llama_cpp 0.8.0 → 0.9.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 8045208b5f7801979212a4f6ed395217e78f06bcfbc2d0362aaaa04c529745cd
-  data.tar.gz: 4011dfe279d8d4041c6c79dc5a6bad199777f83b5f0559f11ccd2f68c957e462
+  metadata.gz: dae7507ce41f18e3fd0fb2d7445275a387a3914068aa9eef922f260de699970a
+  data.tar.gz: d66cc2629aeca3285bc10988f8c410fb8cf5b7f1fe6f835b5dc60e9dcab4be9d
 SHA512:
-  metadata.gz: d15e74da491773961006eca8ca6c6d80b30ffc995c56a9140961be0002eb09134f1a029c4e8ee192497fb7256fe36cf1c3ed928967ce57ece4c7a0904392c8fe
-  data.tar.gz: a863596304ddb9ac5e4be2b2b65bebc7d3913705b8a0f516debfee0ca213f9dca69707edda8d70cfafb15500fcb6e70cffb6d5d1119302d24e05059c50f0da77
+  metadata.gz: 3e3e92aa38413877620947ec7996494cd720a3c211fcdf1973ce0d7a9a7e8803e293e2ce2f601b11e35858c5b4ef6b00d716069e322ea8d6b4c93412990fd746
+  data.tar.gz: 20a1e9e0e5812da9b00787afbf0f3aa0b762c8168f54ce3b7f2f25ff5b61cca5b2e7ab5faa065fbc3e266468d1c5747b8e0779fc7e073cc66240d1f3085e71c7

data/CHANGELOG.md

@@ -1,3 +1,22 @@
+## [[0.9.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.9.0...v0.9.1)] - 2023-11-03
+
+- Bump bundled llama.cpp from b1429 to b1472
+  - Rename `kv_cahe_tokens_rm` method to `kv_cahce_clear` in Context.
+  - Add `sample_min_p method` to Context.
+  - Add `rope_scaling_type`, `rope_freq_base`, `rope_freq_scale`, `yarn_ext_factor`, `yarn_attn_factor`, `yarn_beta_fast`, `yarn_beta_slow`, and `yarn_orig_ctx` to ContextParams.
+  - Add `pure` to ModelQuantizeParams.
+  - Add contstants for RoPE scaling type.
+
+## [[0.9.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.8.0...v0.9.0)] - 2023-10-28
+
+- Fix missing object file for ggml-backend when building with metal and cublas options.
+
+**Breaking Changes**
+- Bump bundled llama.cpp from b1405 to b1429
+  - Move following methods from Context to Model:
+    - text, score, type, token_bos, token_eos, token_nl, token_prefix, token_middle, token_suffix, and token_eos.
+  - Add `sample_repetition_penalties` method, which integrates sample_frequency_and_presence_penalties and sample_repetition_penalty methods.
+
 ## [[0.8.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.7.1...v0.8.0)] - 2023-10-21
 
 **Breaking Changes**

data/examples/chat.rb

@@ -83,10 +83,12 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
         candidates = LLaMACpp::TokenDataArray.new(base_candidates)
 
         last_n_repeat = [last_n_tokens.size, options[:repeat_last_n], n_ctx].min
-        context.sample_repetition_penalty(candidates, last_n_tokens[-last_n_repeat..], penalty: options[:repeat_penalty])
-        context.sample_frequency_and_presence_penalties(
-          candidates, last_n_tokens[-last_n_repeat..],
-          frequency: options[:frequency_penalty], presence: options[:presence_penalty]
+        context.sample_repetition_penalties(
+          candidates,
+          last_n_tokens[-last_n_repeat..],
+          penalty_repeat: options[:repeat_penalty],
+          penalty_freq: options[:frequency_penalty],
+          penalty_present: options[:presence_penalty]
         )
 
         context.sample_top_k(candidates, k: options[:top_k])
@@ -99,8 +101,8 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
         last_n_tokens.shift
         last_n_tokens.push(id)
 
-        if id == context.token_eos
-          id = context.token_nl
+        if id == context.model.token_eos
+          id = context.model.token_nl
           unless antiprompt.empty?
             first_antiprompt = context.model.tokenize(text: antiprompt, add_bos: false)
             embd_input.concat(first_antiprompt)

data/ext/llama_cpp/extconf.rb

@@ -5,7 +5,7 @@ require 'fileutils'
 
 abort 'libstdc++ is not found.' unless have_library('stdc++')
 
-$srcs = %w[ggml.c ggml-backend.c ggml-alloc.c llama.cpp llama_cpp.cpp]
+$srcs = %w[ggml.c ggml-backend.c ggml-alloc.c ggml-quants.c llama.cpp llama_cpp.cpp]
 $srcs << 'ggml-opencl.cpp' if with_config('clblast')
 $srcs << 'ggml-mpi.c' if with_config('mpi')
 $CFLAGS << ' -w -DNDEBUG'
@@ -18,12 +18,6 @@ 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'
@@ -53,16 +47,14 @@ if with_config('metal')
   $CFLAGS << ' -DGGML_USE_METAL'
   $CXXFLAGS << ' -DGGML_USE_METAL'
   $LDFLAGS << ' -framework Foundation -framework Metal -framework MetalKit'
-  $objs = %w[ggml.o ggml-alloc.o ggml-metal.o llama.o llama_cpp.o]
-  $objs << 'k_quants.o' unless with_config('no_k_quants')
+  $objs = %w[ggml.o ggml-backend.o ggml-alloc.o ggml-quants.o ggml-metal.o llama.o llama_cpp.o]
 end
 
 if with_config('cublas')
   $CFLAGS << ' -DGGML_USE_CUBLAS -I/usr/local/cuda/include'
   $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.o ggml-alloc.o ggml-cuda.o llama.o llama_cpp.o]
-  $objs << 'k_quants.o' unless with_config('no_k_quants')
+  $objs = %w[ggml.o ggml-backend.o ggml-alloc.o ggml-quants.o ggml-cuda.o llama.o llama_cpp.o]
 end
 
 if with_config('clblast')

data/ext/llama_cpp/llama_cpp.cpp

@@ -796,10 +796,22 @@ public:
     rb_define_method(rb_cLLaMAContextParams, "n_threads", RUBY_METHOD_FUNC(_llama_context_params_get_n_threads), 0);
     rb_define_method(rb_cLLaMAContextParams, "n_threads_batch=", RUBY_METHOD_FUNC(_llama_context_params_set_n_threads_batch), 1);
     rb_define_method(rb_cLLaMAContextParams, "n_threads_batch", RUBY_METHOD_FUNC(_llama_context_params_get_n_threads_batch), 0);
+    rb_define_method(rb_cLLaMAContextParams, "rope_scaling_type=", RUBY_METHOD_FUNC(_llama_context_params_set_rope_scaling_type), 1);
+    rb_define_method(rb_cLLaMAContextParams, "rope_scaling_type", RUBY_METHOD_FUNC(_llama_context_params_get_rope_scaling_type), 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, "yarn_ext_factor=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_ext_factor), 1);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_ext_factor", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_ext_factor), 0);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_attn_factor=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_attn_factor), 1);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_attn_factor", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_attn_factor), 0);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_beta_fast=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_beta_fast), 1);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_beta_fast", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_beta_fast), 0);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_beta_slow=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_beta_slow), 1);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_beta_slow", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_beta_slow), 0);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_orig_ctx=", RUBY_METHOD_FUNC(_llama_context_params_set_yarn_orig_ctx), 1);
+    rb_define_method(rb_cLLaMAContextParams, "yarn_orig_ctx", RUBY_METHOD_FUNC(_llama_context_params_get_yarn_orig_ctx), 0);
     rb_define_method(rb_cLLaMAContextParams, "mul_mat_q=", RUBY_METHOD_FUNC(_llama_context_params_set_mul_mat_q), 1);
     rb_define_method(rb_cLLaMAContextParams, "mul_mat_q", RUBY_METHOD_FUNC(_llama_context_params_get_mul_mat_q), 0);
     rb_define_method(rb_cLLaMAContextParams, "f16_kv=", RUBY_METHOD_FUNC(_llama_context_params_set_f16_kv), 1);
@@ -883,6 +895,18 @@ private:
     return INT2NUM(ptr->params.n_threads_batch);
   }
 
+  // rope_scaling_type
+  static VALUE _llama_context_params_set_rope_scaling_type(VALUE self, VALUE scaling_type) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.rope_scaling_type = NUM2INT(scaling_type);
+    return INT2NUM(ptr->params.rope_scaling_type);
+  }
+
+  static VALUE _llama_context_params_get_rope_scaling_type(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return INT2NUM(ptr->params.rope_scaling_type);
+  }
+
   // 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);
@@ -907,6 +931,66 @@ private:
     return DBL2NUM(ptr->params.rope_freq_scale);
   }
 
+  // yarn_ext_factor
+  static VALUE _llama_context_params_set_yarn_ext_factor(VALUE self, VALUE yarn_ext_factor) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.yarn_ext_factor = NUM2DBL(yarn_ext_factor);
+    return DBL2NUM(ptr->params.yarn_ext_factor);
+  }
+
+  static VALUE _llama_context_params_get_yarn_ext_factor(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.yarn_ext_factor);
+  }
+
+  // yarn_attn_factor
+  static VALUE _llama_context_params_set_yarn_attn_factor(VALUE self, VALUE yarn_attn_factor) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.yarn_attn_factor = NUM2DBL(yarn_attn_factor);
+    return DBL2NUM(ptr->params.yarn_attn_factor);
+  }
+
+  static VALUE _llama_context_params_get_yarn_attn_factor(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.yarn_attn_factor);
+  }
+
+  // yarn_beta_fast
+  static VALUE _llama_context_params_set_yarn_beta_fast(VALUE self, VALUE yarn_beta_fast) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.yarn_beta_fast = NUM2DBL(yarn_beta_fast);
+    return DBL2NUM(ptr->params.yarn_beta_fast);
+  }
+
+  static VALUE _llama_context_params_get_yarn_beta_fast(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.yarn_beta_fast);
+  }
+
+  // yarn_beta_slow
+  static VALUE _llama_context_params_set_yarn_beta_slow(VALUE self, VALUE yarn_beta_slow) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.yarn_beta_slow = NUM2DBL(yarn_beta_slow);
+    return DBL2NUM(ptr->params.yarn_beta_slow);
+  }
+
+  static VALUE _llama_context_params_get_yarn_beta_slow(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.yarn_beta_slow);
+  }
+
+  // yarn_orig_ctx
+  static VALUE _llama_context_params_set_yarn_orig_ctx(VALUE self, VALUE yarn_orig_ctx) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.yarn_orig_ctx = NUM2UINT(yarn_orig_ctx);
+    return UINT2NUM(ptr->params.yarn_orig_ctx);
+  }
+
+  static VALUE _llama_context_params_get_yarn_orig_ctx(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return UINT2NUM(ptr->params.yarn_orig_ctx);
+  }
+
   // mul_mat_q
   static VALUE _llama_context_params_set_mul_mat_q(VALUE self, VALUE mul_mat_q) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
@@ -1011,6 +1095,8 @@ public:
     rb_define_method(rb_cLLaMAModelQuantizeParams, "quantize_output_tensor", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_quantize_output_tensor), 0);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "only_copy=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_only_copy), 1);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "only_copy", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_only_copy), 0);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "pure=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_pure), 1);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "pure", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_pure), 0);
   }
 
 private:
@@ -1083,6 +1169,18 @@ private:
     LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
     return ptr->params.only_copy ? Qtrue : Qfalse;
   }
+
+  // pure
+  static VALUE _llama_model_quantize_params_set_pure(VALUE self, VALUE pure) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    ptr->params.pure = RTEST(pure) ? true : false;
+    return ptr->params.pure ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_model_quantize_params_get_pure(VALUE self) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    return ptr->params.pure ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAModelQuantizeParams::llama_model_quantize_params_type = {
@@ -1148,6 +1246,16 @@ public:
     rb_define_method(rb_cLLaMAModel, "desc", RUBY_METHOD_FUNC(_llama_model_get_model_desc), 0);
     rb_define_method(rb_cLLaMAModel, "size", RUBY_METHOD_FUNC(_llama_model_get_model_size), 0);
     rb_define_method(rb_cLLaMAModel, "n_params", RUBY_METHOD_FUNC(_llama_model_get_model_n_params), 0);
+    rb_define_method(rb_cLLaMAModel, "text", RUBY_METHOD_FUNC(_llama_model_get_text), 1);
+    rb_define_method(rb_cLLaMAModel, "score", RUBY_METHOD_FUNC(_llama_model_get_score), 1);
+    rb_define_method(rb_cLLaMAModel, "type", RUBY_METHOD_FUNC(_llama_model_get_type), 1);
+    rb_define_method(rb_cLLaMAModel, "token_bos", RUBY_METHOD_FUNC(_llama_model_token_bos), 0);
+    rb_define_method(rb_cLLaMAModel, "token_eos", RUBY_METHOD_FUNC(_llama_model_token_eos), 0);
+    rb_define_method(rb_cLLaMAModel, "token_nl", RUBY_METHOD_FUNC(_llama_model_token_nl), 0);
+    rb_define_method(rb_cLLaMAModel, "token_prefix", RUBY_METHOD_FUNC(_llama_model_token_prefix), 0);
+    rb_define_method(rb_cLLaMAModel, "token_middle", RUBY_METHOD_FUNC(_llama_model_token_middle), 0);
+    rb_define_method(rb_cLLaMAModel, "token_suffix", RUBY_METHOD_FUNC(_llama_model_token_suffix), 0);
+    rb_define_method(rb_cLLaMAModel, "token_eot", RUBY_METHOD_FUNC(_llama_model_token_eot), 0);
   }
 
 private:
@@ -1396,6 +1504,62 @@ private:
     LLaMAModelWrapper* ptr = get_llama_model(self);
     return UINT2NUM(llama_model_n_params(ptr->model));
   }
+
+  static VALUE _llama_model_get_text(VALUE self, VALUE token_) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const llama_token token = NUM2INT(token_);
+    const char* text = llama_token_get_text(ptr->model, token);
+    return rb_utf8_str_new_cstr(text);
+  }
+
+  static VALUE _llama_model_get_score(VALUE self, VALUE token_) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const llama_token token = NUM2INT(token_);
+    const float score = llama_token_get_score(ptr->model, token);
+    return DBL2NUM(score);
+  }
+
+  static VALUE _llama_model_get_type(VALUE self, VALUE token_) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    const llama_token token = NUM2INT(token_);
+    const int type = llama_token_get_type(ptr->model, token);
+    return INT2NUM(type);
+  }
+
+  static VALUE _llama_model_token_bos(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_bos(ptr->model));
+  }
+
+  static VALUE _llama_model_token_eos(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_eos(ptr->model));
+  }
+
+  static VALUE _llama_model_token_nl(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_nl(ptr->model));
+  }
+
+  static VALUE _llama_model_token_prefix(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_prefix(ptr->model));
+  }
+
+  static VALUE _llama_model_token_middle(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_middle(ptr->model));
+  }
+
+  static VALUE _llama_model_token_suffix(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_suffix(ptr->model));
+  }
+
+  static VALUE _llama_model_token_eot(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_token_eot(ptr->model));
+  }
 };
 
 const rb_data_type_t RbLLaMAModel::llama_model_type = {
@@ -1670,22 +1834,12 @@ public:
     rb_define_method(rb_cLLaMAContext, "decode", RUBY_METHOD_FUNC(_llama_context_decode), 1);
     rb_define_method(rb_cLLaMAContext, "logits", RUBY_METHOD_FUNC(_llama_context_logits), 0);
     rb_define_method(rb_cLLaMAContext, "embeddings", RUBY_METHOD_FUNC(_llama_context_embeddings), 0);
-    rb_define_method(rb_cLLaMAContext, "text", RUBY_METHOD_FUNC(_llama_context_text), 1);
-    rb_define_method(rb_cLLaMAContext, "score", RUBY_METHOD_FUNC(_llama_context_score), 1);
-    rb_define_method(rb_cLLaMAContext, "type", RUBY_METHOD_FUNC(_llama_context_type), 1);
-    rb_define_method(rb_cLLaMAContext, "token_bos", RUBY_METHOD_FUNC(_llama_context_token_bos), 0);
-    rb_define_method(rb_cLLaMAContext, "token_eos", RUBY_METHOD_FUNC(_llama_context_token_eos), 0);
-    rb_define_method(rb_cLLaMAContext, "token_nl", RUBY_METHOD_FUNC(_llama_context_token_nl), 0);
-    rb_define_method(rb_cLLaMAContext, "token_prefix", RUBY_METHOD_FUNC(_llama_context_token_prefix), 0);
-    rb_define_method(rb_cLLaMAContext, "token_middle", RUBY_METHOD_FUNC(_llama_context_token_middle), 0);
-    rb_define_method(rb_cLLaMAContext, "token_suffix", RUBY_METHOD_FUNC(_llama_context_token_suffix), 0);
-    rb_define_method(rb_cLLaMAContext, "token_eot", RUBY_METHOD_FUNC(_llama_context_token_eot), 0);
     rb_define_method(rb_cLLaMAContext, "n_ctx", RUBY_METHOD_FUNC(_llama_context_n_ctx), 0);
     rb_define_method(rb_cLLaMAContext, "timings", RUBY_METHOD_FUNC(_llama_context_get_timings), 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, "kv_cache_token_count", RUBY_METHOD_FUNC(_llama_context_kv_cache_token_count), 0);
-    rb_define_method(rb_cLLaMAContext, "kv_cache_tokens_rm", RUBY_METHOD_FUNC(_llama_context_kv_cache_tokens_rm), 2);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_clear", RUBY_METHOD_FUNC(_llama_context_kv_cache_clear), 0);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_rm", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_rm), 3);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_cp", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_cp), 4);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_keep", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_keep), 1);
@@ -1693,12 +1847,12 @@ public:
     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);
     rb_define_method(rb_cLLaMAContext, "save_session_file", RUBY_METHOD_FUNC(_llama_context_save_session_file), -1);
-    rb_define_method(rb_cLLaMAContext, "sample_repetition_penalty", RUBY_METHOD_FUNC(_llama_context_sample_repetition_penalty), -1);
-    rb_define_method(rb_cLLaMAContext, "sample_frequency_and_presence_penalties", RUBY_METHOD_FUNC(_llama_context_sample_frequency_and_presence_penalties), -1);
+    rb_define_method(rb_cLLaMAContext, "sample_repetition_penalties", RUBY_METHOD_FUNC(_llama_context_sample_repetition_penalties), -1);
     rb_define_method(rb_cLLaMAContext, "sample_classifier_free_guidance", RUBY_METHOD_FUNC(_llama_context_sample_classifier_free_guidance), -1);
     rb_define_method(rb_cLLaMAContext, "sample_softmax", RUBY_METHOD_FUNC(_llama_context_sample_softmax), 1);
     rb_define_method(rb_cLLaMAContext, "sample_top_k", RUBY_METHOD_FUNC(_llama_context_sample_top_k), -1);
     rb_define_method(rb_cLLaMAContext, "sample_top_p", RUBY_METHOD_FUNC(_llama_context_sample_top_p), -1);
+    rb_define_method(rb_cLLaMAContext, "sample_min_p", RUBY_METHOD_FUNC(_llama_context_sample_min_p), -1);
     rb_define_method(rb_cLLaMAContext, "sample_tail_free", RUBY_METHOD_FUNC(_llama_context_sample_tail_free), -1);
     rb_define_method(rb_cLLaMAContext, "sample_typical", RUBY_METHOD_FUNC(_llama_context_sample_typical), -1);
     rb_define_method(rb_cLLaMAContext, "sample_temp", RUBY_METHOD_FUNC(_llama_context_sample_temp), -1);
@@ -1927,102 +2081,6 @@ private:
     return output;
   }
 
-  static VALUE _llama_context_text(VALUE self, VALUE token_) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    const llama_token token = NUM2INT(token_);
-    const char* text = llama_token_get_text(ptr->ctx, token);
-    return rb_utf8_str_new_cstr(text);
-  }
-
-  static VALUE _llama_context_score(VALUE self, VALUE token_) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    const llama_token token = NUM2INT(token_);
-    const float score = llama_token_get_score(ptr->ctx, token);
-    return DBL2NUM(score);
-  }
-
-  static VALUE _llama_context_type(VALUE self, VALUE token_) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    const llama_token token = NUM2INT(token_);
-    const int type = llama_token_get_type(ptr->ctx, token);
-    return INT2NUM(type);
-  }
-
-  static VALUE _llama_context_token_bos(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_bos(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_eos(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_eos(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_nl(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_nl(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_prefix(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_prefix(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_middle(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_middle(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_suffix(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_suffix(ptr->ctx));
-  }
-
-  static VALUE _llama_context_token_eot(VALUE self) {
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    return INT2NUM(llama_token_eot(ptr->ctx));
-  }
-
   static VALUE _llama_context_n_ctx(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2073,13 +2131,13 @@ private:
     return INT2NUM(llama_get_kv_cache_token_count(ptr->ctx));
   }
 
-  static VALUE _llama_context_kv_cache_tokens_rm(VALUE self, VALUE c0, VALUE c1) {
+  static VALUE _llama_context_kv_cache_clear(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
       return Qnil;
     }
-    llama_kv_cache_tokens_rm(ptr->ctx, NUM2INT(c0), NUM2INT(c1));
+    llama_kv_cache_clear(ptr->ctx);
     return Qnil;
   }
 
@@ -2231,14 +2289,14 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_context_sample_repetition_penalty(int argc, VALUE* argv, VALUE self) {
+  static VALUE _llama_context_sample_repetition_penalties(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
-    ID kw_table[1] = { rb_intern("penalty") };
-    VALUE kw_values[1] = { Qundef };
+    ID kw_table[3] = { rb_intern("penalty_repeat"), rb_intern("penalty_freq"), rb_intern("penalty_present") };
+    VALUE kw_values[3] = { Qundef, Qundef, Qundef };
     VALUE candidates = Qnil;
     VALUE last_n_tokens = Qnil;
     rb_scan_args(argc, argv, "2:", &candidates, &last_n_tokens, &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
+    rb_get_kwargs(kw_args, kw_table, 3, 0, kw_values);
 
     if (!rb_obj_is_kind_of(candidates, rb_cLLaMATokenDataArray)) {
       rb_raise(rb_eArgError, "candidates must be a TokenDataArray");
@@ -2249,56 +2307,15 @@ private:
       return Qnil;
     }
     if (!RB_FLOAT_TYPE_P(kw_values[0])) {
-      rb_raise(rb_eArgError, "penalty must be a float");
-      return Qnil;
-    }
-
-    const size_t last_tokens_size = RARRAY_LEN(last_n_tokens);
-    std::vector<llama_token> last_n_tokens_data(last_tokens_size);
-    for (size_t i = 0; i < last_tokens_size; i++) {
-      last_n_tokens_data[i] = NUM2INT(rb_ary_entry(last_n_tokens, i));
-    }
-
-    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
-    if (ctx_ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      rb_raise(rb_eArgError, "penalty_repeat must be a float");
       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;
-    }
-    const float penalty = NUM2DBL(kw_values[0]);
-
-    llama_sample_repetition_penalty(ctx_ptr->ctx, &(cnd_ptr->array), last_n_tokens_data.data(), last_tokens_size, penalty);
-
-    return Qnil;
-  }
-
-  static VALUE _llama_context_sample_frequency_and_presence_penalties(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[2] = { rb_intern("frequency"), rb_intern("presence") };
-    VALUE kw_values[2] = { Qundef, Qundef };
-    VALUE candidates = Qnil;
-    VALUE last_n_tokens = Qnil;
-    rb_scan_args(argc, argv, "2:", &candidates, &last_n_tokens, &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
-
-    if (!rb_obj_is_kind_of(candidates, rb_cLLaMATokenDataArray)) {
-      rb_raise(rb_eArgError, "candidates must be a TokenDataArray");
-      return Qnil;
-    }
-    if (!RB_TYPE_P(last_n_tokens, T_ARRAY)) {
-      rb_raise(rb_eArgError, "last_n_tokens must be an Array");
-      return Qnil;
-    }
-    if (!RB_FLOAT_TYPE_P(kw_values[0])) {
-      rb_raise(rb_eArgError, "frequency must be a float");
+    if (!RB_FLOAT_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "penalty_freq must be a float");
       return Qnil;
     }
-    if (!RB_FLOAT_TYPE_P(kw_values[1])) {
-      rb_raise(rb_eArgError, "presence must be a float");
+    if (!RB_FLOAT_TYPE_P(kw_values[2])) {
+      rb_raise(rb_eArgError, "penalty_present must be a float");
       return Qnil;
     }
 
@@ -2318,11 +2335,12 @@ private:
       rb_raise(rb_eRuntimeError, "TokenDataArray is empty");
       return Qnil;
     }
+    const float penalty_repeat = NUM2DBL(kw_values[0]);
+    const float penalty_freq = NUM2DBL(kw_values[1]);
+    const float penalty_present = NUM2DBL(kw_values[2]);
 
-    const float alpha_frequency = NUM2DBL(kw_values[0]);
-    const float alpha_presence = NUM2DBL(kw_values[1]);
-
-    llama_sample_frequency_and_presence_penalties(ctx_ptr->ctx, &(cnd_ptr->array), last_n_tokens_data.data(), last_tokens_size, alpha_frequency, alpha_presence);
+    llama_sample_repetition_penalties(ctx_ptr->ctx, &(cnd_ptr->array), last_n_tokens_data.data(), last_tokens_size,
+        penalty_repeat, penalty_freq, penalty_present);
 
     return Qnil;
   }
@@ -2467,6 +2485,45 @@ private:
     return Qnil;
   }
 
+  static VALUE _llama_context_sample_min_p(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("prob"), rb_intern("min_keep") };
+    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, 1, 1, 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_FLOAT_TYPE_P(kw_values[0])) {
+      rb_raise(rb_eArgError, "prob must be a float");
+      return Qnil;
+    }
+    if (kw_values[1] != Qundef && !RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "min_keep 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;
+    }
+    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;
+    }
+    const float prob = NUM2DBL(kw_values[0]);
+    const size_t min_keep = kw_values[1] != Qundef ? NUM2SIZET(kw_values[1]) : 1;
+
+    llama_sample_min_p(ctx_ptr->ctx, &(cnd_ptr->array), prob, min_keep);
+
+    return Qnil;
+  }
+
   static VALUE _llama_context_sample_tail_free(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
     ID kw_table[2] = { rb_intern("z"), rb_intern("min_keep") };
@@ -2962,6 +3019,12 @@ extern "C" void Init_llama_cpp(void) {
   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));
 
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_UNSPECIFIED", INT2NUM(LLAMA_ROPE_SCALING_UNSPECIFIED));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_NONE", INT2NUM(LLAMA_ROPE_SCALING_NONE));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_LINEAR", INT2NUM(LLAMA_ROPE_SCALING_LINEAR));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_YARN", INT2NUM(LLAMA_ROPE_SCALING_YARN));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_MAX_VALUE", INT2NUM(LLAMA_ROPE_SCALING_MAX_VALUE));
+
   std::stringstream ss_magic;
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGSN;
   rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGSN", rb_str_new2(ss_magic.str().c_str()));