Skip to main content
The PoolingParams class controls how vLLM performs pooling operations for embeddings, classification, and scoring tasks.

​
Constructor

from vllm import PoolingParams

pooling_params = PoolingParams(
    use_activation=True,
    dimensions=768,
)

​
Parameters

​
use_activation
bool | None
default:"None"
Whether to apply activation function to pooler outputs. None uses the model’s default (typically True).
​
dimensions
int | None
default:"None"
Reduce embedding dimensions if the model supports matryoshka representation. Only valid for embedding tasks.
​
task
str | None
default:"None"
The pooling task to perform. Should be one of:
  • "embed" - Generate embeddings
  • "classify" - Classification task
  • "score" - Scoring/ranking task
  • "token_embed" - Token-level embeddings
  • "token_classify" - Token-level classification

​
Task-specific parameters

Different pooling tasks support different parameters:

​
Embedding tasks (embed, token_embed)

  • use_activation: Whether to apply activation
  • dimensions: Output dimensionality (if model supports matryoshka)

​
Classification tasks (classify, token_classify)

  • use_activation: Whether to apply activation

​
Scoring task (score)

  • use_activation: Whether to apply activation

​
Example: Generate embeddings

from vllm import LLM, PoolingParams

# Initialize embedding model
llm = LLM(
    model="sentence-transformers/all-MiniLM-L6-v2",
    runner="pooling",
)

# Configure pooling
pooling_params = PoolingParams(
    use_activation=True,
)

# Generate embeddings
prompts = [
    "Hello world",
    "How are you?",
]

outputs = llm.embed(prompts, pooling_params=pooling_params)

for output in outputs:
    embedding = output.outputs.embedding
    print(f"Embedding dimension: {len(embedding)}")
    print(f"Embedding: {embedding[:5]}...")  # First 5 values

​
Example: Matryoshka embeddings

# Reduce embedding dimensions for a matryoshka model
pooling_params = PoolingParams(
    use_activation=True,
    dimensions=256,  # Reduce from default (e.g., 768) to 256
)

outputs = llm.embed(["Sample text"], pooling_params=pooling_params)
embedding = outputs[0].outputs.embedding
assert len(embedding) == 256

​
Example: Classification

from vllm import LLM, PoolingParams

# Initialize classification model
llm = LLM(
    model="your-classifier-model",
    runner="pooling",
)

pooling_params = PoolingParams(
    use_activation=True,
)

# Classify text
outputs = llm.classify(
    ["This movie is amazing!"],
    pooling_params=pooling_params,
)

for output in outputs:
    probs = output.outputs.probs
    print(f"Classification probabilities: {probs}")

​
Example: Scoring/Reranking

# Score query-document pairs
llm = LLM(
    model="your-reranker-model",
    runner="pooling",
)

pooling_params = PoolingParams(
    use_activation=True,
)

query = "What is machine learning?"
documents = [
    "Machine learning is a subset of AI",
    "Python is a programming language",
    "Deep learning uses neural networks",
]

# Create query-document pairs
pairs = [f"{query} [SEP] {doc}" for doc in documents]

outputs = llm.score(pairs, pooling_params=pooling_params)

for i, output in enumerate(outputs):
    score = output.outputs.score
    print(f"Document {i} score: {score}")

​
Valid parameter combinations

The PoolingParams class validates that only task-appropriate parameters are specified:
TaskValid Parameters
embeduse_activation, dimensions
classifyuse_activation
scoreuse_activation
token_embeduse_activation, dimensions
token_classifyuse_activation
Attempting to use invalid parameters for a task will raise a validation error.
  • LLM - Use PoolingParams with llm.embed(), llm.classify(), or llm.score()
  • SamplingParams - Parameters for text generation
  • Output classes - Output formats for pooling tasks

Build docs developers (and LLMs) love

Get started for freeTalk to us