typedef void * ggml_backend_buffer_type_context_t;
struct ggml_backend_buffer_type_i {
- const char * (*get_name) (ggml_backend_buffer_type_t buft);
- ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
- size_t (*get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
- size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
- bool (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+ const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft);
+ ggml_backend_buffer_t (*GGML_CALL alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
+ size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
+ size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+ bool (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
// check if tensor data is in host memory
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
- bool (*is_host) (ggml_backend_buffer_type_t buft);
+ bool (*GGML_CALL is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
typedef void * ggml_backend_buffer_context_t;
struct ggml_backend_buffer_i {
- const char * (*get_name) (ggml_backend_buffer_t buffer);
- void (*free_buffer)(ggml_backend_buffer_t buffer);
- void * (*get_base) (ggml_backend_buffer_t buffer);
- void (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
- 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);
- bool (*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 (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
- void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+ 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 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
};
struct ggml_backend_buffer {
enum ggml_backend_buffer_usage usage;
};
- ggml_backend_buffer_t ggml_backend_buffer_init(
+ 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,
typedef void * ggml_backend_context_t;
struct ggml_backend_i {
- const char * (*get_name)(ggml_backend_t backend);
+ const char * (*GGML_CALL get_name)(ggml_backend_t backend);
- void (*free)(ggml_backend_t backend);
+ void (*GGML_CALL free)(ggml_backend_t backend);
// buffer allocation
- ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
+ ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
// (optional) asynchronous tensor data access
- 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, const struct ggml_tensor * src, struct ggml_tensor * dst);
+ 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, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) complete all pending operations
- void (*synchronize)(ggml_backend_t backend);
+ void (*GGML_CALL synchronize)(ggml_backend_t backend);
// compute graph with a plan
- 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);
- void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+ 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);
+ void (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph without a plan (async)
- bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+ bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
// check if the backend supports an operation
- bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+ bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
};
struct ggml_backend {
// Backend registry
//
- typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+ typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
- void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+ 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);
#ifdef __cplusplus
}
return buft->iface.get_name(buft);
}
-ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+GGML_CALL 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);
}
return buft->iface.get_alignment(buft);
}
-size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+GGML_CALL 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) {
return buft->iface.get_alloc_size(buft, tensor);
// backend buffer
-ggml_backend_buffer_t ggml_backend_buffer_init(
+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,
return base;
}
-void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+GGML_CALL 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);
}
}
-void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+GGML_CALL 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(tensor->data != NULL && "tensor not allocated");
tensor->buffer->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_CALL 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(tensor->data != NULL && "tensor not allocated");
static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
static size_t ggml_backend_registry_count = 0;
-static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
+GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
-static void ggml_backend_registry_init(void) {
+GGML_CALL static void ggml_backend_registry_init(void) {
static bool initialized = false;
if (initialized) {
// add forward decls here to avoid including the backend headers
#ifdef GGML_USE_CUBLAS
- extern void ggml_backend_cuda_reg_devices(void);
+ extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
ggml_backend_cuda_reg_devices();
#endif
#ifdef GGML_USE_METAL
- extern ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
- extern ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+ extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
+ extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
#endif
}
-void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
+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) {
GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
size_t id = ggml_backend_registry_count;
// backend CPU
-static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
+GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
return "CPU";
GGML_UNUSED(buffer);
}
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
return (void *)buffer->context;
}
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
-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) {
+GGML_CALL 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) {
+GGML_CALL 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) {
+GGML_CALL 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;
GGML_UNUSED(buffer);
}
-static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL 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) {
+GGML_CALL 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: maybe use GGML_ALIGNED_MALLOC?
return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
}
-static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+GGML_CALL 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_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+GGML_CALL static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_cpu(backend);
GGML_UNUSED(buft);
}
-static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+GGML_CALL 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) {
+GGML_CALL 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,
#include <hbwmalloc.h>
-static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL 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) {
+GGML_CALL 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) {
+GGML_CALL 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) {
+GGML_CALL 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);
size_t work_size;
};
-static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
+GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
free(cpu_ctx->work_data);
free(cpu_ctx);
free(backend);
}
-static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+GGML_CALL 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_cgraph cgraph;
};
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+GGML_CALL 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 = malloc(sizeof(struct ggml_backend_plan_cpu));
return cpu_plan;
}
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+GGML_CALL 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;
free(cpu_plan->cplan.work_data);
GGML_UNUSED(backend);
}
-static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+GGML_CALL static void 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;
ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
-static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static bool 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);
return true;
}
-static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
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;
return cpu_backend;
}
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
+GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_cpu_name;
}
ctx->n_threads = n_threads;
}
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
}
-static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
+GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
return ggml_backend_cpu_init();
GGML_UNUSED(params);
//
// buffer type
- 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_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
- GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
- 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_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 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_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+ GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
// buffer
enum ggml_backend_buffer_usage {
GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
};
- 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_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 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 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_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 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);
//
// Backend
GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
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);
- 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 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 void ggml_backend_synchronize(ggml_backend_t 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 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);
// 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_CALL 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_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
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_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+ typedef bool (*GGML_CALL 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);
static bool g_cublas_loaded = false;
-bool ggml_cublas_loaded(void) {
+GGML_CALL bool ggml_cublas_loaded(void) {
return g_cublas_loaded;
}
-void ggml_init_cublas() {
+GGML_CALL void ggml_init_cublas() {
static bool initialized = false;
if (!initialized) {
}
}
-void * ggml_cuda_host_malloc(size_t size) {
+GGML_CALL void * ggml_cuda_host_malloc(size_t size) {
if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
return nullptr;
}
return ptr;
}
-void ggml_cuda_host_free(void * ptr) {
+GGML_CALL void ggml_cuda_host_free(void * ptr) {
CUDA_CHECK(cudaFreeHost(ptr));
}
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
}
-bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+GGML_CALL bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (!g_cublas_loaded) return false;
const int64_t ne10 = src1->ne[0];
return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
}
-static void ggml_cuda_set_main_device(const int main_device) {
+GGML_CALL static void ggml_cuda_set_main_device(const int main_device) {
if (main_device >= g_device_count) {
fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n",
main_device, g_device_count, g_main_device);
}
}
-bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
+GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
if (!g_cublas_loaded) return false;
ggml_cuda_func_t func;
return true;
}
-int ggml_cuda_get_device_count() {
+GGML_CALL int ggml_cuda_get_device_count() {
int device_count;
if (cudaGetDeviceCount(&device_count) != cudaSuccess) {
return 0;
return device_count;
}
-void ggml_cuda_get_device_description(int device, char * description, size_t description_size) {
+GGML_CALL void ggml_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);
}
};
-static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
+GGML_CALL 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();
}
-static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
+GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
}
-static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+GGML_CALL 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;
CUDA_CHECK(cudaFree(ctx->dev_ptr));
delete ctx;
}
-static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
+GGML_CALL 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;
}
-static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+GGML_CALL 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 && tensor->view_offs == 0) {
}
}
-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_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) {
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
CUDA_CHECK(cudaDeviceSynchronize());
}
-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_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) {
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
CUDA_CHECK(cudaDeviceSynchronize());
}
-static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+GGML_CALL 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 *)buffer->context;
return false;
}
-static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+GGML_CALL 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);
};
// cuda buffer type
-
struct ggml_backend_cuda_buffer_type_context {
int device;
std::string name;
};
-static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL static const char * ggml_backend_cuda_buffer_type_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 ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+GGML_CALL 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);
return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
}
-static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 128;
UNUSED(buft);
}
-static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
int64_t row_low = 0;
int64_t row_high = ggml_nrows(tensor);
int64_t nrows_split = row_high - row_low;
UNUSED(buft);
}
-static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
if (!ggml_backend_is_cuda(backend)) {
return false;
}
/* .is_host = */ NULL,
};
-ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
+GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
// FIXME: this is not thread safe
if (device >= ggml_backend_cuda_get_device_count()) {
return nullptr;
std::vector<ggml_tensor_extra_gpu *> tensor_extras;
};
-static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
return GGML_CUDA_NAME "_Split";
UNUSED(buffer);
// return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
//}
-static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+GGML_CALL 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;
}
-static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
+GGML_CALL 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;
UNUSED(buffer);
}
-static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+GGML_CALL 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;
tensor->extra = extra;
}
-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) {
+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) {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
}
}
-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) {
+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) {
// split tensors must always be set in their entirety at once
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
}
}
-static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
UNUSED(buffer);
UNUSED(value);
}
// cuda split buffer type
-static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
return GGML_CUDA_NAME "_Split";
UNUSED(buft);
}
-static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+GGML_CALL 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,
return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
}
-static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 128;
UNUSED(buft);
}
-static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+GGML_CALL 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;
return total_size;
}
-static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_cuda(backend);
UNUSED(buft);
}
-static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return false;
UNUSED(buft);
/* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host,
};
-ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
+GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
// FIXME: this is not thread safe
static std::map<std::array<float, GGML_CUDA_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
// host buffer type
-static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
return GGML_CUDA_NAME "_Host";
UNUSED(buft);
}
-static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
+GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
return GGML_CUDA_NAME "_Host";
UNUSED(buffer);
}
-static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_cuda_host_free(buffer->context);
}
-static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+GGML_CALL 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) {
return buffer;
}
-ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
+GGML_CALL 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,
// backend
-static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
+GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
return cuda_ctx->name.c_str();
}
-static void ggml_backend_cuda_free(ggml_backend_t backend) {
+GGML_CALL 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;
}
-static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
+GGML_CALL 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);
}
-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_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) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
}
-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_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) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
}
-static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
+GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
if (dst->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && ggml_backend_buffer_is_cuda(src->buffer)) {
return false;
}
-static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+GGML_CALL 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(g_cudaStreams[cuda_ctx->device][0]));
UNUSED(backend);
}
-static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+GGML_CALL static bool 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_main_device(cuda_ctx->device);
return true;
}
-static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
/* .supports_op = */ ggml_backend_cuda_supports_op,
};
-ggml_backend_t ggml_backend_cuda_init(int device) {
+GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
ggml_init_cublas(); // TODO: remove from ggml.c
if (device < 0 || device >= ggml_cuda_get_device_count()) {
return cuda_backend;
}
-bool ggml_backend_is_cuda(ggml_backend_t backend) {
+GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_cuda_name;
}
-int ggml_backend_cuda_get_device_count() {
+GGML_CALL int ggml_backend_cuda_get_device_count() {
return ggml_cuda_get_device_count();
}
-void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
+GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
ggml_cuda_get_device_description(device, description, description_size);
}
-void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
+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));
}
// backend registry
-static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
+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;
UNUSED(params);
}
-extern "C" int ggml_backend_cuda_reg_devices();
+extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
-int ggml_backend_cuda_reg_devices() {
+GGML_CALL int ggml_backend_cuda_reg_devices() {
int device_count = ggml_cuda_get_device_count();
//int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
for (int i = 0; i < device_count; i++) {
#define GGML_CUDA_MAX_DEVICES 16
// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
-GGML_API void ggml_init_cublas(void);
+GGML_API GGML_CALL void ggml_init_cublas(void);
// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
-GGML_API bool ggml_cublas_loaded(void);
+GGML_API GGML_CALL bool ggml_cublas_loaded(void);
-GGML_API void * ggml_cuda_host_malloc(size_t size);
-GGML_API void ggml_cuda_host_free(void * ptr);
+GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
+GGML_API GGML_CALL void ggml_cuda_host_free(void * ptr);
-GGML_API bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+GGML_API GGML_CALL bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-GGML_API int ggml_cuda_get_device_count(void);
-GGML_API void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL int ggml_cuda_get_device_count(void);
+GGML_API GGML_CALL void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
// backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_API GGML_CALL 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_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_API GGML_CALL 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_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-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 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);
#ifdef __cplusplus
}
GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+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 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_API GGML_CALL 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
}
}
-static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
+GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
UNUSED(buffer);
}
-static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+GGML_CALL 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++) {
free(ctx);
}
-static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+GGML_CALL 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;
}
-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) {
+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) {
memcpy((char *)tensor->data + offset, data, size);
UNUSED(buffer);
}
-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) {
+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) {
memcpy(data, (const char *)tensor->data + offset, size);
UNUSED(buffer);
}
-static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+GGML_CALL 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;
UNUSED(buffer);
}
-static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+GGML_CALL 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);
// default buffer type
-static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "Metal";
UNUSED(buft);
}
-static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+GGML_CALL 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);
return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
}
-static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 32;
UNUSED(buft);
}
-static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
UNUSED(buft);
}
-static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
UNUSED(buft);
}
-ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
+GGML_CALL 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,
// buffer from ptr
-ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+GGML_CALL 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;
// backend
-static const char * ggml_backend_metal_name(ggml_backend_t backend) {
+GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
return "Metal";
UNUSED(backend);
}
-static void ggml_backend_metal_free(ggml_backend_t backend) {
+GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ggml_metal_free(ctx);
free(backend);
}
-static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
+GGML_CALL 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);
}
-static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_graph_compute(metal_ctx, cgraph);
}
-static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_supports_op(metal_ctx, op);
return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
}
-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); // silence warning
-ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
+GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
return ggml_backend_metal_init();
GGML_UNUSED(params);
GGML_PRINT("%s: --- end ---\n", __func__);
}
-int64_t ggml_nelements(const struct ggml_tensor * tensor) {
+GGML_CALL 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];
}
-int64_t ggml_nrows(const struct ggml_tensor * tensor) {
+GGML_CALL 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];
}
-size_t ggml_nbytes(const struct ggml_tensor * tensor) {
+GGML_CALL 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) {
return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
}
-int ggml_blck_size(enum ggml_type type) {
+GGML_CALL int ggml_blck_size(enum ggml_type type) {
return type_traits[type].blck_size;
}
-size_t ggml_type_size(enum ggml_type type) {
+GGML_CALL size_t ggml_type_size(enum ggml_type type) {
return type_traits[type].type_size;
}
-size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+GGML_CALL 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);
}
return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
}
-const char * ggml_type_name(enum ggml_type type) {
+GGML_CALL const char * ggml_type_name(enum ggml_type type) {
return type_traits[type].type_name;
}
-bool ggml_is_quantized(enum ggml_type type) {
+GGML_CALL bool ggml_is_quantized(enum ggml_type type) {
return type_traits[type].is_quantized;
}
-const char * ggml_op_name(enum ggml_op op) {
+GGML_CALL const char * ggml_op_name(enum ggml_op op) {
return GGML_OP_NAME[op];
}
return GGML_UNARY_OP_NAME[op];
}
-const char * ggml_op_desc(const struct ggml_tensor * t) {
+GGML_CALL 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);
}
}
-size_t ggml_element_size(const struct ggml_tensor * tensor) {
+GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor) {
return ggml_type_size(tensor->type);
}
return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE;
}
-bool ggml_is_transposed(const struct ggml_tensor * tensor) {
+GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor) {
return tensor->nb[0] > tensor->nb[1];
}
-bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
+GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
-bool ggml_is_permuted(const struct ggml_tensor * tensor) {
+GGML_CALL 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];
return (float *)(tensor->data);
}
-enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
+GGML_CALL 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);
}
}
}
-void ggml_rope_yarn_corr_dims(
+GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
) {
// start and end correction dims
# 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)))
GGML_API void ggml_print_object (const struct ggml_object * obj);
GGML_API void ggml_print_objects(const struct ggml_context * ctx);
- 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_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 int 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_API GGML_CALL int 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_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 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 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_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 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 size_t ggml_element_size(const struct ggml_tensor * tensor);
+ GGML_API GGML_CALL size_t ggml_element_size(const struct ggml_tensor * tensor);
- GGML_API bool ggml_is_quantized(enum ggml_type type);
+ GGML_API GGML_CALL 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 bool ggml_is_transposed(const struct ggml_tensor * tensor);
- GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
- GGML_API bool ggml_is_permuted (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_transposed(const struct ggml_tensor * tensor);
+ GGML_API GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+ GGML_API GGML_CALL bool ggml_is_permuted (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_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
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 enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+ GGML_API GGML_CALL 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);
float beta_slow);
// compute correction dims for YaRN RoPE scaling
- void ggml_rope_yarn_corr_dims(
+ GGML_CALL void ggml_rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// xPos RoPE, in-place, returns view(a)
overview / pkg / ggml / sources / whisper.cpp / commitdiff