allow embeddings models to use mmap

expose.h

@@ -256,6 +256,7 @@ struct embeddings_load_model_inputs
     const char * vulkan_info = nullptr;
     const int gpulayers = 0;
     const bool flash_attention = false;
+    const bool use_mmap = false;
     const bool quiet = false;
     const int debugmode = 0;
 };

koboldcpp.py

@@ -346,6 +346,7 @@ class embeddings_load_model_inputs(ctypes.Structure):
                 ("vulkan_info", ctypes.c_char_p),
                 ("gpulayers", ctypes.c_int),
                 ("flash_attention", ctypes.c_bool),
+                ("use_mmap", ctypes.c_bool),
                 ("quiet", ctypes.c_bool),
                 ("debugmode", ctypes.c_int)]
 
@@ -1213,7 +1214,6 @@ def load_model(model_filename):
     inputs.lora_multiplier = args.loramult
     if args.lora:
         inputs.lora_filename = args.lora[0].encode("UTF-8")
-        inputs.use_mmap = False
 
     inputs.draftmodel_filename = args.draftmodel.encode("UTF-8") if args.draftmodel else "".encode("UTF-8")
     inputs.draft_amount = args.draftamount
@@ -1741,6 +1741,7 @@ def embeddings_load_model(model_filename):
     inputs.gpulayers = 0
     inputs.flash_attention = False
     inputs.threads = args.threads
+    inputs.use_mmap = args.usemmap
     inputs = set_backend_props(inputs)
     ret = handle.embeddings_load_model(inputs)
     return ret

otherarch/embeddings_adapter.cpp

@@ -117,7 +117,7 @@ bool embeddingstype_load_model(const embeddings_load_model_inputs inputs)
     llama_model_params model_params = llama_model_default_params();
     llama_context_params ctx_params = llama_context_default_params();
     const int nthreads = inputs.threads;
-    model_params.use_mmap = false;
+    model_params.use_mmap = inputs.use_mmap;
     model_params.use_mlock = false;
     model_params.n_gpu_layers = inputs.gpulayers; //offload if possible
     model_params.split_mode = llama_split_mode::LLAMA_SPLIT_MODE_LAYER;