[Model] Pooling model activation supports per request control by PoolingParams (#20538)

Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-08-05 00:37:00 -07:00
committed by GitHub
parent 811ac13d03
commit 586f286789
21 changed files with 948 additions and 173 deletions
+19 -3
View File
@@ -1189,6 +1189,8 @@ class LLM:
/,
*,
use_tqdm: Union[bool, Callable[..., tqdm]] = True,
pooling_params: Optional[Union[PoolingParams,
Sequence[PoolingParams]]] = None,
lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
) -> list[ClassificationRequestOutput]:
"""
@@ -1207,7 +1209,8 @@ class LLM:
it is used to create the progress bar.
If `False`, no progress bar is created.
lora_request: LoRA request to use for generation, if any.
pooling_params: The pooling parameters for pooling. If None, we
use the default pooling parameters.
Returns:
A list of `ClassificationRequestOutput` objects containing the
embedding vectors in the same order as the input prompts.
@@ -1220,6 +1223,7 @@ class LLM:
items = self.encode(
prompts,
use_tqdm=use_tqdm,
pooling_params=pooling_params,
lora_request=lora_request,
pooling_task="classify",
)
@@ -1272,6 +1276,7 @@ class LLM:
text_2: list[Union[str, TextPrompt, TokensPrompt]],
truncate_prompt_tokens: Optional[int] = None,
use_tqdm: Union[bool, Callable[..., tqdm]] = True,
pooling_params: Optional[PoolingParams] = None,
lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
) -> list[ScoringRequestOutput]:
@@ -1280,6 +1285,7 @@ class LLM:
truncate_prompt_tokens=truncate_prompt_tokens,
use_tqdm=use_tqdm,
lora_request=lora_request,
pooling_params=pooling_params,
pooling_task="embed",
)
@@ -1306,6 +1312,7 @@ class LLM:
data_2: Union[list[str], list[ScoreContentPartParam]],
truncate_prompt_tokens: Optional[int] = None,
use_tqdm: Union[bool, Callable[..., tqdm]] = True,
pooling_params: Optional[PoolingParams] = None,
lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
) -> list[ScoringRequestOutput]:
model_config = self.llm_engine.model_config
@@ -1317,7 +1324,12 @@ class LLM:
if len(data_1) == 1:
data_1 = data_1 * len(data_2)
pooling_params = PoolingParams(task="score")
if pooling_params is None:
pooling_params = PoolingParams(task="score")
model_config = self.llm_engine.model_config
pooling_params.verify("score", model_config)
tokenization_kwargs: dict[str, Any] = {}
_validate_truncation_size(model_config.max_model_len,
@@ -1379,6 +1391,7 @@ class LLM:
*,
truncate_prompt_tokens: Optional[int] = None,
use_tqdm: Union[bool, Callable[..., tqdm]] = True,
pooling_params: Optional[PoolingParams] = None,
lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
) -> list[ScoringRequestOutput]:
"""Generate similarity scores for all pairs `<text,text_pair>` or
@@ -1410,7 +1423,8 @@ class LLM:
it is used to create the progress bar.
If `False`, no progress bar is created.
lora_request: LoRA request to use for generation, if any.
pooling_params: The pooling parameters for pooling. If None, we
use the default pooling parameters.
Returns:
A list of `ScoringRequestOutput` objects containing the
generated scores in the same order as the input prompts.
@@ -1494,6 +1508,7 @@ class LLM:
data_2, # type: ignore[arg-type]
truncate_prompt_tokens,
use_tqdm,
pooling_params,
lora_request)
else:
return self._embedding_score(
@@ -1502,6 +1517,7 @@ class LLM:
data_2, # type: ignore[arg-type]
truncate_prompt_tokens,
use_tqdm,
pooling_params,
lora_request)
def start_profile(self) -> None:
+15 -5
View File
@@ -1274,11 +1274,13 @@ class EmbeddingCompletionRequest(OpenAIBaseModel):
"not set it, a random_uuid will be generated. This id is used "
"through out the inference process and return in response."),
)
normalize: Optional[bool] = None
# --8<-- [end:embedding-extra-params]
def to_pooling_params(self):
return PoolingParams(dimensions=self.dimensions)
return PoolingParams(dimensions=self.dimensions,
normalize=self.normalize)
class EmbeddingChatRequest(OpenAIBaseModel):
@@ -1332,6 +1334,7 @@ class EmbeddingChatRequest(OpenAIBaseModel):
"not set it, a random_uuid will be generated. This id is used "
"through out the inference process and return in response."),
)
normalize: Optional[bool] = None
# --8<-- [end:chat-embedding-extra-params]
@model_validator(mode="before")
@@ -1344,7 +1347,8 @@ class EmbeddingChatRequest(OpenAIBaseModel):
return data
def to_pooling_params(self):
return PoolingParams(dimensions=self.dimensions)
return PoolingParams(dimensions=self.dimensions,
normalize=self.normalize)
EmbeddingRequest = Union[EmbeddingCompletionRequest, EmbeddingChatRequest]
@@ -1375,10 +1379,12 @@ class ScoreRequest(OpenAIBaseModel):
"if the served model does not use priority scheduling."),
)
activation: Optional[bool] = None
# --8<-- [end:score-extra-params]
def to_pooling_params(self):
return PoolingParams()
return PoolingParams(activation=self.activation)
class RerankRequest(OpenAIBaseModel):
@@ -1403,10 +1409,12 @@ class RerankRequest(OpenAIBaseModel):
"if the served model does not use priority scheduling."),
)
activation: Optional[bool] = None
# --8<-- [end:rerank-extra-params]
def to_pooling_params(self):
return PoolingParams()
return PoolingParams(activation=self.activation)
class RerankDocument(BaseModel):
@@ -1553,10 +1561,12 @@ class ClassificationRequest(OpenAIBaseModel):
"if the served model does not use priority scheduling."),
)
activation: Optional[bool] = None
# --8<-- [end:classification-extra-params]
def to_pooling_params(self):
return PoolingParams()
return PoolingParams(activation=self.activation)
class ClassificationData(OpenAIBaseModel):