ggml-backend : add device and backend reg interfaces (ggerganov#9707)

Co-authored-by: Johannes Gäßler <[email protected]>

.github/workflows/bench.yml.disabled

@@ -27,10 +27,10 @@ on:
   push:
     branches:
       - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   schedule:
     -  cron: '04 2 * * *'
 

Makefile

@@ -1054,10 +1054,11 @@ ggml/src/ggml-alloc.o: \
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
 ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.c \
+	ggml/src/ggml-backend.cpp \
+	ggml/src/ggml-backend-impl.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ggml/src/ggml-quants.o: \
 	ggml/src/ggml-quants.c \

Package.swift

@@ -11,7 +11,7 @@ var sources = [
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
     "ggml/src/ggml-alloc.c",
-    "ggml/src/ggml-backend.c",
+    "ggml/src/ggml-backend.cpp",
     "ggml/src/ggml-quants.c",
     "ggml/src/ggml-aarch64.c",
 ]

ggml/include/ggml-blas.h

@@ -9,13 +9,13 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void);
+GGML_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
 
 // number of threads used for conversion to float
 // for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-cann.h

@@ -44,7 +44,7 @@ extern "C" {
  * @param device The index of the device to initialize.
  * @return A pointer to the initialized backend instance, or nullptr on failure.
  */
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
 
 /**
  * @brief Checks if a given backend is a CANN backend.
@@ -55,7 +55,7 @@ GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device);
  * @param backend The backend instance to check.
  * @return True if the backend is a CANN backend, false otherwise.
  */
-GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
 
 /**
  * @brief Retrieves the CANN buffer type for a specified device.
@@ -67,7 +67,7 @@ GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend);
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_API GGML_CALL ggml_backend_buffer_type_t
+GGML_API ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device);
 
 /**
@@ -78,14 +78,14 @@ ggml_backend_cann_buffer_type(int32_t device);
  *
  * @return The number of CANN devices available.
  */
-GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void);
+GGML_API int32_t ggml_backend_cann_get_device_count(void);
 
 /**
  * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
  *
  * @return A pointer to the host buffer type interface.
  */
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
 
 /**
  * @brief Retrieves the description of a specific CANN device.
@@ -97,7 +97,7 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_API GGML_CALL void ggml_backend_cann_get_device_description(
+GGML_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -112,9 +112,9 @@ GGML_API GGML_CALL void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_API GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device,
-                                                            size_t* free,
-                                                            size_t* total);
+GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+                                                  size_t* free,
+                                                  size_t* total);
 
 /**
  * @brief Set the logging callback for GGML.

ggml/include/ggml-cuda.h

@@ -3,6 +3,10 @@
 #include "ggml.h"
 #include "ggml-backend.h"
 
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
@@ -13,35 +17,33 @@
 #define GGML_CUDA_NAME "CUDA"
 #define GGML_CUBLAS_NAME "cuBLAS"
 #endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
 #define GGML_CUDA_MAX_DEVICES       16
 
 // backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_API int  ggml_backend_cuda_get_device_count(void);
+GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
 GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
+GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-metal.h

@@ -1,3 +1,5 @@
+// Note: this description is outdated
+//
 // An interface allowing to compute ggml_cgraph with Metal
 //
 // This is a fully functional interface that extends ggml with GPU support for Apple devices.
@@ -43,11 +45,11 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
 
 GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks

ggml/include/ggml-rpc.h

@@ -10,14 +10,14 @@ extern "C" {
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
-GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
+GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-sycl.h

@@ -23,20 +23,20 @@ GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
+GGML_API int  ggml_backend_sycl_get_device_count();
+GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
-// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-vulkan.h

@@ -13,16 +13,16 @@ extern "C" {
 GGML_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_API int  ggml_backend_vk_get_device_count(void);
+GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -187,16 +187,6 @@
 #    define GGML_API
 #endif
 
-#ifdef GGML_MULTIPLATFORM
-#    if defined(_WIN32)
-#        define GGML_CALL
-#    else
-#        define GGML_CALL __attribute__((__ms_abi__))
-#    endif
-#else
-#    define GGML_CALL
-#endif
-
 // TODO: support for clang
 #ifdef __GNUC__
 #    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
@@ -340,7 +330,7 @@ extern "C" {
     };
 
     // get ggml_status name string
-    GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
+    GGML_API const char * ggml_status_to_string(enum ggml_status status);
 
     // ieee 754-2008 half-precision float16
     // todo: make this not an integral type
@@ -716,46 +706,46 @@ extern "C" {
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
-    GGML_API GGML_CALL int64_t ggml_nelements   (const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL int64_t ggml_nrows       (const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
-    GGML_API           size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
+    GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nrows     (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes    (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes_pad(const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
 
-    GGML_API GGML_CALL int64_t ggml_blck_size(enum ggml_type type);
-    GGML_API GGML_CALL size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
-    GGML_API GGML_CALL size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
+    GGML_API int64_t ggml_blck_size(enum ggml_type type);
+    GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
+    GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
     GGML_DEPRECATED(
     GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
     "use ggml_row_size() instead");
 
-    GGML_API GGML_CALL const char * ggml_type_name(enum ggml_type type);
-    GGML_API GGML_CALL const char * ggml_op_name  (enum ggml_op   op);
-    GGML_API           const char * ggml_op_symbol(enum ggml_op   op);
+    GGML_API const char * ggml_type_name(enum ggml_type type);
+    GGML_API const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
-    GGML_API           const char * ggml_unary_op_name(enum ggml_unary_op op);
-    GGML_API GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 
-    GGML_API GGML_CALL size_t  ggml_element_size(const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
-    GGML_API GGML_CALL bool    ggml_is_quantized(enum ggml_type type);
+    GGML_API bool    ggml_is_quantized(enum ggml_type type);
 
     // TODO: temporary until model loading of ggml examples is refactored
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
-    GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL bool ggml_is_permuted  (const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL bool ggml_is_empty     (const struct ggml_tensor * tensor);
-    GGML_API           bool ggml_is_scalar    (const struct ggml_tensor * tensor);
-    GGML_API           bool ggml_is_vector    (const struct ggml_tensor * tensor);
-    GGML_API           bool ggml_is_matrix    (const struct ggml_tensor * tensor);
-    GGML_API           bool ggml_is_3d        (const struct ggml_tensor * tensor);
-    GGML_API           int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
+    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_empty     (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_3d        (const struct ggml_tensor * tensor);
+    GGML_API int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
-    GGML_API GGML_CALL bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
-    GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
-    GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
+    GGML_API bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
+    GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
+    GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
 
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
@@ -847,7 +837,7 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API GGML_CALL enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
 
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
@@ -1561,7 +1551,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    GGML_CALL void ggml_rope_yarn_corr_dims(
+    void ggml_rope_yarn_corr_dims(
         int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy

ggml/src/CMakeLists.txt

@@ -1325,7 +1325,7 @@ add_library(ggml
             ../include/ggml-backend.h
             ggml.c
             ggml-alloc.c
-            ggml-backend.c
+            ggml-backend.cpp
             ggml-quants.c
             ggml-quants.h
             ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}