Implementation:Ollama Ollama Llama Grammar Types

Knowledge Sources	Ollama
Domains	Grammar, Constrained Generation
Last Updated	2025-02-15 00:00 GMT

Overview

Header declaring the grammar engine types: grammar element types, grammar state structures, GBNF parser, the Ollama-specific vocabulary wrapper, and the internal grammar API.

Description

Defines the llama_gretype enum for grammar element types (end, alt, rule ref, char, char range, char any, token). Declares ollama_vocab struct with token-to-piece mapping and EOG token tracking for Ollama's custom grammar support. Defines llama_grammar_element for individual grammar rules, llama_partial_utf8 for partial UTF-8 state, llama_grammar_candidate for token candidates during filtering, and llama_grammar_parser for parsing GBNF text into rule structures. The llama_grammar struct holds grammar stacks, rules, vocabulary reference, lazy evaluation state, and trigger patterns for conditional grammar activation.

Usage

Include this header when implementing grammar-constrained decoding or working with the grammar engine internals. The ollama_vocab type is an Ollama-specific addition for grammar operations without a full model context.

Code Reference

Source Location

• Repository: Ollama
• File: llama/llama.cpp/src/llama-grammar.h
• Lines: 1-206

Signature

struct ollama_vocab {
    std::map<uint32_t, std::string> token_to_piece_map;
    std::set<uint32_t> special_eog_ids;
    const std::string & token_to_piece(const uint32_t token) const;
    bool is_eog(const uint32_t token) const;
};

enum llama_gretype {
    LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT, LLAMA_GRETYPE_RULE_REF,
    LLAMA_GRETYPE_CHAR, LLAMA_GRETYPE_CHAR_NOT, LLAMA_GRETYPE_CHAR_RNG_UPPER,
    LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_ANY,
    LLAMA_GRETYPE_TOKEN, LLAMA_GRETYPE_TOKEN_NOT,
};

typedef struct llama_grammar_element {
    enum llama_gretype type;
    uint32_t value;
} llama_grammar_element;

struct llama_grammar_parser {
    const llama_vocab * vocab;
    std::map<std::string, uint32_t> symbol_ids;
    llama_grammar_rules rules;
    bool parse(const char * src);
};

struct llama_grammar {
    const llama_vocab * vocab;
    const ollama_vocab * o_vocab;
    const llama_grammar_rules rules;
    llama_grammar_stacks stacks;
    llama_partial_utf8 partial_utf8;
    bool lazy = false;
    bool awaiting_trigger = false;
    std::string trigger_buffer;
    std::vector<llama_token> trigger_tokens;
    std::vector<llama_grammar_trigger_pattern> trigger_patterns;
};

Import

#include "llama-grammar.h"

I/O Contract

Inputs

Name	Type	Required	Description
src	const char *	Yes	GBNF grammar source text to parse
vocab	const llama_vocab *	No	Vocabulary for token-level grammar operations
o_vocab	const ollama_vocab *	No	Ollama vocabulary wrapper for custom grammar support

Outputs

Name	Type	Description
rules	llama_grammar_rules	Parsed grammar rules
stacks	llama_grammar_stacks	Grammar state stacks for tracking position
success	bool	Whether parsing succeeded

Usage Examples

#include "llama-grammar.h"

// Parse a grammar
llama_grammar_parser parser(vocab);
bool ok = parser.parse("root ::= \"hello\" | \"world\"");

// Access parsed rules
const auto & rules = parser.rules;

// Ollama vocab for grammar without full model
ollama_vocab o_vocab;
o_vocab.add_token_pieces(token_ids, n_tokens, pieces);
o_vocab.set_eog_tokens(eog_ids, n_eog);

// Check grammar element types
llama_grammar_element elem;
if (elem.type == LLAMA_GRETYPE_CHAR) {
    // Character match element
}

Related Pages

• Principle:Ollama_Ollama_Structured_Output