llama_cpp 0.9.0 → 0.9.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 683f2d81aff9e82234925ba08cd5b46b56a2283ff8397a6c06ce50d34a95dbfc
-  data.tar.gz: d3005cab273b8d85f47f4cb4314fbab3a540d366a42829e5ec8d2c29576ae09e
+  metadata.gz: dae7507ce41f18e3fd0fb2d7445275a387a3914068aa9eef922f260de699970a
+  data.tar.gz: d66cc2629aeca3285bc10988f8c410fb8cf5b7f1fe6f835b5dc60e9dcab4be9d
 SHA512:
-  metadata.gz: 559f1ba1253a704c38480336decd315c65b4d80e6895ad1dc0faa3b5b81570a1faeaadcb6ec7ee3145f0fff758ab5e38e6cb8163382ce9b693d893deebe9a8f9
-  data.tar.gz: cb3d96b8c3f79cd20d4169a175270e8768c04bcaa24e51cb2c4d7872db88bc6e3349e6b1e93a130b89d21daab8be6e57b5305412059ea722084c7cb7d4a01e93
+  metadata.gz: 3e3e92aa38413877620947ec7996494cd720a3c211fcdf1973ce0d7a9a7e8803e293e2ce2f601b11e35858c5b4ef6b00d716069e322ea8d6b4c93412990fd746
+  data.tar.gz: 20a1e9e0e5812da9b00787afbf0f3aa0b762c8168f54ce3b7f2f25ff5b61cca5b2e7ab5faa065fbc3e266468d1c5747b8e0779fc7e073cc66240d1f3085e71c7

data/CHANGELOG.md

@@ -1,3 +1,12 @@
+## [[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.

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-backend.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-backend.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 = {
@@ -1741,7 +1839,7 @@ public:
     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);
@@ -1754,6 +1852,7 @@ public:
     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);
@@ -2032,13 +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;
   }
 
@@ -2386,6 +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") };
@@ -2881,6 +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()));