[Frontend][2/n] Improve pooling entrypoints | embed. (#36110)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2026-03-09 11:42:19 +08:00
committed by GitHub
parent c4d859c274
commit fff3711a24
20 changed files with 508 additions and 755 deletions
+4 -21
View File
@@ -60,12 +60,6 @@ from vllm.entrypoints.openai.speech_to_text.protocol import (
TranscriptionResponse,
TranslationRequest,
)
from vllm.entrypoints.pooling.embed.protocol import (
EmbeddingBytesResponse,
EmbeddingChatRequest,
EmbeddingCompletionRequest,
EmbeddingResponse,
)
from vllm.entrypoints.pooling.pooling.protocol import (
IOProcessorRequest,
PoolingChatRequest,
@@ -144,17 +138,13 @@ CompletionLikeRequest: TypeAlias = (
CompletionRequest
| TokenizeCompletionRequest
| DetokenizeRequest
| EmbeddingCompletionRequest
| RerankRequest
| ScoreRequest
| PoolingCompletionRequest
)
ChatLikeRequest: TypeAlias = (
ChatCompletionRequest
| TokenizeChatRequest
| EmbeddingChatRequest
| PoolingChatRequest
ChatCompletionRequest | TokenizeChatRequest | PoolingChatRequest
)
SpeechToTextRequest: TypeAlias = TranscriptionRequest | TranslationRequest
@@ -171,8 +161,6 @@ AnyRequest: TypeAlias = (
AnyResponse: TypeAlias = (
CompletionResponse
| ChatCompletionResponse
| EmbeddingResponse
| EmbeddingBytesResponse
| TranscriptionResponse
| TokenizeResponse
| PoolingResponse
@@ -203,8 +191,7 @@ class ServeContext(Generic[RequestT]):
class OpenAIServing:
request_id_prefix: ClassVar[str] = """
A short string prepended to every request’s ID (e.g. "embd")
so you can easily tell “this ID came from Embedding.”
A short string prepended to every request’s ID.
"""
def __init__(
@@ -432,8 +419,7 @@ class OpenAIServing:
ctx: ServeContext,
) -> ErrorResponse | None:
"""
Default preprocessing hook. Subclasses may override
to prepare `ctx` (embedding, etc.).
Default preprocessing hook. Subclasses may override to prepare `ctx`.
"""
return None
@@ -730,13 +716,10 @@ class OpenAIServing:
token_num = len(input_ids)
max_model_len = self.model_config.max_model_len
# Note: EmbeddingRequest,
# and ScoreRequest doesn't have max_tokens
# Note: ScoreRequest doesn't have max_tokens
if isinstance(
request,
(
EmbeddingChatRequest,
EmbeddingCompletionRequest,
ScoreDataRequest,
ScoreTextRequest,
ScoreQueriesDocumentsRequest,