llama_cpp 0.2.2 → 0.3.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e5e221d4831be790a990b121e6ac780d10b4cbfb85b2a9b4284d9c216f6e5604
-  data.tar.gz: fba76ac1a70bfd7b02b8d123c57e4c8096a29ac7f658bb090cda91c6a54752d2
+  metadata.gz: 9e0152eb9e091932225356614b57fad416c2aa96a83316f8585c9ef2872e1504
+  data.tar.gz: 8ea2f00f11be7dd6524bfe69e3181fc63df7c841ed1e2d91b1b2bcafd99d0b66
 SHA512:
-  metadata.gz: 994029383219077e134d170177954251c20ede6d1c83843ecd22c42eeae83584079d124b41702f55add7f3f237e9bdb14382fbd37dde2d0e74f8cffcfed1715b
-  data.tar.gz: ca4e94b6ddf4e4e9ddabbb2b8309cf4b2b06a881df09fdf4ad96e27c4f1f620ca0024ac46f69d9b474849c074a5c9ba9b0440777a0b52a12413bc356457a02f3
+  metadata.gz: a85a4bdd2d1fd575eb406b9bebdf7f388db33dc42f7a2980ba9a7a6b346b539854d9df5515c9b6968727e76f035a23f59d4bc65bc5525df962dfbdf56d8b3b01
+  data.tar.gz: 33641d622102257dbc1358bde0871a03c595928f5d8cedee512e1df414e4aa93433eadfcd082d4db42046320c1ed7f806dfb3aafd7934a1becb33fe275f9435c

data/CHANGELOG.md

@@ -1,3 +1,31 @@
+## [[0.3.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.2.2...v0.3.0)] - 2023-06-30
+
+- Add no_k_quants and qkk_64 config options:
+  ```
+  $ gem install llama_cpp -- --with-no_k_quants
+  ```
+  ```
+  $ gem install llama_cpp -- --with-qkk_64
+  ```
+
+**Breaking Changes**
+- Remove `Client` class to concentrate on developing bindings.
+- Bump bundled llama.cpp from master-7487137 to master-9d23589.
+  - llama_init_from_file and llama_apply_lora_from_file are deprecated.
+- Add `Model` class for wrapping llama_model.
+- Move the `apply_lora_from_file method`, `free`, `load`, and `empty?` methods to `Model` class from `Context` class.
+- Change arguments of initialize method of Context. Its initialize method requires Model object instead of the model's file path.
+  ```ruby
+  requre 'llama_cpp'
+
+  params = LLaMACpp::ContextParams.new
+
+  model = LLaMACpp::Model.new(model_path: '/path/to/quantized-model.bin', params: params)
+  context = LLaMACpp::Context.new(model: model)
+
+  LLaMACpp.generate(context, 'Hello, world.')
+  ```
+
 ## [[0.2.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.2.1...v0.2.2)] - 2023-06-24
 
 - Bump bundled llama.cpp from master-a09f919 to master-7487137.

data/README.md

@@ -20,21 +20,54 @@ If bundler is not being used to manage dependencies, install the gem by executin
 
 ## Usage
 
-Prepare the quantized model by refering to [the usage section on the llama.cpp README](https://github.com/ggerganov/llama.cpp#usage) or
-download the qunatized model, for example [ggml-vicuna-7b-4bit](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5541351), from Hugging Face.
+Prepare the quantized model by refering to [the usage section on the llama.cpp README](https://github.com/ggerganov/llama.cpp#usage).
+For example, preparing the quatization model based on [open_llama_7b](https://huggingface.co/openlm-research/open_llama_7b) is as follows:
+
+```sh
+$ cd ~/
+$ brew install git-lfs
+$ git lfs install
+$ git clone https://github.com/ggerganov/llama.cpp.git
+$ cd llama.cpp
+$ python3 -m pip install -r requirements.txt
+$ cd models
+$ git clone https://huggingface.co/openlm-research/open_llama_7b
+$ cd ../
+$ python3 convert.py models/open_llama_7b
+$ make
+$ ./quantize ./models/open_llama_7b/ggml-model-f16.bin ./models/open_llama_7b/ggml-model-q4_0.bin q4_0
+```
+
+An example of Ruby code that generates sentences with the quantization model is as follows:
 
 ```ruby
 require 'llama_cpp'
 
 params = LLaMACpp::ContextParams.new
-params.seed = 12
+params.seed = 42
 
-context = LLaMACpp::Context.new(model_path: '/path/to/quantized-model.bin', params: params)
+model = LLaMACpp::Model.new(model_path: '/home/user/llama.cpp/models/open_llama_7b/ggml-model-q4_0.bin', params: params)
+context = LLaMACpp::Context.new(model: model)
 
-puts LLaMACpp.generate(context, 'Please tell me the largest city in Japan.', n_threads: 4)
-# => "There are two major cities in Japan, Tokyo and Osaka, which have about 30 million populations."
+puts LLaMACpp.generate(context, 'Hello, World.', n_threads: 4)
 ```
 
+## Examples
+There is a sample program in the [examples](https://github.com/yoshoku/llama_cpp.rb/tree/main/examples) directory that allow interactvie communication like ChatGPT.
+
+```sh
+$ git clone https://github.com/yoshoku/llama_cpp.rb.git
+$ cd examples
+$ bundle install
+$ ruby chat.rb --model /home/user/llama.cpp/models/open_llama_7b/ggml-model-q4_0.bin --seed 2023
+...
+User: Who is the originator of the Ruby programming language?
+Bob: The originator of the Ruby programming language is Mr. Yukihiro Matsumoto.
+User:
+```
+
+![llama_cpp_chat_example](https://github.com/yoshoku/llama_cpp.rb/assets/5562409/374ae3d8-63a6-498f-ae6e-5552b464bdda)
+
 ## Contributing
 
 Bug reports and pull requests are welcome on GitHub at https://github.com/yoshoku/llama_cpp.rb.

data/examples/chat.rb

@@ -35,7 +35,8 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
     params = LLaMACpp::ContextParams.new
     params.seed = options[:seed]
     params.n_gpu_layers = options[:n_gpu_layers]
-    context = LLaMACpp::Context.new(model_path: options[:model], params: params)
+    model = LLaMACpp::Model.new(model_path: options[:model], params: params)
+    context = LLaMACpp::Context.new(model: model)
 
     antiprompt = options[:reverse_prompt] || 'User:'
     start_prompt = read_prompt(options[:file]) || default_prompt(antiprompt)

data/examples/embedding.rb

@@ -16,12 +16,13 @@ class Embedding < Thor # rubocop:disable Style/Documentation
   option :model, type: :string, aliases: '-m', desc: 'path to model file', required: true
   option :prompt, type: :string, aliases: '-p', desc: 'prompt to generate embedding', required: true
   option :n_gpu_layers, type: :numeric, desc: 'number of layers on GPU', default: 0
-  def main # rubocop:disable Metrics/AbcSize
+  def main # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
     params = LLaMACpp::ContextParams.new
     params.seed = options[:seed]
     params.n_gpu_layers = options[:n_gpu_layers]
     params.embedding = true
-    context = LLaMACpp::Context.new(model_path: options[:model], params: params)
+    model = LLaMACpp::Model.new(model_path: options[:model], params: params)
+    context = LLaMACpp::Context.new(model: model)
 
     embd_input = context.tokenize(text: options[:prompt], add_bos: true)
 

data/ext/llama_cpp/extconf.rb

@@ -17,6 +17,17 @@ if RUBY_PLATFORM.match?(/darwin|linux|bsd/) && try_compile('#include <stdio.h>',
   $CXXFLAGS << ' -pthread'
 end
 
+unless with_config('no_k_quants')
+  $CFLAGS << ' -DGGML_USE_K_QUANTS'
+  $CXXFLAGS << ' -DGGML_USE_K_QUANTS'
+  $srcs << 'k_quants.c'
+end
+
+if with_config('qkk_64')
+  $CFLAGS << ' -DGGML_QKK_64'
+  $CXXFLAGS << ' -DGGML_QKK_64'
+end
+
 if with_config('openblas')
   abort 'libopenblas is not found.' unless have_library('openblas')
   abort 'cblas.h is not found.' unless have_header('cblas.h')
@@ -42,6 +53,7 @@ if with_config('metal')
   $CXXFLAGS << ' -DGGML_USE_METAL'
   $LDFLAGS << ' -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders'
   $objs = %w[ggml.o llama.o llama_cpp.o ggml-metal.o]
+  $objs << 'k_quants.o' unless with_config('no_k_quants')
 end
 
 if with_config('cublas')
@@ -49,6 +61,7 @@ if with_config('cublas')
   $CXXFLAGS << ' -DGGML_USE_CUBLAS -I/usr/local/cuda/include'
   $LDFLAGS << ' -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64'
   $objs = %w[ggml-cuda.o ggml.o llama.o llama_cpp.o]
+  $objs << 'k_quants.o' unless with_config('no_k_quants')
 end
 
 if with_config('clblast')

data/ext/llama_cpp/llama_cpp.cpp

@@ -2,6 +2,7 @@
 #include "llama_cpp.h"
 
 VALUE rb_mLLaMACpp;
+VALUE rb_cLLaMAModel;
 VALUE rb_cLLaMAContext;
 VALUE rb_cLLaMAContextParams;
 VALUE rb_cLLaMAModelQuantizeParams;
@@ -610,6 +611,202 @@ const rb_data_type_t RbLLaMAModelQuantizeParams::llama_model_quantize_params_typ
   RUBY_TYPED_FREE_IMMEDIATELY
 };
 
+class LLaMAModelWrapper {
+public:
+  struct llama_model* model;
+
+  LLaMAModelWrapper() : model(NULL){};
+
+  ~LLaMAModelWrapper() {
+    if (model != NULL) {
+      llama_free_model(model);
+    }
+  };
+};
+
+class RbLLaMAModel {
+public:
+  static VALUE llama_model_alloc(VALUE self) {
+    LLaMAModelWrapper* ptr = (LLaMAModelWrapper*)ruby_xmalloc(sizeof(LLaMAModelWrapper));
+    new (ptr) LLaMAModelWrapper();
+    return TypedData_Wrap_Struct(self, &llama_model_type, ptr);
+  }
+
+  static void llama_model_free(void* ptr) {
+    ((LLaMAModelWrapper*)ptr)->~LLaMAModelWrapper();
+    ruby_xfree(ptr);
+  }
+
+  static size_t llama_model_size(const void* ptr) {
+    return sizeof(*((LLaMAModelWrapper*)ptr));
+  }
+
+  static LLaMAModelWrapper* get_llama_model(VALUE self) {
+    LLaMAModelWrapper* ptr;
+    TypedData_Get_Struct(self, LLaMAModelWrapper, &llama_model_type, ptr);
+    return ptr;
+  }
+
+  static void define_class(VALUE outer) {
+    rb_cLLaMAModel = rb_define_class_under(outer, "Model", rb_cObject);
+    rb_define_alloc_func(rb_cLLaMAModel, llama_model_alloc);
+    rb_define_method(rb_cLLaMAModel, "initialize", RUBY_METHOD_FUNC(_llama_model_initialize), -1);
+    rb_define_method(rb_cLLaMAModel, "empty?", RUBY_METHOD_FUNC(_llama_model_empty), 0);
+    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);
+  }
+
+private:
+  static const rb_data_type_t llama_model_type;
+
+  static VALUE _llama_model_initialize(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 0, 2, kw_values);
+
+    if (kw_values[0] == Qundef && kw_values[1] == Qundef) {
+      rb_iv_set(self, "@params", Qnil);
+      return Qnil;
+    }
+
+    if (!RB_TYPE_P(kw_values[0], T_STRING)) {
+      rb_raise(rb_eArgError, "model_path must be a string");
+      return Qnil;
+    }
+    if (!rb_obj_is_kind_of(kw_values[1], rb_cLLaMAContextParams)) {
+      rb_raise(rb_eArgError, "params must be a ContextParams");
+      return Qnil;
+    }
+
+    VALUE filename = kw_values[0];
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
+    LLaMAModelWrapper* model_ptr = get_llama_model(self);
+
+    try {
+      model_ptr->model = llama_load_model_from_file(StringValueCStr(filename), prms_ptr->params);
+    } catch (const std::runtime_error& e) {
+      rb_raise(rb_eRuntimeError, "%s", e.what());
+      return Qnil;
+    }
+
+    if (model_ptr->model == NULL) {
+      rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA model");
+      return Qnil;
+    }
+
+    rb_iv_set(self, "@params", kw_values[1]);
+
+    RB_GC_GUARD(filename);
+    return Qnil;
+  }
+
+  static VALUE _llama_model_empty(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    if (ptr->model != NULL) {
+      return Qfalse;
+    }
+    return Qtrue;
+  }
+
+  static VALUE _llama_model_free(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    if (ptr->model != NULL) {
+      llama_free_model(ptr->model);
+      ptr->model = NULL;
+      rb_iv_set(self, "@params", Qnil);
+    }
+    return Qnil;
+  }
+
+  static VALUE _llama_model_load(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
+    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_STRING)) {
+      rb_raise(rb_eArgError, "model_path must be a string");
+      return Qnil;
+    }
+    if (!rb_obj_is_kind_of(kw_values[1], rb_cLLaMAContextParams)) {
+      rb_raise(rb_eArgError, "params must be a LLaMAContextParams");
+      return Qnil;
+    }
+
+    LLaMAModelWrapper* model_ptr = get_llama_model(self);
+    if (model_ptr->model != NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA model is already loaded");
+      return Qnil;
+    }
+
+    VALUE filename = kw_values[0];
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
+
+    try {
+      model_ptr->model = llama_load_model_from_file(StringValueCStr(filename), prms_ptr->params);
+    } catch (const std::runtime_error& e) {
+      rb_raise(rb_eRuntimeError, "%s", e.what());
+      return Qnil;
+    }
+
+    if (model_ptr->model == NULL) {
+      rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA model");
+      return Qnil;
+    }
+
+    rb_iv_set(self, "@params", kw_values[1]);
+
+    RB_GC_GUARD(filename);
+    return Qnil;
+  }
+
+  static VALUE _llama_model_apply_lora_from_file(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[3] = { rb_intern("lora_path"), rb_intern("base_model_path"), rb_intern("n_threads") };
+    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, "lora_path must be a string");
+      return Qnil;
+    }
+    if (kw_values[1] != Qundef && !RB_TYPE_P(kw_values[1], T_STRING)) {
+      rb_raise(rb_eArgError, "base_model_path must be a string");
+      return Qnil;
+    }
+    if (kw_values[2] != Qundef && !RB_INTEGER_TYPE_P(kw_values[2])) {
+      rb_raise(rb_eArgError, "n_threads must be an integer");
+      return Qnil;
+    }
+
+    const char* lora_path = StringValueCStr(kw_values[0]);
+    const char* base_model_path = kw_values[1] == Qundef ? NULL : StringValueCStr(kw_values[1]);
+    const int n_threads = kw_values[2] == Qundef ? 1 : NUM2INT(kw_values[2]);
+
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    if (llama_model_apply_lora_from_file(ptr->model, lora_path, base_model_path, n_threads) != 0) {
+      rb_raise(rb_eRuntimeError, "Failed to apply LoRA");
+      return Qnil;
+    }
+    return Qnil;
+  };
+};
+
+const rb_data_type_t RbLLaMAModel::llama_model_type = {
+  "RbLLaMAModel",
+  { NULL,
+    RbLLaMAModel::llama_model_free,
+    RbLLaMAModel::llama_model_size },
+  NULL,
+  NULL,
+  RUBY_TYPED_FREE_IMMEDIATELY
+};
+
 class LLaMAContextWrapper {
 public:
   struct llama_context* ctx;
@@ -662,10 +859,6 @@ public:
     rb_define_method(rb_cLLaMAContext, "n_embd", RUBY_METHOD_FUNC(_llama_context_n_embd), 0);
     rb_define_method(rb_cLLaMAContext, "print_timings", RUBY_METHOD_FUNC(_llama_context_print_timings), 0);
     rb_define_method(rb_cLLaMAContext, "reset_timings", RUBY_METHOD_FUNC(_llama_context_reset_timings), 0);
-    rb_define_method(rb_cLLaMAContext, "empty?", RUBY_METHOD_FUNC(_llama_context_empty), 0);
-    rb_define_method(rb_cLLaMAContext, "free", RUBY_METHOD_FUNC(_llama_context_free), 0);
-    rb_define_method(rb_cLLaMAContext, "load", RUBY_METHOD_FUNC(_llama_context_load), -1);
-    rb_define_method(rb_cLLaMAContext, "apply_lora_from_file", RUBY_METHOD_FUNC(_llama_context_apply_lora_from_file), -1);
     rb_define_method(rb_cLLaMAContext, "kv_cache_token_count", RUBY_METHOD_FUNC(_llama_context_kv_cache_token_count), 0);
     rb_define_method(rb_cLLaMAContext, "set_rng_seed", RUBY_METHOD_FUNC(_llama_context_set_rng_seed), 1);
     rb_define_method(rb_cLLaMAContext, "load_session_file", RUBY_METHOD_FUNC(_llama_context_load_session_file), -1);
@@ -689,46 +882,37 @@ private:
 
   static VALUE _llama_context_initialize(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
-    ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
-    VALUE kw_values[2] = { Qundef, Qundef };
+    ID kw_table[1] = { rb_intern("model") };
+    VALUE kw_values[1] = { Qundef };
     rb_scan_args(argc, argv, ":", &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 0, 2, kw_values);
+    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
 
-    if (kw_values[0] == Qundef && kw_values[1] == Qundef) {
-      rb_iv_set(self, "@params", Qnil);
-      rb_iv_set(self, "@has_evaluated", Qfalse);
+    VALUE model = kw_values[0];
+    if (!rb_obj_is_kind_of(model, rb_cLLaMAModel)) {
+      rb_raise(rb_eArgError, "model must be a Model");
       return Qnil;
     }
 
-    if (!RB_TYPE_P(kw_values[0], T_STRING)) {
-      rb_raise(rb_eArgError, "model_path must be a string");
-      return Qnil;
-    }
-    if (!rb_obj_is_kind_of(kw_values[1], rb_cLLaMAContextParams)) {
-      rb_raise(rb_eArgError, "params must be a ContextParams");
+    LLaMAModelWrapper* model_ptr = RbLLaMAModel::get_llama_model(model);
+    if (model_ptr->model == NULL) {
+      rb_raise(rb_eRuntimeError, "Model is empty");
       return Qnil;
     }
 
-    VALUE filename = kw_values[0];
-    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
+    VALUE params = rb_iv_get(model, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
     LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
 
-    try {
-      ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
-    } catch (const std::runtime_error& e) {
-      rb_raise(rb_eRuntimeError, "%s", e.what());
-      return Qnil;
-    }
+    ctx_ptr->ctx = llama_new_context_with_model(model_ptr->model, prms_ptr->params);
 
     if (ctx_ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA context");
       return Qnil;
     }
 
-    rb_iv_set(self, "@params", kw_values[1]);
+    rb_iv_set(self, "@model", model);
     rb_iv_set(self, "@has_evaluated", Qfalse);
 
-    RB_GC_GUARD(filename);
     return Qnil;
   };
 
@@ -873,7 +1057,9 @@ private:
       return Qnil;
     }
 
-    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(rb_iv_get(self, "@params"));
+    VALUE model = rb_iv_get(self, "@model");
+    VALUE params = rb_iv_get(model, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
     const int n_tokens = prms_ptr->params.logits_all ? NUM2INT(rb_iv_get(self, "@n_tokens")) : 1;
     const int n_vocab = llama_n_vocab(ptr->ctx);
     const float* logits = llama_get_logits(ptr->ctx);
@@ -891,7 +1077,9 @@ private:
       rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
       return Qnil;
     }
-    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(rb_iv_get(self, "@params"));
+    VALUE model = rb_iv_get(self, "@model");
+    VALUE params = rb_iv_get(model, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
     if (!prms_ptr->params.embedding) {
       rb_raise(rb_eRuntimeError, "embedding parameter is false");
       return Qnil;
@@ -995,106 +1183,6 @@ private:
     return Qnil;
   };
 
-  static VALUE _llama_context_empty(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx != NULL) {
-      return Qfalse;
-    }
-    return Qtrue;
-  }
-
-  static VALUE _llama_context_free(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx != NULL) {
-      llama_free(ptr->ctx);
-      ptr->ctx = NULL;
-      rb_iv_set(self, "@params", Qnil);
-      rb_iv_set(self, "@has_evaluated", Qfalse);
-    }
-    return Qnil;
-  }
-
-  static VALUE _llama_context_load(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
-    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_STRING)) {
-      rb_raise(rb_eArgError, "model_path must be a string");
-      return Qnil;
-    }
-    if (!rb_obj_is_kind_of(kw_values[1], rb_cLLaMAContextParams)) {
-      rb_raise(rb_eArgError, "params must be a LLaMAContextParams");
-      return Qnil;
-    }
-
-    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
-    if (ctx_ptr->ctx != NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is already loaded");
-      return Qnil;
-    }
-
-    VALUE filename = kw_values[0];
-    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
-
-    try {
-      ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
-    } catch (const std::runtime_error& e) {
-      rb_raise(rb_eRuntimeError, "%s", e.what());
-      return Qnil;
-    }
-
-    if (ctx_ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA context");
-      return Qnil;
-    }
-
-    rb_iv_set(self, "@params", kw_values[1]);
-    rb_iv_set(self, "@has_evaluated", Qfalse);
-
-    RB_GC_GUARD(filename);
-    return Qnil;
-  };
-
-  static VALUE _llama_context_apply_lora_from_file(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[3] = { rb_intern("lora_path"), rb_intern("base_model_path"), rb_intern("n_threads") };
-    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, "lora_path must be a string");
-      return Qnil;
-    }
-    if (kw_values[1] != Qundef && !RB_TYPE_P(kw_values[1], T_STRING)) {
-      rb_raise(rb_eArgError, "base_model_path must be a string");
-      return Qnil;
-    }
-    if (kw_values[2] != Qundef && !RB_INTEGER_TYPE_P(kw_values[2])) {
-      rb_raise(rb_eArgError, "n_threads must be an integer");
-      return Qnil;
-    }
-
-    const char* lora_path = StringValueCStr(kw_values[0]);
-    const char* base_model_path = kw_values[1] == Qundef ? NULL : StringValueCStr(kw_values[1]);
-    const int n_threads = kw_values[2] == Qundef ? 1 : NUM2INT(kw_values[2]);
-
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx != NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is already loaded");
-      return Qnil;
-    }
-
-    if (llama_apply_lora_from_file(ptr->ctx, lora_path, base_model_path, n_threads) != 0) {
-      rb_raise(rb_eRuntimeError, "Failed to apply LoRA");
-      return Qnil;
-    }
-    return Qnil;
-  };
-
   static VALUE _llama_context_kv_cache_token_count(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -1137,7 +1225,9 @@ private:
       return Qnil;
     }
 
-    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(rb_iv_get(self, "@params"));
+    VALUE model = rb_iv_get(self, "@model");
+    VALUE params = rb_iv_get(model, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
     const int n_ctx = prms_ptr->params.n_ctx;
 
     std::vector<llama_token> session_tokens(n_ctx);
@@ -1664,8 +1754,16 @@ const rb_data_type_t RbLLaMAContext::llama_context_type = {
 
 // module functions
 
-static VALUE rb_llama_llama_init_backend(VALUE self) {
-  llama_init_backend();
+static VALUE rb_llama_llama_init_backend(int argc, VALUE* argv, VALUE self) {
+  VALUE kw_args = Qnil;
+  ID kw_table[1] = { rb_intern("numa") };
+  VALUE kw_values[1] = { Qundef };
+  rb_scan_args(argc, argv, ":", &kw_args);
+  rb_get_kwargs(kw_args, kw_table, 0, 1, kw_values);
+
+  const bool numa = kw_values[0] == Qundef ? false : (RTEST ? true : false);
+  llama_init_backend(numa);
+
   return Qnil;
 }
 
@@ -1731,10 +1829,11 @@ extern "C" void Init_llama_cpp(void) {
 
   RbLLaMATokenData::define_class(rb_mLLaMACpp);
   RbLLaMATokenDataArray::define_class(rb_mLLaMACpp);
+  RbLLaMAModel::define_class(rb_mLLaMACpp);
   RbLLaMAContext::define_class(rb_mLLaMACpp);
   RbLLaMAContextParams::define_class(rb_mLLaMACpp);
 
-  rb_define_module_function(rb_mLLaMACpp, "init_backend", rb_llama_llama_init_backend, 0);
+  rb_define_module_function(rb_mLLaMACpp, "init_backend", rb_llama_llama_init_backend, -1);
   rb_define_module_function(rb_mLLaMACpp, "model_quantize", rb_llama_model_quantize, -1);
   rb_define_module_function(rb_mLLaMACpp, "token_bos", rb_llama_token_bos, 0);
   rb_define_module_function(rb_mLLaMACpp, "token_eos", rb_llama_token_eos, 0);