From 33429ec4e1f49ac32a9ab5a0e1bf4ec628ecb0f0 Mon Sep 17 00:00:00 2001
From: Zonghang Li <lizhuestc@gmail.com>
Date: Wed, 22 Jan 2025 11:25:09 +0400
Subject: [PATCH] add option --keep-inp-out-in-metal

---
 common/arg.cpp | 3 ++-
 src/llama.cpp  | 1 +
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/common/arg.cpp b/common/arg.cpp
index 602ad9f3..33464eaa 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -737,7 +737,7 @@ 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
+#ifdef GGML_USE_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"),
@@ -745,6 +745,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.keep_inp_out_in_metal = true;
         }
     ).set_env("LLAMA_ARG_KEEP_INP_OUT_IN_METAL"));
+#endif
     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),
diff --git a/src/llama.cpp b/src/llama.cpp
index 6a94f017..36eb971d 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -19773,6 +19773,7 @@ struct llama_model_params llama_model_default_params() {
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
+        /*.keep_inp_out_in_metal       =*/ false,
     };
 
 #ifdef GGML_USE_METAL