cmake: Add GGML_BACKEND_DIR option (#15074)

* cmake: Add GGML_BACKEND_DIR option

This can be used by distributions to specify where to look for backends
when ggml is built with GGML_BACKEND_DL=ON.

* Fix phrasing

ggml/CMakeLists.txt

@@ -39,8 +39,9 @@ if (WIN32)
     set(CMAKE_SHARED_MODULE_PREFIX  "")
 endif()
 
-option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
-option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+option(BUILD_SHARED_LIBS           "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL             "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
 
 #
 # option list

ggml/cmake/ggml-config.cmake.in

@@ -106,7 +106,7 @@ if(NOT TARGET ggml::ggml)
 
     find_library(GGML_LIBRARY ggml
         REQUIRED
-        HINTS ${GGML_LIB_DIR}
+        HINTS ${GGML_LIB_DIR} ${GGML_BACKEND_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
     add_library(ggml::ggml UNKNOWN IMPORTED)

ggml/src/CMakeLists.txt

@@ -214,6 +214,13 @@ add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
+if (GGML_BACKEND_DIR)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
+    endif()
+    target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
+endif()
+
 target_link_libraries(ggml PUBLIC ggml-base)
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -227,7 +234,11 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
-        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        if (GGML_BACKEND_DIR)
+            install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
+        else()
+            install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        endif()
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})

ggml/src/ggml-backend-reg.cpp

@@ -498,6 +498,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     std::vector<fs::path> search_paths;
     if (user_search_path == nullptr) {
+#ifdef GGML_BACKEND_DIR
+        search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
+#endif
         // default search paths: executable directory, current directory
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());

gguf-py/gguf/scripts/gguf_new_metadata.py

@@ -111,6 +111,7 @@ def main() -> None:
     parser.add_argument("--general-description",                       type=str,  help="The models general.description", metavar='"Description ..."')
     parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
     parser.add_argument("--chat-template-config",                      type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
+    parser.add_argument("--chat-template-file",                        type=Path, help="Jinja file containing chat template", metavar='chat_template.jinja')
     parser.add_argument("--pre-tokenizer",                             type=str,  help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
     parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model", metavar='general.url')
     parser.add_argument("--special-token",        action="append",     type=str,  help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
@@ -134,12 +135,17 @@ def main() -> None:
         new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
 
     if args.chat_template_config:
-        with open(args.chat_template_config, 'r') as fp:
+        with open(args.chat_template_config, 'r', encoding='utf-8') as fp:
             config = json.load(fp)
             template = config.get('chat_template')
             if template:
                 new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
 
+    if args.chat_template_file:
+        with open(args.chat_template_file, 'r', encoding='utf-8') as fp:
+            template = fp.read()
+            new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
+
     if args.pre_tokenizer:
         new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)