]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commitdiff
llama: add default load-mode auto, which avoids mmap on iGPUs (llama/26081)
authorRuben Ortlam <redacted>
Tue, 11 Aug 2026 06:20:46 +0000 (08:20 +0200)
committerGeorgi Gerganov <redacted>
Fri, 14 Aug 2026 19:16:06 +0000 (22:16 +0300)
* 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>
23 files changed:
ggml/include/ggml-backend.h
ggml/src/ggml-backend-meta.cpp
ggml/src/ggml-blas/ggml-blas.cpp
ggml/src/ggml-cann/ggml-cann.cpp
ggml/src/ggml-cpu/ggml-cpu.cpp
ggml/src/ggml-cuda/ggml-cuda.cu
ggml/src/ggml-et/ggml-et.cpp
ggml/src/ggml-hexagon/ggml-hexagon.cpp
ggml/src/ggml-metal/ggml-metal.cpp
ggml/src/ggml-opencl/ggml-opencl.cpp
ggml/src/ggml-openvino/ggml-openvino.cpp
ggml/src/ggml-rpc/ggml-rpc.cpp
ggml/src/ggml-sycl/ggml-sycl.cpp
ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
ggml/src/ggml-virtgpu/ggml-backend-device.cpp
ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
ggml/src/ggml-vulkan/ggml-vulkan.cpp
ggml/src/ggml-webgpu/ggml-webgpu.cpp
ggml/src/ggml-zdnn/ggml-zdnn.cpp
ggml/src/ggml-zendnn/ggml-zendnn.cpp

index 2924fdbe9884df40abf505fd89d277f5281a835b..cc3f8cd36e3549c4f3ed3d6ad5f6b577af4b837b 100644 (file)
@@ -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
index a5a3a58ad054e24a829906253bef5d9aacfe491c..7654ea1f30f61f7b1bb02275f6d5e7759ccfd2fd 100644 (file)
@@ -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;
     }
 }
 
index 9745fa29f5dbd46e5537eb4a9e5e39b97fedd423..e4b5bd254747664def59ae6e434c46583ca48e13 100644 (file)
@@ -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,
     };
 }
 
index 5f51ea3bb3c866324b0b513688084710a69cfc4e..ffa361af4e8b4aae53bd663815629f72932e679b 100644 (file)
@@ -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,
     };
 }
 
index 16cc5116c5451787c6a1dd1988e38b761f20ef12..c0c9aa3cf09c573e308b32d3c90f058c709ce997 100644 (file)
@@ -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,
     };
 }
 
index 05e8d7f7368471a0119f6f691191fce3d238464f..1d4f4dfbd2148729baef7673dfbfe90f6d17980f 100644 (file)
@@ -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,
     };
 }
 
index b30209095672b3bfebf8fa46b997f3f60506b5e8..e8482f734621d25209deca8e43dee3078778fdba 100644 (file)
@@ -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,
     };
 }
 
index bdb8af0820a350af025dbf8c5ba40aa0cb472b3f..f80c60a500bdbddd8a9df2e90ac4594021631399 100644 (file)
@@ -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,
     };
 }
 
index a1003b3acff87dc55c819cf8c296a21a9ec7f72f..ef3c92f27127a1a516c99643273704748c9e88e8 100644 (file)
@@ -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,
     };
 }
 
index 9874ffe8d21ebaee3231013cb6393233c1e60ed2..19dca4ced0b8bb462f0ba96e52844ff8821181e2 100644 (file)
@@ -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,
     };
 }
 
index 0e7501fefe385ba5f0eb78d84bed9f86e8866759..dfe80e6c86e70b2095307e8ef17a04ab93704a03 100644 (file)
@@ -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,
     };
 }
 
index 17c53a5f049ef56f22e99d2cfef338487e5dc61f..e9de0d0aa98afd134dc08b0478637ef590ba7ae4 100644 (file)
@@ -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,
     };
 }
 
index 18d58782ebff5c14ba3d8412d002927b06706bc4..8a19f648bf077d2c04f5aa50fbeb4fa7a58cce0a 100644 (file)
@@ -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,
     };
 }
 
index c7acb8b51ce742601e5243bd2a60beb23e28d2ae..87872df1c7bc82685f8e0c877480d930bd5e1fb7 100644 (file)
@@ -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;
 }
index 6bf97e8a3a243341c920efe35540ecfafef23eca..a5ef3ea476d0220e6c510d896ea9ba14ac1af37a 100644 (file)
@@ -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
 
index 8fa20ff43bd58810ab39ebced8cd68c05309294e..d5bdc993b464279f5ab1b0096b7a36d30073c013 100644 (file)
@@ -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);
index a978812cd90816cdef6bb251a6c1263360ef1247..987ce9dd110c71f656c7bd3c1a5f956defed9909 100644 (file)
@@ -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;
index 9f513c138dd24d331ab7c2500eb06ca010c10c15..864264f213b5378a9693d45d6b9a42e6d17c60f0 100644 (file)
@@ -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);
 
index 44b0ad1ffa1d1b483836bf1a0166d8800f1528d4..da28aa5f90471040280defc0d408b0ecbd841ba1 100644 (file)
@@ -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 */
index a923755f9edd444e14f73e0b8b3171ae58317ee9..45fa97f8129775c6e31947ea0fc22daa11363150 100644 (file)
@@ -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,
     };
 }
 
index 98c7162478f81fd74d3a21d80b6b8a84092a1fff..6741752b36112e0d5bc24fbde9c23ae708cccea1 100644 (file)
@@ -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,
     };
 }
 
index 639b818d128e6e5b7c72b3751687f4314ffe9790..4007ac9dfc7db777004eb07b3b3d59b37fccc44a 100644 (file)
@@ -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,
     };
 }
 
index e6a9b51b7925e19186cecc5b063568d62e2568c5..ec7ce233145a491c3e28dec3cd2c0f4a49caba3f 100644 (file)
@@ -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,
     };
 }