add option --keep-inp-out-in-metal and fix bugs in unmap

common/arg.cpp

@@ -737,6 +737,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.cuda_mem = value; // in GiB
         }
     ).set_env("LLAMA_ARG_CUDA_MEM"));
+    // "--keep-inp-out-in-metal" is a temporary option to keep the input and output in metal
+    add_opt(llama_arg(
+        {"--keep-inp-out-in-metal"},
+        format("whether to keep input and output weight in metal (default: %s)", params.keep_inp_out_in_metal ? "true" : "false"),
+        [](gpt_params & params) {
+            params.keep_inp_out_in_metal = true;
+        }
+    ).set_env("LLAMA_ARG_KEEP_INP_OUT_IN_METAL"));
     add_opt(llama_arg(
         {"-n", "--predict", "--n-predict"}, "N",
         format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict),

common/common.cpp

@@ -1603,6 +1603,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+    mparams.keep_inp_out_in_metal = params.keep_inp_out_in_metal;
     std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), mparams.n_layer_window);
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

common/common.h

@@ -148,6 +148,7 @@ struct gpt_params {
     std::string master_ip         = "localhost"; // ip address of the master node
     std::string next_node_ip      = "localhost"; // ip address of my next node
     bool    unload                = false; // unload layer weights after use or not
+    bool    keep_inp_out_in_metal = false; // whether to keep input/output weight in metal, not by default
     int32_t cuda_mem              = 999.0; // cuda memory to use, in GiB
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =     0; // context size