llama_cpp 0.0.2 → 0.0.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e4b9a70ca3137fb187c1455291828001086e373db7d9189f7f8d45f0d252b0dc
-  data.tar.gz: 22d67fa3d1c71d73569735876aebe953038bb0465a67b07ea991dc8568d11bac
+  metadata.gz: 0e659b4cc27e9ab45b524ec3d341892b72cebdf84ccad823f24ff7e472f2ffa8
+  data.tar.gz: 790f4d2f6dc9ddf211701f6014ae91ca19e0492efd7c64eb881e66729f929544
 SHA512:
-  metadata.gz: 3767e7950004aba7980a27dbffaec2c360a18295e845b58ab647eff4b9f90515e47c646e48e5d75cac261908415602df50908e429fca0637664e93b2efd7dc1a
-  data.tar.gz: b08e00960ab036fe7ac7778dd33a5a72795153cd7c8beea642b5422da41575a19ea41e1b865e25d16f36afe2879ff4b5b3f303d49598c30888a95ecf459501da
+  metadata.gz: 54eb4dd6c4ca9affc7b6f2da1c9dc719f60c8dc3841645cf47b8f0310ff31ad16a5bc841051663f03e962bc7424f56f1d7a1c5c5eabd03e3f5e7b706467bb0ba
+  data.tar.gz: b2a16c6be3d7f117fabea5cc837b9306b0768d9ad99104a6fa2752932d1e1a034312983a87ebfe9e3ccb1bf83257d5ce40520e049a40291c64fb2fab8663882a

data/CHANGELOG.md

@@ -1,8 +1,33 @@
 ## [Unreleased]
 
-## [0.0.2] - 2023-04-02
+## [[0.0.4](https://github.com/yoshoku/llama_cpp.rb/compare/v0.0.3...v0.0.4)] - 2023-04-15
 
-- Bump bundled llama.cpp from master-2a98bc1 to master-5b70e7d
+- Bump bundled llama.cpp from master-698f7b5 to master-c85e03d.
+- Add parameterless constructor to LLaMACpp::Context.
+- Add free and load methods to LLaMACpp::Context.
+  ```ruby
+  require 'llama_cpp'
+
+  context = LLaMACpp::Context.new
+
+  params = LLaMACpp::ContextParams.new
+  context.load(model_path: '/path/to/ggml-model-q4_0.bin', params: params)
+
+  # ...
+
+  context.free
+  ```
+
+## [[0.0.3](https://github.com/yoshoku/llama_cpp.rb/compare/v0.0.2...v0.0.3)] - 2023-04-08
+
+- Bump bundled llama.cpp from master-5b70e7d to master-698f7b5.
+- Add logits method to LLaMACpp::Context.
+- Add type signatures.
+- Add class alias Params for LLaMACpp::ContextParams.
+
+## [[0.0.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.0.1...v0.0.2)] - 2023-04-02
+
+- Bump bundled llama.cpp from master-2a98bc1 to master-5b70e7d.
 - Add n_threads arguments to generate method.
 
 ## [0.0.1] - 2023-04-02

data/README.md

@@ -26,11 +26,12 @@ Prepare a quantized model file by refering to [the usage section on the llama.cp
 require 'llama_cpp'
 
 params = LLaMACpp::ContextParams.new
-params.seed = 123456
+params.seed = 12
 
 context = LLaMACpp::Context.new(model_path: '/path/to/ggml-model-q4_0.bin', params: params)
 
-puts LLaMACpp.generate(context, 'Please tell me the largest city in Japan.')
+puts LLaMACpp.generate(context, 'Please tell me the largest city in Japan.', n_threads: 4)
+# => "There are two major cities in Japan, Tokyo and Osaka, which have about 30 million populations."
 ```
 
 ## Contributing

data/ext/llama_cpp/extconf.rb

@@ -10,4 +10,30 @@ $CXXFLAGS << ' -std=c++11'
 $INCFLAGS << ' -I$(srcdir)/src'
 $VPATH << '$(srcdir)/src'
 
+if RUBY_PLATFORM.match?(/darwin|linux|bsd/) && try_compile('#include <stdio.h>', '-pthread')
+  $CFLAGS << ' -pthread'
+  $CXXFLAGS << ' -pthread'
+end
+
+UNAME_M = RbConfig::CONFIG['build_cpu'] || RbConfig::CONFIG['host_cpu'] || RbConfig::CONFIG['target_cpu']
+
+# rubocop:disable Layout/LineLength
+if UNAME_M.match?(/x86_64|i686/) && try_compile('#include <stdio.h>', '-march=native -mtune=native')
+  $CFLAGS << ' -march=native -mtune=native'
+  $CXXFLAGS << ' -march=native -mtune=native'
+elsif UNAME_M.match?(/aarch64/) && try_compile('#include <stdio.h>', '-mcpu=native')
+  $CFLAGS << ' -mcpu=native'
+  $CXXFLAGS << ' -mcpu=native'
+elsif UNAME_M.match?(/armv6/) && try_compile('#include <stdio.h>', '-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access')
+  $CFLAGS << ' -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access'
+  $CXXFLAGS << ' -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access'
+elsif UNAME_M.match?(/armv7/) && try_compile('#include <stdio.h>', '-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations')
+  $CFLAGS << ' -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations'
+  $CXXFLAGS << ' -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations'
+elsif UNAME_M.match?(/armv8/) && try_compile('#include <stdio.h>', '-mfp16-format=ieee -mno-unaligned-access')
+  $CFLAGS << ' -mfp16-format=ieee -mno-unaligned-access'
+  $CXXFLAGS << ' -mfp16-format=ieee -mno-unaligned-access'
+end
+# rubocop:enable Layout/LineLength
+
 create_makefile('llama_cpp/llama_cpp')

data/ext/llama_cpp/llama_cpp.cpp

@@ -217,7 +217,7 @@ public:
     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, "tokenize", RUBY_METHOD_FUNC(_llama_context_tokenize), -1);
-    // rb_define_method(rb_cLLaMAContext, "logits", RUBY_METHOD_FUNC(_llama_context_logits), 0);
+    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, "token_to_str", RUBY_METHOD_FUNC(_llama_context_token_to_str), 1);
     rb_define_method(rb_cLLaMAContext, "sample_top_p_top_k", RUBY_METHOD_FUNC(_llama_context_sample_top_p_top_k), -1);
@@ -226,6 +226,8 @@ public:
     rb_define_method(rb_cLLaMAContext, "n_embd", RUBY_METHOD_FUNC(_llama_context_n_embd), 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, "free", RUBY_METHOD_FUNC(_llama_context_free), 0);
+    rb_define_method(rb_cLLaMAContext, "load", RUBY_METHOD_FUNC(_llama_context_load), -1);
   };
 
 private:
@@ -236,7 +238,13 @@ private:
     ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
     VALUE kw_values[2] = { Qundef, Qundef };
     rb_scan_args(argc, argv, ":", &kw_args);
-    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
+    rb_get_kwargs(kw_args, kw_table, 0, 2, kw_values);
+
+    if (kw_values[0] == Qundef && kw_values[1] == Qundef) {
+      rb_iv_set(self, "@params", Qnil);
+      rb_iv_set(self, "@has_evaluated", Qfalse);
+      return Qnil;
+    }
 
     if (!RB_TYPE_P(kw_values[0], T_STRING)) {
       rb_raise(rb_eArgError, "model_path must be a string");
@@ -256,8 +264,11 @@ private:
       return Qnil;
     }
 
+    rb_iv_set(self, "@params", kw_values[1]);
+    rb_iv_set(self, "@has_evaluated", Qfalse);
+
     RB_GC_GUARD(filename);
-    return self;
+    return Qnil;
   };
 
   static VALUE _llama_context_eval(int argc, VALUE* argv, VALUE self) {
@@ -305,6 +316,9 @@ private:
       return Qnil;
     }
 
+    rb_iv_set(self, "@n_tokens", INT2NUM(n_tokens));
+    rb_iv_set(self, "@has_evaluated", Qtrue);
+
     return Qnil;
   };
 
@@ -361,12 +375,44 @@ private:
     return str != nullptr ? rb_utf8_str_new_cstr(str) : rb_utf8_str_new_cstr("");
   };
 
+  static VALUE _llama_context_logits(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    if (rb_iv_get(self, "@has_evaluated") != Qtrue) {
+      rb_raise(rb_eRuntimeError, "LLaMA context has not been evaluated");
+      return Qnil;
+    }
+
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(rb_iv_get(self, "@params"));
+    const int n_tokens = prms_ptr->params.logits_all ? NUM2INT(rb_iv_get(self, "@n_tokens")) : 1;
+    const int n_vocab = llama_n_vocab(ptr->ctx);
+    const float* logits = llama_get_logits(ptr->ctx);
+    VALUE output = rb_ary_new();
+    for (int i = 0; i < n_tokens * n_vocab; i++) {
+      rb_ary_push(output, DBL2NUM((double)(logits[i])));
+    }
+
+    return output;
+  };
+
   static VALUE _llama_context_embeddings(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
       return Qnil;
     }
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(rb_iv_get(self, "@params"));
+    if (!prms_ptr->params.embedding) {
+      rb_raise(rb_eRuntimeError, "embedding parameter is false");
+      return Qnil;
+    }
+    if (rb_iv_get(self, "@has_evaluated") != Qtrue) {
+      rb_raise(rb_eRuntimeError, "LLaMA context has not been evaluated");
+      return Qnil;
+    }
 
     const int n_embd = llama_n_embd(ptr->ctx);
     const float* embd = llama_get_embeddings(ptr->ctx);
@@ -454,6 +500,54 @@ private:
     llama_reset_timings(ptr->ctx);
     return Qnil;
   };
+
+  static VALUE _llama_context_free(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx != NULL) {
+      llama_free(ptr->ctx);
+      ptr->ctx = NULL;
+      rb_iv_set(self, "@params", Qnil);
+      rb_iv_set(self, "@has_evaluated", Qfalse);
+    }
+    return Qnil;
+  }
+
+  static VALUE _llama_context_load(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("model_path"), rb_intern("params") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 2, 0, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_STRING)) {
+      rb_raise(rb_eArgError, "model_path must be a string");
+      return Qnil;
+    }
+    if (!rb_obj_is_kind_of(kw_values[1], rb_cLLaMAContextParams)) {
+      rb_raise(rb_eArgError, "params must be a LLaMAContextParams");
+      return Qnil;
+    }
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_ptr->ctx != NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is already loaded");
+      return Qnil;
+    }
+
+    VALUE filename = kw_values[0];
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
+    ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
+    if (ctx_ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA context");
+      return Qnil;
+    }
+
+    rb_iv_set(self, "@params", kw_values[1]);
+    rb_iv_set(self, "@has_evaluated", Qfalse);
+
+    RB_GC_GUARD(filename);
+    return Qnil;
+  };
 };
 
 const rb_data_type_t RbLLaMAContext::llama_context_type = {