Module pooling 

{arch}.pooling_type: how a sequence of hidden states becomes one embedding vector.

This is llama.cpp’s enum llama_pooling_type and the switch in llm_graph_context::build_pooling (src/llama-graph.cpp), transcribed. The key is written by the HF converter (gguf_writer.add_pooling_type) as a uint32 holding that enum’s value, so the wire numbers below are load-bearing and are not ferrox’s own invention.

Its own module rather than a section of crate::bert_encoder because pooling is not a BERT fact: llama-embed and gemma-embedding are decoders that carry the same key, and /v1/embeddings needs to honour it for whatever produced the hidden states.

What is implemented, and what refuses

NONE, MEAN, CLS and LAST are here. RANK is not, and never will be: it is not a pooling rule at all but a classification head — upstream runs cls/cls_out matrices, a tanh, and an optional head norm over the pooled row, and reports the result through /v1/rerank against classifier.output_labels. That head is crate::rank_head and that route exists, but neither is reachable from here: pool sees hidden states and a width, so the head’s matrices are not a thing it could apply. So PoolingType::Rank parses (which is how the refusal can name it) and pool returns PoolingError::Unimplemented rather than quietly handing back a CLS row that means something else. An /v1/embeddings request against a reranker checkpoint refuses here, deliberately, and that is what sends the caller to /v1/rerank.

Enums

PoolingError

PoolingType

enum llama_pooling_type, by its wire values.

Functions

l2_normalize

L2-normalizes in place, which is what every BGE/E5/GTE consumer expects of an embedding and what common_embd_normalize’s default (p == 2) does. A zero vector is left alone, exactly as upstream leaves it (it divides by norm > 0 ? 1/norm : 0, i.e. it zeroes — and a zero vector is already zero).

pool

Pools hidden (n_tokens rows of n_embd floats, in row order) down to one vector — except for PoolingType::None, which returns every row unchanged.