[Frontend][3/n] Improve pooling entrypoints | scoring. (#28631)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2026-03-31 07:52:00 +00:00
committed by GitHub
parent f09daea261
commit d9d21eb8e3
37 changed files with 1256 additions and 1779 deletions
+21 -25
View File
@@ -13,13 +13,16 @@ from vllm.entrypoints.chat_utils import (
ConversationMessage,
)
from vllm.entrypoints.openai.engine.serving import RendererChatRequest, RendererRequest
from vllm.entrypoints.pooling.scoring.typing import ScoringData
from vllm.entrypoints.pooling.typing import (
OfflineInputsContext,
OfflineOutputsContext,
PoolingChatLikeRequest,
PoolingCompletionLikeRequest,
PoolingServeContext,
)
from vllm.inputs import EngineInput, SingletonPrompt
from vllm.renderers import BaseRenderer, merge_kwargs
from vllm.renderers import BaseRenderer, TokenizeParams, merge_kwargs
from vllm.renderers.inputs.preprocess import parse_model_prompt, prompt_to_seq
from vllm.tool_parsers import ToolParser
from vllm.utils.mistral import is_mistral_tokenizer
@@ -96,29 +99,29 @@ class PoolingIOProcessor:
#######################################
# offline APIs
def pre_process_offline(
self,
prompts: PromptType | Sequence[PromptType],
tokenization_kwargs: dict[str, Any] | None = None,
) -> Sequence[EngineInput]:
def pre_process_offline(self, ctx: OfflineInputsContext) -> Sequence[EngineInput]:
assert not isinstance(ctx.prompts, ScoringData)
tok_params = self.renderer.default_cmpl_tok_params.with_kwargs(
**(ctx.tokenization_kwargs or {})
)
return self._preprocess_completion_offline(
prompts=prompts, tokenization_kwargs=tokenization_kwargs
prompts=ctx.prompts, tok_params=tok_params
)
async def pre_process_offline_async(self, *args, **kwargs):
return self.pre_process_offline(*args, **kwargs)
async def pre_process_offline_async(self, ctx: OfflineInputsContext):
return self.pre_process_offline(ctx)
def post_process_offline(
self,
outputs: list[PoolingRequestOutput],
ctx: OfflineOutputsContext,
) -> list[PoolingRequestOutput]:
return outputs
return ctx.outputs
async def post_process_offline_async(
self,
outputs: list[PoolingRequestOutput],
ctx: OfflineOutputsContext,
) -> list[PoolingRequestOutput]:
return self.post_process_offline(outputs)
return self.post_process_offline(ctx)
#######################################
# helpers
@@ -204,28 +207,21 @@ class PoolingIOProcessor:
def _preprocess_completion_offline(
self,
prompts: PromptType | Sequence[PromptType],
tokenization_kwargs: dict[str, Any] | None = None,
tok_params: TokenizeParams,
prompt_extras: dict[str, Any] | None = None,
) -> Sequence[EngineInput]:
renderer = self.renderer
model_config = self.model_config
prompts = prompt_to_seq(prompts)
parsed_prompts = [
(
prompt
if isinstance(prompt, bytes)
else parse_model_prompt(model_config, prompt)
else parse_model_prompt(self.model_config, prompt)
)
for prompt in prompts
]
tok_params = renderer.default_cmpl_tok_params.with_kwargs(
**(tokenization_kwargs or {})
)
return renderer.render_cmpl(
parsed_prompts,
tok_params,
return self.renderer.render_cmpl(
parsed_prompts, tok_params, prompt_extras=prompt_extras
)
def _validate_chat_template(