@novastera-oss/llamarn 0.0.1-alpha.4

package/ios/generated/RNLlamaCppSpecJSI-generated.cpp

@@ -0,0 +1,42 @@
+/**
+ * This code was generated by [react-native-codegen](https://www.npmjs.com/package/react-native-codegen).
+ *
+ * Do not edit this file as changes may cause incorrect behavior and will be lost
+ * once the code is regenerated.
+ *
+ * @generated by codegen project: GenerateModuleCpp.js
+ */
+
+#include "RNLlamaCppSpecJSI.h"
+
+namespace facebook::react {
+
+static jsi::Value __hostFunction_NativeRNLlamaCppCxxSpecJSI_multiply(jsi::Runtime &rt, TurboModule &turboModule, const jsi::Value* args, size_t count) {
+  return static_cast<NativeRNLlamaCppCxxSpecJSI *>(&turboModule)->multiply(
+    rt,
+    count <= 0 ? throw jsi::JSError(rt, "Expected argument in position 0 to be passed") : args[0].asNumber(),
+    count <= 1 ? throw jsi::JSError(rt, "Expected argument in position 1 to be passed") : args[1].asNumber()
+  );
+}
+static jsi::Value __hostFunction_NativeRNLlamaCppCxxSpecJSI_initLlama(jsi::Runtime &rt, TurboModule &turboModule, const jsi::Value* args, size_t count) {
+  return static_cast<NativeRNLlamaCppCxxSpecJSI *>(&turboModule)->initLlama(
+    rt,
+    count <= 0 ? throw jsi::JSError(rt, "Expected argument in position 0 to be passed") : args[0].asObject(rt)
+  );
+}
+static jsi::Value __hostFunction_NativeRNLlamaCppCxxSpecJSI_loadLlamaModelInfo(jsi::Runtime &rt, TurboModule &turboModule, const jsi::Value* args, size_t count) {
+  return static_cast<NativeRNLlamaCppCxxSpecJSI *>(&turboModule)->loadLlamaModelInfo(
+    rt,
+    count <= 0 ? throw jsi::JSError(rt, "Expected argument in position 0 to be passed") : args[0].asString(rt)
+  );
+}
+
+NativeRNLlamaCppCxxSpecJSI::NativeRNLlamaCppCxxSpecJSI(std::shared_ptr<CallInvoker> jsInvoker)
+  : TurboModule("RNLlamaCpp", jsInvoker) {
+  methodMap_["multiply"] = MethodMetadata {2, __hostFunction_NativeRNLlamaCppCxxSpecJSI_multiply};
+  methodMap_["initLlama"] = MethodMetadata {1, __hostFunction_NativeRNLlamaCppCxxSpecJSI_initLlama};
+  methodMap_["loadLlamaModelInfo"] = MethodMetadata {1, __hostFunction_NativeRNLlamaCppCxxSpecJSI_loadLlamaModelInfo};
+}
+
+
+} // namespace facebook::react

package/ios/generated/RNLlamaCppSpecJSI.h

@@ -0,0 +1,336 @@
+/**
+ * This code was generated by [react-native-codegen](https://www.npmjs.com/package/react-native-codegen).
+ *
+ * Do not edit this file as changes may cause incorrect behavior and will be lost
+ * once the code is regenerated.
+ *
+ * @generated by codegen project: GenerateModuleH.js
+ */
+
+#pragma once
+
+#include <ReactCommon/TurboModule.h>
+#include <react/bridging/Bridging.h>
+
+namespace facebook::react {
+
+
+  
+#pragma mark - NativeRNLlamaCppLlamaModelParams
+
+template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9, typename P10, typename P11, typename P12, typename P13, typename P14, typename P15, typename P16, typename P17, typename P18, typename P19, typename P20, typename P21, typename P22, typename P23>
+struct NativeRNLlamaCppLlamaModelParams {
+  P0 model;
+  P1 n_ctx;
+  P2 n_batch;
+  P3 n_ubatch;
+  P4 n_threads;
+  P5 n_keep;
+  P6 n_gpu_layers;
+  P7 use_mmap;
+  P8 use_mlock;
+  P9 vocab_only;
+  P10 embedding;
+  P11 seed;
+  P12 rope_freq_base;
+  P13 rope_freq_scale;
+  P14 yarn_ext_factor;
+  P15 yarn_attn_factor;
+  P16 yarn_beta_fast;
+  P17 yarn_beta_slow;
+  P18 logits_all;
+  P19 chat_template;
+  P20 use_jinja;
+  P21 verbose;
+  P22 lora_adapters;
+  P23 grammar;
+  bool operator==(const NativeRNLlamaCppLlamaModelParams &other) const {
+    return model == other.model && n_ctx == other.n_ctx && n_batch == other.n_batch && n_ubatch == other.n_ubatch && n_threads == other.n_threads && n_keep == other.n_keep && n_gpu_layers == other.n_gpu_layers && use_mmap == other.use_mmap && use_mlock == other.use_mlock && vocab_only == other.vocab_only && embedding == other.embedding && seed == other.seed && rope_freq_base == other.rope_freq_base && rope_freq_scale == other.rope_freq_scale && yarn_ext_factor == other.yarn_ext_factor && yarn_attn_factor == other.yarn_attn_factor && yarn_beta_fast == other.yarn_beta_fast && yarn_beta_slow == other.yarn_beta_slow && logits_all == other.logits_all && chat_template == other.chat_template && use_jinja == other.use_jinja && verbose == other.verbose && lora_adapters == other.lora_adapters && grammar == other.grammar;
+  }
+};
+
+template <typename T>
+struct NativeRNLlamaCppLlamaModelParamsBridging {
+  static T types;
+
+  static T fromJs(
+      jsi::Runtime &rt,
+      const jsi::Object &value,
+      const std::shared_ptr<CallInvoker> &jsInvoker) {
+    T result{
+      bridging::fromJs<decltype(types.model)>(rt, value.getProperty(rt, "model"), jsInvoker),
+      bridging::fromJs<decltype(types.n_ctx)>(rt, value.getProperty(rt, "n_ctx"), jsInvoker),
+      bridging::fromJs<decltype(types.n_batch)>(rt, value.getProperty(rt, "n_batch"), jsInvoker),
+      bridging::fromJs<decltype(types.n_ubatch)>(rt, value.getProperty(rt, "n_ubatch"), jsInvoker),
+      bridging::fromJs<decltype(types.n_threads)>(rt, value.getProperty(rt, "n_threads"), jsInvoker),
+      bridging::fromJs<decltype(types.n_keep)>(rt, value.getProperty(rt, "n_keep"), jsInvoker),
+      bridging::fromJs<decltype(types.n_gpu_layers)>(rt, value.getProperty(rt, "n_gpu_layers"), jsInvoker),
+      bridging::fromJs<decltype(types.use_mmap)>(rt, value.getProperty(rt, "use_mmap"), jsInvoker),
+      bridging::fromJs<decltype(types.use_mlock)>(rt, value.getProperty(rt, "use_mlock"), jsInvoker),
+      bridging::fromJs<decltype(types.vocab_only)>(rt, value.getProperty(rt, "vocab_only"), jsInvoker),
+      bridging::fromJs<decltype(types.embedding)>(rt, value.getProperty(rt, "embedding"), jsInvoker),
+      bridging::fromJs<decltype(types.seed)>(rt, value.getProperty(rt, "seed"), jsInvoker),
+      bridging::fromJs<decltype(types.rope_freq_base)>(rt, value.getProperty(rt, "rope_freq_base"), jsInvoker),
+      bridging::fromJs<decltype(types.rope_freq_scale)>(rt, value.getProperty(rt, "rope_freq_scale"), jsInvoker),
+      bridging::fromJs<decltype(types.yarn_ext_factor)>(rt, value.getProperty(rt, "yarn_ext_factor"), jsInvoker),
+      bridging::fromJs<decltype(types.yarn_attn_factor)>(rt, value.getProperty(rt, "yarn_attn_factor"), jsInvoker),
+      bridging::fromJs<decltype(types.yarn_beta_fast)>(rt, value.getProperty(rt, "yarn_beta_fast"), jsInvoker),
+      bridging::fromJs<decltype(types.yarn_beta_slow)>(rt, value.getProperty(rt, "yarn_beta_slow"), jsInvoker),
+      bridging::fromJs<decltype(types.logits_all)>(rt, value.getProperty(rt, "logits_all"), jsInvoker),
+      bridging::fromJs<decltype(types.chat_template)>(rt, value.getProperty(rt, "chat_template"), jsInvoker),
+      bridging::fromJs<decltype(types.use_jinja)>(rt, value.getProperty(rt, "use_jinja"), jsInvoker),
+      bridging::fromJs<decltype(types.verbose)>(rt, value.getProperty(rt, "verbose"), jsInvoker),
+      bridging::fromJs<decltype(types.lora_adapters)>(rt, value.getProperty(rt, "lora_adapters"), jsInvoker),
+      bridging::fromJs<decltype(types.grammar)>(rt, value.getProperty(rt, "grammar"), jsInvoker)};
+    return result;
+  }
+
+#ifdef DEBUG
+  static jsi::String modelToJs(jsi::Runtime &rt, decltype(types.model) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_ctxToJs(jsi::Runtime &rt, decltype(types.n_ctx) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_batchToJs(jsi::Runtime &rt, decltype(types.n_batch) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_ubatchToJs(jsi::Runtime &rt, decltype(types.n_ubatch) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_threadsToJs(jsi::Runtime &rt, decltype(types.n_threads) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_keepToJs(jsi::Runtime &rt, decltype(types.n_keep) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double n_gpu_layersToJs(jsi::Runtime &rt, decltype(types.n_gpu_layers) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool use_mmapToJs(jsi::Runtime &rt, decltype(types.use_mmap) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool use_mlockToJs(jsi::Runtime &rt, decltype(types.use_mlock) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool vocab_onlyToJs(jsi::Runtime &rt, decltype(types.vocab_only) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool embeddingToJs(jsi::Runtime &rt, decltype(types.embedding) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double seedToJs(jsi::Runtime &rt, decltype(types.seed) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double rope_freq_baseToJs(jsi::Runtime &rt, decltype(types.rope_freq_base) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double rope_freq_scaleToJs(jsi::Runtime &rt, decltype(types.rope_freq_scale) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double yarn_ext_factorToJs(jsi::Runtime &rt, decltype(types.yarn_ext_factor) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double yarn_attn_factorToJs(jsi::Runtime &rt, decltype(types.yarn_attn_factor) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double yarn_beta_fastToJs(jsi::Runtime &rt, decltype(types.yarn_beta_fast) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double yarn_beta_slowToJs(jsi::Runtime &rt, decltype(types.yarn_beta_slow) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool logits_allToJs(jsi::Runtime &rt, decltype(types.logits_all) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static jsi::String chat_templateToJs(jsi::Runtime &rt, decltype(types.chat_template) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static bool use_jinjaToJs(jsi::Runtime &rt, decltype(types.use_jinja) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static double verboseToJs(jsi::Runtime &rt, decltype(types.verbose) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static jsi::Array lora_adaptersToJs(jsi::Runtime &rt, decltype(types.lora_adapters) value) {
+    return bridging::toJs(rt, value);
+  }
+
+  static jsi::String grammarToJs(jsi::Runtime &rt, decltype(types.grammar) value) {
+    return bridging::toJs(rt, value);
+  }
+#endif
+
+  static jsi::Object toJs(
+      jsi::Runtime &rt,
+      const T &value,
+      const std::shared_ptr<CallInvoker> &jsInvoker) {
+    auto result = facebook::jsi::Object(rt);
+    result.setProperty(rt, "model", bridging::toJs(rt, value.model, jsInvoker));
+    if (value.n_ctx) {
+      result.setProperty(rt, "n_ctx", bridging::toJs(rt, value.n_ctx.value(), jsInvoker));
+    }
+    if (value.n_batch) {
+      result.setProperty(rt, "n_batch", bridging::toJs(rt, value.n_batch.value(), jsInvoker));
+    }
+    if (value.n_ubatch) {
+      result.setProperty(rt, "n_ubatch", bridging::toJs(rt, value.n_ubatch.value(), jsInvoker));
+    }
+    if (value.n_threads) {
+      result.setProperty(rt, "n_threads", bridging::toJs(rt, value.n_threads.value(), jsInvoker));
+    }
+    if (value.n_keep) {
+      result.setProperty(rt, "n_keep", bridging::toJs(rt, value.n_keep.value(), jsInvoker));
+    }
+    if (value.n_gpu_layers) {
+      result.setProperty(rt, "n_gpu_layers", bridging::toJs(rt, value.n_gpu_layers.value(), jsInvoker));
+    }
+    if (value.use_mmap) {
+      result.setProperty(rt, "use_mmap", bridging::toJs(rt, value.use_mmap.value(), jsInvoker));
+    }
+    if (value.use_mlock) {
+      result.setProperty(rt, "use_mlock", bridging::toJs(rt, value.use_mlock.value(), jsInvoker));
+    }
+    if (value.vocab_only) {
+      result.setProperty(rt, "vocab_only", bridging::toJs(rt, value.vocab_only.value(), jsInvoker));
+    }
+    if (value.embedding) {
+      result.setProperty(rt, "embedding", bridging::toJs(rt, value.embedding.value(), jsInvoker));
+    }
+    if (value.seed) {
+      result.setProperty(rt, "seed", bridging::toJs(rt, value.seed.value(), jsInvoker));
+    }
+    if (value.rope_freq_base) {
+      result.setProperty(rt, "rope_freq_base", bridging::toJs(rt, value.rope_freq_base.value(), jsInvoker));
+    }
+    if (value.rope_freq_scale) {
+      result.setProperty(rt, "rope_freq_scale", bridging::toJs(rt, value.rope_freq_scale.value(), jsInvoker));
+    }
+    if (value.yarn_ext_factor) {
+      result.setProperty(rt, "yarn_ext_factor", bridging::toJs(rt, value.yarn_ext_factor.value(), jsInvoker));
+    }
+    if (value.yarn_attn_factor) {
+      result.setProperty(rt, "yarn_attn_factor", bridging::toJs(rt, value.yarn_attn_factor.value(), jsInvoker));
+    }
+    if (value.yarn_beta_fast) {
+      result.setProperty(rt, "yarn_beta_fast", bridging::toJs(rt, value.yarn_beta_fast.value(), jsInvoker));
+    }
+    if (value.yarn_beta_slow) {
+      result.setProperty(rt, "yarn_beta_slow", bridging::toJs(rt, value.yarn_beta_slow.value(), jsInvoker));
+    }
+    if (value.logits_all) {
+      result.setProperty(rt, "logits_all", bridging::toJs(rt, value.logits_all.value(), jsInvoker));
+    }
+    if (value.chat_template) {
+      result.setProperty(rt, "chat_template", bridging::toJs(rt, value.chat_template.value(), jsInvoker));
+    }
+    if (value.use_jinja) {
+      result.setProperty(rt, "use_jinja", bridging::toJs(rt, value.use_jinja.value(), jsInvoker));
+    }
+    if (value.verbose) {
+      result.setProperty(rt, "verbose", bridging::toJs(rt, value.verbose.value(), jsInvoker));
+    }
+    if (value.lora_adapters) {
+      result.setProperty(rt, "lora_adapters", bridging::toJs(rt, value.lora_adapters.value(), jsInvoker));
+    }
+    if (value.grammar) {
+      result.setProperty(rt, "grammar", bridging::toJs(rt, value.grammar.value(), jsInvoker));
+    }
+    return result;
+  }
+};
+
+class JSI_EXPORT NativeRNLlamaCppCxxSpecJSI : public TurboModule {
+protected:
+  NativeRNLlamaCppCxxSpecJSI(std::shared_ptr<CallInvoker> jsInvoker);
+
+public:
+  virtual double multiply(jsi::Runtime &rt, double a, double b) = 0;
+  virtual jsi::Value initLlama(jsi::Runtime &rt, jsi::Object params) = 0;
+  virtual jsi::Value loadLlamaModelInfo(jsi::Runtime &rt, jsi::String modelPath) = 0;
+
+};
+
+template <typename T>
+class JSI_EXPORT NativeRNLlamaCppCxxSpec : public TurboModule {
+public:
+  jsi::Value create(jsi::Runtime &rt, const jsi::PropNameID &propName) override {
+    return delegate_.create(rt, propName);
+  }
+
+  std::vector<jsi::PropNameID> getPropertyNames(jsi::Runtime& runtime) override {
+    return delegate_.getPropertyNames(runtime);
+  }
+
+  static constexpr std::string_view kModuleName = "RNLlamaCpp";
+
+protected:
+  NativeRNLlamaCppCxxSpec(std::shared_ptr<CallInvoker> jsInvoker)
+    : TurboModule(std::string{NativeRNLlamaCppCxxSpec::kModuleName}, jsInvoker),
+      delegate_(reinterpret_cast<T*>(this), jsInvoker) {}
+
+
+private:
+  class Delegate : public NativeRNLlamaCppCxxSpecJSI {
+  public:
+    Delegate(T *instance, std::shared_ptr<CallInvoker> jsInvoker) :
+      NativeRNLlamaCppCxxSpecJSI(std::move(jsInvoker)), instance_(instance) {
+
+    }
+
+    double multiply(jsi::Runtime &rt, double a, double b) override {
+      static_assert(
+          bridging::getParameterCount(&T::multiply) == 3,
+          "Expected multiply(...) to have 3 parameters");
+
+      return bridging::callFromJs<double>(
+          rt, &T::multiply, jsInvoker_, instance_, std::move(a), std::move(b));
+    }
+    jsi::Value initLlama(jsi::Runtime &rt, jsi::Object params) override {
+      static_assert(
+          bridging::getParameterCount(&T::initLlama) == 2,
+          "Expected initLlama(...) to have 2 parameters");
+
+      return bridging::callFromJs<jsi::Value>(
+          rt, &T::initLlama, jsInvoker_, instance_, std::move(params));
+    }
+    jsi::Value loadLlamaModelInfo(jsi::Runtime &rt, jsi::String modelPath) override {
+      static_assert(
+          bridging::getParameterCount(&T::loadLlamaModelInfo) == 2,
+          "Expected loadLlamaModelInfo(...) to have 2 parameters");
+
+      return bridging::callFromJs<jsi::Value>(
+          rt, &T::loadLlamaModelInfo, jsInvoker_, instance_, std::move(modelPath));
+    }
+
+  private:
+    friend class NativeRNLlamaCppCxxSpec;
+    T *instance_;
+  };
+
+  Delegate delegate_;
+};
+
+} // namespace facebook::react

package/ios/include/chat.h

@@ -0,0 +1,135 @@
+// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
+
+#pragma once
+
+#include "common.h"
+#include <string>
+#include <vector>
+
+struct common_chat_templates;
+
+struct common_chat_tool_call {
+    std::string name;
+    std::string arguments;
+    std::string id;
+};
+
+struct common_chat_msg_content_part {
+    std::string type;
+    std::string text;
+};
+
+struct common_chat_msg {
+    std::string role;
+    std::string content;
+    std::vector<common_chat_msg_content_part> content_parts = {};
+    std::vector<common_chat_tool_call> tool_calls = {};
+    std::string reasoning_content;
+    std::string tool_name;
+    std::string tool_call_id;
+};
+
+struct common_chat_tool {
+    std::string name;
+    std::string description;
+    std::string parameters;
+};
+
+enum common_chat_tool_choice {
+    COMMON_CHAT_TOOL_CHOICE_AUTO,
+    COMMON_CHAT_TOOL_CHOICE_REQUIRED,
+    COMMON_CHAT_TOOL_CHOICE_NONE,
+};
+
+enum common_chat_format {
+    COMMON_CHAT_FORMAT_CONTENT_ONLY,
+    COMMON_CHAT_FORMAT_GENERIC,
+    COMMON_CHAT_FORMAT_MISTRAL_NEMO,
+    COMMON_CHAT_FORMAT_LLAMA_3_X,
+    COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
+    COMMON_CHAT_FORMAT_DEEPSEEK_R1,
+    COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING,
+    COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
+    COMMON_CHAT_FORMAT_HERMES_2_PRO,
+    COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING,
+    COMMON_CHAT_FORMAT_COMMAND_R7B,
+    COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
+
+    COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
+};
+
+struct common_chat_templates_inputs {
+    std::vector<common_chat_msg> messages;
+    std::string grammar;
+    std::string json_schema;
+    bool add_generation_prompt = true;
+    bool use_jinja = true;
+    // Parameters below only supported when use_jinja is true
+    std::vector<common_chat_tool> tools;
+    common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
+    bool parallel_tool_calls = false;
+    bool extract_reasoning     = true;
+};
+
+struct common_chat_params {
+    common_chat_format                  format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    std::string                         prompt;
+    std::string                         grammar;
+    bool                                grammar_lazy = false;
+    std::vector<common_grammar_trigger> grammar_triggers;
+    std::vector<std::string>            preserved_tokens;
+    std::vector<std::string>            additional_stops;
+};
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
+
+void common_chat_templates_free(struct common_chat_templates * tmpls);
+
+struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } };
+
+typedef std::unique_ptr<struct common_chat_templates, common_chat_templates_deleter> common_chat_templates_ptr;
+
+common_chat_templates_ptr common_chat_templates_init(
+                                    const struct llama_model * model,
+                                           const std::string & chat_template_override,
+                                           const std::string & bos_token_override = "",
+                                           const std::string & eos_token_override = "");
+
+bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
+const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
+
+
+struct common_chat_params      common_chat_templates_apply(
+    const struct common_chat_templates * tmpls,
+    const struct common_chat_templates_inputs & inputs);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string common_chat_format_single(
+        const struct common_chat_templates * tmpls,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
+        bool add_ass,
+        bool use_jinja);
+
+// Returns an example of formatted chat
+std::string common_chat_format_example(
+    const struct common_chat_templates * tmpls,
+    bool use_jinja);
+
+std::string               common_chat_format_name(common_chat_format format);
+common_chat_msg           common_chat_parse(      const std::string & input, common_chat_format format);
+
+common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
+
+// Parses a JSON array of messages in OpenAI's chat completion API format.
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
+template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+
+// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
+template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);