ggml-backend : add device and backend reg interfaces (llama/9707)

ggml/include/ggml-backend.h

@@ -12,43 +12,52 @@ extern "C" {
     typedef struct ggml_backend_event * ggml_backend_event_t;
     typedef struct ggml_backend * ggml_backend_t;
     typedef void * ggml_backend_graph_plan_t;
+    typedef struct ggml_backend_reg * ggml_backend_reg_t;
+    typedef struct ggml_backend_device * ggml_backend_dev_t;
+
 
     //
-    // Backend buffer
+    // Backend buffer type
     //
 
-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+    GGML_API const char *          ggml_backend_buft_name          (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool                  ggml_backend_buft_is_host       (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_dev_t    ggml_backend_buft_get_device    (ggml_backend_buffer_type_t buft);
+
+    //
+    // Backend buffer
+    //
 
-    // buffer
     enum ggml_backend_buffer_usage {
         GGML_BACKEND_BUFFER_USAGE_ANY = 0,
         GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
         GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
     };
 
-    GGML_API           const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
-    GGML_API           ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+    GGML_API const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
+    GGML_API ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
 
     //
-    // Backend
+    // Backend (stream)
     //
 
     GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
@@ -64,9 +73,9 @@ extern "C" {
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
     // "offset" refers to the offset of the tensor data for setting/getting data
-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
 
     GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
 
@@ -76,65 +85,121 @@ extern "C" {
     GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
     GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    // NOTE: will be removed, use device version instead
     GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
     GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
     GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
     // asynchronous copy
     // the copy is performed after all the currently queued operations in backend_src
     // backend_dst will wait for the copy to complete before performing other operations
     // automatic fallback to sync copy if async is not supported
     GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
 
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event);
+    GGML_API ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend);
 
     //
-    // CPU backend
+    // Events
     //
 
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+    GGML_API ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device);
+    GGML_API void                 ggml_backend_event_free(ggml_backend_event_t event);
+    GGML_API void                 ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend);
+    GGML_API void                 ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                 ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event);
 
-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    //
+    // Backend device
+    //
 
-    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    enum ggml_backend_dev_type {
+        GGML_BACKEND_DEVICE_TYPE_CPU,
+        GGML_BACKEND_DEVICE_TYPE_GPU,
+        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
+        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
+        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
+    };
 
-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+    // functionality supported by the device
+    struct ggml_backend_dev_caps {
+        // asynchronous operations
+        bool async;
+        // pinned host buffer
+        bool host_buffer;
+        // event synchronization
+        bool events;
+    };
 
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
+    // all the device properties
+    struct ggml_backend_dev_props {
+        const char * name;
+        const char * description;
+        size_t memory_free;
+        size_t memory_total;
+        enum ggml_backend_dev_type type;
+        struct ggml_backend_dev_caps caps;
+    };
+
+    GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);
+    GGML_API const char *                  ggml_backend_dev_description(ggml_backend_dev_t device);
+    GGML_API void                          ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total);
+    GGML_API enum ggml_backend_dev_type    ggml_backend_dev_type(ggml_backend_dev_t device);
+    GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
+    GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
+    GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
+    GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
+    GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
+    GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);
+
+    GGML_API bool                          ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op);
+    GGML_API bool                          ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft);
+    GGML_API bool                          ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op);
+
+    //
+    // Backend (reg)
+    //
+
+    GGML_API const char *       ggml_backend_reg_name(ggml_backend_reg_t reg);
+    GGML_API size_t             ggml_backend_reg_dev_count(ggml_backend_reg_t reg);
+    GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
+    GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
+    GGML_API void               ggml_backend_reg_set_log_callback(ggml_backend_reg_t reg, ggml_log_callback log_callback, void * user_data);
+
+    // Functions that may be obtained using ggml_backend_reg_get_proc_address
+    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
 
     //
     // Backend registry
     //
 
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+    // Backend (reg) enumeration
+    GGML_API size_t             ggml_backend_reg_count(void);
+    GGML_API ggml_backend_reg_t ggml_backend_reg_get(size_t index);
+    GGML_API ggml_backend_reg_t ggml_backend_reg_by_name(const char * name);
+
+    // Device enumeration
+    GGML_API size_t             ggml_backend_dev_count(void);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_get(size_t index);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_by_name(const char * name);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type);
+
+    // Set the log callback for all registered backends
+    GGML_API void ggml_backend_set_log_callback(ggml_log_callback log_callback, void * user_data);
 
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name); // returns index of backend with name, or SIZE_MAX if not found
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is backend_name:params (params is optional)
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
+    // Direct backend (stream) initialization
+    // = ggml_backend_dev_init(ggml_backend_dev_by_name(name), params)
+    GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
+    GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
+    GGML_API ggml_backend_t ggml_backend_init_best(void);
 
     //
     // Backend scheduler
     //
 
-    // The backend scheduler allows for multiple backends to be used together
+    // The backend scheduler allows for multiple backend devices to be used together
     // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
     // The backends are selected based on:
     // - the backend that supports the operation
@@ -169,9 +234,9 @@ extern "C" {
     }
     */
 
-    struct ggml_backend_sched;
     typedef struct ggml_backend_sched * ggml_backend_sched_t;
 
+    // Evaluation callback for each node in the graph (set with ggml_backend_sched_set_eval_callback)
     // when ask == true, the scheduler wants to know if the user wants to observe this node
     // this allows the scheduler to batch nodes together in order to evaluate them in a single call
     //
@@ -226,7 +291,7 @@ extern "C" {
     GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
     GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
 
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
     GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
@@ -235,6 +300,26 @@ extern "C" {
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
 
 #ifdef  __cplusplus
 }

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
@@ -717,46 +707,46 @@ 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);
@@ -848,7 +838,7 @@ 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);
@@ -1568,7 +1558,7 @@ 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}

ggml/src/ggml-backend-impl.h

@@ -9,145 +9,229 @@ extern "C" {
 #endif
 
     //
-    // Backend buffer
+    // Backend buffer type
     //
 
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
     struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        const char *          (*get_name)      (ggml_backend_buffer_type_t buft);
         // allocate a buffer of this type
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        ggml_backend_buffer_t (*alloc_buffer)  (ggml_backend_buffer_type_t buft, size_t size);
         // tensor alignment
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft);
-        // max buffer size that can be allocated
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft);
-        // data size needed to allocate the tensor, including padding
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // check if tensor data is in host memory
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
+        size_t                (*get_alignment) (ggml_backend_buffer_type_t buft);
+        // (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
+        size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
+        // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
+        size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
+        // (optional) check if tensor data is in host memory (defaults to false)
+        bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
         struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
+        ggml_backend_dev_t device;
+        void * context;
     };
 
-    // buffer
-    typedef void * ggml_backend_buffer_context_t;
+    //
+    // Backend buffer
+    //
 
     struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)      (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)   (ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)      (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL memset_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
-        void         (*GGML_CALL set_tensor)    (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor)    (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor)    (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)         (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)         (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        const char * (*get_name)     (ggml_backend_buffer_t buffer);
+        // (optional) free the buffer
+        void         (*free_buffer)  (ggml_backend_buffer_t buffer);
+        // base address of the buffer
+        void *       (*get_base)     (ggml_backend_buffer_t buffer);
+        // (optional) initialize a tensor in the buffer (eg. add tensor extras)
+        void         (*init_tensor)  (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        // tensor data access
+        void         (*memset_tensor)(ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
+        void         (*set_tensor)   (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*get_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
+        bool         (*cpy_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        // clear the entire buffer
+        void         (*clear)        (ggml_backend_buffer_t buffer, uint8_t value);
+        // (optional) reset any internal state due to tensor initialization, such as tensor extras
+        void         (*reset)        (ggml_backend_buffer_t buffer);
     };
 
     struct ggml_backend_buffer {
         struct ggml_backend_buffer_i  iface;
         ggml_backend_buffer_type_t    buft;
-        ggml_backend_buffer_context_t context;
+        void * context;
         size_t size;
         enum ggml_backend_buffer_usage usage;
     };
 
-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
+    ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t buft,
+            struct ggml_backend_buffer_i      iface,
+                   void *                     context,
+                   size_t                     size);
 
     // do not use directly, use ggml_backend_tensor_copy instead
     bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
+    // multi-buffer
     // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
-    // Backend
+    // Backend (stream)
     //
 
-    typedef void * ggml_backend_context_t;
-
     struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+        const char * (*get_name)(ggml_backend_t backend);
 
-        void (*GGML_CALL free)(ggml_backend_t backend);
+        void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
         // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
         // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
+        void (*synchronize)(ggml_backend_t backend);
 
-        // compute graph with a plan (not used currently)
-        // create a new plan for a graph
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // (optional) compute graph with a plan (not used currently)
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
-        void                      (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
         // compute the graph with the plan
-        enum ggml_status          (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        enum ggml_status          (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph (always async if supported by the backend)
+        enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface
+        //            new backends should implement the device interface instead
 
+        // These functions are being moved to the device interface
         // check if the backend can compute an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+        bool (*supports_op)  (ggml_backend_t backend, const struct ggml_tensor * op);
 
         // check if the backend can use tensors allocated in a buffer type
-        bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+        bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
 
         // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
         // these should be expensive operations with large batch sizes that may benefit from running on this backend
         // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+        bool (*offload_op)   (ggml_backend_t backend, const struct ggml_tensor * op);
 
         // (optional) event synchronization
-        // create a new event that can record events on this backend instance
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        // record an event on the backend instance that created it
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        // wait for an event on on a different backend instance
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        // block until an event is recorded
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+        // record an event on this stream
+        void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
+        // wait for an event on on a different stream
+        void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
     };
 
     struct ggml_backend {
         ggml_guid_t guid;
-
         struct ggml_backend_i iface;
-        ggml_backend_context_t context;
+        ggml_backend_dev_t device;
+        void * context;
     };
 
     struct ggml_backend_event {
-        ggml_backend_t backend;
+        struct ggml_backend_device * device;
         void * context;
     };
 
     //
-    // Backend registry
+    // Backend device
     //
 
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
+    // Note: if additional properties are needed, we should add a struct with all of them
+    //       the current functions to obtain the properties can remain, since they are more convenient for often used properties
+    struct ggml_backend_device_i {
+        // device name: short identifier for this device, such as "CPU" or "CUDA0"
+        const char * (*get_name)(ggml_backend_dev_t dev);
+
+        // device description: short informative description of the device, could be the model name
+        const char * (*get_description)(ggml_backend_dev_t dev);
+
+        // device memory in bytes
+        void         (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);
+
+        // device type
+        enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev);
+
+        // device properties
+        void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props);
+
+        // backend (stream) initialization
+        ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params);
+
+        // preferred buffer type
+        ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev);
+
+        // (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device)
+        ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev);
+
+        // (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries)
+        ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size);
+
+        // check if the backend can compute an operation
+        bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
+        void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
+        void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
+    };
+
+    struct ggml_backend_device {
+        struct ggml_backend_device_i iface;
+        ggml_backend_reg_t reg;
+        void * context;
+    };
+
+    //
+    // Backend (reg)
+    //
+
+    struct ggml_backend_reg_i {
+        const char * (*get_name)(ggml_backend_reg_t reg);
+
+        // enumerate available devices
+        size_t             (*get_device_count)(ggml_backend_reg_t reg);
+        ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index);
+
+        // (optional) get a pointer to a function in the backend
+        // backends can add custom functions that are not part of the standard ggml-backend interface
+        void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
+
+        // (optional) set the log callback for the backend
+        void (*set_log_callback)(ggml_backend_reg_t reg, ggml_log_callback log_callback, void * user_data);
+    };
+
+    struct ggml_backend_reg {
+        // int api_version; // TODO: for dynamic loading
+        struct ggml_backend_reg_i iface;
+        void * context;
+    };
+
 
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+    // Internal backend registry API
+    void ggml_backend_register(ggml_backend_reg_t reg);
+    void ggml_backend_device_register(ggml_backend_dev_t device);
+    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
+    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend.cpp

@@ -0,0 +1,2496 @@
+// Note: porting this file to C++ is a work in progress
+
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <vector>
+
+// backend buffer type
+
+const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name(buft);
+}
+
+ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    return buft->iface.alloc_buffer(buft, size);
+}
+
+size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_alignment(buft);
+}
+
+size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    // get_max_size is optional, defaults to SIZE_MAX
+    if (buft->iface.get_max_size) {
+        return buft->iface.get_max_size(buft);
+    }
+    return SIZE_MAX;
+}
+
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buft->iface.get_alloc_size) {
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
+    }
+    return ggml_nbytes(tensor);
+}
+
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
+ggml_backend_dev_t ggml_backend_buft_get_device(ggml_backend_buffer_type_t buft) {
+    return buft->device;
+}
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+               ggml_backend_buffer_type_t buft,
+        struct ggml_backend_buffer_i      iface,
+               void *                     context,
+               size_t                     size) {
+    ggml_backend_buffer_t buffer = new ggml_backend_buffer {
+        /* .interface = */ iface,
+        /* .buft      = */ buft,
+        /* .context   = */ context,
+        /* .size      = */ size,
+        /* .usage     = */ GGML_BACKEND_BUFFER_USAGE_ANY
+    };
+
+    return buffer;
+}
+
+const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name(buffer);
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    delete buffer;
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // init_tensor is optional
+    if (buffer->iface.init_tensor) {
+        buffer->iface.init_tensor(buffer, tensor);
+    }
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
+}
+
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
+bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
+}
+
+void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    buffer->usage = usage;
+
+    // FIXME: add a generic callback to the buffer interface
+    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
+        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    }
+}
+
+enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
+    return buffer->usage;
+}
+
+ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
+    return buffer->buft;
+}
+
+void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
+    if (buffer->iface.reset) {
+        buffer->iface.reset(buffer);
+    }
+}
+
+bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
+    if (dst_buf->iface.cpy_tensor) {
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
+    }
+    return false;
+}
+
+// backend
+
+ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return NULL;
+    }
+    return backend->guid;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
+    return backend->iface.get_default_buffer_type(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
+}
+
+size_t ggml_backend_get_max_size(ggml_backend_t backend) {
+    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
+}
+
+void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (backend->iface.set_tensor_async == NULL) {
+        ggml_backend_tensor_set(tensor, data, offset, size);
+    } else {
+        backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (backend->iface.get_tensor_async == NULL) {
+        ggml_backend_tensor_get(tensor, data, offset, size);
+    } else {
+        backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.set_tensor(buf, tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
+}
+
+GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not supported by backend buffer");
+
+    buf->iface.memset_tensor(buf, tensor, value, offset, size);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    if (backend->iface.synchronize == NULL) {
+        return;
+    }
+
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend->iface.graph_plan_create != NULL);
+
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_free != NULL);
+
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
+
+    return backend->iface.graph_plan_compute(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
+    ggml_backend_synchronize(backend);
+    return err;
+}
+
+enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_op(backend->device, op);
+    }
+
+    return backend->iface.supports_op(backend, op);
+}
+
+bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_buft(backend->device, buft);
+    }
+
+    return backend->iface.supports_buft(backend, buft);
+}
+
+bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_offload_op(backend->device, op);
+    }
+
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
+}
+
+ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
+    return backend->device;
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+    } else if (ggml_backend_buffer_is_host(dst->buffer)) {
+        ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+    } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
+#endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (backend_dst->iface.cpy_tensor_async != NULL) {
+        if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
+            return;
+        }
+    }
+
+    // an async copy would normally happen after all the queued operations on both backends are completed
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
+}
+
+// events
+
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device) {
+    // null device is allowed for the transition period to the device interface
+    if (device == NULL || device->iface.event_new == NULL) {
+        return NULL;
+    }
+    return device->iface.event_new(device);
+}
+
+void ggml_backend_event_free(ggml_backend_event_t event) {
+    if (event == NULL) {
+        return;
+    }
+    event->device->iface.event_free(event->device, event);
+}
+
+void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend) {
+    GGML_ASSERT(backend->iface.event_record != NULL);
+
+    backend->iface.event_record(backend, event);
+}
+
+void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event->device->iface.event_synchronize);
+
+    event->device->iface.event_synchronize(event->device, event);
+}
+
+void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend->iface.event_wait != NULL);
+
+    backend->iface.event_wait(backend, event);
+}
+
+// Backend device
+
+const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
+    return device->iface.get_name(device);
+}
+
+const char * ggml_backend_dev_description(ggml_backend_dev_t device) {
+    return device->iface.get_description(device);
+}
+
+void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
+    device->iface.get_memory(device, free, total);
+}
+
+enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
+    return device->iface.get_type(device);
+}
+
+void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    device->iface.get_props(device, props);
+}
+
+ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device) {
+    return device->reg;
+}
+
+ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params) {
+    return device->iface.init_backend(device, params);
+}
+
+ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
+    return device->iface.get_buffer_type(device);
+}
+
+ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) {
+    return device->iface.get_host_buffer_type(device);
+}
+
+ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size) {
+    return device->iface.buffer_from_host_ptr(device, ptr, size, max_tensor_size);
+}
+
+bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    return device->iface.supports_op(device, op);
+}
+
+bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) {
+    return device->iface.supports_buft(device, buft);
+}
+
+bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    return device->iface.offload_op(device, op);
+}
+
+// Backend (reg)
+
+const char * ggml_backend_reg_name(ggml_backend_reg_t reg) {
+    return reg->iface.get_name(reg);
+}
+
+size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg) {
+    return reg->iface.get_device_count(reg);
+}
+
+ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index) {
+    return reg->iface.get_device(reg, index);
+}
+
+void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (!reg->iface.get_proc_address) {
+        return NULL;
+    }
+    return reg->iface.get_proc_address(reg, name);
+}
+
+void ggml_backend_reg_set_log_callback(ggml_backend_reg_t reg, ggml_log_callback log_callback, void * user_data) {
+    if (reg->iface.set_log_callback) {
+        reg->iface.set_log_callback(reg, log_callback, user_data);
+    }
+}
+
+// Backend registry
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+
+        register_backend(ggml_backend_cpu_reg());
+
+        // TODO: sycl, metal, vulkan, kompute, cann
+    }
+
+    void register_backend(ggml_backend_reg_t reg) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back(reg);
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index];
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
+            return reg;
+        }
+    }
+    return NULL;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+void ggml_backend_set_log_callback(ggml_log_callback log_callback, void * user_data) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        ggml_backend_reg_set_log_callback(reg, log_callback, user_data);
+    }
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU_FULL);
+    }
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, NULL);
+}
+
+// backend CPU
+
+static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
+
+static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "CPU";
+
+    GGML_UNUSED(buffer);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    uintptr_t data = (uintptr_t)buffer->context;
+
+    // align the buffer
+    if (data % TENSOR_ALIGNMENT != 0) {
+        data = GGML_PAD(data, TENSOR_ALIGNMENT);
+    }
+
+    return (void *)data;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
+    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface   = */ {
+            /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buf);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
+struct ggml_backend_cpu_context {
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    uint8_t *           work_data;
+    size_t              work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    delete[] cpu_ctx->work_data;
+    delete cpu_ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
+        if (cpu_plan->cplan.work_data == NULL) {
+            delete cpu_plan;
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    delete[] cpu_plan->cplan.work_data;
+    delete cpu_plan;
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        delete[] cpu_ctx->work_data;
+        cpu_ctx->work_data = new uint8_t[cplan.work_size];
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+static const struct ggml_backend_i ggml_backend_cpu_i = {
+    /* .get_name                = */ ggml_backend_cpu_get_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ ggml_backend_cpu_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    if (cpu_backend == NULL) {
+        delete ctx;
+        return NULL;
+    }
+
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+    GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
+    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
+}
+
+////////////////////////
+
+static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
+    // TODO
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_CPU_FULL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cpu_device_get_name(dev);
+    props->description = ggml_backend_cpu_device_get_description(dev);
+    props->type        = ggml_backend_cpu_device_get_type(dev);
+    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* async       */ false,
+        /* host_buffer */ false,
+        /* events      */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_cpu_device_init(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_OUT_PROD:
+            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
+        default:
+            return true;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
+    /* .get_name             = */ ggml_backend_cpu_device_get_name,
+    /* .get_description      = */ ggml_backend_cpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_cpu_device_get_type,
+    /* .get_props            = */ ggml_backend_cpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_cpu_device_init,
+    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+////////////////////////
+
+static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
+    return "CPU";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_cpu_device = {
+        /* .iface   = */ ggml_backend_cpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
+    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
+    /* .get_proc_address = */ NULL,
+    /* .set_log_callback = */ NULL,
+};
+
+ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    static struct ggml_backend_reg ggml_backend_cpu_reg = {
+        /* .iface   = */ ggml_backend_cpu_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_reg;
+}
+
+// multi-buffer buffer
+
+struct ggml_backend_multi_buffer_context {
+    ggml_backend_buffer_t * buffers;
+    size_t n_buffers;
+};
+
+static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
+static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_free(ctx->buffers[i]);
+    }
+
+    free(ctx->buffers);
+    free(ctx);
+}
+
+static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_clear(ctx->buffers[i], value);
+    }
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
+    /* .get_name        = */ ggml_backend_multi_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+    /* .get_base        = */ NULL,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ NULL,
+    /* .get_tensor      = */ NULL,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_multi_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+    ctx->n_buffers = n_buffers;
+    ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
+
+    GGML_ASSERT(ctx->buffers != NULL);
+
+    size_t total_size = 0;
+    for (size_t i = 0; i < n_buffers; i++) {
+        ctx->buffers[i] = buffers[i];
+        total_size += ggml_backend_buffer_get_size(buffers[i]);
+    }
+
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_i, ctx, total_size);
+}
+
+bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+}
+
+void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// scheduler
+
+#ifndef GGML_SCHED_MAX_BACKENDS
+#define GGML_SCHED_MAX_BACKENDS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
+#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#endif
+
+#ifndef GGML_SCHED_MAX_COPIES
+#define GGML_SCHED_MAX_COPIES 4
+#endif
+
+struct ggml_backend_sched_split {
+    int backend_id;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    // graph view of this split
+    struct ggml_cgraph graph;
+};
+
+struct ggml_backend_sched {
+    bool is_reset; // true if the scheduler has been reset since the last graph split
+    bool is_alloc;
+
+    int n_backends;
+
+    ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
+    ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
+    ggml_gallocr_t galloc;
+
+    // hash map of the nodes in the graph
+    struct ggml_hash_set  hash_set;
+    int                 * hv_tensor_backend_ids; // [hash_set.size]
+    struct ggml_tensor ** hv_tensor_copies;      // [hash_set.size][n_backends][n_copies]
+
+    int * node_backend_ids; // [graph_size]
+    int * leaf_backend_ids; // [graph_size]
+
+    int * prev_node_backend_ids; // [graph_size]
+    int * prev_leaf_backend_ids; // [graph_size]
+
+    // copy of the graph with modified inputs
+    struct ggml_cgraph graph;
+
+    // graph splits
+    struct ggml_backend_sched_split * splits;
+    int n_splits;
+    int splits_capacity;
+
+    // pipeline parallelism support
+    int n_copies;
+    int cur_copy;
+    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_graph_inputs;
+
+    struct ggml_context * ctx;
+
+    ggml_backend_sched_eval_callback callback_eval;
+    void * callback_eval_user_data;
+
+    char * context_buffer;
+    size_t context_buffer_size;
+
+    bool debug;
+};
+
+#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
+#define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
+#define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
+
+// returns the priority of the backend, lower id is higher priority
+static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return -1;
+}
+
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
+    ggml_backend_buffer_t buffer = tensor->buffer;
+    if (buffer == NULL) {
+        return -1;
+    }
+
+    // find highest prio backend that supports the buffer type and the op
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (ggml_backend_supports_buft(sched->backends[i], buffer->buft) &&
+            ggml_backend_supports_op(sched->backends[i], op)) {
+            return i;
+        }
+    }
+
+#ifndef NDEBUG
+    fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
+        __func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
+#endif
+
+    return -1;
+}
+
+#if 0
+#define GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
+#define GET_CAUSE(node) causes[hash_id(node)]
+#else
+#define SET_CAUSE(node, ...)
+#define GET_CAUSE(node) ""
+#endif
+
+// returns the backend that should be used for the node based on the current locations
+static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
+    // TODO: use supports_op to check if the backend supports the op
+
+    // assign pre-allocated nodes to their backend
+    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
+    if (cur_backend_id != -1) {
+        SET_CAUSE(tensor, "1.dst");
+        return cur_backend_id;
+    }
+
+    // view_src
+    if (tensor->view_src != NULL) {
+        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src, tensor);
+        if (cur_backend_id != -1) {
+            SET_CAUSE(tensor, "1.vsrc");
+            return cur_backend_id;
+        }
+    }
+
+    if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
+        // since the tensor is pre-allocated, it cannot be moved to another backend
+        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+    }
+
+    // graph input
+    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
+        cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
+        SET_CAUSE(tensor, "1.inp");
+        return cur_backend_id;
+    }
+
+    // operations with weights are preferably run on the same backend as the weights
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = tensor->src[i];
+        if (src == NULL) {
+            continue;
+        }
+        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
+            // check if a backend with higher prio wants to offload the op
+            if (src_backend_id == sched->n_backends - 1) {
+                for (int b = 0; b < src_backend_id; b++) {
+                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
+                        SET_CAUSE(tensor, "1.off");
+                        return b;
+                    }
+                }
+            }
+            SET_CAUSE(tensor, "1.wgt%d", i);
+            return src_backend_id;
+        }
+    }
+
+    return -1;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
+    } else {
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
+            ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
+            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+                sched->splits[cur_split].n_inputs);
+            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
+                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+            fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
+                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+static bool ggml_backend_sched_buffer_supported(ggml_backend_sched_t sched, struct ggml_tensor * t, int backend_id) {
+    ggml_backend_buffer_t buf = t->view_src ? t->view_src->buffer : t->buffer;
+    ggml_backend_buffer_type_t buft = NULL;
+
+    if (buf) {
+        // the tensor is already allocated
+        buft = buf->buft;
+    } else {
+        // see if the tensor already has a backend assigned, and use the buffer type of that backend
+        int tensor_backend_id = tensor_backend_id(t);
+        if (tensor_backend_id == -1 && t->view_src) {
+            tensor_backend_id = tensor_backend_id(t->view_src);
+        }
+        if (tensor_backend_id != -1) {
+            buft = sched->bufts[tensor_backend_id];
+        }
+    }
+
+    return buft != NULL && ggml_backend_supports_buft(sched->backends[backend_id], buft);
+}
+
+static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, struct ggml_tensor * node, int cur_backend_id, int * node_backend_id) {
+    if (ggml_backend_supports_op(sched->backends[cur_backend_id], node)) {
+        *node_backend_id = cur_backend_id;
+        SET_CAUSE(node, "2.sup");
+    }
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset splits
+    sched->n_splits = 0;
+    sched->n_graph_inputs = 0;
+    sched->is_reset = false;
+
+    struct ggml_init_params params = {
+        /* .mem_size =   */ sched->context_buffer_size,
+        /* .mem_buffer = */ sched->context_buffer,
+        /* .no_alloc =   */ true
+    };
+
+    ggml_free(sched->ctx);
+
+    sched->ctx = ggml_init(params);
+    if (sched->ctx == NULL) {
+        GGML_ABORT("%s: failed to initialize context\n", __func__);
+    }
+
+    // pass 1: assign backends to ops with pre-allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        int * leaf_backend_id = &tensor_backend_id(leaf);
+        // do not overwrite user assignments
+        if (*leaf_backend_id == -1) {
+            *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * node_backend_id = &tensor_backend_id(node);
+        // do not overwrite user assignments
+        if (*node_backend_id == -1) {
+            *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+
+#if 0
+            // src
+            if (node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                int * src_backend_id = &tensor_backend_id(src);
+                if (*src_backend_id == -1) {
+                    *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+                }
+            }
+#endif
+        }
+    }
+
+    // pass 2: expand current backend assignments
+    // assign the same backend to adjacent nodes
+    // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
+    // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
+    // ops unsupported by the backend being expanded will be left unassigned so that they can be assigned later when the locations of its inputs are known
+    // expand gpu down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand gpu up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+
+    // pass 3: upgrade nodes to higher prio backends with compatible buffer types
+    // if the tensor is already in the same buffer type (*) as another higher priority backend, we should move it there
+    // however, we also need to verify that the sources are in compatible buffer types
+    // (*) the actual requirement is more relaxed, the buffer type of the backend should be supported by all the users of this tensor further down the graph
+    // however, this is slow to verify, so we have a more strict requirement that the buffer type is the same
+    // this is not uncommon since multiple backends can use host memory, with the same buffer type (eg. BLAS and CPU)
+    // additionally, set remaining unassigned nodes to the backend with the most supported inputs
+    // only nodes that could not be assigned during expansion due to the backend not supporting the op should be unassigned at this point
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        int * node_backend_id = &tensor_backend_id(node);
+        if (*node_backend_id == -1) {
+            // unassigned node: find the backend with the most supported inputs
+            int n_supported_best = -1;
+            for (int b = 0; b < sched->n_backends; b++) {
+                if (ggml_backend_supports_op(sched->backends[b], node)) {
+                    int n_supported = 0;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if ((tensor_backend_id(src) != -1 || tensor_backend_id(src->view_src) != -1) && ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            n_supported++;
+                        }
+                    }
+                    if (n_supported > n_supported_best) {
+                        n_supported_best = n_supported;
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.best");
+                    }
+                }
+            }
+        } else {
+            // assigned node: upgrade to higher prio backend if possible
+            for (int b = 0; b < *node_backend_id; b++) {
+                if (sched->bufts[b] == sched->bufts[*node_backend_id] && ggml_backend_supports_op(sched->backends[b], node)) {
+                    bool supported = true;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if (!ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            supported = false;
+                            break;
+                        }
+                    }
+                    if (supported) {
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.upg");
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    // pass 4: assign backends to remaining src from dst and view_src
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * cur_backend_id = &tensor_backend_id(node);
+        if (node->view_src != NULL && *cur_backend_id == -1) {
+            *cur_backend_id = tensor_backend_id(node->view_src);
+            SET_CAUSE(node, "4.vsrc");
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            int * src_backend_id = &tensor_backend_id(src);
+            if (*src_backend_id == -1) {
+                if (src->view_src != NULL) {
+                    // views are always on the same backend as the source
+                    *src_backend_id = tensor_backend_id(src->view_src);
+                    SET_CAUSE(src, "4.vsrc");
+                } else {
+                    *src_backend_id = *cur_backend_id;
+                    SET_CAUSE(src, "4.cur");
+                }
+            }
+        }
+    }
+
+    // pass 5: split graph, find tensors that need to be copied
+    {
+        int i_split = 0;
+        struct ggml_backend_sched_split * split = &sched->splits[0];
+        // find the backend of the first split, skipping view ops
+        int i = 0;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (!ggml_is_view_op(node->op)) {
+                split->backend_id = tensor_backend_id(node);
+                break;
+            }
+        }
+        split->i_start = 0;
+        split->n_inputs = 0;
+        int cur_backend_id = split->backend_id;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+
+            const int node_backend_id = tensor_backend_id(node);
+
+            assert(node_backend_id != -1); // all nodes should be assigned by now
+
+            // check if we should start a new split based on the sources of the current node
+            bool need_new_split = false;
+            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * src = node->src[j];
+                    if (src == NULL) {
+                        continue;
+                    }
+                    // check if a weight is on a different backend
+                    // by starting a new split, the memory of the previously offloaded weights can be reused
+                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                        int src_backend_id = tensor_backend_id(src);
+                        if (src_backend_id != cur_backend_id) {
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                    // check if the split has too many inputs
+                    // FIXME: count the number of inputs instead of only checking when full
+                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                        const size_t id = hash_id(src);
+                        int src_backend_id = sched->hv_tensor_backend_ids[id];
+                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
+                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
+                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (node_backend_id != cur_backend_id || need_new_split) {
+                split->i_end = i;
+                i_split++;
+                if (i_split >= sched->splits_capacity) {
+                    sched->splits_capacity *= 2;
+                    sched->splits = (ggml_backend_sched_split *)
+                        realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
+                    GGML_ASSERT(sched->splits != NULL);
+                }
+                split = &sched->splits[i_split];
+                split->backend_id = node_backend_id;
+                split->i_start = i;
+                split->n_inputs = 0;
+                cur_backend_id = node_backend_id;
+            }
+
+            // find inputs that are not on the same backend
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+
+                size_t src_id = hash_id(src);
+                const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
+                assert(src_backend_id != -1); // all inputs should be assigned by now
+
+                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
+                    if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[src_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy;
+                            if (c == sched->cur_copy) {
+                                tensor_copy = src; // use the original tensor as the current copy
+                            } else {
+                                tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                                ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            }
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_graph_inputs = sched->n_graph_inputs++;
+                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        sched->graph_inputs[n_graph_inputs] = src;
+                    }
+                }
+
+                if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
+                    // create a copy of the input in the split's backend
+                    if (tensor_id_copy(src_id, cur_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[cur_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                            ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_inputs = split->n_inputs++;
+                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        split->inputs[n_inputs] = src;
+                    }
+                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
+                }
+            }
+        }
+        split->i_end = graph->n_nodes;
+        sched->n_splits = i_split + 1;
+    }
+
+    if (sched->debug) {
+        ggml_backend_sched_print_assignments(sched, graph);
+    }
+
+    // swap node_backend_ids and leaf _backend_ids with prevs
+    {
+        int * tmp = sched->node_backend_ids;
+        sched->node_backend_ids = sched->prev_node_backend_ids;
+        sched->prev_node_backend_ids = tmp;
+
+        tmp = sched->leaf_backend_ids;
+        sched->leaf_backend_ids = sched->prev_leaf_backend_ids;
+        sched->prev_leaf_backend_ids = tmp;
+    }
+
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    if (sched->graph.size < graph_size) {
+        sched->graph.size = graph_size;
+        sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = (ggml_tensor **) realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        GGML_ASSERT(sched->graph.nodes != NULL);
+        GGML_ASSERT(sched->graph.leafs != NULL);
+    }
+    sched->graph.n_nodes = 0;
+    sched->graph.n_leafs = 0;
+
+    struct ggml_cgraph * graph_copy = &sched->graph;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            assert(graph_copy->size > (graph_copy->n_nodes + 1));
+
+            struct ggml_tensor * input = split->inputs[j];
+            const size_t input_id = hash_id(input);
+            struct ggml_tensor * input_cpy = tensor_id_copy(input_id, split->backend_id, sched->cur_copy);
+
+            // add a dependency to the input source so that it is not freed before the copy is done
+            struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
+            input_dep->src[0] = input;
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->hv_tensor_backend_ids[input_id];
+            graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
+
+            // add a dependency to the input copy so that it is allocated at the start of the split
+            sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            assert(graph_copy->size > graph_copy->n_nodes);
+            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+
+    if (sched->n_copies > 1) {
+        // add input copies as leafs so that they are allocated first
+        for (int i = 0; i < sched->n_graph_inputs; i++) {
+            struct ggml_tensor * input = sched->graph_inputs[i];
+            size_t id = hash_id(input);
+            int backend_id = tensor_backend_id(input);
+            for (int c = 0; c < sched->n_copies; c++) {
+                struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                assert(graph_copy->size > graph_copy->n_leafs);
+                graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+            }
+        }
+
+        for (int i = 0; i < sched->n_splits; i++) {
+            struct ggml_backend_sched_split * split = &sched->splits[i];
+            int backend_id = split->backend_id;
+            for (int j = 0; j < split->n_inputs; j++) {
+                struct ggml_tensor * input = split->inputs[j];
+                size_t id = hash_id(input);
+                for (int c = 0; c < sched->n_copies; c++) {
+                    struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                    sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    assert(graph_copy->size > graph_copy->n_leafs);
+                    graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+                }
+            }
+        }
+    }
+
+    // add leafs from the original graph
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        assert(graph_copy->size > graph_copy->n_leafs);
+        graph_copy->leafs[graph_copy->n_leafs++] = leaf;
+    }
+}
+
+static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
+    bool backend_ids_changed = false;
+    for (int i = 0; i < sched->graph.n_nodes; i++) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
+            backend_ids_changed = true;
+            break;
+        }
+    }
+    if (!backend_ids_changed) {
+        for (int i = 0; i < sched->graph.n_leafs; i++) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
+                backend_ids_changed = true;
+                break;
+            }
+        }
+    }
+
+    // allocate graph
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+        // the re-allocation may cause the split inputs to be moved to a different address
+        ggml_backend_sched_synchronize(sched);
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        int split_backend_id = split->backend_id;
+        ggml_backend_t split_backend = sched->backends[split_backend_id];
+
+        // copy the input tensors to the split backend
+        for (int j = 0; j < split->n_inputs; j++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
+
+            if (input->flags & GGML_TENSOR_FLAG_INPUT) {
+                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy(input, input_cpy);
+            } else {
+                // wait for the split backend to finish using the input before overwriting it
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
+            }
+        }
+
+        if (!sched->callback_eval) {
+            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
+            if (ec != GGML_STATUS_SUCCESS) {
+                return ec;
+            }
+        } else {
+            // similar to ggml_backend_compare_graph_backend
+            for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
+                struct ggml_tensor * t = split->graph.nodes[j0];
+
+                // check if the user needs data from this node
+                bool need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+
+                int j1 = j0;
+
+                // determine the range [j0, j1] of nodes that can be computed together
+                while (!need && j1 < split->graph.n_nodes - 1) {
+                    t = split->graph.nodes[++j1];
+                    need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+                }
+
+                struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
+
+                enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
+                if (ec != GGML_STATUS_SUCCESS) {
+                    return ec;
+                }
+
+                // TODO: pass backend to the callback, then the user can decide if they want to synchronize
+                ggml_backend_synchronize(split_backend);
+
+                if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
+                    break;
+                }
+
+                j0 = j1;
+            }
+        }
+
+        // record the event of this copy
+        if (split->n_inputs > 0) {
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
+            }
+        }
+    }
+
+    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
+
+    return GGML_STATUS_SUCCESS;
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(
+        ggml_backend_t * backends,
+        ggml_backend_buffer_type_t * bufts,
+        int n_backends,
+        size_t graph_size,
+        bool parallel) {
+    GGML_ASSERT(n_backends > 0);
+    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
+    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+
+    struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
+
+    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    sched->n_backends = n_backends;
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+
+    // initialize hash table
+    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
+    sched->hash_set    = ggml_hash_set_new(graph_size);
+    sched->hv_tensor_backend_ids = (int *) malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = (ggml_tensor **) malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+
+    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
+    sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
+
+    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = (char *) malloc(sched->context_buffer_size);
+
+    const int initial_splits_capacity = 16;
+    sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
+    sched->splits_capacity = initial_splits_capacity;
+
+    for (int b = 0; b < n_backends; b++) {
+        sched->backends[b] = backends[b];
+        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
+        GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
+        if (sched->n_copies > 1) {
+            for (int c = 0; c < sched->n_copies; c++) {
+                sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
+            }
+        }
+    }
+
+    sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+
+    ggml_backend_sched_reset(sched);
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int b = 0; b < sched->n_backends; b++) {
+        for (int c = 0; c < sched->n_copies; c++) {
+            ggml_backend_event_free(sched->events[b][c]);
+        }
+    }
+    ggml_gallocr_free(sched->galloc);
+    ggml_free(sched->ctx);
+    ggml_hash_set_free(&sched->hash_set);
+    free(sched->splits);
+    free(sched->hv_tensor_backend_ids);
+    free(sched->hv_tensor_copies);
+    free(sched->node_backend_ids);
+    free(sched->leaf_backend_ids);
+    free(sched->prev_node_backend_ids);
+    free(sched->prev_leaf_backend_ids);
+    free(sched->context_buffer);
+    free(sched->graph.nodes);
+    free(sched->graph.leafs);
+    free(sched);
+}
+
+void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
+    // reset state for the next run
+    if (!sched->is_reset) {
+        ggml_hash_set_reset(&sched->hash_set);
+        memset(sched->hv_tensor_backend_ids, -1, sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+        memset(sched->hv_tensor_copies,       0, sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+        sched->is_reset = true;
+    }
+    sched->is_alloc = false;
+}
+
+bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
+    if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+        return false;
+    }
+
+    ggml_backend_sched_reset(sched);
+    ggml_backend_sched_synchronize(sched);
+
+    return true;
+}
+
+bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, graph);
+
+
+    if (!ggml_backend_sched_alloc_splits(sched)) {
+        return false;
+    }
+
+    sched->is_alloc = true;
+
+    return true;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
+    ggml_backend_sched_synchronize(sched);
+    return err;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    if (!sched->is_reset && !sched->is_alloc) {
+        ggml_backend_sched_reset(sched);
+    }
+
+    if (!sched->is_alloc) {
+        if (!ggml_backend_sched_alloc_graph(sched, graph)) {
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+    }
+
+    return ggml_backend_sched_compute_splits(sched);
+}
+
+void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_backend_synchronize(sched->backends[i]);
+    }
+}
+
+void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
+    sched->callback_eval = callback;
+    sched->callback_eval_user_data = user_data;
+}
+
+int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
+    return sched->n_splits;
+}
+
+int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    return sched->n_copies;
+}
+
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
+size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
+    return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
+}
+
+void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    tensor_backend_id(node) = backend_index;
+    SET_CAUSE(node, "usr");
+    sched->is_reset = false;
+}
+
+ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    int backend_index = tensor_backend_id(node);
+    if (backend_index == -1) {
+        return NULL;
+    }
+    return sched->backends[backend_index];
+}
+
+// utils
+
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->view_src != NULL);
+    GGML_ASSERT(tensor->view_src->buffer != NULL);
+    GGML_ASSERT(tensor->view_src->data != NULL);
+
+    tensor->buffer = tensor->view_src->buffer;
+    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
+}
+
+void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->data == NULL);
+    GGML_ASSERT(tensor->view_src == NULL);
+    GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
+    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
+                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
+
+    tensor->buffer = buffer;
+    tensor->data = addr;
+    ggml_backend_buffer_init_tensor(buffer, tensor);
+}
+
+static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
+    struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
+
+    GGML_ASSERT(src != NULL);
+    GGML_ASSERT(src->data && "graph must be allocated");
+
+    size_t id = ggml_hash_insert(&hash_set, src);
+    if (id == GGML_HASHSET_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(&hash_set, src)];
+    }
+
+    struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
+    if (src->view_src != NULL) {
+        dst->view_src = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
+        dst->view_offs = src->view_offs;
+    }
+    dst->op = src->op;
+    memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
+    ggml_set_name(dst, src->name);
+
+    // copy src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
+    }
+
+    node_copies[id] = dst;
+    return dst;
+}
+
+static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+    size_t id = ggml_hash_find(hash_set, src);
+    if (node_init[id]) {
+        return;
+    }
+    node_init[id] = true;
+
+    struct ggml_tensor * dst = node_copies[id];
+    if (dst->view_src != NULL) {
+        graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
+        ggml_backend_view_init(dst);
+    }
+    else {
+        ggml_backend_tensor_copy(src, dst);
+    }
+
+    // init src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        graph_copy_init_tensor(hash_set, node_copies, node_init, s);
+    }
+}
+
+struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
+    struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
+    struct ggml_tensor ** node_copies = (ggml_tensor **) calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = (bool *) calloc(hash_set.size, sizeof(node_init[0]));
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true
+    };
+
+    struct ggml_context * ctx_allocated = ggml_init(params);
+    struct ggml_context * ctx_unallocated = ggml_init(params);
+
+    if (ctx_allocated == NULL || ctx_unallocated == NULL) {
+        fprintf(stderr, "failed to allocate context for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    // dup nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
+    }
+
+    // allocate nodes
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
+    if (buffer == NULL) {
+        fprintf(stderr, "failed to allocate buffer for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
+
+    // copy data and init views
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_init_tensor(&hash_set, node_copies, node_init, node);
+    }
+
+    // build graph copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(&hash_set, node)];
+        graph_copy->nodes[i] = node_copy;
+    }
+    graph_copy->n_nodes = graph->n_nodes;
+
+    ggml_hash_set_free(&hash_set);
+    free(node_copies);
+    free(node_init);
+
+    return {
+        /* .buffer           = */ buffer,
+        /* .ctx_allocated    = */ ctx_allocated,
+        /* .ctx_unallocated  = */ ctx_unallocated,
+        /* .graph            = */ graph_copy,
+    };
+}
+
+void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
+    ggml_backend_buffer_free(copy.buffer);
+    ggml_free(copy.ctx_allocated);
+    ggml_free(copy.ctx_unallocated);
+}
+
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
+    if (copy.buffer == NULL) {
+        return false;
+    }
+
+    struct ggml_cgraph * g1 = graph;
+    struct ggml_cgraph * g2 = copy.graph;
+
+    assert(g1->n_nodes == g2->n_nodes);
+
+    for (int i = 0; i < g1->n_nodes; i++) {
+        //printf("eval %d/%d\n", i, g1->n_nodes);
+        struct ggml_tensor * t1 = g1->nodes[i];
+        struct ggml_tensor * t2 = g2->nodes[i];
+
+        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+        ggml_backend_graph_compute(backend1, &g1v);
+        ggml_backend_graph_compute(backend2, &g2v);
+
+        if (ggml_is_view_op(t1->op)) {
+            continue;
+        }
+
+        // compare results, calculate rms etc
+        if (!callback(i, t1, t2, user_data)) {
+            break;
+        }
+    }
+
+    ggml_backend_graph_copy_free(copy);
+
+    return true;
+}

ggml/src/ggml-blas.cpp

@@ -235,25 +235,25 @@ static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct g
 
 // backend interface
 
-GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+static const char * ggml_backend_blas_name(ggml_backend_t backend) {
     return "BLAS";
 
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
+static void ggml_backend_blas_free(ggml_backend_t backend) {
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
     delete ctx;
     delete backend;
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
     return ggml_backend_cpu_buffer_type();
 
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -285,7 +285,7 @@ GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     const struct ggml_tensor * src0 = op->src[0];
     const struct ggml_tensor * src1 = op->src[1];
 
@@ -300,7 +300,7 @@ GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, cons
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     return ggml_backend_buft_is_host(buft);
 
     GGML_UNUSED(backend);
@@ -322,11 +322,8 @@ static struct ggml_backend_i blas_backend_i = {
     /* .supports_op             = */ ggml_backend_blas_supports_op,
     /* .supports_buft           = */ ggml_backend_blas_supports_buft,
     /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_blas_guid(void) {
@@ -340,6 +337,7 @@ ggml_backend_t ggml_backend_blas_init(void) {
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_blas_guid(),
         /* .interface = */ blas_backend_i,
+        /* .device    = */ nullptr,
         /* .context   = */ ctx,
     };
 
@@ -356,7 +354,7 @@ ggml_backend_t ggml_backend_blas_init(void) {
     return backend;
 }
 
-GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
+bool ggml_backend_is_blas(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
 }
 

ggml/src/ggml-cann.cpp

@@ -560,7 +560,7 @@ struct ggml_backend_cann_buffer_context {
  * @return A pointer to a C-string containing the name of the buffer.
  */
 
-GGML_CALL static const char* ggml_backend_cann_buffer_get_name(
+static const char* ggml_backend_cann_buffer_get_name(
     ggml_backend_buffer_t buffer) {
     return "CANN";
 
@@ -576,7 +576,7 @@ GGML_CALL static const char* ggml_backend_cann_buffer_get_name(
  * @param buffer The buffer to check.
  * @return true if the buffer is a CANN buffer, false otherwise.
  */
-GGML_CALL static bool ggml_backend_buffer_is_cann(
+static bool ggml_backend_buffer_is_cann(
     ggml_backend_buffer_t buffer) {
     return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
 }
@@ -589,7 +589,7 @@ GGML_CALL static bool ggml_backend_buffer_is_cann(
  *
  * @param buffer The CANN buffer to free.
  */
-GGML_CALL static void ggml_backend_cann_buffer_free_buffer(
+static void ggml_backend_cann_buffer_free_buffer(
     ggml_backend_buffer_t buffer) {
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
@@ -605,7 +605,7 @@ GGML_CALL static void ggml_backend_cann_buffer_free_buffer(
  * @param buffer The CANN buffer whose base pointer is to be retrieved.
  * @return A pointer to the base of the device memory allocated for the buffer.
  */
-GGML_CALL static void* ggml_backend_cann_buffer_get_base(
+static void* ggml_backend_cann_buffer_get_base(
     ggml_backend_buffer_t buffer) {
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
@@ -625,9 +625,9 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
  * @param dst Pointer to the destination buffer where transformed data will be
  * stored.
  */
-GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
-                                                       const void* src,
-                                                       void* dst) {
+static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -677,7 +677,7 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
  * @param dst Pointer to the destination buffer where the Q4.0 formatted data
  * will be stored.
  */
-GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
+static void ggml_backend_cann_transform_back_q4_0(
     const ggml_tensor* tensor, void* src, void* dst) {
 
     int64_t n_elems = ggml_nelements(tensor);
@@ -726,9 +726,9 @@ GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
  * @param dst Pointer to the destination buffer where transformed data will be
  * stored.
  */
-GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
-                                                       const void* src,
-                                                       void* dst) {
+static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK8_0;
     size_t quant_bytes = n_elems * sizeof(uint8_t);
@@ -760,7 +760,7 @@ GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
  * @param dst Pointer to the destination buffer where the Q8.0 formatted data
  * will be stored.
  */
-GGML_CALL static void ggml_backend_cann_transform_back_q8_0(
+static void ggml_backend_cann_transform_back_q8_0(
     const ggml_tensor* tensor, const void* src, void* dst) {
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK8_0;
@@ -792,8 +792,8 @@ GGML_CALL static void ggml_backend_cann_transform_back_q8_0(
  * @param dst Pointer to the destination buffer where transformed data will be
  * stored.
  */
-GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor,
-                                                  const void* src, void* dst) {
+static void ggml_backend_cann_transform(ggml_tensor* tensor,
+                                        const void* src, void* dst) {
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
             ggml_backend_cann_transform_q4_0(tensor, src, dst);
@@ -818,7 +818,7 @@ GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor,
  * @param dst Pointer to the destination buffer where transformed tensor data
  * will be stored.
  */
-GGML_CALL static void ggml_backend_cann_transform_back(
+static void ggml_backend_cann_transform_back(
     const ggml_tensor* tensor, void* src, void* dst) {
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
@@ -841,7 +841,7 @@ GGML_CALL static void ggml_backend_cann_transform_back(
  * @param type The tensor type to check.
  * @return true if transformation is needed, false otherwise.
  */
-GGML_CALL static bool need_transform(ggml_type type) {
+static bool need_transform(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
@@ -860,7 +860,7 @@ GGML_CALL static bool need_transform(ggml_type type) {
  * @param buffer The CANN buffer from which to initialize the tensor.
  * @param tensor Pointer to the tensor to be initialized.
  */
-GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
+static void ggml_backend_cann_buffer_init_tensor(
     ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
     if (tensor->view_src != NULL && tensor->view_offs == 0) {
         GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
@@ -896,7 +896,7 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
  * @param offset Offset in the source data from where to start copying.
  * @param size Size of the data to be copied, in bytes.
  */
-GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
+static void ggml_backend_cann_buffer_set_tensor(
     ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
     size_t offset, size_t size) {
     ggml_backend_cann_buffer_context *ctx =
@@ -941,7 +941,7 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
  * @param offset Offset in the destination buffer where to start copying.
  * @param size Size of the data to be copied, in bytes.
  */
-GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
+static void ggml_backend_cann_buffer_get_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
     size_t offset, size_t size) {
     ggml_backend_cann_buffer_context* ctx =
@@ -975,7 +975,7 @@ GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
  * @param dst Pointer to the destination tensor where the data will be copied.
  * @return true if the copy operation succeeded, false otherwise.
  */
-GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor(
+static bool ggml_backend_cann_buffer_cpy_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) {
     if (ggml_backend_buffer_is_cann(src->buffer)) {
         ggml_backend_cann_buffer_context* src_ctx =
@@ -1017,7 +1017,7 @@ GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor(
  * @param buffer The CANN buffer to be cleared.
  * @param value The value to which each byte in the buffer will be set.
  */
-GGML_CALL static void ggml_backend_cann_buffer_clear(
+static void ggml_backend_cann_buffer_clear(
     ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
@@ -1065,7 +1065,7 @@ struct ggml_backend_cann_buffer_type_context {
  * @param buft Pointer to the buffer type context.
  * @return Const pointer to the C-style string containing the name.
  */
-GGML_CALL static const char* ggml_backend_cann_buffer_type_name(
+static const char* ggml_backend_cann_buffer_type_name(
     ggml_backend_buffer_type_t buft) {
     return "CANN";
 
@@ -1082,7 +1082,7 @@ GGML_CALL static const char* ggml_backend_cann_buffer_type_name(
  * @param size Size in bytes of the buffer to allocate.
  * @return Pointer to the allocated buffer, or nullptr if allocation fails.
  */
-GGML_CALL static ggml_backend_buffer_t
+static ggml_backend_buffer_t
 ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                            size_t size) {
     ggml_backend_cann_buffer_type_context* buft_ctx =
@@ -1121,7 +1121,7 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
  * @return The alignment requirement in bytes (fixed at 128 bytes for CANN
  * buffers).
  */
-GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment(
+static size_t ggml_backend_cann_buffer_type_get_alignment(
     ggml_backend_buffer_type_t buft) {
     return 128;
 
@@ -1142,7 +1142,7 @@ GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment(
  * @return The total allocation size in bytes required for the tensor in the
  * CANN buffer.
  */
-GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alloc_size(
+static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
@@ -1193,7 +1193,7 @@ static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_CALL ggml_backend_buffer_type_t
+ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
@@ -1231,7 +1231,7 @@ ggml_backend_cann_buffer_type(int32_t device) {
  * @param buft Pointer to the host buffer type context.
  * @return Const pointer to the C-style string containing the name.
  */
-GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
     return "CANN_Host";
 
     GGML_UNUSED(buft);
@@ -1246,7 +1246,7 @@ GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backe
  * @param buft Pointer to the host buffer context.
  * @return Const pointer to the C-style string containing the name.
  */
-GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
     return "CANN_Host";
 
     GGML_UNUSED(buffer);
@@ -1260,7 +1260,7 @@ GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_bu
  *
  * @param buffer The CANN host buffer to free.
  */
-GGML_CALL static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
+static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
     ACL_CHECK(aclrtFreeHost(buffer->context));
 }
 
@@ -1294,7 +1294,7 @@ static void * ggml_cann_host_malloc(size_t size) {
  * @param size Size in bytes of the host buffer to allocate.
  * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
  */
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     void * hostPtr = ggml_cann_host_malloc(size);
 
     if (hostPtr == nullptr) {
@@ -1316,7 +1316,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_
  * Provides function pointers for allocating, querying properties, and managing
  * memory for CANN buffer types in the GGML backend.
  */
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
+ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
         /* .iface    = */ {
             /* .get_name         = */ ggml_backend_cann_host_buffer_type_name,
@@ -1326,6 +1326,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
+        /* .device   = */ nullptr,
         /* .context  = */ nullptr,
     };
 
@@ -1495,7 +1496,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
  * @param backend Pointer to the CANN backend structure.
  * @return A pointer to a constant string representing the backend name.
  */
-GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) {
+static const char* ggml_backend_cann_name(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
 
@@ -1510,7 +1511,7 @@ GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) {
  *
  * @param backend Pointer to the CANN backend structure to be freed.
  */
-GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) {
+static void ggml_backend_cann_free(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
     ACL_CHECK(aclrtSynchronizeDevice());
@@ -1535,7 +1536,7 @@ GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) {
  * @param backend Pointer to the CANN backend structure.
  * @return Pointer to the buffer type structure for the CANN backend.
  */
-GGML_CALL static ggml_backend_buffer_type_t
+static ggml_backend_buffer_type_t
 ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
@@ -1556,11 +1557,11 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
  * @param offset Offset in bytes within the host data.
  * @param size Size of the data to copy in bytes.
  */
-GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                                         ggml_tensor *tensor,
-                                                         const void *data,
-                                                         size_t offset,
-                                                         size_t size) {
+static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
+                                               ggml_tensor *tensor,
+                                               const void *data,
+                                               size_t offset,
+                                               size_t size) {
     ggml_backend_cann_context *cann_ctx =
         (ggml_backend_cann_context *)backend->context;
 
@@ -1587,7 +1588,7 @@ GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
     }
 }
 
-GGML_CALL static void ggml_backend_cann_get_tensor_async(
+static void ggml_backend_cann_get_tensor_async(
     ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
     ggml_backend_cann_context *cann_ctx =
@@ -1626,7 +1627,7 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
  * @param dst Pointer to the destination tensor to copy data to.
  * @return true if the copy operation succeeds, false otherwise.
  */
-GGML_CALL static bool ggml_backend_cann_cpy_tensor_async(
+static bool ggml_backend_cann_cpy_tensor_async(
     ggml_backend_t backend_src, ggml_backend_t backend_dst,
     const ggml_tensor* src, ggml_tensor* dst) {
     GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
@@ -1694,7 +1695,7 @@ GGML_CALL static bool ggml_backend_cann_cpy_tensor_async(
  *
  * @param backend Pointer to the CANN backend structure to synchronize.
  */
-GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
+static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
 
@@ -1715,7 +1716,7 @@ GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
  * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation
  *         completes successfully, otherwise an appropriate error status.
  */
-GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute(
+static enum ggml_status ggml_backend_cann_graph_compute(
     ggml_backend_t backend, ggml_cgraph* cgraph) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
@@ -1753,7 +1754,7 @@ GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute(
  * @return bool Returns true if the operation is supported by the backend,
  *              otherwise false.
  */
-GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                                                     const ggml_tensor* op) {
     switch (op->op) {
         case GGML_OP_UNARY:
@@ -1875,7 +1876,7 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
  * @return bool Returns true if the CANN backend supports the buffer type,
  *              otherwise false.
  */
-GGML_CALL static bool ggml_backend_cann_supports_buft(
+static bool ggml_backend_cann_supports_buft(
     ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     if (ggml_backend_buft_is_cann(buft)) {
         ggml_backend_cann_context * cann_ctx =
@@ -1901,7 +1902,7 @@ GGML_CALL static bool ggml_backend_cann_supports_buft(
  * @return bool Returns true if the operation should be offloaded, otherwise
  * false.
  */
-GGML_CALL static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
                                                    const ggml_tensor* op) {
     const int min_batch_size = 32;
     GGML_UNUSED(backend);
@@ -2021,11 +2022,8 @@ static ggml_backend_i ggml_backend_cann_interface = {
     /* .supports_op             = */ ggml_backend_cann_supports_op,
     /* .supports_buft           = */ ggml_backend_cann_supports_buft,
     /* .offload_op              = */ ggml_backend_cann_offload_op,
-    /* .event_new               = */ ggml_backend_cann_event_new,
-    /* .event_free              = */ ggml_backend_cann_event_free,
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
-    /* .event_synchronize       = */ ggml_backend_cann_event_synchronize,
 };
 
 /**
@@ -2042,7 +2040,7 @@ static ggml_guid_t ggml_backend_cann_guid() {
     return &guid;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
+ggml_backend_t ggml_backend_cann_init(int32_t device) {
     aclInit(nullptr);
     if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
         GGML_CANN_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
@@ -2058,75 +2056,30 @@ GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,
+                         /* .device    = */ nullptr,
                          /* .context   = */ ctx};
 
     return cann_backend;
 }
 
-GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend) {
+bool ggml_backend_is_cann(ggml_backend_t backend) {
     return backend != NULL &&
            ggml_guid_matches(backend->guid, ggml_backend_cann_guid());
 }
 
-GGML_CALL int32_t ggml_backend_cann_get_device_count() {
+int32_t ggml_backend_cann_get_device_count() {
     return ggml_cann_info().device_count;
 }
 
-GGML_CALL void ggml_backend_cann_get_device_description(
+void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size) {
     ggml_cann_set_device(device);
     const char* soc_name = aclrtGetSocName();
     snprintf(description, description_size, "%s", soc_name);
 }
 
-GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
-                                                   size_t* total) {
+void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
+                                         size_t* total) {
     ggml_cann_set_device(device);
     ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
 }
-
-// backend registry
-/**
- * @brief Initializes a CANN backend based on the provided parameters.
- *
- * This function initializes a CANN backend using the device index and then
- * initializes the backend using `ggml_backend_cann_init`.
- *
- * @param params Parameters for initialization (unused in this implementation).
- * @param user_data User data containing the device index to initialize the
- * backend.
- * @return ggml_backend_t The initialized CANN backend.
- */
-GGML_CALL static ggml_backend_t ggml_backend_reg_cann_init(const char* params,
-                                                           void* user_data) {
-    ggml_backend_t cann_backend =
-        ggml_backend_cann_init((int)(intptr_t)user_data);
-    return cann_backend;
-
-    GGML_UNUSED(params);
-}
-
-extern "C" GGML_CALL int ggml_backend_cann_reg_devices();
-
-/**
- * @brief Registers CANN (Ascend) devices as backend options.
- *
- * This function initializes ACL, retrieves the number of available CANN
- * devices, and registers each device as a backend option using
- * `ggml_backend_register`. Each device is given a unique name based on
- * `GGML_CANN_NAME` followed by its index.
- *
- * @return int The number of CANN devices registered.
- */
-GGML_CALL int ggml_backend_cann_reg_devices() {
-    uint32_t device_count = ggml_backend_cann_get_device_count();
-    // initialization
-    for (uint32_t i = 0; i < device_count; i++) {
-        char name[128];
-        snprintf(name, sizeof(name), "CANN%d", i);
-        ggml_backend_register(name, ggml_backend_reg_cann_init,
-                              ggml_backend_cann_buffer_type(i),
-                              (void*)(intptr_t)i);
-    }
-    return device_count;
-}

ggml/src/ggml-cuda.cu

@@ -101,11 +101,11 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
     int id = -1; // in case cudaGetDevice fails
     cudaGetDevice(&id);
 
-    GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
+    GGML_CUDA_LOG_ERROR(GGML_CUDA_NAME " error: %s\n", msg);
     GGML_CUDA_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
     GGML_CUDA_LOG_ERROR("  %s\n", stmt);
-    // abort with GGML_ASSERT to get a stack trace
-    GGML_ABORT("CUDA error");
+    // abort with GGML_ABORT to get a stack trace
+    GGML_ABORT(GGML_CUDA_NAME " error");
 }
 
 // this is faster on Windows
@@ -329,7 +329,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
                 return;
             }
         }
-        GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
+        GGML_CUDA_LOG_WARN(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
         ggml_cuda_set_device(device);
         CUDA_CHECK(cudaFree(ptr));
         pool_size -= size;
@@ -459,26 +459,26 @@ struct ggml_backend_cuda_buffer_context {
     }
 };
 
-GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
     return ctx->name.c_str();
 }
 
-GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
+static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
     return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
     delete ctx;
 }
 
-GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
     return ctx->dev_ptr;
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     if (tensor->view_src != NULL) {
@@ -498,7 +498,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
     }
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
@@ -506,7 +506,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
@@ -514,7 +514,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
@@ -522,7 +522,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
-GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     if (ggml_backend_buffer_is_cuda(src->buffer)) {
         ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
         ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
@@ -543,7 +543,7 @@ GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t
     GGML_UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
@@ -552,7 +552,7 @@ GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffe
     CUDA_CHECK(cudaDeviceSynchronize());
 }
 
-static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
+static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
     /* .get_name        = */ ggml_backend_cuda_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
@@ -571,17 +571,17 @@ struct ggml_backend_cuda_buffer_type_context {
     std::string name;
 };
 
-GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_cuda_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
 
     return ctx->name.c_str();
 }
 
 static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cuda_buffer_type_name;
+    return buft->iface.get_name == ggml_backend_cuda_buffer_type_get_name;
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
 
     ggml_cuda_set_device(buft_ctx->device);
@@ -602,13 +602,13 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
     return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
 }
 
-GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 128;
 
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
@@ -623,8 +623,8 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backen
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_cuda_buffer_type_name,
+static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cuda_buffer_type_get_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
@@ -632,7 +632,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
     /* .is_host          = */ NULL,
 };
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
+ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
 
@@ -645,9 +645,10 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
     static bool ggml_backend_cuda_buffer_type_initialized = false;
 
     if (!ggml_backend_cuda_buffer_type_initialized) {
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
+        for (int i = 0; i < ggml_backend_cuda_get_device_count(); i++) {
             ggml_backend_cuda_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
+                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), i),
                 /* .context  = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)},
             };
         }
@@ -717,7 +718,7 @@ struct ggml_backend_cuda_split_buffer_context {
     std::vector<ggml_tensor_extra_gpu *> tensor_extras;
 };
 
-GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
     return GGML_CUDA_NAME "_Split";
 
     GGML_UNUSED(buffer);
@@ -728,19 +729,19 @@ static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
     GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
 }
 
-GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
     delete ctx;
 }
 
-GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
     // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
     return (void *)0x1000;
 
     GGML_UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
 
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
@@ -788,7 +789,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
     tensor->extra = extra;
 }
 
-GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
     GGML_ASSERT(size == ggml_nbytes(tensor));
@@ -826,7 +827,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buf
     }
 }
 
-GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
     GGML_ASSERT(size == ggml_nbytes(tensor));
@@ -864,12 +865,12 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buf
     }
 }
 
-GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     GGML_UNUSED(buffer);
     GGML_UNUSED(value);
 }
 
-static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
+static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
     /* .get_name        = */ ggml_backend_cuda_split_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_cuda_split_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cuda_split_buffer_get_base,
@@ -884,17 +885,17 @@ static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
 
 // cuda split buffer type
 
-GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_cuda_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return GGML_CUDA_NAME "_Split";
 
     GGML_UNUSED(buft);
 }
 
 static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_name;
+    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_get_name;
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
     // instead, we allocate them for each tensor separately in init_tensor
     // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
@@ -904,13 +905,13 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc
     return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
 }
 
-GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 128;
 
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
 
     size_t total_size = 0;
@@ -937,14 +938,14 @@ GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_
     return total_size;
 }
 
-GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_cuda_split_buffer_type_name,
+static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cuda_split_buffer_type_get_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
@@ -952,7 +953,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
+ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
 
@@ -981,6 +982,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const f
 
     struct ggml_backend_buffer_type buft {
         /* .iface   = */ ggml_backend_cuda_split_buffer_type_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
         /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr},
     };
 
@@ -990,19 +992,19 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const f
 
 // host buffer type
 
-GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
     return GGML_CUDA_NAME "_Host";
 
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
     return GGML_CUDA_NAME "_Host";
 
     GGML_UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
 
@@ -1024,7 +1026,7 @@ static void * ggml_cuda_host_malloc(size_t size) {
     return ptr;
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     void * ptr = ggml_cuda_host_malloc(size);
 
     if (ptr == nullptr) {
@@ -1040,7 +1042,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_
     return buffer;
 }
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
+ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
         /* .iface    = */ {
             /* .get_name         = */ ggml_backend_cuda_host_buffer_type_name,
@@ -1050,6 +1052,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
         /* .context  = */ nullptr,
     };
 
@@ -2383,26 +2386,26 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
 
 // backend
 
-GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
+static const char * ggml_backend_cuda_get_name(ggml_backend_t backend) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     return cuda_ctx->name.c_str();
 }
 
-GGML_CALL static void ggml_backend_cuda_free(ggml_backend_t backend) {
+static void ggml_backend_cuda_free(ggml_backend_t backend) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     delete cuda_ctx;
     delete backend;
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     return ggml_backend_cuda_buffer_type(cuda_ctx->device);
 }
 
-GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -2411,7 +2414,7 @@ GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend,
     CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
 }
 
-GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -2420,7 +2423,7 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
     CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
 }
 
-GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
+static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
@@ -2475,7 +2478,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
     return true;
 }
 
-GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream()));
@@ -2534,7 +2537,7 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     return true;
 }
 
-GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
     ggml_cuda_set_device(cuda_ctx->device);
@@ -2806,8 +2809,187 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
     return GGML_STATUS_SUCCESS;
 }
 
-GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
+static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream()));
+}
+
+static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    if (ggml_backend_is_cuda(backend)) {
+        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0));
+    } else {
+#if 0
+        // untested
+        auto wait_fn = [](void * user_data) {
+            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
+            ggml_backend_event_synchronize(event);
+        };
+
+        CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
+#endif
+        GGML_ABORT("fatal error");
+    }
+}
+
+static const ggml_backend_i ggml_backend_cuda_interface = {
+    /* .get_name                = */ ggml_backend_cuda_get_name,
+    /* .free                    = */ ggml_backend_cuda_free,
+    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
+    /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
+    /* .synchronize             = */ ggml_backend_cuda_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
+    /* .event_record            = */ ggml_backend_cuda_event_record,
+    /* .event_wait              = */ ggml_backend_cuda_event_wait,
+};
+
+static ggml_guid_t ggml_backend_cuda_guid() {
+    static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
+    return &guid;
+}
+
+bool ggml_backend_is_cuda(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
+}
+
+int ggml_backend_cuda_get_device_count() {
+    return ggml_cuda_info().device_count;
+}
+
+void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
+    snprintf(description, description_size, "%s", prop.name);
+}
+
+void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
+    ggml_cuda_set_device(device);
+
+    CUDA_CHECK(cudaMemGetInfo(free, total));
+}
+
+bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
+    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
+        return false;
+    }
+
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
+    cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+
+        GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
+        return false;
+    }
+    return true;
+#else
+    return false;
+#endif
+}
+
+void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
+    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
+        return;
+    }
+
+    cudaError_t err = cudaHostUnregister(buffer);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+    }
+}
+
+
+// backend device
+
+struct ggml_backend_cuda_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemGetInfo(free, total));
+}
+
+static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
+}
+
+static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cuda_device_get_name(dev);
+    props->description = ggml_backend_cuda_device_get_description(dev);
+    props->type        = ggml_backend_cuda_device_get_type(dev);
+    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
+#ifdef GGML_CUDA_NO_PEER_COPY
+    bool events = false;
+#else
+    bool events = true;
+#endif
+
+    props->caps = {
+        /* async       */ true,
+        /* host_buffer */ host_buffer,
+        /* events      */ events,
+    };
+}
+
+static ggml_backend_t ggml_backend_cuda_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ggml_backend_cuda_init(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ggml_backend_cuda_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cuda_host_buffer_type();
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    GGML_UNUSED(dev);
+    GGML_UNUSED(ptr);
+    GGML_UNUSED(size);
+    GGML_UNUSED(max_tensor_size);
+    return nullptr;
+}
+
+// TODO: move these functions here
+static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -3021,7 +3203,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
                 return true;
             }
-            const int cc = ggml_cuda_info().devices[cuda_ctx->device].cc;
+            const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
             return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }
         case GGML_OP_CROSS_ENTROPY_LOSS:
@@ -3031,115 +3213,170 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         default:
             return false;
     }
-
-    GGML_UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_cuda_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
     if (ggml_backend_buft_is_cuda_split(buft)) {
         return true;
     }
 
     if (ggml_backend_buft_is_cuda(buft)) {
-        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+        ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
         ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
-        return buft_ctx->device == cuda_ctx->device;
+        return buft_ctx->device == dev_ctx->device;
     }
 
     return false;
 }
 
-GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
     return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
            (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
 
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 }
 
-static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) {
+static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
 #ifdef GGML_CUDA_NO_PEER_COPY
     return nullptr;
 #else
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
 
-    ggml_cuda_set_device(cuda_ctx->device);
+    ggml_cuda_set_device(dev_ctx->device);
 
     cudaEvent_t event;
     CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
 
     return new ggml_backend_event {
-        /* .backend = */ backend,
+        /* .device  = */ dev,
         /* .context = */ event,
     };
 #endif
 }
 
-static void ggml_backend_cuda_event_free(ggml_backend_event_t event) {
-    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
+static void ggml_backend_cuda_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    GGML_UNUSED(dev);
 
+    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
     delete event;
 }
 
-static void ggml_backend_cuda_event_record(ggml_backend_event_t event) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context;
+static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    GGML_UNUSED(dev);
+    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+}
+
+static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+    /* .get_name                = */ ggml_backend_cuda_device_get_name,
+    /* .get_description         = */ ggml_backend_cuda_device_get_description,
+    /* .get_memory              = */ ggml_backend_cuda_device_get_memory,
+    /* .get_type                = */ ggml_backend_cuda_device_get_type,
+    /* .get_props               = */ ggml_backend_cuda_device_get_props,
+    /* .init_backend            = */ ggml_backend_cuda_device_init,
+    /* .get_buffer_type         = */ ggml_backend_cuda_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cuda_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ ggml_backend_cuda_device_buffer_from_host_ptr,
+    /* .supports_op             = */ ggml_backend_cuda_device_supports_op,
+    /* .supports_buft           = */ ggml_backend_cuda_device_supports_buft,
+    /* .offload_op              = */ ggml_backend_cuda_device_offload_op,
+    /* .event_new               = */ ggml_backend_cuda_device_event_new,
+    /* .event_free              = */ ggml_backend_cuda_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
+};
 
-    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream()));
+// backend reg
+
+struct ggml_backend_cuda_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cuda_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CUDA_NAME;
 }
 
-static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+static size_t ggml_backend_cuda_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
+    return ctx->devices.size();
+}
 
-    if (ggml_backend_is_cuda(event->backend)) {
-        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0));
-    } else {
-#if 0
-        // untested
-        auto wait_fn = [](void * user_data) {
-            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
-            ggml_backend_event_synchronize(event);
-        };
+static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
 
-        CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
-#endif
-        GGML_ABORT("fatal error");
+static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
+        return (void *)ggml_backend_cuda_split_buffer_type;
+    }
+    if (strcmp(name, "ggml_backend_register_host_buffer") == 0) {
+        return (void *)ggml_backend_cuda_register_host_buffer;
     }
+    if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
+        return (void *)ggml_backend_cuda_unregister_host_buffer;
+    }
+    return nullptr;
 }
 
-static void ggml_backend_cuda_event_synchronize(ggml_backend_event_t event) {
-    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+static void ggml_backend_cuda_reg_set_log_callback(ggml_backend_reg_t reg, ggml_log_callback log_callback, void * user_data) {
+    GGML_UNUSED(reg);
+    ggml_backend_cuda_log_set_callback(log_callback, user_data);
 }
 
-static ggml_backend_i ggml_backend_cuda_interface = {
-    /* .get_name                = */ ggml_backend_cuda_name,
-    /* .free                    = */ ggml_backend_cuda_free,
-    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
-    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
-    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
-    /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
-    /* .synchronize             = */ ggml_backend_cuda_synchronize,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
-    /* .supports_op             = */ ggml_backend_cuda_supports_op,
-    /* .supports_buft           = */ ggml_backend_cuda_supports_buft,
-    /* .offload_op              = */ ggml_backend_cuda_offload_op,
-    /* .event_new               = */ ggml_backend_cuda_event_new,
-    /* .event_free              = */ ggml_backend_cuda_event_free,
-    /* .event_record            = */ ggml_backend_cuda_event_record,
-    /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .event_synchronize       = */ ggml_backend_cuda_event_synchronize,
+static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
+    /* .get_name          = */ ggml_backend_cuda_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,
+    /* .get_device_get    = */ ggml_backend_cuda_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
+    /* .set_log_callback  = */ ggml_backend_cuda_reg_set_log_callback,
 };
 
-static ggml_guid_t ggml_backend_cuda_guid() {
-    static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
-    return &guid;
+// backend registry
+ggml_backend_reg_t ggml_backend_cuda_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
+
+            for (int i = 0; i < ggml_cuda_info().device_count; i++) {
+                ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
+
+                ggml_cuda_set_device(i);
+                cudaDeviceProp prop;
+                CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
+                dev_ctx->description = prop.name;
+
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .interface = */ ggml_backend_cuda_device_interface,
+                    /* .reg       = */ &reg,
+                    /* .context   = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .interface = */ ggml_backend_cuda_reg_interface,
+                /* .context   = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
+ggml_backend_t ggml_backend_cuda_init(int device) {
     if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
         GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
         return nullptr;
@@ -3154,82 +3391,9 @@ GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
     ggml_backend_t cuda_backend = new ggml_backend {
         /* .guid      = */ ggml_backend_cuda_guid(),
         /* .interface = */ ggml_backend_cuda_interface,
-        /* .context   = */ ctx
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
+        /* .context   = */ ctx,
     };
 
     return cuda_backend;
 }
-
-GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
-}
-
-GGML_CALL int ggml_backend_cuda_get_device_count() {
-    return ggml_cuda_info().device_count;
-}
-
-GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
-    cudaDeviceProp prop;
-    CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
-    snprintf(description, description_size, "%s", prop.name);
-}
-
-GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
-    ggml_cuda_set_device(device);
-
-    CUDA_CHECK(cudaMemGetInfo(free, total));
-}
-
-GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
-    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
-        return false;
-    }
-
-#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
-    cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
-    if (err != cudaSuccess) {
-        // clear the error
-        cudaGetLastError();
-
-        GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
-                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
-        return false;
-    }
-    return true;
-#else
-    return false;
-#endif
-}
-
-GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
-    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
-        return;
-    }
-
-    cudaError_t err = cudaHostUnregister(buffer);
-    if (err != cudaSuccess) {
-        // clear the error
-        cudaGetLastError();
-    }
-}
-
-// backend registry
-GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
-    ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data);
-    return cuda_backend;
-
-    GGML_UNUSED(params);
-}
-
-extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
-
-GGML_CALL int ggml_backend_cuda_reg_devices() {
-    int device_count = ggml_backend_cuda_get_device_count();
-    //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
-    for (int i = 0; i < device_count; i++) {
-        char name[128];
-        snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i);
-        ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i);
-    }
-    return device_count;
-}

ggml/src/ggml-kompute.cpp

@@ -1921,6 +1921,7 @@ ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
         for (const auto & dev : devices) {
             vec.push_back({
                 /* .iface   = */ ggml_backend_kompute_buffer_type_interface,
+                /* .device  = */ nullptr,
                 /* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc)
             });
         }
@@ -1989,11 +1990,8 @@ static struct ggml_backend_i kompute_backend_i = {
     /* .supports_op             = */ ggml_backend_kompute_supports_op,
     /* .supports_buft           = */ ggml_backend_kompute_supports_buft,
     /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_kompute_guid() {
@@ -2008,6 +2006,7 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
     ggml_backend_t kompute_backend = new ggml_backend {
         /* .guid      = */ ggml_backend_kompute_guid(),
         /* .interface = */ kompute_backend_i,
+        /* .device    = */ nullptr,
         /* .context   = */ s_kompute_context,
     };
 
@@ -2017,23 +2016,3 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
 bool ggml_backend_is_kompute(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
 }
-
-static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {
-    GGML_UNUSED(params);
-    return ggml_backend_kompute_init(intptr_t(user_data));
-}
-
-extern "C" int ggml_backend_kompute_reg_devices();
-
-int ggml_backend_kompute_reg_devices() {
-    auto devices = ggml_vk_available_devices_internal(0);
-    for (const auto & device : devices) {
-        ggml_backend_register(
-            ggml_kompute_format_name(device.index).c_str(),
-            ggml_backend_reg_kompute_init,
-            ggml_backend_kompute_buffer_type(device.index),
-            reinterpret_cast<void *>(intptr_t(device.index))
-        );
-    }
-    return devices.size();
-}

ggml/src/ggml-metal.m

@@ -3201,13 +3201,13 @@ static void ggml_backend_metal_free_device(void) {
     }
 }
 
-GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
     return "Metal";
 
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
     for (int i = 0; i < ctx->n_buffers; i++) {
@@ -3226,25 +3226,25 @@ GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_
     free(ctx);
 }
 
-GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
     return ctx->all_data;
 }
 
-GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     memcpy((char *)tensor->data + offset, data, size);
 
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     memcpy(data, (const char *)tensor->data + offset, size);
 
     UNUSED(buffer);
 }
 
-GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
     if (ggml_backend_buffer_is_host(src->buffer)) {
         memcpy(dst->data, src->data, ggml_nbytes(src));
         return true;
@@ -3254,7 +3254,7 @@ GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
     memset(ctx->all_data, value, ctx->all_size);
@@ -3275,7 +3275,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
 
 // default buffer type
 
-GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return "Metal";
 
     UNUSED(buft);
@@ -3306,7 +3306,7 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t s
     UNUSED(size_aligned);
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
 
     const size_t size_page = sysconf(_SC_PAGESIZE);
@@ -3348,12 +3348,12 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
     return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
 
-GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 32;
     UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
     id<MTLDevice> device = ggml_backend_metal_get_device();
     size_t max_size = device.maxBufferLength;
     ggml_backend_metal_free_device();
@@ -3363,13 +3363,13 @@ GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend
     UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return true;
 
     UNUSED(buft);
 }
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
+ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
         /* .iface = */ {
             /* .get_name         = */ ggml_backend_metal_buffer_type_get_name,
@@ -3379,6 +3379,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
         },
+        /* .device  = */ NULL,
         /* .context = */ NULL,
     };
 
@@ -3387,7 +3388,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
 
 // buffer from ptr
 
-GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
     struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
 
     ctx->all_data = data;
@@ -3467,37 +3468,37 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
 
 // backend
 
-GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
+static const char * ggml_backend_metal_name(ggml_backend_t backend) {
     return "Metal";
 
     UNUSED(backend);
 }
 
-GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
+static void ggml_backend_metal_free(ggml_backend_t backend) {
     struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
     ggml_metal_free(ctx);
     free(backend);
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
     return ggml_backend_metal_buffer_type();
 
     UNUSED(backend);
 }
 
-GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
-GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
 
     return ggml_metal_supports_op(metal_ctx, op);
 }
 
-GGML_CALL static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name;
 
     UNUSED(backend);
@@ -3538,11 +3539,8 @@ static struct ggml_backend_i ggml_backend_metal_i = {
     /* .supports_op             = */ ggml_backend_metal_supports_op,
     /* .supports_buft           = */ ggml_backend_metal_supports_buft,
     /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
 void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -3567,6 +3565,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
     *backend = (struct ggml_backend) {
         /* .guid      = */ ggml_backend_metal_guid(),
         /* .interface = */ ggml_backend_metal_i,
+        /* .device    = */ NULL,
         /* .context   = */ ctx,
     };
 
@@ -3603,9 +3602,9 @@ void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
     ctx->capture_next_compute = true;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
 
-GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
     return ggml_backend_metal_init();
 
     GGML_UNUSED(params);

ggml/src/ggml-rpc.cpp

@@ -319,12 +319,12 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     return sock;
 }
 
-GGML_CALL static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     return ctx->name.c_str();
 }
 
-GGML_CALL static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     // input serialization format: | remote_ptr (8 bytes) |
     std::vector<uint8_t> input(sizeof(uint64_t), 0);
@@ -337,7 +337,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t
     delete ctx;
 }
 
-GGML_CALL static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) {
         return ctx->base_cache[buffer];
@@ -388,7 +388,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     return result;
 }
 
-GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     UNUSED(buffer);
     if (ggml_is_quantized(tensor->type)) {
         // TODO: this check is due to MATRIX_ROW_PADDING in CUDA and should be generalized
@@ -396,7 +396,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t
     }
 }
 
-GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
     size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
@@ -410,7 +410,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t b
     GGML_ASSERT(status);
 }
 
-GGML_CALL static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     // input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
     int input_size = sizeof(rpc_tensor) + 2*sizeof(uint64_t);
@@ -427,7 +427,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t b
     memcpy(data, output.data(), size);
 }
 
-GGML_CALL static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     // check if src and dst are on the same server
     ggml_backend_buffer_t src_buffer = src->buffer;
     ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context;
@@ -452,7 +452,7 @@ GGML_CALL static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t b
     return output[0];
 }
 
-GGML_CALL static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     // serialization format: | bufptr (8 bytes) | value (1 byte) |
     int input_size = sizeof(uint64_t) + sizeof(uint8_t);
@@ -477,12 +477,12 @@ static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
     /* .reset           = */ NULL,
 };
 
-GGML_CALL static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) {
     ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
     return buft_ctx->name.c_str();
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
     // input serialization format: | size (8 bytes) |
     int input_size = sizeof(uint64_t);
@@ -522,7 +522,7 @@ static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
     return alignment;
 }
 
-GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
     return buft_ctx->alignment;
 }
@@ -540,12 +540,12 @@ static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
     return max_size;
 }
 
-GGML_CALL static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
     ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
     return buft_ctx->max_size;
 }
 
-GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     UNUSED(buft);
     return ggml_nbytes(tensor);
 }
@@ -559,24 +559,24 @@ static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
     /* .is_host          = */ NULL,
 };
 
-GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
+static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
     ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
 
     return rpc_ctx->name.c_str();
 }
 
-GGML_CALL static void ggml_backend_rpc_free(ggml_backend_t backend) {
+static void ggml_backend_rpc_free(ggml_backend_t backend) {
     ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
     delete rpc_ctx;
     delete backend;
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
     ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
     return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
 }
 
-GGML_CALL static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
+static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
     UNUSED(backend);
     // this is no-op because we don't have any async operations
 }
@@ -618,7 +618,7 @@ static void serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & o
     memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor));
 }
 
-GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
     std::vector<uint8_t> input;
     serialize_graph(cgraph, input);
@@ -630,14 +630,14 @@ GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t
     return (enum ggml_status)output[0];
 }
 
-GGML_CALL static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
     UNUSED(backend);
     UNUSED(op);
     //TODO: call the remote backend and cache the results
     return true;
 }
 
-GGML_CALL static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) {
         return false;
     }
@@ -662,14 +662,11 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .supports_op             = */ ggml_backend_rpc_supports_op,
     /* .supports_buft           = */ ggml_backend_rpc_supports_buft,
     /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
-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) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
     // NOTE: buffer types are allocated and never freed; this is by design
@@ -694,13 +691,14 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const
 
     ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type {
         /* .iface   = */ ggml_backend_rpc_buffer_type_interface,
+        /* .device  = */ nullptr,
         /* .context = */ buft_ctx
     };
     buft_map[endpoint] = buft;
     return buft;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
+ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
         /* .endpoint  = */ endpoint,
         /* .name      = */ "RPC[" + std::string(endpoint) + "]",
@@ -709,12 +707,13 @@ GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_rpc_guid(),
         /* .interface = */ ggml_backend_rpc_interface,
+        /* .device    = */ nullptr,
         /* .context   = */ ctx
     };
     return backend;
 }
 
-GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend) {
+GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
 }
 
@@ -734,7 +733,7 @@ static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * f
     *total = total_mem;
 }
 
-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) {
     auto sock = get_socket(endpoint);
     if (sock == nullptr) {
         *free = 0;

ggml/src/ggml-sycl.cpp

@@ -4038,7 +4038,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
     return true;
 }
 
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len) try {
+GGML_API void   ggml_sycl_get_gpu_list(int *id_list, int max_len) try {
     GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_gpu_list\n");
     for(int i=0;i<max_len;i++) id_list[i] = -1;
 
@@ -4068,7 +4068,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description,
+GGML_API void ggml_sycl_get_device_description(int device, char *description,
                                       size_t description_size) try {
     GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_device_description\n");
     dpct::device_info prop;
@@ -4082,7 +4082,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free,
+void ggml_backend_sycl_get_device_memory(int device, size_t *free,
                                                    size_t *total) try {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_memory\n");
     ggml_sycl_set_device(device);
@@ -4135,12 +4135,12 @@ struct ggml_backend_sycl_buffer_context {
     }
 };
 
-GGML_CALL static const char * ggml_backend_sycl_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_sycl_buffer_get_name(ggml_backend_buffer_t buffer) {
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
     return ctx->name.c_str();
 }
 
-GGML_CALL static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) {
+static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) {
     return buffer->iface.get_name == ggml_backend_sycl_buffer_get_name;
 }
 
@@ -4162,7 +4162,7 @@ static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) {
     return ctx->dev_ptr;
 }
 
-GGML_CALL static void
+static void
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
@@ -4237,7 +4237,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static bool
+static bool
 ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                     const ggml_tensor *src,
                                     ggml_tensor *dst) try {
@@ -4339,12 +4339,12 @@ struct ggml_backend_sycl_buffer_type_context {
     queue_ptr stream = nullptr;
 };
 
-GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) {
     ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
 
     return ctx->name.c_str();
 }
-GGML_CALL static ggml_backend_buffer_t
+static ggml_backend_buffer_t
 ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                            size_t size) try {
     ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
@@ -4368,7 +4368,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 128;
     UNUSED(buft);
 }
@@ -4379,7 +4379,7 @@ static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_typ
     UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
@@ -4424,6 +4424,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
             queue_ptr stream = &(device_i.default_queue());
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
+                /* .device   = */ nullptr,
                 /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), stream},
             };
         }
@@ -4449,6 +4450,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_conte
         for (int i = 0; i < ggml_sycl_info().device_count; i++) {
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
+                /* .device   = */ nullptr,
                 /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), ctx->stream(i, 0)},
             };
         }
@@ -4513,7 +4515,7 @@ struct ggml_backend_sycl_split_buffer_context {
     std::vector<queue_ptr> streams;
 };
 
-GGML_CALL static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) {
     return GGML_SYCL_NAME "_Split";
 
     UNUSED(buffer);
@@ -4523,19 +4525,19 @@ static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
    return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name;
 }
 
-GGML_CALL static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
     delete ctx;
 }
 
-GGML_CALL static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
     // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
     return (void *)0x1000;
 
     UNUSED(buffer);
 }
 
-GGML_CALL static void
+static void
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
@@ -4618,7 +4620,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static void
+static void
 ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           ggml_tensor *tensor, const void *data,
                                           size_t offset, size_t size) try {
@@ -4671,7 +4673,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static void
+static void
 ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                           const ggml_tensor *tensor, void *data,
                                           size_t offset, size_t size) try {
@@ -4724,7 +4726,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     UNUSED(buffer);
     UNUSED(value);
 }
@@ -4742,13 +4744,13 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = {
     /* .reset           = */ NULL,
 };
 
-GGML_CALL static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
     return GGML_SYCL_NAME "_Split";
 
     UNUSED(buft);
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
     // instead, we allocate them for each tensor separately in init_tensor
     // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
@@ -4758,12 +4760,12 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc
     return ggml_backend_buffer_init(buft, ggml_backend_sycl_split_buffer_interface, ctx, size);
 }
 
-GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 128;
     UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     ggml_backend_sycl_split_buffer_type_context * ctx = (ggml_backend_sycl_split_buffer_type_context *)buft->context;
 
     size_t total_size = 0;
@@ -4790,7 +4792,7 @@ GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_
     return total_size;
 }
 
-GGML_CALL static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 
     UNUSED(buft);
@@ -4805,7 +4807,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface
     /* .is_host          = */ ggml_backend_sycl_split_buffer_type_is_host,
 };
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
+ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
 
@@ -4837,6 +4839,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const f
 
     struct ggml_backend_buffer_type buft {
         /* .iface   = */ ggml_backend_sycl_split_buffer_type_interface,
+        /* .device  = */ nullptr,
         /* .context = */ new ggml_backend_sycl_split_buffer_type_context{tensor_split_arr},
     };
 
@@ -4846,13 +4849,13 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const f
 
 // host buffer type
 
-GGML_CALL static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
     return GGML_SYCL_NAME "_Host";
 
     UNUSED(buft);
 }
 
-GGML_CALL static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) {
     return GGML_SYCL_NAME "_Host";
 
     UNUSED(buffer);
@@ -4890,6 +4893,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
+        /* .device   = */ nullptr,
         /* .context  = */ nullptr,
     };
 
@@ -4898,14 +4902,14 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
 
 // backend
 
-GGML_CALL static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
+static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
 
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     return sycl_ctx->name.c_str();
 }
 
-GGML_CALL static void ggml_backend_sycl_free(ggml_backend_t backend) {
+static void ggml_backend_sycl_free(ggml_backend_t backend) {
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     delete sycl_ctx;
@@ -4913,12 +4917,12 @@ GGML_CALL static void ggml_backend_sycl_free(ggml_backend_t backend) {
 }
 
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     return ggml_backend_sycl_buffer_type(sycl_ctx->device);
 }
 
-GGML_CALL static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
+static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                ggml_tensor *tensor,
                                                const void *data, size_t offset,
                                                size_t size) try {
@@ -4936,7 +4940,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
+static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                const ggml_tensor *tensor,
                                                void *data, size_t offset,
                                                size_t size) try {
@@ -4954,9 +4958,9 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
-                                                         const ggml_tensor *src,
-                                                         ggml_tensor *dst) try {
+static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
+                                               const ggml_tensor *src,
+                                               ggml_tensor *dst) try {
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     if (dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && ggml_backend_buffer_is_sycl(src->buffer)) {
         /*
@@ -4991,7 +4995,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_sycl_set_main_device(sycl_ctx->device);
 
@@ -5019,7 +5023,7 @@ GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t back
     return GGML_STATUS_SUCCESS;
 }
 
-GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
@@ -5166,13 +5170,13 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
     return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS && op->op != GGML_OP_MUL_MAT_ID;
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     if (buft->iface.get_name != ggml_backend_sycl_buffer_type_name) {
         return false;
     }
@@ -5197,11 +5201,8 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .supports_op             = */ ggml_backend_sycl_supports_op,
     /* .supports_buft           = */ ggml_backend_sycl_supports_buft,
     /* .offload_op              = */ ggml_backend_sycl_offload_op,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -5209,7 +5210,7 @@ static ggml_guid_t ggml_backend_sycl_guid() {
     return &guid;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) {
+ggml_backend_t ggml_backend_sycl_init(int device) {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_init\n");
     ggml_check_sycl();
 
@@ -5224,6 +5225,7 @@ GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) {
     ggml_backend_t sycl_backend = new ggml_backend {
         /* .guid      = */ ggml_backend_sycl_guid(),
         /* .interface = */ ggml_backend_sycl_interface,
+        /* .device    = */ nullptr,
         /* .context   = */ ctx
     };
 
@@ -5234,26 +5236,7 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_sycl_guid());
 }
 
-GGML_CALL int ggml_backend_sycl_get_device_count() {
+int ggml_backend_sycl_get_device_count() {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n");
     return ggml_sycl_info().device_count;
 }
-
-GGML_CALL static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) {
-    ggml_backend_t sycl_backend = ggml_backend_sycl_init((int) (intptr_t) user_data);
-    return sycl_backend;
-
-    UNUSED(params);
-}
-
-extern "C" int ggml_backend_sycl_reg_devices();
-
-int ggml_backend_sycl_reg_devices() {
-    assert(ggml_sycl_info().device_count>0);
-    for (int i = 0; i < ggml_sycl_info().device_count; i++) {
-        char name[128];
-        snprintf(name, sizeof(name), "%s%d", GGML_SYCL_NAME, i);
-        ggml_backend_register(name, ggml_backend_reg_sycl_init, ggml_backend_sycl_buffer_type(i), (void *) (intptr_t) i);
-    }
-    return ggml_sycl_info().device_count;
-}

ggml/src/ggml-vulkan.cpp

@@ -119,11 +119,11 @@ struct ggml_backend_vk_buffer_type_context {
     vk_device device;
 };
 
-GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
-GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft);
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor);
+static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
+static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
+static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft);
+static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor);
 static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_vk_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_vk_buffer_type_alloc_buffer,
@@ -622,7 +622,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor);
 
 typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
-GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend);
+static void ggml_backend_vk_free(ggml_backend_t backend);
 
 // variables to track number of compiles in progress
 static uint32_t compile_count = 0;
@@ -1953,6 +1953,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->buffer_type = {
             /* .iface    = */ ggml_backend_vk_buffer_type_interface,
+            /* .device   = */ nullptr,
             /* .context  = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
         };
 
@@ -6147,13 +6148,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     ctx->device->device.destroyFence(ctx->fence);
 }
 
-GGML_CALL static int ggml_vk_get_device_count() {
+static int ggml_vk_get_device_count() {
     ggml_vk_instance_init();
 
     return vk_instance.device_indices.size();
 }
 
-GGML_CALL static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
+static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
     ggml_vk_instance_init();
 
     std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
@@ -6170,36 +6171,36 @@ GGML_CALL static void ggml_vk_get_device_description(int device, char * descript
 
 // device backend
 
-GGML_CALL static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     return ctx->name.c_str();
 }
 
-GGML_CALL static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
+static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
     return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
 }
 
-GGML_CALL static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     VK_LOG_MEMORY("ggml_backend_vk_buffer_free_buffer()");
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     ggml_vk_destroy_buffer(ctx->dev_buffer);
     delete ctx;
 }
 
-GGML_CALL static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
+static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
     return vk_ptr_base;
 
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")");
     if (tensor->view_src != nullptr) {
         GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
     }
 }
 
-GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     VK_LOG_DEBUG("ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     vk_buffer buf = buf_ctx->dev_buffer;
@@ -6207,7 +6208,7 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
     ggml_vk_buffer_write(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
-GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     VK_LOG_DEBUG("ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
@@ -6216,7 +6217,7 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
     ggml_vk_buffer_read(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
-GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     if (ggml_backend_buffer_is_vk(src->buffer)) {
         ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
         ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
@@ -6233,7 +6234,7 @@ GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t bu
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
     ggml_vk_buffer_memset(ctx->dev_buffer, 0, value, buffer->size);
@@ -6253,13 +6254,13 @@ static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
 };
 
 // vk buffer type
-GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) {
     ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
 
     return ctx->name.c_str();
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     VK_LOG_MEMORY("ggml_backend_vk_buffer_type_alloc_buffer(" << size << ")");
     ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
 
@@ -6275,23 +6276,23 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(
     return ggml_backend_buffer_init(buft, ggml_backend_vk_buffer_interface, bufctx, size);
 }
 
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
     return ctx->device->properties.limits.minStorageBufferOffsetAlignment;
 }
 
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
     ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
     return ctx->device->max_memory_allocation_size;
 }
 
-GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(buft);
 }
 
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) {
+ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) {
     ggml_vk_instance_init();
 
     VK_LOG_DEBUG("ggml_backend_vk_buffer_type(" << dev_num << ")");
@@ -6303,24 +6304,24 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num)
 
 // host buffer type
 
-GGML_CALL static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
     return GGML_VK_NAME "_Host";
 
     UNUSED(buft);
 }
 
-GGML_CALL static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) {
+static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) {
     return GGML_VK_NAME "_Host";
 
     UNUSED(buffer);
 }
 
-GGML_CALL static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
     ggml_vk_host_free(vk_instance.devices[0], buffer->context);
 }
 
-GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     VK_LOG_MEMORY("ggml_backend_vk_host_buffer_type_alloc_buffer(" << size << ")");
 
     size += 32;  // Behave like the CPU buffer type
@@ -6344,7 +6345,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_bu
     UNUSED(buft);
 }
 
-GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return vk_instance.devices[0]->properties.limits.minMemoryMapAlignment;
 
     UNUSED(buft);
@@ -6352,7 +6353,7 @@ GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_back
 
 // Should be changed to return device-specific host buffer type
 // but that probably requires changes in llama.cpp
-GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
+ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_vk_buffer_type_host = {
         /* .iface    = */ {
             /* .get_name         = */ ggml_backend_vk_host_buffer_type_name,
@@ -6362,6 +6363,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
+        /* .device   = */ nullptr,
         /* .context  = */ nullptr,
     };
 
@@ -6375,13 +6377,13 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
 
 // backend
 
-GGML_CALL static const char * ggml_backend_vk_name(ggml_backend_t backend) {
+static const char * ggml_backend_vk_name(ggml_backend_t backend) {
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
     return ctx->name.c_str();
 }
 
-GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) {
+static void ggml_backend_vk_free(ggml_backend_t backend) {
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     VK_LOG_DEBUG("ggml_backend_vk_free(" << ctx->name << ")");
 
@@ -6391,13 +6393,13 @@ GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) {
     delete backend;
 }
 
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) {
+static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) {
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
     return &ctx->device->buffer_type;
 }
 
-GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     VK_LOG_DEBUG("ggml_backend_vk_set_tensor_async(" << size << ")");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
@@ -6420,7 +6422,7 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
     ggml_vk_buffer_write_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
-GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     VK_LOG_DEBUG("ggml_backend_vk_get_tensor_async(" << size << ")");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
@@ -6443,7 +6445,7 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
     ggml_vk_buffer_read_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
-GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
+static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async()");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     if ((dst->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || dst->buffer->buft == ggml_backend_vk_host_buffer_type()) && ggml_backend_buffer_is_vk(src->buffer)) {
@@ -6471,7 +6473,7 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c
     return false;
 }
 
-GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
+static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
     VK_LOG_DEBUG("ggml_backend_vk_synchronize()");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     if(ctx->transfer_ctx.expired()) {
@@ -6501,7 +6503,7 @@ static bool ggml_vk_is_empty(ggml_tensor * node) {
     return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
 }
 
-GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
@@ -6564,7 +6566,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
     // ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
 
     switch (op->op) {
@@ -6687,7 +6689,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
     return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
@@ -6696,7 +6698,7 @@ GGML_CALL static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const g
     UNUSED(backend);
 }
 
-GGML_CALL static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
     if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
         return false;
     }
@@ -6724,11 +6726,8 @@ static ggml_backend_i ggml_backend_vk_interface = {
     /* .supports_op             = */ ggml_backend_vk_supports_op,
     /* .supports_buft           = */ ggml_backend_vk_supports_buft,
     /* .offload_op              = */ ggml_backend_vk_offload_op,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_vk_guid() {
@@ -6736,7 +6735,7 @@ static ggml_guid_t ggml_backend_vk_guid() {
     return &guid;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
+ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
     VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
 
     ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
@@ -6745,25 +6744,26 @@ GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
     ggml_backend_t vk_backend = new ggml_backend {
         /* .guid      = */ ggml_backend_vk_guid(),
         /* .interface = */ ggml_backend_vk_interface,
+        /* .device    = */ nullptr,
         /* .context   = */ ctx,
     };
 
     return vk_backend;
 }
 
-GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend) {
+bool ggml_backend_is_vk(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
 }
 
-GGML_CALL int ggml_backend_vk_get_device_count() {
+int ggml_backend_vk_get_device_count() {
     return ggml_vk_get_device_count();
 }
 
-GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
+void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
     ggml_vk_get_device_description(device, description, description_size);
 }
 
-GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
+void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
     GGML_ASSERT(device < (int) vk_instance.device_indices.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
@@ -6779,27 +6779,6 @@ GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size
     }
 }
 
-// backend registry
-GGML_CALL static ggml_backend_t ggml_backend_reg_vk_init(const char * params, void * user_data) {
-    ggml_backend_t vk_backend = ggml_backend_vk_init((int) (intptr_t) user_data);
-    return vk_backend;
-
-    UNUSED(params);
-}
-
-extern "C" GGML_CALL int ggml_backend_vk_reg_devices();
-
-GGML_CALL int ggml_backend_vk_reg_devices() {
-    ggml_vk_instance_init();
-
-    for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
-        char name[128];
-        snprintf(name, sizeof(name), "%s%ld", GGML_VK_NAME, i);
-        ggml_backend_register(name, ggml_backend_reg_vk_init, ggml_backend_vk_buffer_type(i), (void *) (intptr_t) i);  // NOLINT
-    }
-    return vk_instance.device_indices.size();
-}
-
 // Extension availability
 static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
 #ifdef GGML_VULKAN_VALIDATE

ggml/src/ggml.c

@@ -461,7 +461,7 @@ struct ggml_arm_arch_features_type {
 } ggml_arm_arch_features = {-1, -1, -1, 0};
 #endif
 
-GGML_CALL const char * ggml_status_to_string(enum ggml_status status) {
+const char * ggml_status_to_string(enum ggml_status status) {
     switch (status) {
         case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
         case GGML_STATUS_FAILED:       return "GGML status: error (operation failed)";
@@ -3384,19 +3384,19 @@ void ggml_print_objects(const struct ggml_context * ctx) {
     GGML_PRINT("%s: --- end ---\n", __func__);
 }
 
-GGML_CALL int64_t ggml_nelements(const struct ggml_tensor * tensor) {
+int64_t ggml_nelements(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[0]*tensor->ne[1]*tensor->ne[2]*tensor->ne[3];
 }
 
-GGML_CALL int64_t ggml_nrows(const struct ggml_tensor * tensor) {
+int64_t ggml_nrows(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[1]*tensor->ne[2]*tensor->ne[3];
 }
 
-GGML_CALL size_t ggml_nbytes(const struct ggml_tensor * tensor) {
+size_t ggml_nbytes(const struct ggml_tensor * tensor) {
     size_t nbytes;
     size_t blck_size = ggml_blck_size(tensor->type);
     if (blck_size == 1) {
@@ -3419,15 +3419,15 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
     return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
 }
 
-GGML_CALL int64_t ggml_blck_size(enum ggml_type type) {
+int64_t ggml_blck_size(enum ggml_type type) {
     return type_traits[type].blck_size;
 }
 
-GGML_CALL size_t ggml_type_size(enum ggml_type type) {
+size_t ggml_type_size(enum ggml_type type) {
     return type_traits[type].type_size;
 }
 
-GGML_CALL size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+size_t ggml_row_size(enum ggml_type type, int64_t ne) {
     assert(ne % ggml_blck_size(type) == 0);
     return ggml_type_size(type)*ne/ggml_blck_size(type);
 }
@@ -3436,15 +3436,15 @@ double ggml_type_sizef(enum ggml_type type) {
     return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
 }
 
-GGML_CALL const char * ggml_type_name(enum ggml_type type) {
+const char * ggml_type_name(enum ggml_type type) {
     return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
 }
 
-GGML_CALL bool ggml_is_quantized(enum ggml_type type) {
+bool ggml_is_quantized(enum ggml_type type) {
     return type_traits[type].is_quantized;
 }
 
-GGML_CALL const char * ggml_op_name(enum ggml_op op) {
+const char * ggml_op_name(enum ggml_op op) {
     return GGML_OP_NAME[op];
 }
 
@@ -3456,7 +3456,7 @@ const char * ggml_unary_op_name(enum ggml_unary_op op) {
     return GGML_UNARY_OP_NAME[op];
 }
 
-GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t) {
+const char * ggml_op_desc(const struct ggml_tensor * t) {
     if (t->op == GGML_OP_UNARY) {
         enum ggml_unary_op uop = ggml_get_unary_op(t);
         return ggml_unary_op_name(uop);
@@ -3464,7 +3464,7 @@ GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t) {
     return ggml_op_name(t->op);
 }
 
-GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor) {
+size_t ggml_element_size(const struct ggml_tensor * tensor) {
     return ggml_type_size(tensor->type);
 }
 
@@ -3557,7 +3557,7 @@ size_t ggml_tensor_overhead(void) {
     return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE;
 }
 
-GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor) {
+bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
 
@@ -3583,23 +3583,23 @@ static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) {
     return true;
 }
 
-GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
+bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
     return ggml_is_contiguous_0(tensor);
 }
 
-GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
+bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
     return ggml_is_contiguous_n(tensor, 0);
 }
 
-GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
+bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
     return ggml_is_contiguous_n(tensor, 1);
 }
 
-GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
+bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
     return ggml_is_contiguous_n(tensor, 2);
 }
 
-GGML_CALL bool ggml_is_permuted(const struct ggml_tensor * tensor) {
+bool ggml_is_permuted(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->nb[0] > tensor->nb[1] || tensor->nb[1] > tensor->nb[2] || tensor->nb[2] > tensor->nb[3];
@@ -3614,7 +3614,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
-GGML_CALL bool ggml_is_empty(const struct ggml_tensor * tensor) {
+bool ggml_is_empty(const struct ggml_tensor * tensor) {
     for (int i = 0; i < GGML_MAX_DIMS; ++i) {
         if (tensor->ne[i] == 0) {
             // empty if any dimension has no elements
@@ -4634,7 +4634,7 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) {
     return (float *)(tensor->data);
 }
 
-GGML_CALL enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
+enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->op == GGML_OP_UNARY);
     return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0);
 }
@@ -12834,6 +12834,10 @@ static void ggml_compute_forward_out_prod_f32(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -14163,7 +14167,7 @@ static void ggml_rope_cache_init(
     }
 }
 
-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]
 ) {
     // start and end correction dims