llama_cpp 0.25.6 → 0.28.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: fe3f4ccdf08a754d7b331a32a5c2544c672e9da31b9f5eb29e8e6503376391ab
-  data.tar.gz: 73d9ad537f7557232778e882edce50847333867b2b7f2b7f39420a68aeef723f
+  metadata.gz: 002371b1977caa3b67be951da5098f478c540b0b85c02e49e0263579626e983b
+  data.tar.gz: da1b91e34d7c494f3a4661b85a3c6bda909389b9c40a223aac9675330619132e
 SHA512:
-  metadata.gz: ca6b8855194be0c835f4a672600fcff0e6a602afff49175c5668bd62e15c8f721c6034b1919879b56c777c77dd88f73f9e2a6f92db8397c1fe6be874d2a6e188
-  data.tar.gz: 5a4eab3bef52c269e355934a7e8da85cfac17e73befc2eed1a6e6eda183d21e324dac0309c33c3f296f61031e27aab315803b0fbaaf571d910440ca036dfc741
+  metadata.gz: 53a2f89d9dbc2e3fe73de60dcd8c09218f9c4b7e49c89916619d74df3b7c245cd0252a3aead4f5caf918e86b21da5eb7a2630a56fe40a2fb6f01567b2a11bc92
+  data.tar.gz: 3f784cd9cfd6fa24d26ceafd643013b8b70e5c5610366719bea147f62d3febfee24b54ad1099027f279454f8ced4d0b8d0fd0859b54175a2222e311786898f00

data/CHANGELOG.md

@@ -1,3 +1,42 @@
+## [[0.28.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.27.0...v0.28.0)] - 2026-08-23
+
+- Change supported llama.cpp version to v0.2.0.
+  - Add `LLAMA_LOAD_MODE_AUTO` constant value.
+  - Add `n_outputs_max_per_seq accessor` to `LlamaContextParams`.
+  - Add `llama_version` module function to `LlamaCpp`.
+  - Add `llama_sampler_copy` module function to `LlamaCpp`.
+
+## [[0.27.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.26.0...v0.27.0)] - 2026-08-08
+
+- Change supported llama.cpp version to b10310.
+  - Add `n_vocab` argument to `llama_sampler_init_penaltie` module function.
+  - Add `load_mtp` accessor to LlamaModelParams.
+
+## [[0.26.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.25.7...v0.26.0)] - 2026-08-01
+
+- Change supported llama.cpp version to b10210.
+  - Remove `use_mmap` accessor from LlamaModelParams.
+  - Remove `use_direct_io` accessor from LlamaModelParams.
+  - Remove `use_mlock` accessor from LlamaModelParams.
+  - Add `LLAMA_LOAD_MODE_NONE` constant value.
+  - Add `LLAMA_LOAD_MODE_MMAP` constant value.
+  - Add `LLAMA_LOAD_MODE_MLOCK` constant value.
+  - Add `LLAMA_LOAD_MODE_MMAP_MLOCK` constant value.
+  - Add `LLAMA_LOAD_MODE_DIRECT_IO` constant value.
+  - Add `llama_load_mode_name` module function to `LlamaCpp`.
+  - Add `llama_load_mode_from_str` module function to `LlamaCpp`.
+
+## [[0.25.7](https://github.com/yoshoku/llama_cpp.rb/compare/v0.25.6...v0.25.7)] - 2026-07-11
+
+- Change supported llama.cpp version to b9950.
+  - Add `LLAMA_FTYPE_MOSTLY_Q2_0` constant value.
+
+## [[0.25.6](https://github.com/yoshoku/llama_cpp.rb/compare/v0.25.5...v0.25.6)] - 2026-07-04
+
+- Change supported llama.cpp version to b9860.
+  - Add `llama_model_ftype` module function to `LlamaCpp`.
+  - Add `llama_ftype_name` module function to `LlamaCpp`.
+
 ## [[0.25.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.25.4...v0.25.5)] - 2026-06-27
 
 - Change supported llama.cpp version to b9820.

data/ext/llama_cpp/llama_cpp.c

@@ -498,39 +498,6 @@ static VALUE llama_model_params_set_vocab_only(VALUE self, VALUE vocab_only) {
   return vocab_only;
 }
 
-static VALUE llama_model_params_get_use_mmap(VALUE self) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  return data->use_mmap ? Qtrue : Qfalse;
-}
-
-static VALUE llama_model_params_set_use_mmap(VALUE self, VALUE use_mmap) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  data->use_mmap = RTEST(use_mmap) ? true : false;
-  return use_mmap;
-}
-
-static VALUE llama_model_params_get_use_direct_io(VALUE self) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  return data->use_direct_io ? Qtrue : Qfalse;
-}
-
-static VALUE llama_model_params_set_use_direct_io(VALUE self, VALUE use_direct_io) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  data->use_direct_io = RTEST(use_direct_io) ? true : false;
-  return use_direct_io;
-}
-
-static VALUE llama_model_params_get_use_mlock(VALUE self) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  return data->use_mlock ? Qtrue : Qfalse;
-}
-
-static VALUE llama_model_params_set_use_mlock(VALUE self, VALUE use_mlock) {
-  struct llama_model_params* data = get_llama_model_params(self);
-  data->use_mlock = RTEST(use_mlock) ? true : false;
-  return use_mlock;
-}
-
 static VALUE llama_model_params_get_check_tensors(VALUE self) {
   struct llama_model_params* data = get_llama_model_params(self);
   return data->check_tensors ? Qtrue : Qfalse;
@@ -575,6 +542,17 @@ static VALUE llama_model_params_set_no_alloc(VALUE self, VALUE no_alloc) {
   return no_alloc;
 }
 
+static VALUE llama_model_params_get_load_mtp(VALUE self) {
+  struct llama_model_params* data = get_llama_model_params(self);
+  return data->load_mtp ? Qtrue : Qfalse;
+}
+
+static VALUE llama_model_params_set_load_mtp(VALUE self, VALUE load_mtp) {
+  struct llama_model_params* data = get_llama_model_params(self);
+  data->load_mtp = RTEST(load_mtp) ? true : false;
+  return load_mtp;
+}
+
 /* struct llama_context_params */
 static void llama_context_params_free(void *ptr) {
   if (ptr) {
@@ -674,6 +652,17 @@ static VALUE llama_context_params_set_n_outputs_max(VALUE self, VALUE n_outputs_
   return n_outputs_max;
 }
 
+static VALUE llama_context_params_get_n_outputs_max_per_seq(VALUE self) {
+  struct llama_context_params* data = get_llama_context_params(self);
+  return UINT2NUM(data->n_outputs_max_per_seq);
+}
+
+static VALUE llama_context_params_set_n_outputs_max_per_seq(VALUE self, VALUE n_outputs_max_per_seq) {
+  struct llama_context_params* data = get_llama_context_params(self);
+  data->n_outputs_max_per_seq = NUM2UINT(n_outputs_max_per_seq);
+  return n_outputs_max_per_seq;
+}
+
 static VALUE llama_context_params_get_n_threads(VALUE self) {
   struct llama_context_params* data = get_llama_context_params(self);
   return INT2NUM(data->n_threads);
@@ -1329,6 +1318,15 @@ static llama_adapter_lora_wrapper* get_llama_adapter_lora_wrapper(VALUE self) {
   return data;
 }
 
+/**
+ * @overload llama_version
+ *   @return [String]
+ */
+static VALUE rb_llama_version(VALUE self) {
+  const char* version = llama_version();
+  return rb_utf8_str_new_cstr(version);
+}
+
 /**
  * @overload llama_backend_init
  *   @return [NilClass]
@@ -3585,6 +3583,29 @@ static VALUE rb_llama_sampler_clone(VALUE self, VALUE sampler) {
   return TypedData_Wrap_Struct(rb_cLlamaSampler, &llama_sampler_data_type, clone);
 }
 
+/**
+ * @overload llama_sampler_copy(src, dst)
+ *  @param [LlamaSampler] src
+ *  @param [LlamaSampler] dst
+ *  @return [NilClass]
+ */
+static VALUE rb_llama_sampler_copy(VALUE self, VALUE src, VALUE dst) {
+  if (!rb_obj_is_kind_of(src, rb_cLlamaSampler)) {
+    rb_raise(rb_eArgError, "src must be a LlamaSampler");
+    return Qnil;
+  }
+  if (!rb_obj_is_kind_of(dst, rb_cLlamaSampler)) {
+    rb_raise(rb_eArgError, "dst must be a LlamaSampler");
+    return Qnil;
+  }
+  struct llama_sampler* src_ = get_llama_sampler(src);
+  struct llama_sampler* dst_ = get_llama_sampler(dst);
+  llama_sampler_copy(src_, dst_);
+  RB_GC_GUARD(src);
+  RB_GC_GUARD(dst);
+  return Qnil;
+}
+
 /**
  * @overload llama_sampler_free(sampler)
  *  @param [LlamaSampler] sampler
@@ -3964,13 +3985,18 @@ static VALUE rb_llama_sampler_init_grammar(VALUE self, VALUE vocab, VALUE gramma
 
 /**
  * @overload llama_sampler_init_penalties(penalty_last_n, penalty_repeat, penalty_freq, penalty_present)
+ *  @param [Integer] n_vocab
  *  @param [Integer] penalty_last_n
  *  @param [Float] penalty_repeat
  *  @param [Float] penalty_freq
  *  @param [Float] penalty_present
  *  @return [LlamaSampler]
  */
-static VALUE rb_llama_sampler_init_penalties(VALUE self, VALUE penalty_last_n, VALUE penalty_repeat, VALUE penalty_freq, VALUE penalty_present) {
+static VALUE rb_llama_sampler_init_penalties(VALUE self, VALUE n_vocab, VALUE penalty_last_n, VALUE penalty_repeat, VALUE penalty_freq, VALUE penalty_present) {
+  if (!RB_INTEGER_TYPE_P(n_vocab)) {
+    rb_raise(rb_eArgError, "n_vocab must be an Integer");
+    return Qnil;
+  }
   if (!RB_INTEGER_TYPE_P(penalty_last_n)) {
     rb_raise(rb_eArgError, "penalty_last_n must be an Integer");
     return Qnil;
@@ -3987,7 +4013,7 @@ static VALUE rb_llama_sampler_init_penalties(VALUE self, VALUE penalty_last_n, V
     rb_raise(rb_eArgError, "penalty_present must be a Float");
     return Qnil;
   }
-  struct llama_sampler* sampler = llama_sampler_init_penalties(NUM2INT(penalty_last_n), NUM2DBL(penalty_repeat), NUM2DBL(penalty_freq), NUM2DBL(penalty_present));
+  struct llama_sampler* sampler = llama_sampler_init_penalties(NUM2INT(n_vocab), NUM2INT(penalty_last_n), NUM2DBL(penalty_repeat), NUM2DBL(penalty_freq), NUM2DBL(penalty_present));
   return TypedData_Wrap_Struct(rb_cLlamaSampler, &llama_sampler_data_type, sampler);
 }
 
@@ -4356,6 +4382,31 @@ static VALUE rb_llama_flash_attn_type_name(VALUE self, VALUE flash_attn_type) {
   return rb_utf8_str_new_cstr(name);
 }
 
+/**
+ * @overload llama_load_mode_name(load_mode)
+ *  @param [Integer] load_mode
+ *  @return [String]
+ */
+static VALUE rb_llama_load_mode_name(VALUE self, VALUE load_mode) {
+  if (!RB_INTEGER_TYPE_P(load_mode)) {
+    rb_raise(rb_eArgError, "load_mode must be an Integer");
+    return Qnil;
+  }
+  const char* name = llama_load_mode_name((enum llama_load_mode)NUM2INT(load_mode));
+  return rb_utf8_str_new_cstr(name);
+}
+
+static VALUE rb_llama_load_mode_from_str(VALUE self, VALUE str) {
+  if (!RB_TYPE_P(str, T_STRING)) {
+    rb_raise(rb_eArgError, "str must be a String");
+    return Qnil;
+  }
+  const char* str_ = StringValueCStr(str);
+  enum llama_load_mode load_mode = llama_load_mode_from_str(str_);
+  RB_GC_GUARD(str);
+  return INT2NUM(load_mode);
+}
+
 /* MAIN */
 void Init_llama_cpp(void) {
   char tmp[12];
@@ -4487,6 +4538,7 @@ void Init_llama_cpp(void) {
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_MXFP4_MOE", INT2NUM(LLAMA_FTYPE_MOSTLY_MXFP4_MOE));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_NVFP4", INT2NUM(LLAMA_FTYPE_MOSTLY_NVFP4));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_Q1_0", INT2NUM(LLAMA_FTYPE_MOSTLY_Q1_0));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_Q2_0", INT2NUM(LLAMA_FTYPE_MOSTLY_Q2_0));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 
   rb_define_module_function(rb_mLlamaCpp, "llama_ftype_name", rb_llama_ftype_name, 1);
@@ -4523,6 +4575,14 @@ void Init_llama_cpp(void) {
   rb_define_const(rb_mLlamaCpp, "LLAMA_SPLIT_MODE_LAYER", INT2NUM(LLAMA_SPLIT_MODE_LAYER));
   rb_define_const(rb_mLlamaCpp, "LLAMA_SPLIT_MODE_ROW", INT2NUM(LLAMA_SPLIT_MODE_ROW));
   rb_define_const(rb_mLlamaCpp, "LLAMA_SPLIT_MODE_TENSOR", INT2NUM(LLAMA_SPLIT_MODE_TENSOR));
+  /* llama_load_type */
+  /* Document-const: LlamaCpp::LLAMA_LOAD_MODE_NONE */
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_NONE", INT2NUM(LLAMA_LOAD_MODE_NONE));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_AUTO", INT2NUM(LLAMA_LOAD_MODE_AUTO));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_MMAP", INT2NUM(LLAMA_LOAD_MODE_MMAP));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_MLOCK", INT2NUM(LLAMA_LOAD_MODE_MLOCK));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_MMAP_MLOCK", INT2NUM(LLAMA_LOAD_MODE_MMAP_MLOCK));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_LOAD_MODE_DIRECT_IO", INT2NUM(LLAMA_LOAD_MODE_DIRECT_IO));
   /* llama_context_type */
   /* Document-const: LlamaCpp::LLAMA_CONTEXT_TYPE_DEFAULT */
   rb_define_const(rb_mLlamaCpp, "LLAMA_CONTEXT_TYPE_DEFAULT", INT2NUM(LLAMA_CONTEXT_TYPE_DEFAULT));
@@ -4729,39 +4789,6 @@ void Init_llama_cpp(void) {
    * @return [Boolean]
    */
   rb_define_method(rb_cLlamaModelParams, "vocab_only=", RUBY_METHOD_FUNC(llama_model_params_set_vocab_only), 1);
-  /**
-   * Document-method: use_mmap
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_mmap", RUBY_METHOD_FUNC(llama_model_params_get_use_mmap), 0);
-  /**
-   * Document-method: use_mmap=
-   * @param [Boolean] use_mmap
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_mmap=", RUBY_METHOD_FUNC(llama_model_params_set_use_mmap), 1);
-  /**
-   * Document-method: use_direct_io
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_direct_io", RUBY_METHOD_FUNC(llama_model_params_get_use_direct_io), 0);
-  /**
-   * Document-method: use_direct_io=
-   * @param [Boolean] use_direct_io
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_direct_io=", RUBY_METHOD_FUNC(llama_model_params_set_use_direct_io), 1);
-  /**
-   * Document-method: use_mlock
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_mlock", RUBY_METHOD_FUNC(llama_model_params_get_use_mlock), 0);
-  /**
-   * Document-method: use_mlock=
-   * @param [Boolean] use_mlock
-   * @return [Boolean]
-   */
-  rb_define_method(rb_cLlamaModelParams, "use_mlock=", RUBY_METHOD_FUNC(llama_model_params_set_use_mlock), 1);
   /**
    * Document-method: check_tensors
    * @return [Boolean]
@@ -4806,6 +4833,17 @@ void Init_llama_cpp(void) {
    * @return [Boolean]
    */
   rb_define_method(rb_cLlamaModelParams, "no_alloc=", RUBY_METHOD_FUNC(llama_model_params_set_no_alloc), 1);
+  /**
+   * Document-method: load_mtp
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaModelParams, "load_mtp", RUBY_METHOD_FUNC(llama_model_params_get_load_mtp), 0);
+  /**
+   * Document-method: load_mtp=
+   * @param [Boolean] load_mtp_
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaModelParams, "load_mtp=", RUBY_METHOD_FUNC(llama_model_params_set_load_mtp), 1);
 
   /* TODO: struct llama_sampler_seq_config */
 
@@ -4881,6 +4919,17 @@ void Init_llama_cpp(void) {
    * @return [Integer]
    */
   rb_define_method(rb_cLlamaContextParams, "n_outputs_max=", RUBY_METHOD_FUNC(llama_context_params_set_n_outputs_max), 1);
+  /**
+   * Document-method: n_outputs_max_per_seq
+   * @return [Integer]
+   */
+  rb_define_method(rb_cLlamaContextParams, "n_outputs_max_per_seq", RUBY_METHOD_FUNC(llama_context_params_get_n_outputs_max_per_seq), 0);
+  /**
+   * Document-method: n_outputs_max_per_seq=
+   * @param [Integer] n_outputs_max_per_seq
+   * @return [Integer]
+   */
+  rb_define_method(rb_cLlamaContextParams, "n_outputs_max_per_seq=", RUBY_METHOD_FUNC(llama_context_params_set_n_outputs_max_per_seq), 1);
   /**
    * Document-method: n_threads
    * @return [Integer]
@@ -5363,6 +5412,9 @@ void Init_llama_cpp(void) {
   rb_cLlamaAdapterLora = rb_define_class_under(rb_mLlamaCpp, "LlamaAdapterLora", rb_cObject);
   rb_define_alloc_func(rb_cLlamaAdapterLora, llama_adapter_lora_wrapper_alloc);
 
+  /* llama_version */
+  rb_define_module_function(rb_mLlamaCpp, "llama_version", rb_llama_version, 0);
+
   /* llama_backend_init */
   rb_define_module_function(rb_mLlamaCpp, "llama_backend_init", rb_llama_backend_init, 0);
 
@@ -5716,6 +5768,8 @@ void Init_llama_cpp(void) {
   /* llama_vocab_get_add_sep */
   rb_define_module_function(rb_mLlamaCpp, "llama_vocab_get_add_sep", rb_llama_vocab_get_add_sep, 1);
 
+  /* TODO: llama_vocab_get_suppress_tokens */
+
   /* llama_vocab_fim_pre */
   rb_define_module_function(rb_mLlamaCpp, "llama_vocab_fim_pre", rb_llama_vocab_fim_pre, 1);
 
@@ -5777,6 +5831,9 @@ void Init_llama_cpp(void) {
   /* llama_sampler_clone */
   rb_define_module_function(rb_mLlamaCpp, "llama_sampler_clone", rb_llama_sampler_clone, 1);
 
+  /* llama_sampler_copy */
+  rb_define_module_function(rb_mLlamaCpp, "llama_sampler_copy", rb_llama_sampler_copy, 2);
+
   /* llama_sampler_free */
   rb_define_module_function(rb_mLlamaCpp, "llama_sampler_free", rb_llama_sampler_free, 1);
 
@@ -5837,7 +5894,7 @@ void Init_llama_cpp(void) {
   /* TODO: llama_sampler_init_grammar_lazy_patterns */
 
   /* llama_sampler_init_penalties */
-  rb_define_module_function(rb_mLlamaCpp, "llama_sampler_init_penalties", rb_llama_sampler_init_penalties, 4);
+  rb_define_module_function(rb_mLlamaCpp, "llama_sampler_init_penalties", rb_llama_sampler_init_penalties, 5);
 
   /* TODO: llama_sampler_init_dry */
 
@@ -5942,6 +5999,12 @@ void Init_llama_cpp(void) {
   /* llama_perf_sampler_reset */
   rb_define_module_function(rb_mLlamaCpp, "llama_perf_sampler_reset", rb_llama_perf_sampler_reset, 1);
 
+  /* llama_load_mode_name */
+  rb_define_module_function(rb_mLlamaCpp, "llama_load_mode_name", rb_llama_load_mode_name, 1);
+
+  /* llama_load_mode_from_str */
+  rb_define_module_function(rb_mLlamaCpp, "llama_load_mode_from_str", rb_llama_load_mode_from_str, 1);
+
   /* TODO: typedef bool (*llama_opt_param_filter) */
   /* TODO: bool llama_opt_param_filter_all */
   /* TODO: struct llama_opt_params */

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LlamaCpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.25.6'
+  VERSION = '0.28.0'
 
   # The supported version of llama.cpp.
-  LLAMA_CPP_VERSION = 'b9860'
+  LLAMA_CPP_VERSION = 'v0.2.0'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.25.6
+  version: 0.28.0
 platform: ruby
 authors:
 - yoshoku
@@ -33,7 +33,7 @@ metadata:
   homepage_uri: https://github.com/yoshoku/llama_cpp.rb
   source_code_uri: https://github.com/yoshoku/llama_cpp.rb
   changelog_uri: https://github.com/yoshoku/llama_cpp.rb/blob/main/CHANGELOG.md
-  documentation_uri: https://gemdocs.org/gems/llama_cpp/0.25.6/
+  documentation_uri: https://gemdocs.org/gems/llama_cpp/0.28.0/
   rubygems_mfa_required: 'true'
 rdoc_options: []
 require_paths:
@@ -49,7 +49,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 4.0.13
+rubygems_version: 4.0.16
 specification_version: 4
 summary: Ruby bindings for the llama.cpp.
 test_files: []