Skip to content

Commit 9fdfcda

Browse files
slarenslaren
authored
rpc : use backend registry, support dl backends (#13304)
1 parent 6eb7d25 commit 9fdfcda

File tree

4 files changed

+111
-93
lines changed

4 files changed

+111
-93
lines changed

‎ggml/src/ggml-cpu/ggml-cpu.cpp‎

Lines changed: 29 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -11,24 +11,26 @@
1111
#include <vector>
1212

1313
#ifdef GGML_USE_CPU_HBM
14-
#include "ggml-cpu-hbm.h"
14+
# include "ggml-cpu-hbm.h"
1515
#endif
1616

1717
#ifdef GGML_USE_CPU_KLEIDIAI
18-
#include "kleidiai/kleidiai.h"
19-
#endif
20-
21-
#if defined(__APPLE__)
22-
#include <sys/types.h>
23-
#include <sys/sysctl.h>
18+
# include "kleidiai/kleidiai.h"
2419
#endif
2520

2621
#if defined(_WIN32)
27-
#define WIN32_LEAN_AND_MEAN
28-
#ifndef NOMINMAX
29-
#define NOMINMAX
22+
# define WIN32_LEAN_AND_MEAN
23+
# ifndef NOMINMAX
24+
# define NOMINMAX
25+
# endif
26+
# include <windows.h>
27+
#else
28+
# include <unistd.h>
3029
#endif
31-
#include <windows.h>
30+
31+
#if defined(__APPLE__)
32+
# include <sys/sysctl.h>
33+
# include <sys/types.h>
3234
#endif
3335

3436
// ggml-backend interface
@@ -70,8 +72,10 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_ty
7072
}
7173

7274
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
73-
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
74-
if (extra && extra == buft) return true;
75+
for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
76+
if (extra && extra == buft) {
77+
return true;
78+
}
7579
}
7680
return false;
7781
}
@@ -330,9 +334,18 @@ static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t d
330334
}
331335

332336
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
333-
// TODO
334-
*free = 0;
335-
*total = 0;
337+
#ifdef _WIN32
338+
MEMORYSTATUSEX status;
339+
status.dwLength = sizeof(status);
340+
GlobalMemoryStatusEx(&status);
341+
*total = status.ullTotalPhys;
342+
*free = status.ullAvailPhys;
343+
#else
344+
long pages = sysconf(_SC_PHYS_PAGES);
345+
long page_size = sysconf(_SC_PAGE_SIZE);
346+
*total = pages * page_size;
347+
*free = *total;
348+
#endif
336349

337350
GGML_UNUSED(dev);
338351
}

‎ggml/src/ggml-rpc/ggml-rpc.cpp‎

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1594,6 +1594,14 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
15941594
void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
15951595
const char * cache_dir,
15961596
size_t free_mem, size_t total_mem) {
1597+
printf("Starting RPC server v%d.%d.%d\n",
1598+
RPC_PROTO_MAJOR_VERSION,
1599+
RPC_PROTO_MINOR_VERSION,
1600+
RPC_PROTO_PATCH_VERSION);
1601+
printf(" endpoint : %s\n", endpoint);
1602+
printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a");
1603+
printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024));
1604+
15971605
std::string host;
15981606
int port;
15991607
if (!parse_endpoint(endpoint, host, port)) {
@@ -1753,6 +1761,9 @@ static void * ggml_backend_rpc_get_proc_address(ggml_backend_reg_t reg, const ch
17531761
if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) {
17541762
return (void *)ggml_backend_rpc_add_device;
17551763
}
1764+
if (std::strcmp(name, "ggml_backend_rpc_start_server") == 0) {
1765+
return (void *)ggml_backend_rpc_start_server;
1766+
}
17561767
return NULL;
17571768

17581769
GGML_UNUSED(reg);

‎tools/CMakeLists.txt‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -28,12 +28,12 @@ else()
2828
add_subdirectory(tokenize)
2929
add_subdirectory(tts)
3030
add_subdirectory(llava)
31+
if (GGML_RPC)
32+
add_subdirectory(rpc)
33+
endif()
3134
if (NOT GGML_BACKEND_DL)
3235
# these examples use the backends directly and cannot be built with dynamic loading
3336
add_subdirectory(cvector-generator)
3437
add_subdirectory(export-lora)
35-
if (GGML_RPC)
36-
add_subdirectory(rpc)
37-
endif()
3838
endif()
3939
endif()

‎tools/rpc/rpc-server.cpp‎

Lines changed: 68 additions & 74 deletions
Original file line numberDiff line numberDiff line change
@@ -2,24 +2,6 @@
22
#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
33
#endif
44

5-
#include "ggml-cpu.h"
6-
7-
#ifdef GGML_USE_CUDA
8-
#include "ggml-cuda.h"
9-
#endif
10-
11-
#ifdef GGML_USE_METAL
12-
#include "ggml-metal.h"
13-
#endif
14-
15-
#ifdef GGML_USE_VULKAN
16-
#include "ggml-vulkan.h"
17-
#endif
18-
19-
#ifdef GGML_USE_SYCL
20-
#include "ggml-sycl.h"
21-
#endif
22-
235
#include "ggml-rpc.h"
246
#ifdef _WIN32
257
# define NOMINMAX
@@ -154,13 +136,15 @@ struct rpc_server_params {
154136
size_t backend_mem = 0;
155137
bool use_cache = false;
156138
int n_threads = std::max(1U, std::thread::hardware_concurrency()/2);
139+
std::string device;
157140
};
158141

159142
static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
160143
fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
161144
fprintf(stderr, "options:\n");
162145
fprintf(stderr, " -h, --help show this help message and exit\n");
163146
fprintf(stderr, " -t, --threads number of threads for the CPU backend (default: %d)\n", params.n_threads);
147+
fprintf(stderr, " -d DEV, --device device to use\n");
164148
fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str());
165149
fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port);
166150
fprintf(stderr, " -m MEM, --mem MEM backend memory size (in MB)\n");
@@ -186,6 +170,22 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
186170
fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
187171
return false;
188172
}
173+
} else if (arg == "-d" || arg == "--device") {
174+
if (++i >= argc) {
175+
return false;
176+
}
177+
params.device = argv[i];
178+
if (ggml_backend_dev_by_name(params.device.c_str()) == nullptr) {
179+
fprintf(stderr, "error: unknown device: %s\n", params.device.c_str());
180+
fprintf(stderr, "available devices:\n");
181+
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
182+
auto * dev = ggml_backend_dev_get(i);
183+
size_t free, total;
184+
ggml_backend_dev_memory(dev, &free, &total);
185+
printf(" %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
186+
}
187+
return false;
188+
}
189189
} else if (arg == "-p" || arg == "--port") {
190190
if (++i >= argc) {
191191
return false;
@@ -214,66 +214,53 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
214214
}
215215

216216
static ggml_backend_t create_backend(const rpc_server_params & params) {
217-
ggml_backend_t backend = NULL;
218-
#ifdef GGML_USE_CUDA
219-
fprintf(stderr, "%s: using CUDA backend\n", __func__);
220-
backend = ggml_backend_cuda_init(0); // init device 0
221-
if (!backend) {
222-
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
223-
}
224-
#elif GGML_USE_METAL
225-
fprintf(stderr, "%s: using Metal backend\n", __func__);
226-
backend = ggml_backend_metal_init();
227-
if (!backend) {
228-
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
217+
ggml_backend_t backend = nullptr;
218+
219+
if (!params.device.empty()) {
220+
ggml_backend_dev_t dev = ggml_backend_dev_by_name(params.device.c_str());
221+
if (dev) {
222+
backend = ggml_backend_dev_init(dev, nullptr);
223+
if (!backend) {
224+
fprintf(stderr, "Failed to create backend for device %s\n", params.device.c_str());
225+
return nullptr;
226+
}
227+
}
229228
}
230-
#elif GGML_USE_VULKAN
231-
fprintf(stderr, "%s: using Vulkan backend\n", __func__);
232-
backend = ggml_backend_vk_init(0); // init device 0
229+
230+
// try to initialize a GPU backend first
233231
if (!backend) {
234-
fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
232+
backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
235233
}
236-
#elif GGML_USE_SYCL
237-
fprintf(stderr, "%s: using SYCL backend\n", __func__);
238-
backend = ggml_backend_sycl_init(0); // init device 0
234+
235+
// if there aren't GPU backends fallback to CPU backend
239236
if (!backend) {
240-
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
237+
backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
241238
}
242-
#endif
243239

244-
// if there aren't GPU Backends fallback to CPU backend
245-
if (!backend) {
246-
fprintf(stderr, "%s: using CPU backend\n", __func__);
247-
backend = ggml_backend_cpu_init();
248-
ggml_backend_cpu_set_n_threads(backend, params.n_threads);
240+
fprintf(stderr, "%s: using %s backend\n", __func__, ggml_backend_name(backend));
241+
242+
// set the number of threads
243+
ggml_backend_dev_t dev = ggml_backend_get_device(backend);
244+
ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
245+
if (reg) {
246+
auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
247+
if (ggml_backend_set_n_threads_fn) {
248+
ggml_backend_set_n_threads_fn(backend, params.n_threads);
249+
}
249250
}
251+
250252
return backend;
251253
}
252254

253-
static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
254-
#ifdef GGML_USE_CUDA
255-
ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
256-
#elif GGML_USE_VULKAN
257-
ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
258-
#elif GGML_USE_SYCL
259-
ggml_backend_sycl_get_device_memory(0, free_mem, total_mem);
260-
#else
261-
#ifdef _WIN32
262-
MEMORYSTATUSEX status;
263-
status.dwLength = sizeof(status);
264-
GlobalMemoryStatusEx(&status);
265-
*total_mem = status.ullTotalPhys;
266-
*free_mem = status.ullAvailPhys;
267-
#else
268-
long pages = sysconf(_SC_PHYS_PAGES);
269-
long page_size = sysconf(_SC_PAGE_SIZE);
270-
*total_mem = pages * page_size;
271-
*free_mem = *total_mem;
272-
#endif
273-
#endif
255+
static void get_backend_memory(ggml_backend_t backend, size_t * free_mem, size_t * total_mem) {
256+
ggml_backend_dev_t dev = ggml_backend_get_device(backend);
257+
GGML_ASSERT(dev != nullptr);
258+
ggml_backend_dev_memory(dev, free_mem, total_mem);
274259
}
275260

276261
int main(int argc, char * argv[]) {
262+
ggml_backend_load_all();
263+
277264
rpc_server_params params;
278265
if (!rpc_server_params_parse(argc, argv, params)) {
279266
fprintf(stderr, "Invalid parameters\n");
@@ -301,7 +288,7 @@ int main(int argc, char * argv[]) {
301288
free_mem = params.backend_mem;
302289
total_mem = params.backend_mem;
303290
} else {
304-
get_backend_memory(&free_mem, &total_mem);
291+
get_backend_memory(backend, &free_mem, &total_mem);
305292
}
306293
const char * cache_dir = nullptr;
307294
std::string cache_dir_str;
@@ -313,14 +300,21 @@ int main(int argc, char * argv[]) {
313300
}
314301
cache_dir = cache_dir_str.c_str();
315302
}
316-
printf("Starting RPC server v%d.%d.%d\n",
317-
RPC_PROTO_MAJOR_VERSION,
318-
RPC_PROTO_MINOR_VERSION,
319-
RPC_PROTO_PATCH_VERSION);
320-
printf(" endpoint : %s\n", endpoint.c_str());
321-
printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a");
322-
printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024));
323-
ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
303+
304+
ggml_backend_reg_t reg = ggml_backend_reg_by_name("RPC");
305+
if (!reg) {
306+
fprintf(stderr, "Failed to find RPC backend\n");
307+
return 1;
308+
}
309+
310+
auto start_server_fn = (decltype(ggml_backend_rpc_start_server)*) ggml_backend_reg_get_proc_address(reg, "ggml_backend_rpc_start_server");
311+
if (!start_server_fn) {
312+
fprintf(stderr, "Failed to obtain RPC backend start server function\n");
313+
return 1;
314+
}
315+
316+
start_server_fn(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
317+
324318
ggml_backend_free(backend);
325319
return 0;
326320
}

0 commit comments

Comments
 (0)