llama_cpp 0.1.1 → 0.1.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 33b146badd1bebdf9588e48c0adac1f9924a0653aa5ec806fdf5dd288ef665d8
-  data.tar.gz: 134606db2b9fb10b51fc82f410d6653a6481b828d9fd05390b1570d6e198526a
+  metadata.gz: 1fe968c9231c20e614fafe89bc521c313ab68401fedd2d803743b18ccc234a28
+  data.tar.gz: a4916ec0f52b3e131175141f30bd3a70f37859207e732948d2fe7baac98a4b0c
 SHA512:
-  metadata.gz: 462d9e00121408c7af3934b0a663b29f99d5ad28f60a3471155509463bf26a14792c484d1fdc6054460941ae011d39b510774e225ad4ec03d60ce20a1dfef667
-  data.tar.gz: 4bf447ac55bba2b62d204dc975528de6664fe53af89df8ba4aa4172d4dbff709ac5b14a944326be5c71d64baa2cde00b60f7ba5e916e1fb68123c595f74ce24f
+  metadata.gz: fa99138a7a591a7e602e6aa040ccec057dcad09e52c6646edd0def9c0e3ea1aee6796bc32fa05dc9c384af1b8c72a3f5c2077de918d2e0a229901c97732023c1
+  data.tar.gz: 1e4399f4b75fcbe69da61ce23d2cf45594e5502e7d6ea6f9b7f0930ca155bcfb4481f81944496031e79c8ef0e48be20a6797d8f9b41967404e2a54330a93c261

data/CHANGELOG.md

@@ -1,5 +1,12 @@
 ## [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

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

@@ -524,7 +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;
@@ -788,7 +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;
@@ -1407,6 +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") };
@@ -1477,6 +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);
@@ -1494,11 +1514,45 @@ extern "C" void Init_llama_cpp(void) {
   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);
@@ -15,6 +16,7 @@ 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
 }