llama_cpp 0.1.0 → 0.1.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c4058abcb7afa897554fc75bb368caeea0e77429e01fb5f3a1949191c50f4de5
-  data.tar.gz: 9929e94c02b5d9c21379a9275f08668e835f91d3d7be3570a2da9ab4ecbe6ad1
+  metadata.gz: 1fe968c9231c20e614fafe89bc521c313ab68401fedd2d803743b18ccc234a28
+  data.tar.gz: a4916ec0f52b3e131175141f30bd3a70f37859207e732948d2fe7baac98a4b0c
 SHA512:
-  metadata.gz: ab267defd1769e7bf4599da199f50a7c5cc2355d2281ab7fd2ccd1a5ef196b716350cf8df9522a9185d02c8c3ad6a5d0f46f271fad0951440ab9b3fab4019932
-  data.tar.gz: 16727a2ac2c68f7913749b656c26523e9eee0118b69ff06bbc0935f899eac1874f16395d9e72ed2caa853e9c61fb9f614ad5913fca623e356aa249308b2f3dda
+  metadata.gz: fa99138a7a591a7e602e6aa040ccec057dcad09e52c6646edd0def9c0e3ea1aee6796bc32fa05dc9c384af1b8c72a3f5c2077de918d2e0a229901c97732023c1
+  data.tar.gz: 1e4399f4b75fcbe69da61ce23d2cf45594e5502e7d6ea6f9b7f0930ca155bcfb4481f81944496031e79c8ef0e48be20a6797d8f9b41967404e2a54330a93c261

data/CHANGELOG.md

@@ -1,5 +1,22 @@
 ## [Unreleased]
 
+## [[0.1.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.1...v0.1.2)] - 2023-05-22
+
+**Breaking Changes**
+
+- Bump bundled llama.cpp from master-6986c78 to master-265db98
+  - bump LLAMA_FILE_VERSION to 3
+
+## [[0.1.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.1.0...v0.1.1)] - 2023-05-21
+
+- Add load_session_file method to Context
+- Add save_session_file method to Context
+
+**Breaking Changes**
+
+- Bump bundled llama.cpp from master-173d0e6 to master-6986c78
+  - bump LLAMA_FILE_VERSION to 2
+
 ## [[0.1.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.0.7...v0.1.0)] - 2023-05-20
 
 **Breaking Changes**

data/ext/llama_cpp/extconf.rb

@@ -24,6 +24,13 @@ if with_config('openblas')
   $CFLAGS << ' -DGGML_USE_OPENBLAS'
 end
 
+if with_config('blis')
+  abort 'libblis is not found.' unless have_library('blis')
+  abort 'cblas.h is not found.' unless have_header('cblas.h')
+
+  $CFLAGS << ' -DGGML_USE_OPENBLAS'
+end
+
 if with_config('accelerate')
   abort 'Accelerate framework is not found.' unless have_framework('Accelerate')
 

data/ext/llama_cpp/llama_cpp.cpp

@@ -292,8 +292,6 @@ public:
     // rb_define_method(rb_cLLaMAContextParams, "initialize", RUBY_METHOD_FUNC(_llama_context_params_init), 0);
     rb_define_method(rb_cLLaMAContextParams, "n_ctx=", RUBY_METHOD_FUNC(_llama_context_params_set_n_ctx), 1);
     rb_define_method(rb_cLLaMAContextParams, "n_ctx", RUBY_METHOD_FUNC(_llama_context_params_get_n_ctx), 0);
-    rb_define_method(rb_cLLaMAContextParams, "n_parts=", RUBY_METHOD_FUNC(_llama_context_params_set_n_parts), 1);
-    rb_define_method(rb_cLLaMAContextParams, "n_parts", RUBY_METHOD_FUNC(_llama_context_params_get_n_parts), 0);
     rb_define_method(rb_cLLaMAContextParams, "seed=", RUBY_METHOD_FUNC(_llama_context_params_set_seed), 1);
     rb_define_method(rb_cLLaMAContextParams, "seed", RUBY_METHOD_FUNC(_llama_context_params_get_seed), 0);
     rb_define_method(rb_cLLaMAContextParams, "f16_kv=", RUBY_METHOD_FUNC(_llama_context_params_set_f16_kv), 1);
@@ -331,18 +329,6 @@ private:
     return INT2NUM(ptr->params.n_ctx);
   };
 
-  // n_parts
-  static VALUE _llama_context_params_set_n_parts(VALUE self, VALUE n_parts) {
-    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    ptr->params.n_parts = NUM2INT(n_parts);
-    return INT2NUM(ptr->params.n_parts);
-  };
-
-  static VALUE _llama_context_params_get_n_parts(VALUE self) {
-    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
-    return INT2NUM(ptr->params.n_parts);
-  };
-
   // seed
   static VALUE _llama_context_params_set_seed(VALUE self, VALUE seed) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
@@ -494,6 +480,8 @@ public:
     rb_define_method(rb_cLLaMAContext, "apply_lora_from_file", RUBY_METHOD_FUNC(_llama_context_apply_lora_from_file), -1);
     rb_define_method(rb_cLLaMAContext, "kv_cache_token_count", RUBY_METHOD_FUNC(_llama_context_kv_cache_token_count), 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_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_softmax", RUBY_METHOD_FUNC(_llama_context_sample_softmax), 1);
@@ -536,7 +524,14 @@ private:
     VALUE filename = kw_values[0];
     LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(kw_values[1]);
     LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
-    ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
+
+    try {
+      ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
+    } catch (const std::runtime_error& e) {
+      rb_raise(rb_eRuntimeError, "%s", e.what());
+      return Qnil;
+    }
+
     if (ctx_ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA context");
       return Qnil;
@@ -800,7 +795,14 @@ private:
 
     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);
+
+    try {
+      ctx_ptr->ctx = llama_init_from_file(StringValueCStr(filename), prms_ptr->params);
+    } catch (const std::runtime_error& e) {
+      rb_raise(rb_eRuntimeError, "%s", e.what());
+      return Qnil;
+    }
+
     if (ctx_ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "Failed to initialize LLaMA context");
       return Qnil;
@@ -870,6 +872,97 @@ private:
     return Qnil;
   };
 
+  static VALUE _llama_context_load_session_file(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[1] = { rb_intern("session_path") };
+    VALUE kw_values[1] = { Qundef };
+    VALUE candidates = Qnil;
+    VALUE last_n_tokens = Qnil;
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_STRING)) {
+      rb_raise(rb_eArgError, "session_path must be a String");
+      return Qnil;
+    }
+
+    VALUE filename = kw_values[0];
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_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"));
+    const int n_ctx = prms_ptr->params.n_ctx;
+
+    std::vector<llama_token> session_tokens(n_ctx);
+    size_t n_token_count_out = 0;
+
+    try {
+      bool res = llama_load_session_file(ctx_ptr->ctx, StringValueCStr(filename), session_tokens.data(), session_tokens.capacity(), &n_token_count_out);
+      if (!res) {
+        rb_raise(rb_eRuntimeError, "Failed to load session file");
+        return Qnil;
+      }
+      session_tokens.resize(n_token_count_out);
+    } catch (const std::runtime_error& e) {
+      rb_raise(rb_eRuntimeError, "%s", e.what());
+      return Qnil;
+    }
+
+    VALUE ary_session_tokens = rb_ary_new2(n_token_count_out);
+    for (size_t i = 0; i < n_token_count_out; i++) {
+      rb_ary_store(ary_session_tokens, i, INT2NUM(session_tokens[i]));
+    }
+
+    RB_GC_GUARD(filename);
+    return ary_session_tokens;
+  }
+
+  static VALUE _llama_context_save_session_file(int argc, VALUE* argv, VALUE self) {
+    VALUE kw_args = Qnil;
+    ID kw_table[2] = { rb_intern("session_path"), rb_intern("session_tokens") };
+    VALUE kw_values[2] = { Qundef, Qundef };
+    VALUE candidates = Qnil;
+    VALUE last_n_tokens = Qnil;
+    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, "session_path must be a String");
+      return Qnil;
+    }
+    if (!RB_TYPE_P(kw_values[1], T_ARRAY)) {
+      rb_raise(rb_eArgError, "session_tokens must be an Array");
+      return Qnil;
+    }
+
+    VALUE filename = kw_values[0];
+    const size_t sz_session_tokens = RARRAY_LEN(kw_values[1]);
+    std::vector<llama_token> session_tokens(sz_session_tokens);
+    for (size_t i = 0; i < sz_session_tokens; i++) {
+      session_tokens[i] = NUM2INT(rb_ary_entry(kw_values[1], i));
+    }
+
+    LLaMAContextWrapper* ctx_ptr = get_llama_context(self);
+    if (ctx_ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+
+    bool res = llama_save_session_file(ctx_ptr->ctx, StringValueCStr(filename), session_tokens.data(), sz_session_tokens);
+
+    if (!res) {
+      rb_raise(rb_eRuntimeError, "Failed to save session file");
+      return Qnil;
+    }
+
+    RB_GC_GUARD(filename);
+    return Qnil;
+  }
+
   static VALUE _llama_context_sample_repetition_penalty(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
     ID kw_table[1] = { rb_intern("penalty") };
@@ -1328,6 +1421,11 @@ const rb_data_type_t RbLLaMAContext::llama_context_type = {
 
 // module functions
 
+static VALUE rb_llama_llama_init_backend(VALUE self) {
+  llama_init_backend();
+  return Qnil;
+}
+
 static VALUE rb_llama_model_quantize(int argc, VALUE* argv, VALUE self) {
   VALUE kw_args = Qnil;
   ID kw_table[4] = { rb_intern("input_path"), rb_intern("output_path"), rb_intern("ftype"), rb_intern("n_threads") };
@@ -1398,6 +1496,7 @@ extern "C" void Init_llama_cpp(void) {
   RbLLaMAContext::define_class(rb_mLLaMACpp);
   RbLLaMAContextParams::define_class(rb_mLLaMACpp);
 
+  rb_define_module_function(rb_mLLaMACpp, "init_backend", rb_llama_llama_init_backend, 0);
   rb_define_module_function(rb_mLLaMACpp, "model_quantize", rb_llama_model_quantize, -1);
   rb_define_module_function(rb_mLLaMACpp, "token_bos", rb_llama_token_bos, 0);
   rb_define_module_function(rb_mLLaMACpp, "token_eos", rb_llama_token_eos, 0);
@@ -1411,16 +1510,49 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q4_0", INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_0));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q4_1", INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_1));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16", INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16));
-  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q4_2", INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_2));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q8_0", INT2NUM(LLAMA_FTYPE_MOSTLY_Q8_0));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q5_0", INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_0));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q5_1", INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_1));
 
-  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_VERSION", rb_str_new2(std::to_string(LLAMA_FILE_VERSION).c_str()));
   std::stringstream ss_magic;
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGJT;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGJT", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGLA;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGLA", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGMF;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGMF", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGML;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGML", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  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()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC;
   rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC", rb_str_new2(ss_magic.str().c_str()));
-  std::stringstream ss_magic_unversioned;
-  ss_magic_unversioned << std::showbase << std::hex << LLAMA_FILE_MAGIC_UNVERSIONED;
-  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_UNVERSIONED", rb_str_new2(ss_magic_unversioned.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_UNVERSIONED;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_UNVERSIONED", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
+  ss_magic << std::showbase << std::hex << LLAMA_SESSION_MAGIC;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_SESSION_MAGIC", rb_str_new2(ss_magic.str().c_str()));
+
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_VERSION", rb_str_new2(std::to_string(LLAMA_FILE_VERSION).c_str()));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_SESSION_VERSION", rb_str_new2(std::to_string(LLAMA_SESSION_VERSION).c_str()));
 }

data/ext/llama_cpp/src/ggml-cuda.h

@@ -6,6 +6,7 @@ extern "C" {
 
 void   ggml_init_cublas(void);
 
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
@@ -14,6 +15,9 @@ void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
 void * ggml_cuda_host_malloc(size_t size);
 void   ggml_cuda_host_free(void * ptr);
 
+void ggml_cuda_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensors, size_t offset);
+
 #ifdef  __cplusplus
 }
 #endif