Change load mode to auto, benchmarks from RPi5.

1 files changed, 5 insertions, 2 deletions

diff --git a/llm.c b/llm.c
index 23264db..4d3dc58 100644
--- a/llm.c
+++ b/llm.c
@@ -91,8 +91,11 @@ struct llm_ctx *llm_init(const char *model_path)
 
 	mparams = llama_model_default_params();
 	mparams.n_gpu_layers = 0;   /* force all layers onto CPU */
-	mparams.load_mode = LLAMA_LOAD_MODE_MLOCK;  /* Force model to RAM, 
-	                                             * don't swap/compress */ 
+
+	/* LLAMA_LOAD_MODE_MLOCK (force model to RAM, no swap/compression)
+	 * would have been preferrable. But my Raspberry Pi5 doesn't have
+	 * enough memory. */
+	mparams.load_mode = LLAMA_LOAD_MODE_NONE;
 
 	llm = MALLOC(sizeof(*llm));
 	llm->model = llama_model_load_from_file(model_path, mparams);