llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)

* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used

* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp

Co-authored-by: Max Krasnyansky <redacted>
* set mmap_support to false on OpenCL backend

* fix order of load modes

* use -1 for auto

* resolve load mode auto earlier to correctly pick gpu host or cpu memory

* add load mode auto to llama-bench

* bump virtgpu api version, regenerate docs

---------

Co-authored-by: Piotr Wilkin (ilintar) <redacted>
Co-authored-by: Max Krasnyansky <redacted>
Co-authored-by: Georgi Gerganov <redacted>

diff --git a/common/arg.cpp b/common/arg.cpp
index c37d5cd0aa4b4178505beec0e37dde0b7593dee5..cb314eee70442c1c576383267d37185162e5fd7e 100644 (file)
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2605,14 +2605,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_DIO"));
     add_opt(common_arg(
         {"-lm", "--load-mode"}, "MODE",
-        "model loading mode (default: mmap)\n"
+        "model loading mode (default: auto)\n"
+        "- auto: mmap, unless a device does not support it\n"
         "- none: no special loading mode\n"
         "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
         "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
         "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
         "- dio: use DirectIO if available\n",
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            /**/ if (value == "auto")       { params.load_mode = LLAMA_LOAD_MODE_AUTO;       }
+            else if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
             else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
             else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
             else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }

diff --git a/common/common.h b/common/common.h
index 878534dccf50905974ce549603b8210ea05a5337..d485d4fb41f7b59ce86ebc7647a8a5d212301835 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -473,7 +473,7 @@ struct common_params {
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
+    enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
 
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;

diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 2924fdbe9884df40abf505fd89d277f5281a835b..cc3f8cd36e3549c4f3ed3d6ad5f6b577af4b837b 100644 (file)
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -154,6 +154,8 @@ extern "C" {
         bool buffer_from_host_ptr;
         // event synchronization
         bool events;
+        // mmap is supported for loading
+        bool mmap_support;
     };
 
     // all the device properties

diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
index a5a3a58ad054e24a829906253bef5d9aacfe491c..7654ea1f30f61f7b1bb02275f6d5e7759ccfd2fd 100644 (file)
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/ggml/src/ggml-backend-meta.cpp
@@ -132,6 +132,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ false, // Not implemented.
         /* .buffer_from_host_ptr  = */ false, // Not implemented.
         /* .events                = */ false, // Not implemented.
+        /* .mmap_support          = */ true,
     };
     for (ggml_backend_dev_t simple_dev : meta_dev_ctx->simple_devs) {
         ggml_backend_dev_props tmp_props;
@@ -140,6 +141,7 @@ static void ggml_backend_meta_device_get_props(ggml_backend_dev_t dev, ggml_back
         props->caps.host_buffer          = props->caps.host_buffer          && tmp_props.caps.host_buffer;
         props->caps.buffer_from_host_ptr = props->caps.buffer_from_host_ptr && tmp_props.caps.buffer_from_host_ptr;
         props->caps.events               = props->caps.events               && tmp_props.caps.events;
+        props->caps.mmap_support         = props->caps.mmap_support         && tmp_props.caps.mmap_support;
     }
 }
 

diff --git a/ggml/src/ggml-blas/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp
index 9745fa29f5dbd46e5537eb4a9e5e39b97fedd423..e4b5bd254747664def59ae6e434c46583ca48e13 100644 (file)
--- a/ggml/src/ggml-blas/ggml-blas.cpp
+++ b/ggml/src/ggml-blas/ggml-blas.cpp
@@ -367,6 +367,7 @@ static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct gg
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ true,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 5f51ea3bb3c866324b0b513688084710a69cfc4e..ffa361af4e8b4aae53bd663815629f72932e679b 100644 (file)
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2815,6 +2815,7 @@ static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ true,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
index 16cc5116c5451787c6a1dd1988e38b761f20ef12..c0c9aa3cf09c573e308b32d3c90f058c709ce997 100644 (file)
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -397,6 +397,7 @@ static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggm
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ true,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 05e8d7f7368471a0119f6f691191fce3d238464f..1d4f4dfbd2148729baef7673dfbfe90f6d17980f 100644 (file)
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -4801,6 +4801,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ events,
+        /* .mmap_support          = */ props->type != GGML_BACKEND_DEVICE_TYPE_IGPU,
     };
 }
 

diff --git a/ggml/src/ggml-et/ggml-et.cpp b/ggml/src/ggml-et/ggml-et.cpp
index b30209095672b3bfebf8fa46b997f3f60506b5e8..e8482f734621d25209deca8e43dee3078778fdba 100644 (file)
--- a/ggml/src/ggml-et/ggml-et.cpp
+++ b/ggml/src/ggml-et/ggml-et.cpp
@@ -1646,6 +1646,7 @@ static void ggml_backend_et_device_get_props(ggml_backend_dev_t dev, struct ggml
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
index bdb8af0820a350af025dbf8c5ba40aa0cb472b3f..f80c60a500bdbddd8a9df2e90ac4594021631399 100644 (file)
--- a/ggml/src/ggml-hexagon/ggml-hexagon.cpp
+++ b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
@@ -3930,6 +3930,7 @@ static void ggml_backend_hexagon_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ (bool) opt_hostbuf,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ false,
     };
 }
 

diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
index a1003b3acff87dc55c819cf8c296a21a9ec7f72f..ef3c92f27127a1a516c99643273704748c9e88e8 100644 (file)
--- a/ggml/src/ggml-metal/ggml-metal.cpp
+++ b/ggml/src/ggml-metal/ggml-metal.cpp
@@ -681,6 +681,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ true,
         /* .events               = */ true,
+        /* .mmap_support         = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 9874ffe8d21ebaee3231013cb6393233c1e60ed2..19dca4ced0b8bb462f0ba96e52844ff8821181e2 100644 (file)
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -10795,6 +10795,7 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ false,
     };
 }
 

diff --git a/ggml/src/ggml-openvino/ggml-openvino.cpp b/ggml/src/ggml-openvino/ggml-openvino.cpp
index 0e7501fefe385ba5f0eb78d84bed9f86e8866759..dfe80e6c86e70b2095307e8ef17a04ab93704a03 100644 (file)
--- a/ggml/src/ggml-openvino/ggml-openvino.cpp
+++ b/ggml/src/ggml-openvino/ggml-openvino.cpp
@@ -763,6 +763,7 @@ static void ggml_backend_openvino_device_get_props(ggml_backend_dev_t dev, ggml_
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index 17c53a5f049ef56f22e99d2cfef338487e5dc61f..e9de0d0aa98afd134dc08b0478637ef590ba7ae4 100644 (file)
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -1881,6 +1881,7 @@ static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggm
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 18d58782ebff5c14ba3d8412d002927b06706bc4..8a19f648bf077d2c04f5aa50fbeb4fa7a58cce0a 100644 (file)
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -5649,6 +5649,7 @@ static void ggml_backend_sycl_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .host_buffer           = */ host_buffer,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ events,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
index c7acb8b51ce742601e5243bd2a60beb23e28d2ae..87872df1c7bc82685f8e0c877480d930bd5e1fb7 100644 (file)
--- a/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
@@ -111,6 +111,7 @@ uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_
     apir_encode_bool_t(enc, &props.caps.host_buffer);
     apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr);
     apir_encode_bool_t(enc, &props.caps.events);
+    apir_encode_bool_t(enc, &props.caps.mmap_support);
 
     return 0;
 }

diff --git a/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
index 6bf97e8a3a243341c920efe35540ecfafef23eca..a5ef3ea476d0220e6c510d896ea9ba14ac1af37a 100644 (file)
--- a/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
+++ b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
@@ -7,7 +7,7 @@
 #include <cstdint>
 
 #define APIR_PROTOCOL_MAJOR 0
-#define APIR_PROTOCOL_MINOR 1
+#define APIR_PROTOCOL_MINOR 2
 
 #define APIR_HANDSHAKE_MAGIC 0xab1e
 

diff --git a/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
index 8fa20ff43bd58810ab39ebced8cd68c05309294e..d5bdc993b464279f5ab1b0096b7a36d30073c013 100644 (file)
--- a/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
+++ b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
@@ -11,9 +11,9 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
 
     context->gpu = gpu;
 
-    bool async__unused, host_buffer__unused, events__unused;
+    bool async__unused, host_buffer__unused, events__unused, mmap_support__unused;
     bool buffer_from_host_ptr;
-    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused);
+    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused, &mmap_support__unused);
 
     if (buffer_from_host_ptr) {
         context->apir_context = apir_device_buffer_from_ptr(gpu, size, size);

diff --git a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
index a978812cd90816cdef6bb251a6c1263360ef1247..987ce9dd110c71f656c7bd3c1a5f956defed9909 100644 (file)
--- a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
+++ b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
@@ -65,7 +65,7 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
 
     virtgpu * gpu = DEV_TO_GPU(dev);
     apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr,
-                          &props->caps.events);
+                          &props->caps.events, &props->caps.mmap_support);
 
     props->caps.buffer_from_host_ptr = false;
     props->caps.async                = false;

diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
index 9f513c138dd24d331ab7c2500eb06ca010c10c15..864264f213b5378a9693d45d6b9a42e6d17c60f0 100644 (file)
--- a/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
@@ -144,7 +144,8 @@ void apir_device_get_props(virtgpu * gpu,
                            bool *    async,
                            bool *    host_buffer,
                            bool *    buffer_from_host_ptr,
-                           bool *    events) {
+                           bool *    events,
+                           bool *    mmap_support) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -157,6 +158,7 @@ void apir_device_get_props(virtgpu * gpu,
     apir_decode_bool_t(decoder, host_buffer);
     apir_decode_bool_t(decoder, buffer_from_host_ptr);
     apir_decode_bool_t(decoder, events);
+    apir_decode_bool_t(decoder, mmap_support);
 
     remote_call_finish(gpu, encoder, decoder);
 

diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
index 44b0ad1ffa1d1b483836bf1a0166d8800f1528d4..da28aa5f90471040280defc0d408b0ecbd841ba1 100644 (file)
--- a/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
@@ -13,7 +13,8 @@ void                           apir_device_get_props(struct virtgpu * gpu,
                                                      bool *           async,
                                                      bool *           host_buffer,
                                                      bool *           buffer_from_host_ptr,
-                                                     bool *           events);
+                                                     bool *           events,
+                                                     bool *           mmap_support);
 apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
 
 /* buffer-type */

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index a923755f9edd444e14f73e0b8b3171ae58317ee9..45fa97f8129775c6e31947ea0fc22daa11363150 100644 (file)
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -17891,6 +17891,7 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
         /* .host_buffer           = */ true,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ true,
+        /* .mmap_support          = */ !ctx->is_integrated_gpu,
     };
 }
 

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index 98c7162478f81fd74d3a21d80b6b8a84092a1fff..6741752b36112e0d5bc24fbde9c23ae708cccea1 100644 (file)
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -3942,6 +3942,7 @@ static void ggml_backend_webgpu_device_get_props(ggml_backend_dev_t dev, struct
         /* .host_buffer           = */ false,
         /* .buffer_from_host_ptr  = */ false,
         /* .events                = */ false,
+        /* .mmap_support          = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-zdnn/ggml-zdnn.cpp b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
index 639b818d128e6e5b7c72b3751687f4314ffe9790..4007ac9dfc7db777004eb07b3b3d59b37fccc44a 100644 (file)
--- a/ggml/src/ggml-zdnn/ggml-zdnn.cpp
+++ b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
@@ -487,7 +487,8 @@ static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_back
         /* .async                = */ false,
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ false,
-        /* .events               = */ false
+        /* .events               = */ false,
+        /* .mmap_support         = */ true,
     };
 }
 

diff --git a/ggml/src/ggml-zendnn/ggml-zendnn.cpp b/ggml/src/ggml-zendnn/ggml-zendnn.cpp
index e6a9b51b7925e19186cecc5b063568d62e2568c5..ec7ce233145a491c3e28dec3cd2c0f4a49caba3f 100644 (file)
--- a/ggml/src/ggml-zendnn/ggml-zendnn.cpp
+++ b/ggml/src/ggml-zendnn/ggml-zendnn.cpp
@@ -654,7 +654,8 @@ static void ggml_backend_zendnn_device_get_props(ggml_backend_dev_t dev, struct
         /* .async                = */ false,
         /* .host_buffer          = */ false,
         /* .buffer_from_host_ptr = */ true,
-        /* .events               = */ false
+        /* .events               = */ false,
+        /* .mmap_support         = */ true,
     };
 }
 

diff --git a/include/llama.h b/include/llama.h
index 5a0b66dc25f6534980f1798d05c9d80b9c7f254d..bfef0e1d1129c8bf642ff94f56d8e26080fdc99c 100644 (file)
--- a/include/llama.h
+++ b/include/llama.h
@@ -203,11 +203,12 @@ extern "C" {
     };
 
     enum llama_load_mode {
-        LLAMA_LOAD_MODE_NONE       = 0, // no special loading mode
-        LLAMA_LOAD_MODE_MMAP       = 1, // memory map the model
-        LLAMA_LOAD_MODE_MLOCK      = 2, // force system to keep model in RAM rather than swapping or compressing
-        LLAMA_LOAD_MODE_MMAP_MLOCK = 3, // mmap + force system to keep model in RAM rather than swapping or compressing
-        LLAMA_LOAD_MODE_DIRECT_IO  = 4, // use direct I/O if available
+        LLAMA_LOAD_MODE_AUTO       = -1, // auto-detect based on device capabilities
+        LLAMA_LOAD_MODE_NONE       =  0, // no special loading mode
+        LLAMA_LOAD_MODE_MMAP       =  1, // memory map the model
+        LLAMA_LOAD_MODE_MLOCK      =  2, // force system to keep model in RAM rather than swapping or compressing
+        LLAMA_LOAD_MODE_MMAP_MLOCK =  3, // mmap + force system to keep model in RAM rather than swapping or compressing
+        LLAMA_LOAD_MODE_DIRECT_IO  =  4, // use direct I/O if available
     };
 
     LLAMA_API const char * llama_load_mode_name(enum llama_load_mode load_mode);

diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 3d50f8a1cb1cda8eeff09ed496c8049720225c45..51ba0543968212fb91da5c3bf5b6207f9b58f6a5 100644 (file)
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -543,7 +543,7 @@ llama_model_loader::llama_model_loader(
 
     tensor_buft_overrides = param_tensor_buft_overrides_p;
 
-    this->use_mmap      = load_mode == LLAMA_LOAD_MODE_MMAP || load_mode == LLAMA_LOAD_MODE_MMAP_MLOCK;
+    this->use_mmap      = load_mode == LLAMA_LOAD_MODE_MMAP || load_mode == LLAMA_LOAD_MODE_MMAP_MLOCK || load_mode == LLAMA_LOAD_MODE_AUTO;
     this->use_direct_io = load_mode == LLAMA_LOAD_MODE_DIRECT_IO;
 
     if (!fname.empty()) {

diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 3bf3a22f26af954771813cec2ab515e0ee10ca7a..cc1917b6da7731029db9d121250ced9e9b6a0f62 100644 (file)
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -1271,8 +1271,23 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
 
     this->ml = &ml; // to be used by create_tensor() and load_arch_tensors()
 
+    if (ml.use_mmap && params.load_mode == LLAMA_LOAD_MODE_AUTO) {
+        for (const auto & dev : devices) {
+            ggml_backend_dev_props props;
+            ggml_backend_dev_get_props(dev.dev, &props);
+            if (!props.caps.mmap_support) {
+                ml.use_mmap = false;
+                break;
+            }
+        }
+    }
+
+    const char * load_mode_name = params.load_mode == LLAMA_LOAD_MODE_AUTO
+        ? llama_load_mode_name(ml.use_mmap ? LLAMA_LOAD_MODE_MMAP : LLAMA_LOAD_MODE_NONE)
+        : llama_load_mode_name(params.load_mode);
+
     LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (load_mode = %s)\n",
-        __func__, llama_load_mode_name(params.load_mode));
+        __func__, load_mode_name);
 
     // build a list of buffer types for the CPU and GPU devices
     pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
@@ -2452,7 +2467,7 @@ llama_model_params llama_model_default_params() {
         /*.tensor_buft_overrides       =*/ nullptr,
         /*.n_gpu_layers                =*/ -1,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
-        /*.load_mode                   =*/ LLAMA_LOAD_MODE_MMAP,
+        /*.load_mode                   =*/ LLAMA_LOAD_MODE_AUTO,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
         /*.progress_callback           =*/ nullptr,

diff --git a/src/llama.cpp b/src/llama.cpp
index d6e0bbfefa729329fe6b83e46e603a85dab0f2e3..94c8f60e0c439cd9b55d909004f9f6bd03c25b04 100644 (file)
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -48,6 +48,8 @@ const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_ty
 
 const char * llama_load_mode_name(enum llama_load_mode load_mode) {
     switch (load_mode) {
+        case LLAMA_LOAD_MODE_AUTO:
+            return "auto";
         case LLAMA_LOAD_MODE_NONE:
             return "none";
         case LLAMA_LOAD_MODE_MMAP:
@@ -63,11 +65,12 @@ const char * llama_load_mode_name(enum llama_load_mode load_mode) {
 }
 
 enum llama_load_mode llama_load_mode_from_str(const char * str) {
-    if (std::strcmp(str, "none") == 0)       { return LLAMA_LOAD_MODE_NONE;       }
-    if (std::strcmp(str, "mmap") == 0)       { return LLAMA_LOAD_MODE_MMAP;       }
-    if (std::strcmp(str, "mlock") == 0)      { return LLAMA_LOAD_MODE_MLOCK;      }
+    if (std::strcmp(str, "auto")       == 0) { return LLAMA_LOAD_MODE_AUTO;       }
+    if (std::strcmp(str, "none")       == 0) { return LLAMA_LOAD_MODE_NONE;       }
+    if (std::strcmp(str, "mmap")       == 0) { return LLAMA_LOAD_MODE_MMAP;       }
+    if (std::strcmp(str, "mlock")      == 0) { return LLAMA_LOAD_MODE_MLOCK;      }
     if (std::strcmp(str, "mmap+mlock") == 0) { return LLAMA_LOAD_MODE_MMAP_MLOCK; }
-    if (std::strcmp(str, "dio") == 0)        { return LLAMA_LOAD_MODE_DIRECT_IO;  }
+    if (std::strcmp(str, "dio")        == 0) { return LLAMA_LOAD_MODE_DIRECT_IO;  }
     throw std::invalid_argument(std::string("unknown load mode: ") + str);
 }
 

diff --git a/tools/cli/README.md b/tools/cli/README.md
index b42b2e5343ffbf820f6338464122d7ca9a06c066..880c4a5408385e0d7beb20e1b5181fd2107dbaf7 100644 (file)
--- a/tools/cli/README.md
+++ b/tools/cli/README.md
@@ -58,7 +58,7 @@
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |

diff --git a/tools/completion/README.md b/tools/completion/README.md
index 552a0c6abf4e300fda1245436609cf04a5b2e245..c2e52ac066e74a543441a3d4d1c842c31e17f9ee 100644 (file)
--- a/tools/completion/README.md
+++ b/tools/completion/README.md
@@ -141,7 +141,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |

diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp
index c17a27b54019016f12251e89ad5c584786d36793..7c495afe20efc11725f399b196d410060ecc36e2 100644 (file)
--- a/tools/llama-bench/llama-bench.cpp
+++ b/tools/llama-bench/llama-bench.cpp
@@ -384,7 +384,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_gpu_layers         */ { -1 },
     /* n_cpu_moe            */ { 0 },
     /* split_mode           */ { LLAMA_SPLIT_MODE_LAYER },
-    /* load_mode            */ { LLAMA_LOAD_MODE_MMAP },
+    /* load_mode            */ { LLAMA_LOAD_MODE_AUTO },
     /* main_gpu             */ { 0 },
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { LLAMA_FLASH_ATTN_TYPE_AUTO },
@@ -459,7 +459,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -nkvo, --no-kv-offload <0|1>                      (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -fa, --flash-attn <on|off|auto>                   (default: %s)\n", join(transform_to_str(cmd_params_defaults.flash_attn, llama_flash_attn_type_name), ",").c_str());
     printf("  -dev, --device <dev0/dev1/...>                    (default: auto)\n");
-    printf("  -lm, --load-mode <none|mmap|mlock|mmap+mlock|dio> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
+    printf("  -lm, --load-mode <auto|none|mmap|mlock|mmap+mlock|dio> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
     printf("  -mmp, --mmap <0|1>                                (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -dio, --direct-io <0|1>                           (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -embd, --embeddings <0|1>                         (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
@@ -764,7 +764,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 std::vector<llama_load_mode> modes;
                 for (const auto & m : p) {
                     llama_load_mode mode;
-                    if (m == "none") {
+                    if (m == "auto") {
+                        mode = LLAMA_LOAD_MODE_AUTO;
+                    } else if (m == "none") {
                         mode = LLAMA_LOAD_MODE_NONE;
                     } else if (m == "mmap") {
                         mode = LLAMA_LOAD_MODE_MMAP;

diff --git a/tools/server/README.md b/tools/server/README.md
index 6927caddbb6c3357b025aa8043a88de0d43bf16a..a2ab872b4c1a5c54d403224cd77b288ae1dde2b9 100644 (file)
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -75,7 +75,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
-| `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
+| `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |