llama_cpp 0.12.7 → 0.14.0

data/ext/llama_cpp/llama_cpp.cpp

@@ -952,6 +952,8 @@ public:
     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, "pooling_type=", RUBY_METHOD_FUNC(_llama_context_params_set_pooling_type), 1);
+    rb_define_method(rb_cLLaMAContextParams, "pooling_type", RUBY_METHOD_FUNC(_llama_context_params_get_pooling_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);
@@ -966,20 +968,18 @@ public:
     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, "defrag_thold=", RUBY_METHOD_FUNC(_llama_context_params_set_defrag_thold), 1);
+    rb_define_method(rb_cLLaMAContextParams, "defrag_thold", RUBY_METHOD_FUNC(_llama_context_params_get_defrag_thold), 0);
     rb_define_method(rb_cLLaMAContextParams, "type_k=", RUBY_METHOD_FUNC(_llama_context_params_set_type_k), 1);
     rb_define_method(rb_cLLaMAContextParams, "type_k", RUBY_METHOD_FUNC(_llama_context_params_get_type_k), 0);
     rb_define_method(rb_cLLaMAContextParams, "type_v=", RUBY_METHOD_FUNC(_llama_context_params_set_type_v), 1);
     rb_define_method(rb_cLLaMAContextParams, "type_v", RUBY_METHOD_FUNC(_llama_context_params_get_type_v), 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, "logits_all=", RUBY_METHOD_FUNC(_llama_context_params_set_logits_all), 1);
     rb_define_method(rb_cLLaMAContextParams, "logits_all", RUBY_METHOD_FUNC(_llama_context_params_get_logits_all), 0);
-    rb_define_method(rb_cLLaMAContextParams, "embedding=", RUBY_METHOD_FUNC(_llama_context_params_set_embedding), 1);
-    rb_define_method(rb_cLLaMAContextParams, "embedding", RUBY_METHOD_FUNC(_llama_context_params_get_embedding), 0);
+    rb_define_method(rb_cLLaMAContextParams, "embeddings=", RUBY_METHOD_FUNC(_llama_context_params_set_embeddings), 1);
+    rb_define_method(rb_cLLaMAContextParams, "embeddings", RUBY_METHOD_FUNC(_llama_context_params_get_embeddings), 0);
     rb_define_method(rb_cLLaMAContextParams, "offload_kqv=", RUBY_METHOD_FUNC(_llama_context_params_set_offload_kqv), 1);
     rb_define_method(rb_cLLaMAContextParams, "offload_kqv", RUBY_METHOD_FUNC(_llama_context_params_get_offload_kqv), 0);
-    rb_define_method(rb_cLLaMAContextParams, "do_pooling=", RUBY_METHOD_FUNC(_llama_context_params_set_do_pooling), 1);
-    rb_define_method(rb_cLLaMAContextParams, "do_pooling", RUBY_METHOD_FUNC(_llama_context_params_get_do_pooling), 0);
   }
 
 private:
@@ -1058,7 +1058,7 @@ private:
   // 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);
+    ptr->params.rope_scaling_type = static_cast<enum llama_rope_scaling_type>(NUM2INT(scaling_type));
     return INT2NUM(ptr->params.rope_scaling_type);
   }
 
@@ -1067,6 +1067,18 @@ private:
     return INT2NUM(ptr->params.rope_scaling_type);
   }
 
+  // pooling_type
+  static VALUE _llama_context_params_set_pooling_type(VALUE self, VALUE scaling_type) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.pooling_type = static_cast<enum llama_pooling_type>(NUM2INT(scaling_type));
+    return INT2NUM(ptr->params.pooling_type);
+  }
+
+  static VALUE _llama_context_params_get_pooling_type(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return INT2NUM(ptr->params.pooling_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);
@@ -1146,6 +1158,18 @@ private:
     return UINT2NUM(ptr->params.yarn_orig_ctx);
   }
 
+  // defrag_thold
+  static VALUE _llama_context_params_set_defrag_thold(VALUE self, VALUE defrag_thold) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.defrag_thold = NUM2DBL(defrag_thold);
+    return DBL2NUM(ptr->params.defrag_thold);
+  }
+
+  static VALUE _llama_context_params_get_defrag_thold(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return DBL2NUM(ptr->params.defrag_thold);
+  }
+
   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);
@@ -1175,18 +1199,6 @@ private:
     return INT2NUM(ptr->params.type_v);
   }
 
-  // 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);
-    ptr->params.mul_mat_q = RTEST(mul_mat_q) ? true : false;
-    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
-  }
-
-  static VALUE _llama_context_params_get_mul_mat_q(VALUE self) {
-    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return ptr->params.mul_mat_q ? Qtrue : Qfalse;
-  }
-
   // logits_all
   static VALUE _llama_context_params_set_logits_all(VALUE self, VALUE logits_all) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
@@ -1199,16 +1211,16 @@ private:
     return ptr->params.logits_all ? Qtrue : Qfalse;
   }
 
-  // embedding
-  static VALUE _llama_context_params_set_embedding(VALUE self, VALUE embedding) {
+  // embeddings
+  static VALUE _llama_context_params_set_embeddings(VALUE self, VALUE embeddings) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    ptr->params.embedding = RTEST(embedding) ? true : false;
-    return ptr->params.embedding ? Qtrue : Qfalse;
+    ptr->params.embeddings = RTEST(embeddings) ? true : false;
+    return ptr->params.embeddings ? Qtrue : Qfalse;
   }
 
-  static VALUE _llama_context_params_get_embedding(VALUE self) {
+  static VALUE _llama_context_params_get_embeddings(VALUE self) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return ptr->params.embedding ? Qtrue : Qfalse;
+    return ptr->params.embeddings ? Qtrue : Qfalse;
   }
 
   // offload_kqv
@@ -1222,18 +1234,6 @@ private:
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
     return ptr->params.offload_kqv ? Qtrue : Qfalse;
   }
-
-  // do_pooling
-  static VALUE _llama_context_params_set_do_pooling(VALUE self, VALUE do_pooling) {
-    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    ptr->params.do_pooling = RTEST(do_pooling) ? true : false;
-    return ptr->params.do_pooling ? Qtrue : Qfalse;
-  }
-
-  static VALUE _llama_context_params_get_do_pooling(VALUE self) {
-    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return ptr->params.do_pooling ? Qtrue : Qfalse;
-  }
 };
 
 const rb_data_type_t RbLLaMAContextParams::llama_context_params_type = {
@@ -1433,7 +1433,8 @@ public:
     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);
+    rb_define_method(rb_cLLaMAModel, "vocab_type", RUBY_METHOD_FUNC(_llama_model_get_model_vocab_type), 0);
+    rb_define_method(rb_cLLaMAModel, "rope_type", RUBY_METHOD_FUNC(_llama_model_get_model_rope_type), 0);
     rb_define_method(rb_cLLaMAModel, "n_vocab", RUBY_METHOD_FUNC(_llama_model_get_model_n_vocab), 0);
     rb_define_method(rb_cLLaMAModel, "n_ctx_train", RUBY_METHOD_FUNC(_llama_model_get_model_n_ctx_train), 0);
     rb_define_method(rb_cLLaMAModel, "n_embd", RUBY_METHOD_FUNC(_llama_model_get_model_n_embd), 0);
@@ -1559,41 +1560,14 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_model_apply_lora_from_file(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[4] = { rb_intern("lora_path"), rb_intern("base_model_path"), rb_intern("n_threads"), rb_intern("scale") };
-    VALUE kw_values[4] = { Qundef, Qundef, Qundef, Qundef };
-    rb_scan_args(argc, argv, ":", &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 1, 3, 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;
-    }
-    if (kw_values[3] != Qundef && !RB_FLOAT_TYPE_P(kw_values[3])) {
-      rb_raise(rb_eArgError, "scale must be a float");
-      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]);
-    const float scale = kw_values[3] == Qundef ? 1.0 : NUM2DBL(kw_values[3]);
+  static VALUE _llama_model_get_model_vocab_type(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return INT2NUM(llama_vocab_type(ptr->model));
+  }
 
+  static VALUE _llama_model_get_model_rope_type(VALUE self) {
     LLaMAModelWrapper* ptr = get_llama_model(self);
-    if (llama_model_apply_lora_from_file(ptr->model, lora_path, scale, base_model_path, n_threads) != 0) {
-      rb_raise(rb_eRuntimeError, "Failed to apply LoRA");
-      return Qnil;
-    }
-    return Qnil;
+    return INT2NUM(llama_rope_type(ptr->model));
   }
 
   static VALUE _llama_model_get_model_n_vocab(VALUE self) {
@@ -2038,12 +2012,11 @@ public:
     rb_define_alloc_func(rb_cLLaMAContext, llama_context_alloc);
     rb_define_attr(rb_cLLaMAContext, "model", 1, 0);
     rb_define_method(rb_cLLaMAContext, "initialize", RUBY_METHOD_FUNC(_llama_context_initialize), -1);
-    rb_define_method(rb_cLLaMAContext, "eval", RUBY_METHOD_FUNC(_llama_context_eval), -1);
-    rb_define_method(rb_cLLaMAContext, "eval_embd", RUBY_METHOD_FUNC(_llama_context_eval_embd), -1);
     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, "embeddings_ith", RUBY_METHOD_FUNC(_llama_context_embeddings_ith), 1);
+    rb_define_method(rb_cLLaMAContext, "embeddings_seq", RUBY_METHOD_FUNC(_llama_context_embeddings_seq), 1);
     rb_define_method(rb_cLLaMAContext, "n_ctx", RUBY_METHOD_FUNC(_llama_context_n_ctx), 0);
     rb_define_method(rb_cLLaMAContext, "n_batch", RUBY_METHOD_FUNC(_llama_context_n_batch), 0);
     rb_define_method(rb_cLLaMAContext, "timings", RUBY_METHOD_FUNC(_llama_context_get_timings), 0);
@@ -2054,14 +2027,16 @@ public:
     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);
-    rb_define_method(rb_cLLaMAContext, "kv_cache_seq_shift", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_shift), 4);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_seq_add", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_add), 4);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_div", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_div), 4);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_seq_pos_max", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_pos_max), 1);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_kv_cache_defrag", RUBY_METHOD_FUNC(_llama_context_kv_cache_defrag), 0);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_kv_cache_update", RUBY_METHOD_FUNC(_llama_context_kv_cache_update), 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);
     rb_define_method(rb_cLLaMAContext, "save_session_file", RUBY_METHOD_FUNC(_llama_context_save_session_file), -1);
     rb_define_method(rb_cLLaMAContext, "sample_repetition_penalties", RUBY_METHOD_FUNC(_llama_context_sample_repetition_penalties), -1);
     rb_define_method(rb_cLLaMAContext, "sample_apply_guidance", RUBY_METHOD_FUNC(_llama_context_sample_apply_guidance), -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);
@@ -2070,7 +2045,6 @@ public:
     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);
     rb_define_method(rb_cLLaMAContext, "sample_entropy", RUBY_METHOD_FUNC(_llama_context_sample_entropy), -1);
-    rb_define_method(rb_cLLaMAContext, "sample_temperature", RUBY_METHOD_FUNC(_llama_context_sample_temperature), -1);
     rb_define_method(rb_cLLaMAContext, "sample_token_mirostat", RUBY_METHOD_FUNC(_llama_context_sample_token_mirostat), -1);
     rb_define_method(rb_cLLaMAContext, "sample_token_mirostat_v2", RUBY_METHOD_FUNC(_llama_context_sample_token_mirostat_v2), -1);
     rb_define_method(rb_cLLaMAContext, "sample_token_greedy", RUBY_METHOD_FUNC(_llama_context_sample_token_greedy), 1);
@@ -2122,110 +2096,6 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_context_eval(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[3] = { rb_intern("tokens"), rb_intern("n_past"), rb_intern("n_tokens") };
-    VALUE kw_values[3] = { Qundef, Qundef, Qundef };
-    rb_scan_args(argc, argv, ":", &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 2, 2, kw_values);
-
-    rb_warn("eval is deprecated. Use decode instead.");
-
-    if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
-      rb_raise(rb_eArgError, "tokens must be an Array");
-      return Qnil;
-    }
-    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
-      rb_raise(rb_eArgError, "n_past must be an integer");
-      return Qnil;
-    }
-    if (kw_values[2] != Qundef && !RB_INTEGER_TYPE_P(kw_values[2])) {
-      rb_raise(rb_eArgError, "n_tokens must be an integer");
-      return Qnil;
-    }
-
-    const size_t tokens_len = RARRAY_LEN(kw_values[0]);
-    std::vector<llama_token> embd(tokens_len);
-    for (size_t i = 0; i < tokens_len; i++) {
-      VALUE token = rb_ary_entry(kw_values[0], i);
-      if (!RB_INTEGER_TYPE_P(token)) {
-        rb_raise(rb_eArgError, "tokens must be an array of integers");
-        return Qnil;
-      }
-      embd[i] = NUM2INT(token);
-    }
-
-    const int n_tokens = kw_values[2] == Qundef ? (int)tokens_len : NUM2INT(kw_values[2]);
-    const int n_past = NUM2INT(kw_values[1]);
-
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    if (llama_eval(ptr->ctx, embd.data(), n_tokens, n_past) != 0) {
-      rb_raise(rb_eRuntimeError, "Failed to evaluate");
-      return Qnil;
-    }
-
-    rb_iv_set(self, "@n_tokens", INT2NUM(n_tokens));
-    rb_iv_set(self, "@has_evaluated", Qtrue);
-
-    return Qnil;
-  }
-
-  static VALUE _llama_context_eval_embd(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[3] = { rb_intern("embd"), rb_intern("n_past"), rb_intern("n_tokens") };
-    VALUE kw_values[3] = { Qundef, Qundef, Qundef };
-    rb_scan_args(argc, argv, ":", &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 2, 2, kw_values);
-
-    rb_warn("eval_embd is deprecated. Use decode instead.");
-
-    if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
-      rb_raise(rb_eArgError, "tokens must be an Array");
-      return Qnil;
-    }
-    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
-      rb_raise(rb_eArgError, "n_past must be an integer");
-      return Qnil;
-    }
-    if (kw_values[2] != Qundef && !RB_INTEGER_TYPE_P(kw_values[2])) {
-      rb_raise(rb_eArgError, "n_tokens must be an integer");
-      return Qnil;
-    }
-
-    const size_t tokens_len = RARRAY_LEN(kw_values[0]);
-    std::vector<float> embd(tokens_len);
-    for (size_t i = 0; i < tokens_len; i++) {
-      VALUE el = rb_ary_entry(kw_values[0], i);
-      if (!RB_FLOAT_TYPE_P(el)) {
-        rb_raise(rb_eArgError, "embd must be an array of floats");
-        return Qnil;
-      }
-      embd[i] = NUM2DBL(el);
-    }
-
-    const int n_tokens = kw_values[2] == Qundef ? (int)tokens_len : NUM2INT(kw_values[2]);
-    const int n_past = NUM2INT(kw_values[1]);
-
-    LLaMAContextWrapper* ptr = get_llama_context(self);
-    if (ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
-      return Qnil;
-    }
-    if (llama_eval_embd(ptr->ctx, embd.data(), n_tokens, n_past) != 0) {
-      rb_raise(rb_eRuntimeError, "Failed to evaluate");
-      return Qnil;
-    }
-
-    rb_iv_set(self, "@n_tokens", INT2NUM(n_tokens));
-    rb_iv_set(self, "@has_evaluated", Qtrue);
-
-    return Qnil;
-  }
-
   static VALUE _llama_context_decode(VALUE self, VALUE batch) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2282,7 +2152,7 @@ private:
     LLaMAModelWrapper* model_ptr = RbLLaMAModel::get_llama_model(model);
     VALUE params = rb_iv_get(self, "@params");
     LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
-    if (!prms_ptr->params.embedding) {
+    if (!prms_ptr->params.embeddings) {
       rb_raise(rb_eRuntimeError, "embedding parameter is false");
       return Qnil;
     }
@@ -2291,10 +2161,11 @@ private:
       return Qnil;
     }
 
+    const int n_tokens = NUM2INT(rb_iv_get(self, "@n_tokens"));
     const int n_embd = llama_n_embd(model_ptr->model);
     const float* embd = llama_get_embeddings(ptr->ctx);
     VALUE output = rb_ary_new();
-    for (int i = 0; i < n_embd; i++) {
+    for (int i = 0; i < n_tokens * n_embd; i++) {
       rb_ary_push(output, DBL2NUM((double)(embd[i])));
     }
 
@@ -2313,7 +2184,7 @@ private:
     }
     VALUE params = rb_iv_get(self, "@params");
     LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
-    if (!prms_ptr->params.embedding) {
+    if (!prms_ptr->params.embeddings) {
       rb_raise(rb_eRuntimeError, "embedding parameter is false");
       return Qnil;
     }
@@ -2331,6 +2202,36 @@ private:
     return output;
   }
 
+  static VALUE _llama_context_embeddings_seq(VALUE self, VALUE seq_id) {
+    if (!RB_INTEGER_TYPE_P(seq_id)) {
+      rb_raise(rb_eArgError, "seq_id must be an integer");
+      return Qnil;
+    }
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    VALUE params = rb_iv_get(self, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
+    if (!prms_ptr->params.embeddings) {
+      rb_raise(rb_eRuntimeError, "embedding parameter is false");
+      return Qnil;
+    }
+
+    VALUE model = rb_iv_get(self, "@model");
+    LLaMAModelWrapper* model_ptr = RbLLaMAModel::get_llama_model(model);
+    const int n_embd = llama_n_embd(model_ptr->model);
+
+    VALUE output = rb_ary_new();
+    const float* embd = llama_get_embeddings_seq(ptr->ctx, NUM2INT(seq_id));
+    for (int i = 0; i < n_embd; i++) {
+      rb_ary_push(output, DBL2NUM((double)(embd[i])));
+    }
+
+    return output;
+  }
+
   static VALUE _llama_context_n_ctx(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2430,13 +2331,13 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_context_kv_cache_seq_shift(VALUE self, VALUE seq_id, VALUE p0, VALUE p1, VALUE delta) {
+  static VALUE _llama_context_kv_cache_seq_add(VALUE self, VALUE seq_id, VALUE p0, VALUE p1, VALUE delta) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
       rb_raise(rb_eArgError, "LLaMA context is not initialized");
       return Qnil;
     }
-    llama_kv_cache_seq_shift(ptr->ctx, NUM2INT(seq_id), NUM2INT(p0), NUM2INT(p1), NUM2INT(delta));
+    llama_kv_cache_seq_add(ptr->ctx, NUM2INT(seq_id), NUM2INT(p0), NUM2INT(p1), NUM2INT(delta));
     return Qnil;
   }
 
@@ -2450,6 +2351,35 @@ private:
     return Qnil;
   }
 
+  static VALUE _llama_context_kv_cache_seq_pos_max(VALUE self, VALUE seq_id) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eArgError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    return INT2NUM(llama_kv_cache_seq_pos_max(ptr->ctx, NUM2INT(seq_id)));
+  }
+
+  static VALUE _llama_context_kv_cache_defrag(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_defrag(ptr->ctx);
+    return Qnil;
+  }
+
+  static VALUE _llama_context_kv_cache_update(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_update(ptr->ctx);
+    return Qnil;
+  }
+
   static VALUE _llama_context_set_rng_seed(VALUE self, VALUE seed_) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2659,46 +2589,6 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_context_sample_classifier_free_guidance(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[2] = { rb_intern("guidance"), rb_intern("scale") };
-    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, 2, 0, kw_values);
-
-    if (!rb_obj_is_kind_of(kw_values[0], rb_cLLaMAContext)) {
-      rb_raise(rb_eArgError, "guidance must be a Context");
-      return Qnil;
-    }
-    if (!RB_FLOAT_TYPE_P(kw_values[1])) {
-      rb_raise(rb_eArgError, "scale must be a float");
-      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;
-    }
-
-    LLaMAContextWrapper* guidance_ptr = get_llama_context(kw_values[0]);
-    if (guidance_ptr->ctx == NULL) {
-      rb_raise(rb_eRuntimeError, "guidance context is not initialized");
-      return Qnil;
-    }
-    const float scale = NUM2DBL(kw_values[1]);
-
-    llama_sample_classifier_free_guidance(ctx_ptr->ctx, &(cnd_ptr->array), guidance_ptr->ctx, scale);
-
-    return Qnil;
-  }
-
   static VALUE _llama_context_sample_softmax(VALUE self, VALUE candidates) {
     if (!rb_obj_is_kind_of(candidates, rb_cLLaMATokenDataArray)) {
       rb_raise(rb_eArgError, "argument must be a TokenDataArray");
@@ -2994,42 +2884,6 @@ private:
     return Qnil;
   }
 
-  static VALUE _llama_context_sample_temperature(int argc, VALUE* argv, VALUE self) {
-    VALUE kw_args = Qnil;
-    ID kw_table[1] = { rb_intern("temperature") };
-    VALUE kw_values[1] = { Qundef };
-    VALUE candidates = Qnil;
-    rb_scan_args(argc, argv, "1:", &candidates, &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
-
-    rb_warn("sample_temperature is deprecated. Use sample_temp instead.");
-
-    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, "temperature must be a float");
-      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 temperature = NUM2DBL(kw_values[0]);
-
-    llama_sample_temperature(ctx_ptr->ctx, &(cnd_ptr->array), temperature);
-
-    return Qnil;
-  }
-
   static VALUE _llama_context_sample_token_mirostat(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
     ID kw_table[4] = { rb_intern("tau"), rb_intern("eta"), rb_intern("m"), rb_intern("mu") };
@@ -3307,16 +3161,6 @@ static VALUE rb_llama_time_us(VALUE self) {
   return LONG2NUM(llama_time_us());
 }
 
-static VALUE rb_llama_mmap_supported(VALUE self) {
-  rb_warn("mmap_supported? is deprecated. Use supports_mmap? instead.");
-  return llama_mmap_supported() ? Qtrue : Qfalse;
-}
-
-static VALUE rb_llama_mlock_supported(VALUE self) {
-  rb_warn("mlock_supported? is deprecated. Use supports_mlock? instead.");
-  return llama_mlock_supported() ? Qtrue : Qfalse;
-}
-
 static VALUE rb_llama_max_devices(VALUE self) {
   return SIZET2NUM(llama_max_devices());
 }
@@ -3355,8 +3199,6 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_module_function(rb_mLLaMACpp, "model_quantize", rb_llama_model_quantize, -1);
   rb_define_module_function(rb_mLLaMACpp, "print_system_info", rb_llama_print_system_info, 0);
   rb_define_module_function(rb_mLLaMACpp, "time_us", rb_llama_time_us, 0);
-  rb_define_module_function(rb_mLLaMACpp, "mmap_supported?", rb_llama_mmap_supported, 0);
-  rb_define_module_function(rb_mLLaMACpp, "mlock_supported?", rb_llama_mlock_supported, 0);
   rb_define_module_function(rb_mLLaMACpp, "max_devices", rb_llama_max_devices, 0);
   rb_define_module_function(rb_mLLaMACpp, "supports_mmap?", rb_llama_supports_mmap, 0);
   rb_define_module_function(rb_mLLaMACpp, "supports_mlock?", rb_llama_supports_mlock, 0);
@@ -3394,16 +3236,16 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ2_XXS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ2_XXS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ2_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ2_XS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q2_K_S", INT2NUM(LLAMA_FTYPE_MOSTLY_Q2_K_S));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q3_K_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_Q3_K_XS));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_XS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_XXS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_XXS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ1_S", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ1_S));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ4_NL", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ4_NL));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 
-  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_INT", INT2NUM(LLAMA_KV_OVERRIDE_INT));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_FLOAT", INT2NUM(LLAMA_KV_OVERRIDE_FLOAT));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_BOOL", INT2NUM(LLAMA_KV_OVERRIDE_BOOL));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_TYPE_INT", INT2NUM(LLAMA_KV_OVERRIDE_TYPE_INT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_TYPE_FLOAT", INT2NUM(LLAMA_KV_OVERRIDE_TYPE_FLOAT));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_KV_OVERRIDE_TYPE_BOOL", INT2NUM(LLAMA_KV_OVERRIDE_TYPE_BOOL));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_END", INT2NUM(LLAMA_GRETYPE_END));
   rb_define_const(rb_mLLaMACpp, "LLAMA_GRETYPE_ALT", INT2NUM(LLAMA_GRETYPE_ALT));
@@ -3413,19 +3255,20 @@ 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));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED", INT2NUM(LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_TYPE_NONE", INT2NUM(LLAMA_ROPE_SCALING_TYPE_NONE));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_TYPE_LINEAR", INT2NUM(LLAMA_ROPE_SCALING_TYPE_LINEAR));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_TYPE_YARN", INT2NUM(LLAMA_ROPE_SCALING_TYPE_YARN));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_TYPE_MAX_VALUE", INT2NUM(LLAMA_ROPE_SCALING_TYPE_MAX_VALUE));
 
-  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_NONE", INT2NUM(LLAMA_POOLING_NONE));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_MEAN", INT2NUM(LLAMA_POOLING_MEAN));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_CLS", INT2NUM(LLAMA_POOLING_CLS));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_TYPE_UNSPECIFIED", INT2NUM(LLAMA_POOLING_TYPE_UNSPECIFIED));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_TYPE_NONE", INT2NUM(LLAMA_POOLING_TYPE_NONE));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_TYPE_MEAN", INT2NUM(LLAMA_POOLING_TYPE_MEAN));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_TYPE_CLS", INT2NUM(LLAMA_POOLING_TYPE_CLS));
 
-  rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_NONE", INT2NUM(LLAMA_SPLIT_NONE));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_LAYER", INT2NUM(LLAMA_SPLIT_LAYER));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_ROW", INT2NUM(LLAMA_SPLIT_ROW));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_MODE_NONE", INT2NUM(LLAMA_SPLIT_MODE_NONE));
+  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));
 
   std::stringstream ss_magic;
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGLA;

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 = '0.12.7'
+  VERSION = '0.14.0'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2249'
+  LLAMA_CPP_VERSION = 'b2361'
 end