grammar.hpp

Go to the documentation of this file.

#pragma once
 
// SPDX-License-Identifier: Apache-2.0
// Copyright 2026 Lloyal Labs
 
 
#include "common.hpp"
#include "tokenizer.hpp"
#include <llama/llama.h>
#include <json-schema-to-grammar.h>  // llama.cpp common library
#include <nlohmann/json.hpp>
#include <stdexcept>
#include <string>
#include <vector>
 
namespace lloyal::grammar {
 
inline std::string from_json_schema(const std::string &schema_json) {
  LLOYAL_LOG_DEBUG(
      "[grammar::from_json_schema] Converting JSON schema (%zu bytes)",
      schema_json.size());
 
  try {
    // Parse JSON schema
    nlohmann::ordered_json schema = nlohmann::ordered_json::parse(schema_json);
 
    LLOYAL_LOG_DEBUG("[grammar::from_json_schema] Schema parsed, calling "
                     "json_schema_to_grammar");
 
    // Call json_schema_to_grammar from llama.cpp common library
    // Parameters: (schema, force_gbnf)
    // force_gbnf=false allows EBNF optimization when possible
    std::string grammar = json_schema_to_grammar(schema, false);
 
    if (grammar.empty()) {
      LLOYAL_LOG_DEBUG("[grammar::from_json_schema] ERROR: Conversion produced "
                       "empty grammar");
      throw std::runtime_error("Grammar conversion produced empty result");
    }
 
    LLOYAL_LOG_DEBUG(
        "[grammar::from_json_schema] Generated GBNF grammar (%zu bytes)",
        grammar.size());
    return grammar;
 
  } catch (const nlohmann::json::parse_error &e) {
    std::string errorMsg = std::string("JSON parse error: ") + e.what();
    LLOYAL_LOG_DEBUG("[grammar::from_json_schema] ERROR: %s", errorMsg.c_str());
    throw std::runtime_error(errorMsg);
  } catch (const std::exception &e) {
    std::string errorMsg =
        std::string("Grammar conversion failed: ") + e.what();
    LLOYAL_LOG_DEBUG("[grammar::from_json_schema] ERROR: %s", errorMsg.c_str());
    throw std::runtime_error(errorMsg);
  }
}
 
// ===== SAMPLER INITIALIZATION =====
 
inline llama_sampler *init_sampler(const llama_model *model,
                                   const std::string &grammar_str,
                                   const std::string &root_rule = "root") {
  if (!model) {
    LLOYAL_LOG_DEBUG("[grammar::init_sampler] ERROR: model is null");
    return nullptr;
  }
 
  if (grammar_str.empty()) {
    LLOYAL_LOG_DEBUG("[grammar::init_sampler] ERROR: grammar_str is empty");
    return nullptr;
  }
 
  const llama_vocab *vocab = tokenizer::get_vocab(model);
  if (!vocab) {
    LLOYAL_LOG_DEBUG("[grammar::init_sampler] ERROR: get_vocab returned null");
    return nullptr;
  }
 
  LLOYAL_LOG_DEBUG("[grammar::init_sampler] Initializing grammar sampler "
                   "(grammar: %zu bytes, root: %s)",
                   grammar_str.size(), root_rule.c_str());
 
  llama_sampler *sampler =
      llama_sampler_init_grammar(vocab, grammar_str.c_str(), root_rule.c_str());
 
  if (!sampler) {
    LLOYAL_LOG_DEBUG("[grammar::init_sampler] ERROR: "
                     "llama_sampler_init_grammar returned null");
  }
 
  return sampler;
}
 
inline llama_sampler *init_lazy_sampler(
    const llama_model *model,
    const std::string &grammar_str,
    const std::vector<std::string> &trigger_patterns,
    const std::vector<llama_token> &trigger_tokens,
    const std::string &root_rule = "root") {
  if (!model) {
    LLOYAL_LOG_DEBUG("[grammar::init_lazy_sampler] ERROR: model is null");
    return nullptr;
  }
 
  if (grammar_str.empty()) {
    LLOYAL_LOG_DEBUG("[grammar::init_lazy_sampler] ERROR: grammar_str is empty");
    return nullptr;
  }
 
  const llama_vocab *vocab = tokenizer::get_vocab(model);
  if (!vocab) {
    LLOYAL_LOG_DEBUG("[grammar::init_lazy_sampler] ERROR: get_vocab returned null");
    return nullptr;
  }
 
  std::vector<const char *> patterns_c;
  patterns_c.reserve(trigger_patterns.size());
  for (const auto &p : trigger_patterns) patterns_c.push_back(p.c_str());
 
  LLOYAL_LOG_DEBUG("[grammar::init_lazy_sampler] Initializing lazy grammar "
                   "(grammar: %zu bytes, %zu patterns, %zu tokens, root: %s)",
                   grammar_str.size(), trigger_patterns.size(),
                   trigger_tokens.size(), root_rule.c_str());
 
  llama_sampler *sampler = llama_sampler_init_grammar_lazy_patterns(
      vocab, grammar_str.c_str(), root_rule.c_str(),
      patterns_c.data(), patterns_c.size(),
      trigger_tokens.data(), trigger_tokens.size());
 
  if (!sampler) {
    LLOYAL_LOG_DEBUG("[grammar::init_lazy_sampler] ERROR: "
                     "llama_sampler_init_grammar_lazy_patterns returned null");
  }
 
  return sampler;
}
 
inline llama_sampler *clone_sampler(llama_sampler *smpl) {
  if (!smpl) {
    LLOYAL_LOG_DEBUG("[grammar::clone_sampler] Input is null, returning null");
    return nullptr;
  }
 
  llama_sampler *cloned = llama_sampler_clone(smpl);
 
  if (!cloned) {
    LLOYAL_LOG_DEBUG("[grammar::clone_sampler] ERROR: llama_sampler_clone failed");
  } else {
    LLOYAL_LOG_DEBUG("[grammar::clone_sampler] Cloned sampler successfully");
  }
 
  return cloned;
}
 
inline void free_sampler(llama_sampler* smpl) {
  if (smpl) {
    llama_sampler_free(smpl);
  }
}
 
inline void apply(llama_sampler* smpl, llama_token_data_array* cur_p) {
  if (smpl && cur_p) {
    llama_sampler_apply(smpl, cur_p);
  }
}
 
inline void accept(llama_sampler* smpl, llama_token token) {
  if (smpl) {
    llama_sampler_accept(smpl, token);
  }
}
 
} // namespace lloyal::grammar