Implementation:Ollama Ollama Llama Model T5 Encoder

Knowledge Sources	Ollama
Domains	LLM Inference, Model Architecture
Last Updated	2025-02-15 00:00 GMT

Overview

Implements the ggml computation graph builder for the T5 encoder architecture, fundamentally different from the decoder-only models that dominate the codebase.

Description

The llm_build_t5_enc constructor builds a T5 encoder graph that uses relative position bias (via bucketed position embeddings from build_inp_pos_bucket_enc) instead of RoPE, no-cache attention (since the encoder processes the full input at once via build_attn_inp_no_cache), encoder-specific weight matrices (wq_enc, wk_enc, wv_enc, wo_enc), and RMS normalization. Produces encoder output embeddings rather than token logits.

Usage

Enables Ollama to run T5 encoder-decoder models by implementing the encoder half of the T5 architecture.

Code Reference

Source Location

• Repository: Ollama
• File: llama/llama.cpp/src/models/t5-enc.cpp
• Lines: 1-96

Signature

llm_build_t5_enc::llm_build_t5_enc(
    const llama_model & model,
    const llm_graph_params & params) : llm_graph_context(params);

Import

#include "models.h"

I/O Contract

Inputs

Name	Type	Required	Description
model	const llama_model &	Yes	Loaded model with T5 encoder weights
params	const llm_graph_params &	Yes	Graph construction parameters

Outputs

Name	Type	Description
ggml graph	ggml_cgraph	T5 encoder computation graph producing embeddings

Usage Examples

auto builder = llm_build_t5_enc(model, params);
// Produces encoder embeddings, not token logits

Related Pages

• Principle:Ollama_Ollama_LLM_Inference_Pipeline