Enum PoolingStrategy 

pub enum PoolingStrategy {
    MeanVisualTokens,
    LastToken,
}

How to collapse a prefill’s per-position hidden states into one fixed-size embedding vector.

Retrieval quality with the base Qwen3.5-0.8B instruct checkpoint is unvalidated. GME-style pooled embeddings normally come from a checkpoint that has been contrastively fine-tuned for retrieval (image-text matching, hard-negative mining); the base instruct checkpoint was never trained for that objective. What this module provides — and what is tested — is the extraction machinery: pooling over the verifiably correct positions, deterministically, into a unit-norm vector. Picking (or fine-tuning) a checkpoint for retrieval quality is a separate, later decision.

Variants

MeanVisualTokens

Mean over the hidden states at the request’s <|image_pad|> positions (the visual tokens). For a text-only request — no image runs, so no pad tokens are present — this degrades to a mean over every position, i.e. an ordinary mean-pooled text embedding rather than an error.

LastToken

The hidden state at the last physical position — GME’s text-embedding convention. Well-defined for both image and text-only requests.

Trait Implementations

impl Clone for PoolingStrategy

fn clone(&self) -> PoolingStrategy

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for PoolingStrategy

impl Debug for PoolingStrategy

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Eq for PoolingStrategy

impl PartialEq for PoolingStrategy

fn eq(&self, other: &PoolingStrategy) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for PoolingStrategy