||||
|
||||
@router.post(
|
||||
"/rerank",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
@with_cancellation
|
||||
@load_aware_call
|
||||
async def do_rerank(request: RerankRequest, raw_request: Request):
|
||||
handler = rerank(raw_request)
|
||||
if handler is None:
|
||||
raise NotImplementedError("The model does not support Rerank (Score) API")
|
||||
|
||||
return await handler(request, raw_request)
|
||||
|
||||
|
||||
@router.post(
|
||||
"/v1/rerank",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
@with_cancellation
|
||||
async def do_rerank_v1(request: RerankRequest, raw_request: Request):
|
||||
logger.warning_once(
|
||||
"To indicate that the rerank API is not part of the standard OpenAI"
|
||||
" API, we have located it at `/rerank`. Please update your client "
|
||||
"accordingly. (Note: Conforms to JinaAI rerank API)"
|
||||
)
|
||||
|
||||
return await do_rerank(request, raw_request)
|
||||
|
||||
|
||||
@router.post(
|
||||
"/v2/rerank",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
@with_cancellation
|
||||
async def do_rerank_v2(request: RerankRequest, raw_request: Request):
|
||||
return await do_rerank(request, raw_request)
|
||||
@@ -0,0 +1,419 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import time
|
||||
from collections.abc import Sequence
|
||||
from typing import Any, TypeAlias, cast
|
||||
|
||||
import torch.nn.functional as F
|
||||
|
||||
from vllm import PoolingParams, PoolingRequestOutput, TokensPrompt
|
||||
from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor
|
||||
from vllm.entrypoints.pooling.typing import (
|
||||
OfflineInputsContext,
|
||||
OfflineOutputsContext,
|
||||
PoolingServeContext,
|
||||
)
|
||||
from vllm.inputs import EngineInput
|
||||
from vllm.renderers import TokenizeParams
|
||||
from vllm.renderers.hf import safe_apply_chat_template
|
||||
from vllm.tasks import PoolingTask, ScoreType
|
||||
from vllm.utils.mistral import is_mistral_tokenizer
|
||||
|
||||
from ...chat_utils import ChatTemplateResolutionError
|
||||
from .protocol import RerankRequest, ScoreRequest, ScoringRequest
|
||||
from .typing import ScoreData, ScoreInput, ScoringData
|
||||
from .utils import (
|
||||
compress_token_type_ids,
|
||||
compute_maxsim_score,
|
||||
parse_score_data,
|
||||
score_data_to_prompts,
|
||||
validate_score_input,
|
||||
)
|
||||
|
||||
ScoringServeContext: TypeAlias = PoolingServeContext[ScoringRequest]
|
||||
|
||||
|
||||
class ScoringIOProcessor(PoolingIOProcessor):
|
||||
name: ScoreType
|
||||
pooling_task: PoolingTask
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
self.tokenizer = self.renderer.get_tokenizer()
|
||||
self.architecture = self.model_config.architecture
|
||||
self.is_multimodal_model = self.model_config.is_multimodal_model
|
||||
self.pad_token_id = self.tokenizer.pad_token_id
|
||||
|
||||
def create_pooling_params(self, request):
|
||||
return request.to_pooling_params(self.pooling_task)
|
||||
|
||||
def valid_inputs(
|
||||
self,
|
||||
data_1: ScoreInput | list[ScoreInput],
|
||||
data_2: ScoreInput | list[ScoreInput],
|
||||
) -> ScoringData:
|
||||
scoring_data = validate_score_input(
|
||||
data_1,
|
||||
data_2,
|
||||
is_multimodal_model=self.is_multimodal_model,
|
||||
architecture=self.architecture,
|
||||
)
|
||||
return scoring_data
|
||||
|
||||
|
||||
class BiEncoderIOProcessor(ScoringIOProcessor):
|
||||
name: ScoreType = "bi-encoder"
|
||||
pooling_task: PoolingTask = "embed"
|
||||
|
||||
#######################################
|
||||
# online APIs
|
||||
|
||||
def pre_process_online(self, ctx: ScoringServeContext):
|
||||
request = ctx.request
|
||||
|
||||
if isinstance(request, ScoreRequest):
|
||||
data_1 = request.data_1
|
||||
data_2 = request.data_2
|
||||
elif isinstance(request, RerankRequest):
|
||||
data_1 = request.query
|
||||
data_2 = request.documents
|
||||
else:
|
||||
raise ValueError(f"Invalid {self.name} request type")
|
||||
|
||||
scoring_data = self.valid_inputs(data_1, data_2)
|
||||
tok_params = request.build_tok_params(self.model_config)
|
||||
engine_inputs = self._pre_process(
|
||||
scoring_data,
|
||||
tok_params,
|
||||
prompt_extras={
|
||||
k: v
|
||||
for k in ("mm_processor_kwargs", "cache_salt")
|
||||
if (v := getattr(request, k, None)) is not None
|
||||
},
|
||||
)
|
||||
|
||||
ctx.engine_inputs = engine_inputs
|
||||
ctx.intermediates = len(scoring_data.data_1)
|
||||
|
||||
def post_process_online(
|
||||
self,
|
||||
ctx: ScoringServeContext,
|
||||
):
|
||||
if ctx.final_res_batch is None:
|
||||
raise ValueError("Final response batch not available")
|
||||
|
||||
if ctx.intermediates is None:
|
||||
raise ValueError("data_1 len not available")
|
||||
|
||||
ctx.final_res_batch = self._post_process(
|
||||
outputs=ctx.final_res_batch, offset=cast(int, ctx.intermediates)
|
||||
)
|
||||
|
||||
#######################################
|
||||
# offline APIs
|
||||
|
||||
def pre_process_offline(self, ctx: OfflineInputsContext) -> Sequence[EngineInput]:
|
||||
assert isinstance(ctx.prompts, ScoringData)
|
||||
tok_params = self.renderer.default_cmpl_tok_params.with_kwargs(
|
||||
**(ctx.tokenization_kwargs or {})
|
||||
)
|
||||
return self._pre_process(ctx.prompts, tok_params)
|
||||
|
||||
def post_process_offline(
|
||||
self,
|
||||
ctx: OfflineOutputsContext,
|
||||
) -> list[PoolingRequestOutput]:
|
||||
assert ctx.offset is not None
|
||||
return self._post_process(outputs=ctx.outputs, offset=ctx.offset)
|
||||
|
||||
#######################################
|
||||
# helpers
|
||||
|
||||
def _pre_process(
|
||||
self,
|
||||
scoring_data: ScoringData,
|
||||
tok_params: TokenizeParams,
|
||||
prompt_extras: dict[str, Any] | None = None,
|
||||
) -> Sequence[EngineInput]:
|
||||
data_1 = score_data_to_prompts(scoring_data.data_1, "query", self.model_config)
|
||||
data_2 = score_data_to_prompts(
|
||||
scoring_data.data_2, "document", self.model_config
|
||||
)
|
||||
|
||||
return self._preprocess_completion_offline(
|
||||
prompts=data_1 + data_2, tok_params=tok_params, prompt_extras=prompt_extras
|
||||
)
|
||||
|
||||
def _post_process(self, outputs: list[PoolingRequestOutput], offset: int):
|
||||
emb_data_1 = outputs[:offset]
|
||||
emb_data_2 = outputs[offset:]
|
||||
|
||||
if len(emb_data_1) == 1:
|
||||
emb_data_1 = emb_data_1 * len(emb_data_2)
|
||||
|
||||
final_res_batch: list[PoolingRequestOutput] = []
|
||||
for emb_1, emb_2 in zip(emb_data_1, emb_data_2):
|
||||
pair_score = F.cosine_similarity(
|
||||
emb_1.outputs.data.float(), emb_2.outputs.data.float(), dim=0
|
||||
)
|
||||
|
||||
padding: list[int] = []
|
||||
if self.pad_token_id is not None:
|
||||
padding = [self.pad_token_id]
|
||||
|
||||
tokens = emb_1.prompt_token_ids + padding + emb_2.prompt_token_ids
|
||||
|
||||
final_res_batch.append(
|
||||
PoolingRequestOutput(
|
||||
request_id=f"{emb_1.request_id}_{emb_2.request_id}",
|
||||
outputs=pair_score,
|
||||
prompt_token_ids=tokens,
|
||||
num_cached_tokens=emb_1.num_cached_tokens + emb_2.num_cached_tokens,
|
||||
finished=True,
|
||||
)
|
||||
)
|
||||
return final_res_batch
|
||||
|
||||
|
||||
class LateInteractionIOProcessor(BiEncoderIOProcessor):
|
||||
name: ScoreType = "late-interaction"
|
||||
pooling_task: PoolingTask = "token_embed"
|
||||
|
||||
def _post_process(self, outputs: list[PoolingRequestOutput], offset: int):
|
||||
# Split into query and document embeddings
|
||||
emb_data_1 = outputs[:offset]
|
||||
emb_data_2 = outputs[offset:]
|
||||
|
||||
# Expand queries if 1:N scoring
|
||||
if len(emb_data_1) == 1:
|
||||
emb_data_1 = emb_data_1 * len(emb_data_2)
|
||||
|
||||
final_res_batch: list[PoolingRequestOutput] = []
|
||||
padding: list[int] = []
|
||||
if (pad_token_id := self.pad_token_id) is not None:
|
||||
padding = [pad_token_id]
|
||||
|
||||
# Compute MaxSim scores
|
||||
for emb_1, emb_2 in zip(emb_data_1, emb_data_2):
|
||||
# emb_1.outputs.data: [query_len, dim]
|
||||
# emb_2.outputs.data: [doc_len, dim]
|
||||
q_emb = emb_1.outputs.data
|
||||
d_emb = emb_2.outputs.data
|
||||
|
||||
maxsim_score = compute_maxsim_score(q_emb, d_emb)
|
||||
|
||||
tokens = emb_1.prompt_token_ids + padding + emb_2.prompt_token_ids
|
||||
|
||||
final_res_batch.append(
|
||||
PoolingRequestOutput(
|
||||
request_id=f"{emb_1.request_id}_{emb_2.request_id}",
|
||||
outputs=maxsim_score,
|
||||
prompt_token_ids=tokens,
|
||||
num_cached_tokens=emb_1.num_cached_tokens + emb_2.num_cached_tokens,
|
||||
finished=True,
|
||||
)
|
||||
)
|
||||
return final_res_batch
|
||||
|
||||
|
||||
class CrossEncoderIOProcessor(ScoringIOProcessor):
|
||||
name: ScoreType = "cross-encoder"
|
||||
pooling_task: PoolingTask = "classify"
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
if is_mistral_tokenizer(self.tokenizer):
|
||||
raise ValueError("MistralTokenizer not supported for cross-encoding")
|
||||
|
||||
from vllm.model_executor.model_loader import get_model_cls
|
||||
from vllm.model_executor.models.interfaces import supports_score_template
|
||||
|
||||
model = get_model_cls(self.model_config)
|
||||
self.supports_score_template = supports_score_template(model)
|
||||
self.model = model if self.supports_score_template else None
|
||||
self.use_sep_token = self.model_config.use_sep_token
|
||||
|
||||
#######################################
|
||||
# online APIs
|
||||
|
||||
def pre_process_online(self, ctx: ScoringServeContext):
|
||||
request = ctx.request
|
||||
|
||||
if isinstance(request, ScoreRequest):
|
||||
data_1 = request.data_1
|
||||
data_2 = request.data_2
|
||||
elif isinstance(request, RerankRequest):
|
||||
data_1 = request.query
|
||||
data_2 = request.documents
|
||||
else:
|
||||
raise ValueError(f"Invalid {self.name} request type")
|
||||
|
||||
scoring_data = self.valid_inputs(data_1, data_2)
|
||||
tok_params = request.build_tok_params(self.model_config)
|
||||
pooling_params = self.create_pooling_params(request)
|
||||
|
||||
engine_inputs, pooling_params_list = self._pre_process(
|
||||
scoring_data,
|
||||
tok_params,
|
||||
pooling_params,
|
||||
chat_template=self.chat_template,
|
||||
prompt_extras={
|
||||
k: v
|
||||
for k in ("mm_processor_kwargs", "cache_salt")
|
||||
if (v := getattr(request, k, None)) is not None
|
||||
},
|
||||
)
|
||||
|
||||
ctx.engine_inputs = engine_inputs
|
||||
ctx.pooling_params = pooling_params_list
|
||||
|
||||
#######################################
|
||||
# offline APIs
|
||||
|
||||
def pre_process_offline(self, ctx: OfflineInputsContext) -> Sequence[EngineInput]:
|
||||
assert isinstance(ctx.prompts, ScoringData)
|
||||
assert not isinstance(ctx.pooling_params, list)
|
||||
|
||||
tok_params = self.renderer.default_cmpl_tok_params.with_kwargs(
|
||||
**(ctx.tokenization_kwargs or {})
|
||||
)
|
||||
engine_inputs, pooling_params_list = self._pre_process(
|
||||
ctx.prompts, tok_params, ctx.pooling_params, ctx.chat_template
|
||||
)
|
||||
ctx.pooling_params = pooling_params_list
|
||||
return engine_inputs
|
||||
|
||||
#######################################
|
||||
# helpers
|
||||
|
||||
def _pre_process(
|
||||
self,
|
||||
scoring_data: ScoringData,
|
||||
tok_params: TokenizeParams,
|
||||
pooling_params: PoolingParams | None,
|
||||
chat_template: str | None = None,
|
||||
prompt_extras: dict[str, Any] | None = None,
|
||||
) -> tuple[Sequence[EngineInput], list[PoolingParams]]:
|
||||
# todo: support prompt_extras
|
||||
arrival_time = time.time()
|
||||
|
||||
data_1 = scoring_data.data_1
|
||||
data_2 = scoring_data.data_2
|
||||
|
||||
if len(data_1) == 1:
|
||||
data_1 = data_1 * len(data_2)
|
||||
|
||||
if pooling_params is None:
|
||||
pooling_params = PoolingParams(task="classify")
|
||||
|
||||
pooling_params_list = list[PoolingParams]()
|
||||
engine_inputs = list[EngineInput]()
|
||||
for q, d in zip(data_1, data_2):
|
||||
_, engine_prompt = self.get_score_prompt(
|
||||
data_1=q,
|
||||
data_2=d,
|
||||
encode_kwargs=tok_params.get_encode_kwargs(),
|
||||
chat_template=chat_template,
|
||||
)
|
||||
|
||||
if token_type_ids := engine_prompt.pop("token_type_ids", None):
|
||||
params = pooling_params.clone()
|
||||
compressed = compress_token_type_ids(token_type_ids)
|
||||
params.extra_kwargs = {"compressed_token_type_ids": compressed}
|
||||
pooling_params_list.append(params)
|
||||
else:
|
||||
pooling_params_list.append(pooling_params)
|
||||
|
||||
tok_params.apply_post_tokenization(self.tokenizer, engine_prompt)
|
||||
engine_inputs.append(
|
||||
self.renderer.process_for_engine(engine_prompt, arrival_time)
|
||||
)
|
||||
return engine_inputs, pooling_params_list
|
||||
|
||||
def get_score_prompt(
|
||||
self,
|
||||
data_1: ScoreData,
|
||||
data_2: ScoreData,
|
||||
encode_kwargs: dict[str, Any],
|
||||
chat_template: str | None = None,
|
||||
):
|
||||
model_config = self.model_config
|
||||
tokenizer = self.tokenizer
|
||||
|
||||
prompt_1, prompt_2, mm_data, mm_uuids = parse_score_data(
|
||||
data_1,
|
||||
data_2,
|
||||
model_config,
|
||||
)
|
||||
|
||||
def default_tokenizer_encode():
|
||||
if self.supports_score_template:
|
||||
assert self.model is not None
|
||||
full_prompt = self.model.get_score_template(prompt_1, prompt_2)
|
||||
if full_prompt is None:
|
||||
raise ValueError("Get empty score template from model")
|
||||
|
||||
prompt_inputs = tokenizer(full_prompt, **encode_kwargs)
|
||||
else:
|
||||
if self.use_sep_token:
|
||||
# cross_encoder models defaults to using separating token.
|
||||
prompt_inputs = tokenizer(
|
||||
text=prompt_1, text_pair=prompt_2, **encode_kwargs
|
||||
)
|
||||
full_prompt = tokenizer.decode(prompt_inputs["input_ids"])
|
||||
else:
|
||||
# `llm as reranker` defaults to not using separating token.
|
||||
full_prompt = prompt_1 + prompt_2
|
||||
prompt_inputs = tokenizer(text=full_prompt, **encode_kwargs)
|
||||
return full_prompt, prompt_inputs
|
||||
|
||||
# FIXME: For now, we only apply a template when one is explicitly provided.
|
||||
# We cannot rely on the tokenizer's chat template because many models
|
||||
# inherit junk templates from their base LLM, which breaks both the models
|
||||
# and the tests that use them.
|
||||
if chat_template is None:
|
||||
full_prompt, prompt_inputs = default_tokenizer_encode()
|
||||
else:
|
||||
# FIXME:
|
||||
# Try applying a score template from the CLI arg or tokenizer_config.json
|
||||
# If that fails because there is no such template,
|
||||
# fall back to the default implementation.
|
||||
try:
|
||||
full_prompt = safe_apply_chat_template(
|
||||
model_config,
|
||||
tokenizer,
|
||||
[
|
||||
{"role": "query", "content": prompt_1},
|
||||
{"role": "document", "content": prompt_2},
|
||||
],
|
||||
chat_template=chat_template,
|
||||
tools=None,
|
||||
tokenize=False,
|
||||
)
|
||||
prompt_inputs = tokenizer(full_prompt, **encode_kwargs)
|
||||
except ChatTemplateResolutionError:
|
||||
full_prompt, prompt_inputs = default_tokenizer_encode()
|
||||
|
||||
engine_prompt = TokensPrompt(prompt_token_ids=prompt_inputs["input_ids"])
|
||||
|
||||
if (token_type_ids := prompt_inputs.get("token_type_ids")) is not None:
|
||||
engine_prompt["token_type_ids"] = token_type_ids
|
||||
|
||||
if self.model is not None:
|
||||
self.model.post_process_tokens(engine_prompt)
|
||||
|
||||
if mm_data is not None:
|
||||
engine_prompt["multi_modal_data"] = mm_data
|
||||
if mm_uuids is not None:
|
||||
engine_prompt["multi_modal_uuids"] = mm_uuids
|
||||
|
||||
return full_prompt, engine_prompt
|
||||
|
||||
|
||||
ScoringIOProcessors: dict[ScoreType, type[ScoringIOProcessor]] = {
|
||||
"bi-encoder": BiEncoderIOProcessor,
|
||||
"late-interaction": LateInteractionIOProcessor,
|
||||
"cross-encoder": CrossEncoderIOProcessor,
|
||||
}
|
||||
@@ -0,0 +1,159 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import time
|
||||
from typing import TypeAlias
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
from vllm import PoolingParams
|
||||
from vllm.config import ModelConfig
|
||||
from vllm.entrypoints.openai.engine.protocol import OpenAIBaseModel, UsageInfo
|
||||
from vllm.entrypoints.pooling.base.protocol import (
|
||||
ClassifyRequestMixin,
|
||||
PoolingBasicRequestMixin,
|
||||
)
|
||||
from vllm.renderers import TokenizeParams
|
||||
from vllm.tasks import PoolingTask
|
||||
from vllm.utils import random_uuid
|
||||
|
||||
from .typing import ScoreContentPartParam, ScoreInput
|
||||
|
||||
|
||||
class ScoreRequestMixin(PoolingBasicRequestMixin, ClassifyRequestMixin):
|
||||
def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams:
|
||||
encoder_config = model_config.encoder_config or {}
|
||||
|
||||
return TokenizeParams(
|
||||
max_total_tokens=model_config.max_model_len,
|
||||
max_output_tokens=0,
|
||||
truncate_prompt_tokens=self.truncate_prompt_tokens,
|
||||
truncation_side=self.truncation_side,
|
||||
do_lower_case=encoder_config.get("do_lower_case", False),
|
||||
max_total_tokens_param="max_model_len",
|
||||
)
|
||||
|
||||
def to_pooling_params(self, task: PoolingTask = "classify"):
|
||||
return PoolingParams(
|
||||
task=task,
|
||||
use_activation=self.use_activation,
|
||||
)
|
||||
|
||||
|
||||
class ScoreDataRequest(ScoreRequestMixin):
|
||||
data_1: ScoreInput | list[ScoreInput]
|
||||
data_2: ScoreInput | list[ScoreInput]
|
||||
|
||||
|
||||
class ScoreQueriesDocumentsRequest(ScoreRequestMixin):
|
||||
queries: ScoreInput | list[ScoreInput]
|
||||
documents: ScoreInput | list[ScoreInput]
|
||||
|
||||
@property
|
||||
def data_1(self):
|
||||
return self.queries
|
||||
|
||||
@property
|
||||
def data_2(self):
|
||||
return self.documents
|
||||
|
||||
|
||||
class ScoreQueriesItemsRequest(ScoreRequestMixin):
|
||||
queries: ScoreInput | list[ScoreInput]
|
||||
items: ScoreInput | list[ScoreInput]
|
||||
|
||||
@property
|
||||
def data_1(self):
|
||||
return self.queries
|
||||
|
||||
@property
|
||||
def data_2(self):
|
||||
return self.items
|
||||
|
||||
|
||||
class ScoreTextRequest(ScoreRequestMixin):
|
||||
text_1: ScoreInput | list[ScoreInput]
|
||||
text_2: ScoreInput | list[ScoreInput]
|
||||
|
||||
@property
|
||||
def data_1(self):
|
||||
return self.text_1
|
||||
|
||||
@property
|
||||
def data_2(self):
|
||||
return self.text_2
|
||||
|
||||
|
||||
ScoreRequest: TypeAlias = (
|
||||
ScoreQueriesDocumentsRequest
|
||||
| ScoreQueriesItemsRequest
|
||||
| ScoreDataRequest
|
||||
| ScoreTextRequest
|
||||
)
|
||||
|
||||
|
||||
class RerankRequest(PoolingBasicRequestMixin, ClassifyRequestMixin):
|
||||
query: ScoreInput
|
||||
documents: ScoreInput | list[ScoreInput]
|
||||
top_n: int = Field(default_factory=lambda: 0)
|
||||
|
||||
def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams:
|
||||
encoder_config = model_config.encoder_config or {}
|
||||
|
||||
return TokenizeParams(
|
||||
max_total_tokens=model_config.max_model_len,
|
||||
max_output_tokens=0,
|
||||
truncate_prompt_tokens=self.truncate_prompt_tokens,
|
||||
truncation_side=self.truncation_side,
|
||||
do_lower_case=encoder_config.get("do_lower_case", False),
|
||||
max_total_tokens_param="max_model_len",
|
||||
)
|
||||
|
||||
def to_pooling_params(self, task: PoolingTask = "classify"):
|
||||
return PoolingParams(
|
||||
task=task,
|
||||
use_activation=self.use_activation,
|
||||
)
|
||||
|
||||
|
||||
ScoringRequest: TypeAlias = ScoreRequest | RerankRequest
|
||||
|
||||
|
||||
class RerankDocument(BaseModel):
|
||||
text: str | None = None
|
||||
multi_modal: list[ScoreContentPartParam] | None = None
|
||||
|
||||
|
||||
class RerankResult(BaseModel):
|
||||
index: int
|
||||
document: RerankDocument
|
||||
relevance_score: float
|
||||
|
||||
|
||||
class RerankUsage(BaseModel):
|
||||
prompt_tokens: int
|
||||
total_tokens: int
|
||||
|
||||
|
||||
class RerankResponse(OpenAIBaseModel):
|
||||
id: str
|
||||
model: str
|
||||
usage: RerankUsage
|
||||
results: list[RerankResult]
|
||||
|
||||
|
||||
class ScoreResponseData(OpenAIBaseModel):
|
||||
index: int
|
||||
object: str = "score"
|
||||
score: float
|
||||
|
||||
|
||||
class ScoreResponse(OpenAIBaseModel):
|
||||
id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
|
||||
object: str = "list"
|
||||
created: int = Field(default_factory=lambda: int(time.time()))
|
||||
model: str
|
||||
data: list[ScoreResponseData]
|
||||
usage: UsageInfo
|
||||
|
||||
|
||||
ScoringResponse: TypeAlias = RerankResponse | ScoreResponse
|
||||
@@ -0,0 +1,160 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from fastapi.responses import JSONResponse
|
||||
|
||||
from vllm.config import ModelConfig
|
||||
from vllm.entrypoints.chat_utils import ChatTemplateConfig
|
||||
from vllm.entrypoints.openai.engine.protocol import UsageInfo
|
||||
from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor
|
||||
from vllm.entrypoints.pooling.base.serving import PoolingServing
|
||||
from vllm.logger import init_logger
|
||||
from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput
|
||||
from vllm.renderers import BaseRenderer
|
||||
|
||||
from .io_processor import ScoringIOProcessors, ScoringServeContext
|
||||
from .protocol import (
|
||||
RerankDocument,
|
||||
RerankRequest,
|
||||
RerankResponse,
|
||||
RerankResult,
|
||||
RerankUsage,
|
||||
ScoreRequest,
|
||||
ScoreResponse,
|
||||
ScoreResponseData,
|
||||
)
|
||||
from .typing import ScoreInput
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class ServingScores(PoolingServing):
|
||||
request_id_prefix = "score"
|
||||
|
||||
def init_io_processor(
|
||||
self,
|
||||
model_config: ModelConfig,
|
||||
renderer: BaseRenderer,
|
||||
chat_template_config: ChatTemplateConfig,
|
||||
) -> PoolingIOProcessor:
|
||||
score_type = model_config.score_type
|
||||
assert score_type in ScoringIOProcessors
|
||||
processor_cls = ScoringIOProcessors[score_type]
|
||||
return processor_cls(
|
||||
model_config=model_config,
|
||||
renderer=renderer,
|
||||
chat_template_config=chat_template_config,
|
||||
)
|
||||
|
||||
async def _build_response(
|
||||
self,
|
||||
ctx: ScoringServeContext,
|
||||
) -> JSONResponse:
|
||||
final_res_batch = ctx.final_res_batch
|
||||
request_id = ctx.request_id
|
||||
created_time = ctx.created_time
|
||||
model_name = self.models.model_name()
|
||||
|
||||
if isinstance(ctx.request, ScoreRequest):
|
||||
return self._request_output_to_score_response(
|
||||
final_res_batch,
|
||||
request_id,
|
||||
created_time,
|
||||
model_name,
|
||||
)
|
||||
elif isinstance(ctx.request, RerankRequest):
|
||||
return self._request_output_to_rerank_response(
|
||||
final_res_batch,
|
||||
request_id,
|
||||
model_name,
|
||||
ctx.request.documents,
|
||||
ctx.request.top_n if ctx.request.top_n > 0 else len(final_res_batch),
|
||||
)
|
||||
else:
|
||||
raise NotImplementedError("")
|
||||
|
||||
def _request_output_to_score_response(
|
||||
self,
|
||||
final_res_batch: list[PoolingRequestOutput],
|
||||
request_id: str,
|
||||
created_time: int,
|
||||
model_name: str,
|
||||
) -> JSONResponse:
|
||||
items: list[ScoreResponseData] = []
|
||||
num_prompt_tokens = 0
|
||||
|
||||
for idx, final_res in enumerate(final_res_batch):
|
||||
classify_res = ScoringRequestOutput.from_base(final_res)
|
||||
|
||||
item = ScoreResponseData(
|
||||
index=idx,
|
||||
score=classify_res.outputs.score,
|
||||
)
|
||||
prompt_token_ids = final_res.prompt_token_ids
|
||||
|
||||
items.append(item)
|
||||
num_prompt_tokens += len(prompt_token_ids)
|
||||
|
||||
usage = UsageInfo(
|
||||
prompt_tokens=num_prompt_tokens,
|
||||
total_tokens=num_prompt_tokens,
|
||||
)
|
||||
|
||||
response = ScoreResponse(
|
||||
id=request_id,
|
||||
created=created_time,
|
||||
model=model_name,
|
||||
data=items,
|
||||
usage=usage,
|
||||
)
|
||||
|
||||
return JSONResponse(content=response.model_dump())
|
||||
|
||||
def _request_output_to_rerank_response(
|
||||
self,
|
||||
final_res_batch: list[PoolingRequestOutput],
|
||||
request_id: str,
|
||||
model_name: str,
|
||||
documents: ScoreInput | list[ScoreInput],
|
||||
top_n: int,
|
||||
) -> JSONResponse:
|
||||
if not isinstance(documents, list):
|
||||
documents = [documents]
|
||||
|
||||
results: list[RerankResult] = []
|
||||
num_prompt_tokens = 0
|
||||
for idx, final_res in enumerate(final_res_batch):
|
||||
classify_res = ScoringRequestOutput.from_base(final_res)
|
||||
|
||||
document = documents[idx]
|
||||
if isinstance(document, str):
|
||||
rerank_document = RerankDocument(text=document)
|
||||
else:
|
||||
rerank_document = RerankDocument(
|
||||
multi_modal=document.get("content", [])
|
||||
)
|
||||
|
||||
result = RerankResult(
|
||||
index=idx,
|
||||
document=rerank_document,
|
||||
relevance_score=classify_res.outputs.score,
|
||||
)
|
||||
results.append(result)
|
||||
prompt_token_ids = final_res.prompt_token_ids
|
||||
num_prompt_tokens += len(prompt_token_ids)
|
||||
|
||||
# sort by relevance, then return the top n if set
|
||||
results.sort(key=lambda x: x.relevance_score, reverse=True)
|
||||
if top_n < len(documents):
|
||||
results = results[:top_n]
|
||||
|
||||
response = RerankResponse(
|
||||
id=request_id,
|
||||
model=model_name,
|
||||
results=results,
|
||||
usage=RerankUsage(
|
||||
total_tokens=num_prompt_tokens, prompt_tokens=num_prompt_tokens
|
||||
),
|
||||
)
|
||||
|
||||
return JSONResponse(content=response.model_dump())
|
||||
@@ -0,0 +1,46 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from dataclasses import dataclass
|
||||
from typing import TypeAlias
|
||||
|
||||
from typing_extensions import Required, TypedDict
|
||||
|
||||
from vllm.entrypoints.chat_utils import (
|
||||
ChatCompletionContentPartImageEmbedsParam,
|
||||
ChatCompletionContentPartImageParam,
|
||||
ChatCompletionContentPartTextParam,
|
||||
ChatCompletionContentPartVideoParam,
|
||||
)
|
||||
|
||||
ScoreContentPartParam: TypeAlias = (
|
||||
ChatCompletionContentPartImageParam
|
||||
| ChatCompletionContentPartImageEmbedsParam
|
||||
| ChatCompletionContentPartTextParam
|
||||
| ChatCompletionContentPartVideoParam
|
||||
)
|
||||
|
||||
|
||||
class ScoreMultiModalParam(TypedDict, total=False):
|
||||
"""
|
||||
A specialized parameter type for scoring multimodal content
|
||||
|
||||
The reasons why don't reuse `CustomChatCompletionMessageParam` directly:
|
||||
1. Score tasks don't need the 'role' field (user/assistant/system) that's required in chat completions
|
||||
2. Including chat-specific fields would confuse users about their purpose in scoring
|
||||
3. This is a more focused interface that only exposes what's needed for scoring
|
||||
""" # noqa: E501
|
||||
|
||||
content: Required[list[ScoreContentPartParam]]
|
||||
"""The multimodal contents"""
|
||||
|
||||
|
||||
# Raw input data with content key in ScoreMultiModalParam.
|
||||
ScoreInput = str | ScoreMultiModalParam
|
||||
# Score data without content key.
|
||||
ScoreData = str | list[ScoreContentPartParam]
|
||||
|
||||
|
||||
@dataclass
|
||||
class ScoringData:
|
||||
data_1: list[ScoreData]
|
||||
data_2: list[ScoreData]
|
||||
@@ -0,0 +1,233 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from collections.abc import Iterable
|
||||
from typing import cast
|
||||
|
||||
import torch
|
||||
|
||||
from vllm import PromptType, TextPrompt
|
||||
from vllm.config import ModelConfig
|
||||
from vllm.entrypoints.chat_utils import (
|
||||
BaseMultiModalItemTracker,
|
||||
ChatCompletionContentPartParam,
|
||||
ChatCompletionContentPartTextParam,
|
||||
ConversationMessage,
|
||||
MultiModalItemTracker,
|
||||
_parse_chat_message_content_parts,
|
||||
)
|
||||
from vllm.inputs import MultiModalDataDict, MultiModalUUIDDict
|
||||
|
||||
from .typing import (
|
||||
ScoreContentPartParam,
|
||||
ScoreData,
|
||||
ScoreInput,
|
||||
ScoringData,
|
||||
)
|
||||
|
||||
|
||||
def compute_maxsim_score(q_emb: torch.Tensor, d_emb: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Compute ColBERT MaxSim score.
|
||||
|
||||
Args:
|
||||
q_emb: Query token embeddings [query_len, dim]
|
||||
d_emb: Document token embeddings [doc_len, dim]
|
||||
|
||||
Returns:
|
||||
MaxSim score (sum over query tokens of max similarity to any doc token)
|
||||
"""
|
||||
# [query_len, doc_len]
|
||||
token_scores = torch.matmul(q_emb, d_emb.T)
|
||||
# Max over document tokens, sum over query tokens
|
||||
return token_scores.amax(dim=-1).sum()
|
||||
|
||||
|
||||
def _validate_mm_score_input(
|
||||
data: list[ScoreInput],
|
||||
is_multimodal_model: bool,
|
||||
architecture: str,
|
||||
) -> list[ScoreData]:
|
||||
out: list[ScoreData] = []
|
||||
for d in data:
|
||||
if isinstance(d, str):
|
||||
out.append(d)
|
||||
else:
|
||||
if not is_multimodal_model:
|
||||
raise ValueError(f"MultiModalParam is not supported for {architecture}")
|
||||
content = cast(list[ScoreContentPartParam], d.get("content", []))
|
||||
out.append(content)
|
||||
return out
|
||||
|
||||
|
||||
def _validate_score_input_lens(
|
||||
data_1: list[ScoreData],
|
||||
data_2: list[ScoreData],
|
||||
):
|
||||
len_1 = len(data_1)
|
||||
len_2 = len(data_2)
|
||||
|
||||
if len_1 > 1 and len_1 != len_2:
|
||||
raise ValueError("Input lengths must be either 1:1, 1:N or N:N")
|
||||
if len_1 == 0:
|
||||
raise ValueError("At least one text element must be given")
|
||||
if len_2 == 0:
|
||||
raise ValueError("At least one text_pair element must be given")
|
||||
|
||||
|
||||
def validate_score_input(
|
||||
data_1: ScoreInput | list[ScoreInput],
|
||||
data_2: ScoreInput | list[ScoreInput],
|
||||
is_multimodal_model: bool,
|
||||
architecture: str,
|
||||
) -> ScoringData:
|
||||
if not isinstance(data_1, list):
|
||||
data_1 = [data_1]
|
||||
|
||||
if not isinstance(data_2, list):
|
||||
data_2 = [data_2]
|
||||
|
||||
score_input_1 = _validate_mm_score_input(data_1, is_multimodal_model, architecture)
|
||||
score_input_2 = _validate_mm_score_input(data_2, is_multimodal_model, architecture)
|
||||
_validate_score_input_lens(score_input_1, score_input_2)
|
||||
return ScoringData(data_1=score_input_1, data_2=score_input_2)
|
||||
|
||||
|
||||
def score_data_to_prompts(
|
||||
data_list: list[ScoreData],
|
||||
role: str,
|
||||
model_config: ModelConfig,
|
||||
) -> list[PromptType]:
|
||||
"""Convert a list of ScoreData into PromptType objects.
|
||||
|
||||
For plain text inputs, returns the string directly.
|
||||
For multimodal inputs (list of content parts), parses them into
|
||||
a :class:`TextPrompt` with attached ``multi_modal_data`` /
|
||||
``multi_modal_uuids``.
|
||||
|
||||
This is used by late-interaction scoring where each query/document
|
||||
is encoded independently.
|
||||
"""
|
||||
prompts: list[PromptType] = []
|
||||
for data in data_list:
|
||||
if isinstance(data, str):
|
||||
prompts.append(data)
|
||||
else:
|
||||
text, mm_data, mm_uuids = parse_score_data_single(data, role, model_config)
|
||||
prompt: TextPrompt = TextPrompt(prompt=text)
|
||||
if mm_data is not None:
|
||||
prompt["multi_modal_data"] = mm_data
|
||||
if mm_uuids is not None:
|
||||
prompt["multi_modal_uuids"] = mm_uuids
|
||||
prompts.append(prompt)
|
||||
return prompts
|
||||
|
||||
|
||||
def _ensure_str(content: list[ConversationMessage]) -> str:
|
||||
"""Extract a single string prompt from parsed conversation content."""
|
||||
assert len(content) == 1
|
||||
prompt = content[0]["content"]
|
||||
if prompt is not None and isinstance(prompt, str):
|
||||
return cast(str, prompt)
|
||||
raise ValueError(f"Only string content is supported, but got {content}.")
|
||||
|
||||
|
||||
def _parse_score_content(
|
||||
role: str,
|
||||
data: ScoreData,
|
||||
mm_tracker: BaseMultiModalItemTracker,
|
||||
) -> list[ConversationMessage]:
|
||||
parts: Iterable[ChatCompletionContentPartParam]
|
||||
if isinstance(data, str):
|
||||
parts = [ChatCompletionContentPartTextParam(type="text", text=data)]
|
||||
else:
|
||||
parts = cast(Iterable[ChatCompletionContentPartParam], data)
|
||||
|
||||
mm_parser = mm_tracker.create_parser()
|
||||
|
||||
parse_res = _parse_chat_message_content_parts(
|
||||
role=role,
|
||||
parts=parts,
|
||||
mm_tracker=mm_tracker,
|
||||
wrap_dicts=False,
|
||||
interleave_strings=False,
|
||||
)
|
||||
|
||||
if parse_res:
|
||||
return parse_res
|
||||
|
||||
mm_placeholder_storage = mm_parser.mm_placeholder_storage()
|
||||
|
||||
if (
|
||||
len(mm_placeholder_storage) != 1
|
||||
or len(next(iter(mm_placeholder_storage.values()))) != 1
|
||||
):
|
||||
raise ValueError("Only one multi-modal item is supported")
|
||||
|
||||
return next(iter(mm_placeholder_storage.values()))[0]
|
||||
|
||||
|
||||
def parse_score_data_single(
|
||||
data: ScoreData,
|
||||
role: str,
|
||||
model_config: ModelConfig,
|
||||
) -> tuple[str, MultiModalDataDict | None, MultiModalUUIDDict | None]:
|
||||
"""Parse **one** ScoreData into a text prompt and its own multi-modal
|
||||
data.
|
||||
|
||||
Unlike :func:`parse_score_data`, each call creates an **independent**
|
||||
:class:`MultiModalItemTracker` so multi-modal items are kept separate.
|
||||
This is the correct behaviour for late-interaction scoring, where
|
||||
query and document are encoded independently.
|
||||
"""
|
||||
mm_tracker = MultiModalItemTracker(model_config)
|
||||
content = _parse_score_content(role, data, mm_tracker)
|
||||
|
||||
prompt = _ensure_str(content)
|
||||
mm_items, mm_uuids = mm_tracker.resolve_items()
|
||||
return prompt, mm_items, mm_uuids
|
||||
|
||||
|
||||
def parse_score_data(
|
||||
data_1: ScoreData,
|
||||
data_2: ScoreData,
|
||||
model_config: ModelConfig,
|
||||
) -> tuple[str, str, MultiModalDataDict | None, MultiModalUUIDDict | None]:
|
||||
"""Parse a query-document pair into text prompts and shared multi-modal
|
||||
data.
|
||||
|
||||
Uses a **single** :class:`MultiModalItemTracker` so that multi-modal
|
||||
items from both inputs are merged into one ``mm_data`` dict. This is
|
||||
the correct behaviour for cross-encoder scoring, where query and
|
||||
document are concatenated into a single model prompt.
|
||||
"""
|
||||
mm_tracker = MultiModalItemTracker(model_config)
|
||||
|
||||
content_1 = _parse_score_content("query", data_1, mm_tracker)
|
||||
content_2 = _parse_score_content("document", data_2, mm_tracker)
|
||||
|
||||
prompt_1 = _ensure_str(content_1)
|
||||
prompt_2 = _ensure_str(content_2)
|
||||
mm_items, mm_uuids = mm_tracker.resolve_items()
|
||||
|
||||
return prompt_1, prompt_2, mm_items, mm_uuids
|
||||
|
||||
|
||||
def compress_token_type_ids(token_type_ids: list[int]) -> int:
|
||||
"""
|
||||
Return position of the first 1 or the length of the list
|
||||
if not found.
|
||||
"""
|
||||
first_one = len(token_type_ids)
|
||||
err_msg = (
|
||||
"Token type ids are expected to be a sequence"
|
||||
" of zeros followed by a sequence of ones"
|
||||
)
|
||||
for i, type_id in enumerate(token_type_ids):
|
||||
if type_id == 0 and first_one < i:
|
||||
raise ValueError(err_msg)
|
||||
elif type_id == 1 and first_one > i:
|
||||
first_one = i
|
||||
elif type_id > 1:
|
||||
raise ValueError(err_msg)
|
||||
|
||||
return first_one
|
||||
Reference in New Issue
Block a user