Implementation:Ggml org Llama cpp Lookup Stats

Knowledge Sources	Ggml_org_Llama_cpp
Domains	Speculative_Decoding, Analysis
Last Updated	2026-02-15 00:00 GMT

Overview

Evaluates the effectiveness of prompt lookup decoding by computing acceptance statistics for drafted tokens against the actual model output.

Description

Loads static and dynamic n-gram caches, tokenizes the input, then processes the tokens in context-sized chunks simulating sequential generation. For each position, it drafts candidate tokens using n-gram lookup and compares them against the known-correct next tokens, tracking the number of drafted and accepted tokens. Reports overall acceptance rates and timing statistics for the draft process.

Usage

Use this diagnostic tool to evaluate how well prompt lookup decoding would perform for a specific text and model combination, guiding parameter tuning and cache construction strategies.

Code Reference

Source Location

• Repository: Ggml_org_Llama_cpp
• File: examples/lookup/lookup-stats.cpp
• Lines: 1-157

Signature

int main(int argc, char ** argv);

Import

#include "arg.h"
#include "common.h"
#include "log.h"
#include "ngram-cache.h"
#include "llama.h"
#include "ggml.h"

#include <cstdint>
#include <cstdio>
#include <cinttypes>
#include <fstream>
#include <string>
#include <vector>

I/O Contract

Inputs

Name	Type	Required	Description
-m	string	Yes	Path to the GGUF model file
-p	string	Yes	Prompt text to evaluate lookup decoding against
--lookup-cache-static	string	No	Path to a static n-gram cache file
--lookup-cache-dynamic	string	No	Path to a dynamic n-gram cache file
-n	int	No	Maximum number of draft tokens per step

Outputs

Name	Type	Description
stdout	text	Acceptance rate statistics: number of drafted tokens, accepted tokens, and timing
return	int	Exit code: 0 on success, 1 on failure

Usage Examples

# Evaluate lookup stats with static cache
./build/bin/llama-lookup-stats \
  -m model.gguf \
  -p "Your input text here..." \
  --lookup-cache-static static_cache.bin

Related Pages

• Principle:Ggml_org_Llama_cpp_Speculative_Decoding