[Frontend][1/n] Improve pooling entrypoints | classify. (#35604)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
wang.yuqi
2026-03-03 06:05:36 -08:00
committed by GitHub
co-authored by Cyrus Leung
parent 440f0e7dc6
commit ea463978bb
12 changed files with 889 additions and 170 deletions
+4 -17
View File
@@ -62,11 +62,6 @@ from vllm.entrypoints.openai.speech_to_text.protocol import (
TranscriptionResponse,
TranslationRequest,
)
from vllm.entrypoints.pooling.classify.protocol import (
ClassificationChatRequest,
ClassificationCompletionRequest,
ClassificationResponse,
)
from vllm.entrypoints.pooling.embed.protocol import (
EmbeddingBytesResponse,
EmbeddingChatRequest,
@@ -161,7 +156,6 @@ CompletionLikeRequest: TypeAlias = (
| TokenizeCompletionRequest
| DetokenizeRequest
| EmbeddingCompletionRequest
| ClassificationCompletionRequest
| RerankRequest
| ScoreRequest
| PoolingCompletionRequest
@@ -171,7 +165,6 @@ ChatLikeRequest: TypeAlias = (
ChatCompletionRequest
| TokenizeChatRequest
| EmbeddingChatRequest
| ClassificationChatRequest
| PoolingChatRequest
)
@@ -194,12 +187,10 @@ AnyResponse: TypeAlias = (
| TranscriptionResponse
| TokenizeResponse
| PoolingResponse
| ClassificationResponse
| ScoreResponse
| GenerateResponse
)
RequestT = TypeVar("RequestT", bound=AnyRequest)
@@ -223,8 +214,8 @@ class ServeContext(Generic[RequestT]):
class OpenAIServing:
request_id_prefix: ClassVar[str] = """
A short string prepended to every request’s ID (e.g. "embd", "classify")
so you can easily tell “this ID came from Embedding vs Classification.”
A short string prepended to every request’s ID (e.g. "embd")
so you can easily tell “this ID came from Embedding.”
"""
def __init__(
@@ -456,7 +447,7 @@ class OpenAIServing:
) -> ErrorResponse | None:
"""
Default preprocessing hook. Subclasses may override
to prepare `ctx` (classification, embedding, etc.).
to prepare `ctx` (embedding, etc.).
"""
return None
@@ -817,7 +808,7 @@ class OpenAIServing:
token_num = len(input_ids)
max_model_len = self.model_config.max_model_len
# Note: EmbeddingRequest, ClassificationRequest,
# Note: EmbeddingRequest,
# and ScoreRequest doesn't have max_tokens
if isinstance(
request,
@@ -828,8 +819,6 @@ class OpenAIServing:
ScoreTextRequest,
ScoreQueriesDocumentsRequest,
RerankRequest,
ClassificationCompletionRequest,
ClassificationChatRequest,
),
):
# Note: input length can be up to the entire model context length
@@ -839,8 +828,6 @@ class OpenAIServing:
ScoreDataRequest: "score",
ScoreTextRequest: "score",
ScoreQueriesDocumentsRequest: "score",
ClassificationCompletionRequest: "classification",
ClassificationChatRequest: "classification",
}
operation = operations.get(type(request), "embedding generation")
raise VLLMValidationError(