llama_cpp 0.3.3 → 0.3.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: cf337091019bb773e47cf206ff2ff30ed0bef963094494e6493455cad7c59840
-  data.tar.gz: fdbae8e08a6b87d49c5658d5c1857f20bf8efdf5a5371906630dccf4eb0f1159
+  metadata.gz: 991df3df6b16ec98a203a6c6565794988eec04697ccb963faab976f436e1bfcc
+  data.tar.gz: dafd3b8274640eb79353e11056f497a02392fef332a46dcc1717878c836f62bd
 SHA512:
-  metadata.gz: f0fee68294960c5ab9f56ebfe7256a00f9330e55f4954f2b016e07cbc023570298fa8f8b578f3e187fe9183b869769085311931122f93a033c6c21158b4e9485
-  data.tar.gz: 7eec8c98ae9ec1a56fa4bdb4e83a2dc2bdea407fc037af8d1b8f09a30c0d1246333d410707f4d66f3f473bf73574757cf12e56a86a0cb47074501f63f65f0c02
+  metadata.gz: 0a54fdf18c5be5273f01d4d991b59975a8e8b6a0a8f54087fb90df3f1a8a7ebad557d01fb119f3d61cafa8f6c59fb81641624779fda27b732eea2868cb4642e8
+  data.tar.gz: 6574064078070502e36ad933bd5efb2f479c94f47a1260286fe485e48d35a7b3985274943c6163189326891349b47b1d9815d77c600b015fe111cc3842179392

data/CHANGELOG.md

@@ -1,3 +1,34 @@
+## [[0.3.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.3.4...v0.3.5)] - 2023-07-29
+
+- Bump bundled llama.cpp from master-d924522 to master-1a94186.
+  - Add `GrammarElement` and `Grammar` classes.
+  - Add `sample_grammar` method to Context.
+  - Add `grammar_accept_token method` method to Context.
+
+## [[0.3.4](https://github.com/yoshoku/llama_cpp.rb/compare/v0.3.3...v0.3.4)] - 2023-07-23
+
+- Bump bundled llama.cpp from master-32c5411 to master-d924522.
+  - Add `rope_freq_base` and `rope_freq_scale` options to ContextParams.
+  - Add `max_devices` module function to LLaMACpp.
+  - Add `n_vocab`, `n_ctx`, and `n_embd` methods to Model.
+  - Add `vocab`, `tokenize`, and `token_to_str` methods to Model.
+  ```ruby
+  require 'llama_cpp'
+
+  params = LLaMACpp::ContextParams.new
+  model = LLaMACpp::Model.new(model_path: '/path/to/model.bin', params: params)
+
+  p model.tokenize(text: 'hello, world')
+  # => [12199, 29892, 3186]
+
+  p model.token_to_str(12199)
+  # => "hello"
+  ```
+
+**Breaking Changes**
+- Fix to automatically call `backend_free` method when Ruby script exits.
+- Remove `smooth_factor` argument from `sample_classifier_free_guidance methos` on Context.
+
 ## [[0.3.3](https://github.com/yoshoku/llama_cpp.rb/compare/v0.3.2...v0.3.3)] - 2023-07-15
 
 - Bump bundled llama.cpp from master-481f793 to master-32c5411.

data/ext/llama_cpp/extconf.rb

@@ -85,6 +85,7 @@ if with_config('mpi')
   $CXXFLAGS << ' -DGGML_USE_MPI -Wno-cast-qual'
 end
 
+# @!visibility private
 UNAME_M = RbConfig::CONFIG['build_cpu'] || RbConfig::CONFIG['host_cpu'] || RbConfig::CONFIG['target_cpu']
 
 # rubocop:disable Layout/LineLength

data/ext/llama_cpp/llama_cpp.cpp

@@ -8,6 +8,8 @@ VALUE rb_cLLaMAContextParams;
 VALUE rb_cLLaMAModelQuantizeParams;
 VALUE rb_cLLaMATokenData;
 VALUE rb_cLLaMATokenDataArray;
+VALUE rb_cLLaMAGrammarElement;
+VALUE rb_cLLaMAGrammar;
 
 class LLaMATokenDataWrapper {
 public:
@@ -406,6 +408,10 @@ public:
     rb_define_method(rb_cLLaMAContextParams, "main_gpu=", RUBY_METHOD_FUNC(_llama_context_params_set_main_gpu), 1);
     rb_define_method(rb_cLLaMAContextParams, "main_gpu", RUBY_METHOD_FUNC(_llama_context_params_get_main_gpu), 0);
     rb_define_method(rb_cLLaMAContextParams, "tensor_split", RUBY_METHOD_FUNC(_llama_context_params_get_tensor_split), 0);
+    rb_define_method(rb_cLLaMAContextParams, "rope_freq_base=", RUBY_METHOD_FUNC(_llama_context_params_set_rope_freq_base), 1);
+    rb_define_method(rb_cLLaMAContextParams, "rope_freq_base", RUBY_METHOD_FUNC(_llama_context_params_get_rope_freq_base), 0);
+    rb_define_method(rb_cLLaMAContextParams, "rope_freq_scale=", RUBY_METHOD_FUNC(_llama_context_params_set_rope_freq_scale), 1);
+    rb_define_method(rb_cLLaMAContextParams, "rope_freq_scale", RUBY_METHOD_FUNC(_llama_context_params_get_rope_freq_scale), 0);
     rb_define_method(rb_cLLaMAContextParams, "low_vram=", RUBY_METHOD_FUNC(_llama_context_params_set_low_vram), 1);
     rb_define_method(rb_cLLaMAContextParams, "low_vram", RUBY_METHOD_FUNC(_llama_context_params_get_low_vram), 0);
     rb_define_method(rb_cLLaMAContextParams, "seed=", RUBY_METHOD_FUNC(_llama_context_params_set_seed), 1);
@@ -494,6 +500,30 @@ private:
     return ret;
   }
 
+  // rope_freq_base
+  static VALUE _llama_context_params_set_rope_freq_base(VALUE self, VALUE rope_freq_base) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.rope_freq_base = NUM2DBL(rope_freq_base);
+    return DBL2NUM(ptr->params.rope_freq_base);
+  }
+
+  static VALUE _llama_context_params_get_rope_freq_base(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.rope_freq_base);
+  }
+
+  // rope_freq_scale
+  static VALUE _llama_context_params_set_rope_freq_scale(VALUE self, VALUE rope_freq_scale) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.rope_freq_scale = NUM2DBL(rope_freq_scale);
+    return DBL2NUM(ptr->params.rope_freq_scale);
+  }
+
+  static VALUE _llama_context_params_get_rope_freq_scale(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.rope_freq_scale);
+  }
+
   // low_vram
   static VALUE _llama_context_params_set_low_vram(VALUE self, VALUE low_vram) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
@@ -764,6 +794,12 @@ public:
     rb_define_method(rb_cLLaMAModel, "free", RUBY_METHOD_FUNC(_llama_model_free), 0);
     rb_define_method(rb_cLLaMAModel, "load", RUBY_METHOD_FUNC(_llama_model_load), -1);
     rb_define_method(rb_cLLaMAModel, "apply_lora_from_file", RUBY_METHOD_FUNC(_llama_model_apply_lora_from_file), -1);
+    rb_define_method(rb_cLLaMAModel, "n_vocab", RUBY_METHOD_FUNC(_llama_model_get_n_vocab_from_model), 0);
+    rb_define_method(rb_cLLaMAModel, "n_ctx", RUBY_METHOD_FUNC(_llama_model_get_n_ctx_from_model), 0);
+    rb_define_method(rb_cLLaMAModel, "n_embd", RUBY_METHOD_FUNC(_llama_model_get_n_embd_from_model), 0);
+    rb_define_method(rb_cLLaMAModel, "vocab", RUBY_METHOD_FUNC(_llama_model_get_vocab_from_model), -1);
+    rb_define_method(rb_cLLaMAModel, "token_to_str", RUBY_METHOD_FUNC(_llama_model_token_to_str_with_model), 1);
+    rb_define_method(rb_cLLaMAModel, "tokenize", RUBY_METHOD_FUNC(_llama_model_tokenize_with_model), -1);
   }
 
 private:
@@ -908,6 +944,109 @@ private:
     }
     return Qnil;
   }
+
+  static VALUE _llama_model_get_n_vocab_from_model(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_n_vocab_from_model(ptr->model));
+  }
+
+  static VALUE _llama_model_get_n_ctx_from_model(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_n_ctx_from_model(ptr->model));
+  }
+
+  static VALUE _llama_model_get_n_embd_from_model(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_n_embd_from_model(ptr->model));
+  }
+
+  static VALUE _llama_model_get_vocab_from_model(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[1] = { rb_intern("capacity") };
+    VALUE kw_values[1] = { Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
+
+    if (!RB_INTEGER_TYPE_P(kw_values[0])) {
+      rb_raise(rb_eArgError, "capacity must be an integer");
+      return Qnil;
+    }
+
+    const int capacity = NUM2INT(kw_values[0]);
+
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const int n = std::min(capacity, llama_n_vocab_from_model(ptr->model));
+    const char** vocabs = ALLOCA_N(const char*, n);
+    float* scores = ALLOCA_N(float, n);
+
+    llama_get_vocab_from_model(ptr->model, vocabs, scores, capacity);
+
+    VALUE vocabs_ary = rb_ary_new();
+    VALUE scores_ary = rb_ary_new();
+
+    for (int i = 0; i < n; i++) {
+      rb_ary_push(vocabs_ary, rb_str_new_cstr(vocabs[i]));
+      rb_ary_push(scores_ary, DBL2NUM(scores[i]));
+    }
+
+    VALUE ret = rb_ary_new3(2, vocabs_ary, scores_ary);
+
+    return ret;
+  }
+
+  static VALUE _llama_model_token_to_str_with_model(VALUE self, VALUE token_) {
+    if (!RB_INTEGER_TYPE_P(token_)) {
+      rb_raise(rb_eArgError, "token must be an integer");
+      return Qnil;
+    }
+    const llama_token token = NUM2INT(token_);
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const char* str = llama_token_to_str_with_model(ptr->model, token);
+    return rb_str_new_cstr(str);
+  }
+
+  static VALUE _llama_model_tokenize_with_model(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[3] = { rb_intern("text"), rb_intern("n_max_tokens"), rb_intern("add_bos") };
+    VALUE kw_values[3] = { Qundef, Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 1, 2, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_STRING)) {
+      rb_raise(rb_eArgError, "text must be a String");
+      return Qnil;
+    }
+    if (kw_values[1] != Qundef && !RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "n_max_tokens must be an integer");
+      return Qnil;
+    }
+    if (kw_values[2] != Qundef && (kw_values[2] != Qtrue && kw_values[2] != Qfalse)) {
+      rb_raise(rb_eArgError, "add_bos must be a boolean");
+      return Qnil;
+    }
+
+    VALUE text_ = kw_values[0];
+    std::string text = StringValueCStr(text_);
+    const bool add_bos = kw_values[2] == Qtrue ? true : false;
+    const int n_max_tokens = kw_values[1] != Qundef ? NUM2INT(kw_values[1]) : text.size() + (add_bos ? 1 : 0);
+
+    llama_token* tokens = ALLOCA_N(llama_token, n_max_tokens);
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const int n_tokens = llama_tokenize_with_model(ptr->model, text.c_str(), tokens, n_max_tokens, add_bos);
+
+    if (n_tokens < 0) {
+      rb_raise(rb_eRuntimeError, "failed to tokenize. The numebr of tokens (%d) is greater than n_max_tokens.", -n_tokens);
+      return Qnil;
+    }
+
+    VALUE ret = rb_ary_new2(n_tokens);
+    for (int i = 0; i < n_tokens; i++) {
+      rb_ary_store(ret, i, INT2NUM(tokens[i]));
+    }
+
+    RB_GC_GUARD(text_);
+    return ret;
+  }
 };
 
 const rb_data_type_t RbLLaMAModel::llama_model_type = {
@@ -920,6 +1059,222 @@ const rb_data_type_t RbLLaMAModel::llama_model_type = {
   RUBY_TYPED_FREE_IMMEDIATELY
 };
 
+class LLaMAGrammarElementWrapper {
+public:
+  llama_grammar_element element;
+
+  LLaMAGrammarElementWrapper() {
+    element.type = LLAMA_GRETYPE_END;
+    element.value = 0;
+  }
+
+  ~LLaMAGrammarElementWrapper() {}
+};
+
+class RbLLaMAGrammarElement {
+public:
+  static VALUE llama_grammar_element_alloc(VALUE self) {
+    LLaMAGrammarElementWrapper* ptr = (LLaMAGrammarElementWrapper*)ruby_xmalloc(sizeof(LLaMAGrammarElementWrapper));
+    new (ptr) LLaMAGrammarElementWrapper();
+    return TypedData_Wrap_Struct(self, &llama_grammar_element_type, ptr);
+  }
+
+  static void llama_grammar_element_free(void* ptr) {
+    ((LLaMAGrammarElementWrapper*)ptr)->~LLaMAGrammarElementWrapper();
+    ruby_xfree(ptr);
+  }
+
+  static size_t llama_grammar_element_size(const void* ptr) {
+    return sizeof(*((LLaMAGrammarElementWrapper*)ptr));
+  }
+
+  static LLaMAGrammarElementWrapper* get_llama_grammar_element(VALUE self) {
+    LLaMAGrammarElementWrapper* ptr;
+    TypedData_Get_Struct(self, LLaMAGrammarElementWrapper, &llama_grammar_element_type, ptr);
+    return ptr;
+  }
+
+  static void define_class(VALUE outer) {
+    rb_cLLaMAGrammarElement = rb_define_class_under(outer, "GrammarElement", rb_cObject);
+    rb_define_alloc_func(rb_cLLaMAGrammarElement, llama_grammar_element_alloc);
+    rb_define_method(rb_cLLaMAGrammarElement, "initialize", RUBY_METHOD_FUNC(_llama_grammar_element_init), -1);
+    rb_define_method(rb_cLLaMAGrammarElement, "type=", RUBY_METHOD_FUNC(_llama_grammar_element_set_type), 1);
+    rb_define_method(rb_cLLaMAGrammarElement, "type", RUBY_METHOD_FUNC(_llama_grammar_element_get_type), 0);
+    rb_define_method(rb_cLLaMAGrammarElement, "value=", RUBY_METHOD_FUNC(_llama_grammar_element_set_value), 1);
+    rb_define_method(rb_cLLaMAGrammarElement, "value", RUBY_METHOD_FUNC(_llama_grammar_element_get_value), 0);
+  }
+
+private:
+  static const rb_data_type_t llama_grammar_element_type;
+
+  static VALUE _llama_grammar_element_init(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("type"), rb_intern("value") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    VALUE arr = Qnil;
+    rb_scan_args(argc, argv, ":", &arr, &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 0, 2, kw_values);
+
+    if (kw_values[0] != Qundef && !RB_INTEGER_TYPE_P(kw_values[0])) {
+      rb_raise(rb_eArgError, "type must be an integer");
+      return Qnil;
+    }
+    if (kw_values[1] != Qundef && !RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "value must be an integer");
+      return Qnil;
+    }
+
+    LLaMAGrammarElementWrapper* ptr = get_llama_grammar_element(self);
+    new (ptr) LLaMAGrammarElementWrapper();
+
+    if (kw_values[0] != Qundef) {
+      ptr->element.type = (enum llama_gretype)NUM2INT(kw_values[0]);
+    }
+    if (kw_values[1] != Qundef) {
+      ptr->element.value = NUM2INT(kw_values[1]);
+    }
+
+    return self;
+  }
+
+  // type
+  static VALUE _llama_grammar_element_set_type(VALUE self, VALUE type) {
+    LLaMAGrammarElementWrapper* ptr = get_llama_grammar_element(self);
+    ptr->element.type = (enum llama_gretype)NUM2INT(type);
+    return INT2NUM(ptr->element.type);
+  }
+
+  static VALUE _llama_grammar_element_get_type(VALUE self) {
+    LLaMAGrammarElementWrapper* ptr = get_llama_grammar_element(self);
+    return INT2NUM(ptr->element.type);
+  }
+
+  // value
+  static VALUE _llama_grammar_element_set_value(VALUE self, VALUE type) {
+    LLaMAGrammarElementWrapper* ptr = get_llama_grammar_element(self);
+    ptr->element.value = NUM2INT(type);
+    return INT2NUM(ptr->element.value);
+  }
+
+  static VALUE _llama_grammar_element_get_value(VALUE self) {
+    LLaMAGrammarElementWrapper* ptr = get_llama_grammar_element(self);
+    return INT2NUM(ptr->element.value);
+  }
+};
+
+const rb_data_type_t RbLLaMAGrammarElement::llama_grammar_element_type = {
+  "RbLLaMAGrammarElement",
+  { NULL,
+    RbLLaMAGrammarElement::llama_grammar_element_free,
+    RbLLaMAGrammarElement::llama_grammar_element_size },
+  NULL,
+  NULL,
+  RUBY_TYPED_FREE_IMMEDIATELY
+};
+
+class LLaMAGrammarWrapper {
+public:
+  struct llama_grammar* grammar;
+
+  LLaMAGrammarWrapper() : grammar(nullptr) {}
+
+  ~LLaMAGrammarWrapper() {
+    if (grammar) {
+      llama_grammar_free(grammar);
+    }
+  }
+};
+
+class RbLLaMAGrammar {
+public:
+  static VALUE llama_grammar_alloc(VALUE self) {
+    LLaMAGrammarWrapper* ptr = (LLaMAGrammarWrapper*)ruby_xmalloc(sizeof(LLaMAGrammarWrapper));
+    new (ptr) LLaMAGrammarWrapper();
+    return TypedData_Wrap_Struct(self, &llama_grammar_type, ptr);
+  }
+
+  static void llama_grammar_free(void* ptr) {
+    ((LLaMAGrammarWrapper*)ptr)->~LLaMAGrammarWrapper();
+    ruby_xfree(ptr);
+  }
+
+  static size_t llama_grammar_size(const void* ptr) {
+    return sizeof(*((LLaMAGrammarWrapper*)ptr));
+  }
+
+  static LLaMAGrammarWrapper* get_llama_grammar(VALUE self) {
+    LLaMAGrammarWrapper* ptr;
+    TypedData_Get_Struct(self, LLaMAGrammarWrapper, &llama_grammar_type, ptr);
+    return ptr;
+  }
+
+  static void define_class(VALUE outer) {
+    rb_cLLaMAGrammar = rb_define_class_under(outer, "Grammar", rb_cObject);
+    rb_define_alloc_func(rb_cLLaMAGrammar, llama_grammar_alloc);
+    rb_define_method(rb_cLLaMAGrammar, "initialize", RUBY_METHOD_FUNC(_llama_grammar_init), -1);
+  }
+
+private:
+  static const rb_data_type_t llama_grammar_type;
+
+  static VALUE _llama_grammar_init(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("rules"), rb_intern("start_rule_index") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
+      rb_raise(rb_eArgError, "rules must be an array");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "start_rule_index must be an integer");
+      return Qnil;
+    }
+
+    const int n_rules = RARRAY_LEN(kw_values[0]);
+    llama_grammar_element** rules = ALLOCA_N(llama_grammar_element*, n_rules);
+    for (int i = 0; i < n_rules; ++i) {
+      VALUE rule = rb_ary_entry(kw_values[0], i);
+      if (!RB_TYPE_P(rule, T_ARRAY)) {
+        rb_raise(rb_eArgError, "element of rules must be an array");
+        return Qnil;
+      }
+      const int n_elements = RARRAY_LEN(rule);
+      llama_grammar_element* elements = ALLOCA_N(llama_grammar_element, n_elements);
+      for (int j = 0; j < n_elements; ++j) {
+        VALUE element = rb_ary_entry(rule, j);
+        if (!rb_obj_is_kind_of(element, rb_cLLaMAGrammarElement)) {
+          rb_raise(rb_eArgError, "element of rule must be an instance of GrammarElement");
+          return Qnil;
+        }
+        LLaMAGrammarElementWrapper* ptr = RbLLaMAGrammarElement::get_llama_grammar_element(element);
+        elements[j] = ptr->element;
+      }
+      rules[i] = elements;
+    }
+
+    const size_t start_rule_index = NUM2SIZET(kw_values[1]);
+
+    LLaMAGrammarWrapper* ptr = get_llama_grammar(self);
+    new (ptr) LLaMAGrammarWrapper();
+    ptr->grammar = llama_grammar_init((const llama_grammar_element**)rules, n_rules, start_rule_index);
+
+    return self;
+  }
+};
+
+const rb_data_type_t RbLLaMAGrammar::llama_grammar_type = {
+  "RbLLaMAGrammar",
+  { NULL,
+    RbLLaMAGrammar::llama_grammar_free,
+    RbLLaMAGrammar::llama_grammar_size },
+  NULL,
+  NULL,
+  RUBY_TYPED_FREE_IMMEDIATELY
+};
+
 class LLaMAContextWrapper {
 public:
   struct llama_context* ctx;
@@ -991,6 +1346,8 @@ public:
     rb_define_method(rb_cLLaMAContext, "sample_token_mirostat_v2", RUBY_METHOD_FUNC(_llama_context_sample_token_mirostat_v2), -1);
     rb_define_method(rb_cLLaMAContext, "sample_token_greedy", RUBY_METHOD_FUNC(_llama_context_sample_token_greedy), 1);
     rb_define_method(rb_cLLaMAContext, "sample_token", RUBY_METHOD_FUNC(_llama_context_sample_token), 1);
+    rb_define_method(rb_cLLaMAContext, "sample_grammar", RUBY_METHOD_FUNC(_llama_context_sample_grammar), -1);
+    rb_define_method(rb_cLLaMAContext, "grammar_accept_token", RUBY_METHOD_FUNC(_llama_context_grammar_accept_token), -1);
   }
 
 private:
@@ -1581,11 +1938,11 @@ private:
 
   static VALUE _llama_context_sample_classifier_free_guidance(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
-    ID kw_table[3] = { rb_intern("guidance"), rb_intern("scale"), rb_intern("smooth_factor") };
-    VALUE kw_values[3] = { Qundef, Qundef, Qundef };
+    ID kw_table[2] = { rb_intern("guidance"), rb_intern("scale") };
+    VALUE kw_values[2] = { Qundef, Qundef };
     VALUE candidates = Qnil;
     rb_scan_args(argc, argv, "1:", &candidates, &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 3, 0, kw_values);
+    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
 
     if (!rb_obj_is_kind_of(kw_values[0], rb_cLLaMAContext)) {
       rb_raise(rb_eArgError, "guidance must be a Context");
@@ -1595,10 +1952,6 @@ private:
       rb_raise(rb_eArgError, "scale must be a float");
       return Qnil;
     }
-    if (!RB_FLOAT_TYPE_P(kw_values[2])) {
-      rb_raise(rb_eArgError, "smooth_factor must be a float");
-      return Qnil;
-    }
 
     LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
     if (ctx_ptr->ctx == NULL) {
@@ -1617,9 +1970,8 @@ private:
       return Qnil;
     }
     const float scale = NUM2DBL(kw_values[1]);
-    const float smooth_factor = NUM2DBL(kw_values[2]);
 
-    llama_sample_classifier_free_guidance(ctx_ptr->ctx, &(cnd_ptr->array), guidance_ptr->ctx, scale, smooth_factor);
+    llama_sample_classifier_free_guidance(ctx_ptr->ctx, &(cnd_ptr->array), guidance_ptr->ctx, scale);
 
     return Qnil;
   }
@@ -1972,6 +2324,69 @@ private:
     llama_token id = llama_sample_token(ctx_ptr->ctx, &(cnd_ptr->array));
     return INT2NUM(id);
   }
+
+  static VALUE _llama_context_sample_grammar(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[1] = { rb_intern("grammar") };
+    VALUE kw_values[1] = { Qundef };
+    VALUE candidates = Qnil;
+    rb_scan_args(argc, argv, "1:", &candidates, &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
+
+    if (!rb_obj_is_kind_of(candidates, rb_cLLaMATokenDataArray)) {
+      rb_raise(rb_eArgError, "1st argument must be a TokenDataArray");
+      return Qnil;
+    }
+    if (!rb_obj_is_kind_of(kw_values[0], rb_cLLaMAGrammar)) {
+      rb_raise(rb_eArgError, "grammar must be a Grammar");
+      return Qnil;
+    }
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    LLaMATokenDataArrayWrapper* cnd_ptr = RbLLaMATokenDataArray::get_llama_token_data_array(candidates);
+    if (cnd_ptr->array.data == nullptr) {
+      rb_raise(rb_eRuntimeError, "TokenDataArray is empty");
+      return Qnil;
+    }
+    LLaMAGrammarWrapper* grm_ptr = RbLLaMAGrammar::get_llama_grammar(kw_values[0]);
+
+    llama_sample_grammar(ctx_ptr->ctx, &(cnd_ptr->array), grm_ptr->grammar);
+
+    return Qnil;
+  }
+
+  static VALUE _llama_context_grammar_accept_token(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("grammar"), rb_intern("token") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
+
+    if (!rb_obj_is_kind_of(kw_values[0], rb_cLLaMAGrammar)) {
+      rb_raise(rb_eArgError, "grammar must be a Grammar");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "token must be an Integer");
+      return Qnil;
+    }
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    LLaMAGrammarWrapper* grm_ptr = RbLLaMAGrammar::get_llama_grammar(kw_values[0]);
+    llama_token token = NUM2INT(kw_values[1]);
+
+    llama_grammar_accept_token(ctx_ptr->ctx, grm_ptr->grammar, token);
+
+    return Qnil;
+  }
 };
 
 const rb_data_type_t RbLLaMAContext::llama_context_type = {
@@ -2062,6 +2477,10 @@ static VALUE rb_llama_mlock_supported(VALUE self) {
   return llama_mlock_supported() ? Qtrue : Qfalse;
 }
 
+static VALUE rb_llama_max_devices(VALUE self) {
+  return INT2NUM(llama_max_devices());
+}
+
 extern "C" void Init_llama_cpp(void) {
   rb_mLLaMACpp = rb_define_module("LLaMACpp");
 
@@ -2072,6 +2491,8 @@ extern "C" void Init_llama_cpp(void) {
   RbLLaMAContext::define_class(rb_mLLaMACpp);
   RbLLaMAContextParams::define_class(rb_mLLaMACpp);
   RbLLaMAModelQuantizeParams::define_class(rb_mLLaMACpp);
+  RbLLaMAGrammarElement::define_class(rb_mLLaMACpp);
+  RbLLaMAGrammar::define_class(rb_mLLaMACpp);
 
   rb_define_module_function(rb_mLLaMACpp, "backend_init", rb_llama_llama_backend_init, -1);
   rb_define_module_function(rb_mLLaMACpp, "backend_free", rb_llama_llama_backend_free, 0);
@@ -2082,6 +2503,7 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_module_function(rb_mLLaMACpp, "print_system_info", rb_llama_print_system_info, 0);
   rb_define_module_function(rb_mLLaMACpp, "mmap_supported?", rb_llama_mmap_supported, 0);
   rb_define_module_function(rb_mLLaMACpp, "mlock_supported?", rb_llama_mlock_supported, 0);
+  rb_define_module_function(rb_mLLaMACpp, "max_devices", rb_llama_max_devices, 0);
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_MAX_DEVICES", INT2NUM(LLAMA_MAX_DEVICES));
 
@@ -2103,6 +2525,14 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q5_K_M", INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_K_M));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q6_K", INT2NUM(LLAMA_FTYPE_MOSTLY_Q6_K));
 
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_END", INT2NUM(LLAMA_GRETYPE_END));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_ALT", INT2NUM(LLAMA_GRETYPE_ALT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_RULE_REF", INT2NUM(LLAMA_GRETYPE_RULE_REF));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_CHAR", INT2NUM(LLAMA_GRETYPE_CHAR));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_CHAR_NOT", INT2NUM(LLAMA_GRETYPE_CHAR_NOT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_CHAR_RNG_UPPER", INT2NUM(LLAMA_GRETYPE_CHAR_RNG_UPPER));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_CHAR_ALT", INT2NUM(LLAMA_GRETYPE_CHAR_ALT));
+
   std::stringstream ss_magic;
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGJT;
   rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGJT", rb_str_new2(ss_magic.str().c_str()));