llama_cpp 0.21.0 → 0.21.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 691ff52591a63387090485f1d322726b8dc5dc89630cdac4fe2bfcb3372da50d
-  data.tar.gz: c27f3a43878e787f32eaeaa2538d03dca4e2e3de0484ad2e920ec14826cf6ba5
+  metadata.gz: 5f3d80e415240f21df0ba7e314118712f6f2d2d8cb052eef41127e7de7ae7e51
+  data.tar.gz: 67fff55bf83ef4d97f24b0c3c450eb493464763c23595fcb5b60c248fd1f551e
 SHA512:
-  metadata.gz: df049a84a78bb2d95cd4fe1f63f05c1bb9f965c0e5c3bfaca9668e98fca0db2eb8d80560ffa819540fa25b0a099a0b2a3feafc85a38d73632f3db79445d19a07
-  data.tar.gz: d9138045ba1d37dbaab919ea973ef81306dde30dd6d67048660108dd9f90fe1ffe671becd7f295afa2b01150d73269ac045474373c98ff5805ce35c582dd28d4
+  metadata.gz: f6a129952bc812c130a235743beeb0a5ebcafcbe627d7d13d625e8cf5b143e2cb41cd4d78cfacc98c00a1d0857dc68bb8a6b30d486da5f22e8bcbc6a63ac5489
+  data.tar.gz: c944f2a37727bd3e295dc83d9d9c5c5748a04d7f867aad86895c312a100fa65b7c3fa9f3973bf08adc47c26c2a0e73586f7aa92b92230bee72c86b10dc4693fc

data/CHANGELOG.md

@@ -1,3 +1,18 @@
+## [[0.21.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.21.1...v0.21.2)] - 2025-08-09
+
+- Change supported llama.cpp version to b6100.
+  - Add `LLAMA_FTYPE_MOSTLY_MXFP4_MOE` constant.
+  - Add `use_extra_bufts` accessor to `LlamaModelParams`.
+  - Add `llama_model_is_diffusion?` module function.
+
+## [[0.21.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.21.0...v0.21.1)] - 2025-07-19
+
+- Change supported llama.cpp version to b5930.
+  - Add `n_reused` reader to `LlamaPerfContextData`.
+  - Add `llama_vocab_mask` module function.
+  - Add `kv_unified` accessor to `LlamaContextParams`.
+  - Add `LLAMA_VOCAB_TYPE_PLAMO2` constant.
+
 ## [[0.21.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.20.4...v0.21.0)] - 2025-07-12
 
 - Change supported llama.cpp version to b5870.

data/ext/llama_cpp/llama_cpp.c

@@ -530,6 +530,17 @@ static VALUE llama_model_params_set_check_tensors(VALUE self, VALUE check_tensor
   return check_tensors;
 }
 
+static VALUE llama_model_params_get_use_extra_bufts(VALUE self) {
+  struct llama_model_params* data = get_llama_model_params(self);
+  return data->use_extra_bufts ? Qtrue : Qfalse;
+}
+
+static VALUE llama_model_params_set_use_extra_bufts(VALUE self, VALUE use_extra_bufts) {
+  struct llama_model_params* data = get_llama_model_params(self);
+  data->use_extra_bufts = RTEST(use_extra_bufts) ? true : false;
+  return use_extra_bufts;
+}
+
 /* struct llama_context_params */
 static void llama_context_params_free(void *ptr) {
   if (ptr) {
@@ -838,6 +849,17 @@ static VALUE llama_context_params_set_swa_full(VALUE self, VALUE swa_full) {
   return swa_full;
 }
 
+static VALUE llama_context_params_get_kv_unified(VALUE self) {
+  struct llama_context_params* data = get_llama_context_params(self);
+  return data->kv_unified ? Qtrue : Qfalse;
+}
+
+static VALUE llama_context_params_set_kv_unified(VALUE self, VALUE kv_unified) {
+  struct llama_context_params* data = get_llama_context_params(self);
+  data->kv_unified = RTEST(kv_unified) ? true : false;
+  return kv_unified;
+}
+
 /* llama_model_quantize_params */
 static void llama_model_quantize_params_free(void *ptr) {
   if (ptr) {
@@ -1763,6 +1785,20 @@ static VALUE rb_llama_model_is_recurrent(VALUE self, VALUE model) {
   return llama_model_is_recurrent(model_wrapper->model) ? Qtrue : Qfalse;
 }
 
+/**
+ * @overload llama_model_is_diffusion?(model)
+ *  @param [LlamaModel] model
+ *  @return [Boolean]
+ */
+static VALUE rb_llama_model_is_diffusion(VALUE self, VALUE model) {
+  if (!rb_obj_is_kind_of(model, rb_cLlamaModel)) {
+    rb_raise(rb_eArgError, "model must be a LlamaModel");
+    return Qnil;
+  }
+  llama_model_wrapper* model_wrapper = get_llama_model_wrapper(model);
+  return llama_model_is_diffusion(model_wrapper->model) ? Qtrue : Qfalse;
+}
+
 /**
  * @overload llama_model_quantize(fname_inp, fname_out, params)
  *  @param [String] fname_inp
@@ -2940,6 +2976,22 @@ static VALUE rb_llama_vocab_pad(VALUE self, VALUE vocab) {
   return INT2NUM(token);
 }
 
+/**
+ * @overload llama_vocab_mask(vocab)
+ *  @param [LlamaVocab] vocab
+ *  @return [Integer]
+ */
+static VALUE rb_llama_vocab_mask(VALUE self, VALUE vocab) {
+  if (!rb_obj_is_kind_of(vocab, rb_cLlamaVocab)) {
+    rb_raise(rb_eArgError, "vocab must be a LlamaVocab");
+    return Qnil;
+  }
+  llama_vocab_wrapper* vocab_wrapper = get_llama_vocab_wrapper(vocab);
+  const int32_t token = llama_vocab_mask(vocab_wrapper->vocab);
+  RB_GC_GUARD(vocab);
+  return INT2NUM(token);
+}
+
 /**
  * @overload llama_vocab_get_add_bos
  *  @param [LlamaVocab] vocab
@@ -3914,6 +3966,7 @@ static VALUE llama_perf_context_data_alloc(VALUE self) {
   data->t_eval_ms = 0.0;
   data->n_p_eval = 0;
   data->n_eval = 0;
+  data->n_reused = 0;
   return TypedData_Wrap_Struct(self, &llama_perf_context_data_type, data);
 }
 
@@ -3953,6 +4006,11 @@ static VALUE llama_perf_context_data_get_n_eval(VALUE self) {
   return INT2NUM(data->n_eval);
 }
 
+static VALUE llama_perf_context_data_get_n_reused(VALUE self) {
+  struct llama_perf_context_data* data = get_llama_perf_context_data(self);
+  return INT2NUM(data->n_reused);
+}
+
 /* struct llama_perf_sampler_data */
 static void llama_perf_sampler_data_free(void* ptr) {
   if (ptr) {
@@ -4154,6 +4212,7 @@ void Init_llama_cpp(void) {
   rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_TYPE_WPM", INT2NUM(LLAMA_VOCAB_TYPE_WPM));
   rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_TYPE_UGM", INT2NUM(LLAMA_VOCAB_TYPE_UGM));
   rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_TYPE_RWKV", INT2NUM(LLAMA_VOCAB_TYPE_RWKV));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_TYPE_PLAMO2", INT2NUM(LLAMA_VOCAB_TYPE_PLAMO2));
   /* llama_rope_type */
   /* Document-const: LlamaCpp::LLAMA_ROPE_TYPE_NONE */
   rb_define_const(rb_mLlamaCpp, "LLAMA_ROPE_TYPE_NONE", INT2NUM(LLAMA_ROPE_TYPE_NONE));
@@ -4217,6 +4276,7 @@ void Init_llama_cpp(void) {
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_BF16", INT2NUM(LLAMA_FTYPE_MOSTLY_BF16));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_TQ1_0", INT2NUM(LLAMA_FTYPE_MOSTLY_TQ1_0));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_TQ2_0", INT2NUM(LLAMA_FTYPE_MOSTLY_TQ2_0));
+  rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_MOSTLY_MXFP4_MOE", INT2NUM(LLAMA_FTYPE_MOSTLY_MXFP4_MOE));
   rb_define_const(rb_mLlamaCpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
   /* llama_rope_scaling_type */
   /* Document-const: LlamaCpp::LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED */
@@ -4462,6 +4522,17 @@ void Init_llama_cpp(void) {
    * @return [Boolean]
    */
   rb_define_method(rb_cLlamaModelParams, "check_tensors=", RUBY_METHOD_FUNC(llama_model_params_set_check_tensors), 1);
+  /**
+   * Document-method: use_extra_bufts
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaModelParams, "use_extra_bufts", RUBY_METHOD_FUNC(llama_model_params_get_use_extra_bufts), 0);
+  /**
+   * Document-method: use_extra_bufts=
+   * @param [Boolean] use_extra_bufts
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaModelParams, "use_extra_bufts=", RUBY_METHOD_FUNC(llama_model_params_set_use_extra_bufts), 1);
 
   /**
    * Document-class: LlamaCpp::LlamaContextParams
@@ -4746,6 +4817,17 @@ void Init_llama_cpp(void) {
    * @return [Boolean]
    */
   rb_define_method(rb_cLlamaContextParams, "swa_full=", RUBY_METHOD_FUNC(llama_context_params_set_swa_full), 1);
+  /**
+   * Document-method: kv_unified
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaContextParams, "kv_unified", RUBY_METHOD_FUNC(llama_context_params_get_kv_unified), 0);
+  /**
+   * Document-method: kv_unified=
+   * @param [Boolean] kv_unified
+   * @return [Boolean]
+   */
+  rb_define_method(rb_cLlamaContextParams, "kv_unified=", RUBY_METHOD_FUNC(llama_context_params_set_kv_unified), 1);
   /* TODO: ggml_abort_callback abort_callback */
   /* TODO: void* abort_callback_data */
 
@@ -5061,6 +5143,9 @@ void Init_llama_cpp(void) {
   /* llama_model_is_recurrent */
   rb_define_module_function(rb_mLlamaCpp, "llama_model_is_recurrent?", rb_llama_model_is_recurrent, 1);
 
+  /* llama_model_is_diffusion */
+  rb_define_module_function(rb_mLlamaCpp, "llama_model_is_diffusion?", rb_llama_model_is_diffusion, 1);
+
   /* llama_model_quantize */
   rb_define_module_function(rb_mLlamaCpp, "llama_model_quantize", rb_llama_model_quantize, 3);
 
@@ -5248,6 +5333,9 @@ void Init_llama_cpp(void) {
   /* llama_vocab_pad */
   rb_define_module_function(rb_mLlamaCpp, "llama_vocab_pad", rb_llama_vocab_pad, 1);
 
+  /* llama_vocab_mask */
+  rb_define_module_function(rb_mLlamaCpp, "llama_vocab_mask", rb_llama_vocab_mask, 1);
+
   /* llama_vocab_get_add_bos */
   rb_define_module_function(rb_mLlamaCpp, "llama_vocab_get_add_bos", rb_llama_vocab_get_add_bos, 1);
 
@@ -5431,6 +5519,11 @@ void Init_llama_cpp(void) {
    * @return [Integer]
    */
   rb_define_method(rb_cLlamaPerfContextData, "n_eval", RUBY_METHOD_FUNC(llama_perf_context_data_get_n_eval), 0);
+  /**
+   * Document-method: n_reused
+   * @return [Integer]
+   */
+  rb_define_method(rb_cLlamaPerfContextData, "n_reused", RUBY_METHOD_FUNC(llama_perf_context_data_get_n_reused), 0);
 
   /**
    * Document-class: LlamaCpp::LlamaPerfSamplerData

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 = '0.21.0'
+  VERSION = '0.21.2'
 
   # The supported version of llama.cpp.
-  LLAMA_CPP_VERSION = 'b5870'
+  LLAMA_CPP_VERSION = 'b6100'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.21.0
+  version: 0.21.2
 platform: ruby
 authors:
 - yoshoku
@@ -49,7 +49,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.6.9
+rubygems_version: 3.7.0
 specification_version: 4
 summary: Ruby bindings for the llama.cpp.
 test_files: []