llama_cpp 0.7.0 → 0.7.1

data/ext/llama_cpp/src/ggml-cuda.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
@@ -42,6 +43,9 @@ GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, s
 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);
 
+// backend API
+GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+
 #ifdef  __cplusplus
 }
 #endif

data/ext/llama_cpp/src/ggml-metal.h

@@ -20,6 +20,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #include <stddef.h>
 #include <stdbool.h>
@@ -35,10 +36,15 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+//
+// internal API
+// temporary exposed to user-code
+//
 
 struct ggml_metal_context;
 
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 // number of command buffers to use
 struct ggml_metal_context * ggml_metal_init(int n_cb);
 void ggml_metal_free(struct ggml_metal_context * ctx);
@@ -83,6 +89,17 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
 #ifdef __cplusplus
 }
 #endif