feat(模型加载): 添加sync_to_temp选项支持临时目录加载

添加sync_to_temp参数控制是否将模型文件复制到临时目录后再加载

common/arg.cpp

@@ -2039,6 +2039,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.use_mmap = value;
         }
     ).set_env("LLAMA_ARG_MMAP"));
+    add_opt(common_arg(
+        {"--sync-to-temp"},
+        {"--no-sync-to-temp"},
+        string_format("whether to copy model to temporary directory before loading (default: %s)", params.sync_to_temp ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.sync_to_temp = value;
+        }
+    ).set_env("LLAMA_ARG_SYNC_TO_TEMP"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"

common/common.cpp

@@ -1353,6 +1353,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;
     mparams.no_host         = params.no_host;
+    mparams.sync_to_temp    = params.sync_to_temp;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

common/common.h

@@ -430,6 +430,7 @@ struct common_params {
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
     bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
     bool no_host           = false; // bypass host buffer allowing extra buffers to be used
+    bool sync_to_temp      = false; // copy model to temporary directory before loading
 
     bool single_turn       = false; // single turn chat conversation