llama_cpp 0.20.4 → 0.21.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 6acca29477a43c9703c7035e53acd69450de7103b2d7f242506c7e2016f1a261
-  data.tar.gz: 77c108b2f1ea33588a9dbe6c0538e289f90ea5be2090adbf8f663fff8b8b0221
+  metadata.gz: 64aab047f44a9002c5c388f6a774448671d7be9618170310d2b7dd1091a99670
+  data.tar.gz: 63f0d908b99a45865a9b9d81ae595adb0d9a1f258976267da3ce5b5df747ae0d
 SHA512:
-  metadata.gz: 69d453a3cf9c23ab3aaa60c6a20d80a7fc75424cb762c631a06712a2134fc7cf6830168241a3d689fd0b7b621804e27b6461415977fb51c096c10fbb2aa0e922
-  data.tar.gz: bc8de61663616ffd40c90e34df71095a7c85e3aa3373fc0c395fe101ded4f38e1670af42aeb11c6380c6932d2939d137d299d377db3ce9dc0510fd56e9d8b7a5
+  metadata.gz: 4f5344b36ff0c0bbd812f25f2b7a897893cb8ffb0700e3f8c23dfbd81cc8b1c4ee1893d5198c74df625fbc3c4c925f32e4c868c60b6ed13e7a5412eef3d9b2b6
+  data.tar.gz: dbc8d279a9284ba62786d3c5d267352445814289cb1f8a2268d2bb2ff72a4bc7a2534cb2c4582f945e13a68a645cec51101085203ff6dac1a2b7df6885412398

data/CHANGELOG.md

@@ -1,3 +1,16 @@
+## [[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` contant.
+
+## [[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.
+  - Remove constants for `llama_voca_pre_type` such as `LLAMA_VOCAB_PRE_TYPE_DEFAULT` and `LLAMA_VOCAB_PRE_TYPE_LLAMA3`.
+
 ## [[0.20.4](https://github.com/yoshoku/llama_cpp.rb/compare/v0.20.3...v0.20.4)] - 2025-06-21
 
 - Change supported llama.cpp version to b5720.
@@ -5,12 +18,10 @@
 
 ## [[0.20.3](https://github.com/yoshoku/llama_cpp.rb/compare/v0.20.2...v0.20.3)] - 2025-06-14
 
-
 - Change supported llama.cpp version to b5650
   - Add `data` argument to `llama_memory_clear` module function.
 - Fix llama_memory_t wrapper by removing unnecessary struct keyword and pointer symbol.
 
-
 ## [[0.20.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.20.1...v0.20.2)] - 2025-06-07
 
 - Change supported llama.cpp version to b5600

data/ext/llama_cpp/llama_cpp.c

@@ -838,6 +838,17 @@ 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) {
@@ -2940,6 +2951,22 @@ 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 +3941,7 @@ 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 +3981,11 @@ 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) {
@@ -4154,44 +4187,7 @@ 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));
-  /* llama_vocab_pre_type */
-  /* Document-const: LlamaCpp::LLAMA_VOCAB_PRE_TYPE_DEFAULT */
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_DEFAULT", INT2NUM(LLAMA_VOCAB_PRE_TYPE_DEFAULT));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_LLAMA3", INT2NUM(LLAMA_VOCAB_PRE_TYPE_LLAMA3));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM", INT2NUM(LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER", INT2NUM(LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_FALCON", INT2NUM(LLAMA_VOCAB_PRE_TYPE_FALCON));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_MPT", INT2NUM(LLAMA_VOCAB_PRE_TYPE_MPT));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_STARCODER", INT2NUM(LLAMA_VOCAB_PRE_TYPE_STARCODER));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_GPT2", INT2NUM(LLAMA_VOCAB_PRE_TYPE_GPT2));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_REFACT", INT2NUM(LLAMA_VOCAB_PRE_TYPE_REFACT));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_COMMAND_R", INT2NUM(LLAMA_VOCAB_PRE_TYPE_COMMAND_R));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_STABLELM2", INT2NUM(LLAMA_VOCAB_PRE_TYPE_STABLELM2));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_QWEN2", INT2NUM(LLAMA_VOCAB_PRE_TYPE_QWEN2));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_OLMO", INT2NUM(LLAMA_VOCAB_PRE_TYPE_OLMO));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_DBRX", INT2NUM(LLAMA_VOCAB_PRE_TYPE_DBRX));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_SMAUG", INT2NUM(LLAMA_VOCAB_PRE_TYPE_SMAUG));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_PORO", INT2NUM(LLAMA_VOCAB_PRE_TYPE_PORO));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_CHATGLM3", INT2NUM(LLAMA_VOCAB_PRE_TYPE_CHATGLM3));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_CHATGLM4", INT2NUM(LLAMA_VOCAB_PRE_TYPE_CHATGLM4));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_VIKING", INT2NUM(LLAMA_VOCAB_PRE_TYPE_VIKING));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_JAIS", INT2NUM(LLAMA_VOCAB_PRE_TYPE_JAIS));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_TEKKEN", INT2NUM(LLAMA_VOCAB_PRE_TYPE_TEKKEN));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_SMOLLM", INT2NUM(LLAMA_VOCAB_PRE_TYPE_SMOLLM));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_CODESHELL", INT2NUM(LLAMA_VOCAB_PRE_TYPE_CODESHELL));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_BLOOM", INT2NUM(LLAMA_VOCAB_PRE_TYPE_BLOOM));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH", INT2NUM(LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_EXAONE", INT2NUM(LLAMA_VOCAB_PRE_TYPE_EXAONE));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_CHAMELEON", INT2NUM(LLAMA_VOCAB_PRE_TYPE_CHAMELEON));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_MINERVA", INT2NUM(LLAMA_VOCAB_PRE_TYPE_MINERVA));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM", INT2NUM(LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_GPT4O", INT2NUM(LLAMA_VOCAB_PRE_TYPE_GPT4O));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_SUPERBPE", INT2NUM(LLAMA_VOCAB_PRE_TYPE_SUPERBPE));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_TRILLION", INT2NUM(LLAMA_VOCAB_PRE_TYPE_TRILLION));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_BAILINGMOE", INT2NUM(LLAMA_VOCAB_PRE_TYPE_BAILINGMOE));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_LLAMA4", INT2NUM(LLAMA_VOCAB_PRE_TYPE_LLAMA4));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_PIXTRAL", INT2NUM(LLAMA_VOCAB_PRE_TYPE_PIXTRAL));
-  rb_define_const(rb_mLlamaCpp, "LLAMA_VOCAB_PRE_TYPE_SEED_CODER", INT2NUM(LLAMA_VOCAB_PRE_TYPE_SEED_CODER));
+  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));
@@ -4784,6 +4780,17 @@ 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 */
 
@@ -4895,6 +4902,7 @@ void Init_llama_cpp(void) {
   /* TODO: void* imatrix */
   /* TODO: void* kv_overrides */
   /* TODO: void* tensor_types */
+  /* TODO: void* prune_layers */
 
   /**
    * Document-class: LlamaCpp::LlamaLogitBias
@@ -5285,6 +5293,9 @@ 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);
 
@@ -5468,6 +5479,11 @@ 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 = '0.20.4'
+  VERSION = '0.21.1'
 
   # The supported version of llama.cpp.
-  LLAMA_CPP_VERSION = 'b5720'
+  LLAMA_CPP_VERSION = 'b5930'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.20.4
+  version: 0.21.1
 platform: ruby
 authors:
 - yoshoku