Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

metal : support virtual devices (#18919)

Browse Source 
* metal : support virtual devices

* cont : manage buffer type context memory

* metal : add events

* cont : implement cpy_tensor_async
This commit is contained in:
Georgi Gerganov
2026-02-02 14:29:44 +02:00
committed by GitHub
parent 6156ae5111
commit 6fdddb4987
7 changed files with 508 additions and 127 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-device.h
+13 -3
View File
@@ -205,7 +205,9 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
//
struct ggml_metal_device_props {
int device;
char name[128];
char desc[128];
size_t max_buffer_size;
size_t max_working_set_size;
@@ -224,11 +226,15 @@ struct ggml_metal_device_props {
int op_offload_min_batch_size;
};
ggml_metal_device_t ggml_metal_device_init(void);
typedef struct ggml_metal_event * ggml_metal_event_t;
void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
void ggml_metal_event_encode_wait (ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
ggml_metal_device_t ggml_metal_device_init(int device);
void ggml_metal_device_free(ggml_metal_device_t dev);
// return a singleton that is automatically destroyed when the program exits
ggml_metal_device_t ggml_metal_device_get(void);
ggml_metal_device_t ggml_metal_device_get(int device);
void * ggml_metal_device_get_obj (ggml_metal_device_t dev); // id<MTLDevice>
void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQueue>
@@ -240,6 +246,10 @@ void ggml_metal_device_rsets_rm (ggml_metal_device_t dev, ggml_metal_rset_t rset
void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev);
ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev);
void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev);
void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev);
void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total);
bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_tensor * op);