tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

rpc : use backend registry, support dl backends (#13304)

Browse Source
This commit is contained in:
Diego Devesa
2025-05-04 21:25:43 +02:00
committed by GitHub
parent 6eb7d25c70
commit 9fdfcdaedd
4 changed files with 116 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -18,17 +18,19 @@
# include "kleidiai/kleidiai.h" # include "kleidiai/kleidiai.h"
#endif #endif
#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
#endif
#if defined(_WIN32) #if defined(_WIN32)
# define WIN32_LEAN_AND_MEAN # define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX # ifndef NOMINMAX
# define NOMINMAX # define NOMINMAX
# endif # endif
# include <windows.h> # include <windows.h>
#else
# include <unistd.h>
#endif
#if defined(__APPLE__)
# include <sys/sysctl.h>
# include <sys/types.h>
#endif #endif
// ggml-backend interface // ggml-backend interface
@ -70,8 +72,10 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_ty
} }
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) { static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra == buft) return true; if (extra && extra == buft) {
return true;
}
} }
return false; return false;
} }
@ -330,9 +334,18 @@ static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t d
} }
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO #ifdef _WIN32
*free = 0; MEMORYSTATUSEX status;
*total = 0; status.dwLength = sizeof(status);
GlobalMemoryStatusEx(&status);
*total = status.ullTotalPhys;
*free = status.ullAvailPhys;
#else
long pages = sysconf(_SC_PHYS_PAGES);
long page_size = sysconf(_SC_PAGE_SIZE);
*total = pages * page_size;
*free = *total;
#endif
GGML_UNUSED(dev); GGML_UNUSED(dev);
} }

11
ggml/src/ggml-rpc/ggml-rpc.cpp
View File

@ -1594,6 +1594,14 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
const char * cache_dir, const char * cache_dir,
size_t free_mem, size_t total_mem) { size_t free_mem, size_t total_mem) {
printf("Starting RPC server v%d.%d.%d\n",
RPC_PROTO_MAJOR_VERSION,
RPC_PROTO_MINOR_VERSION,
RPC_PROTO_PATCH_VERSION);
printf(" endpoint : %s\n", endpoint);
printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a");
printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024));
std::string host; std::string host;
int port; int port;
if (!parse_endpoint(endpoint, host, port)) { 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
if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) { if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) {
return (void *)ggml_backend_rpc_add_device; return (void *)ggml_backend_rpc_add_device;
} }
if (std::strcmp(name, "ggml_backend_rpc_start_server") == 0) {
return (void *)ggml_backend_rpc_start_server;
}
return NULL; return NULL;
GGML_UNUSED(reg); GGML_UNUSED(reg);

6
tools/CMakeLists.txt
View File

@ -28,12 +28,12 @@ else()
add_subdirectory(tokenize) add_subdirectory(tokenize)
add_subdirectory(tts) add_subdirectory(tts)
add_subdirectory(llava) add_subdirectory(llava)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (NOT GGML_BACKEND_DL) if (NOT GGML_BACKEND_DL)
# these examples use the backends directly and cannot be built with dynamic loading # these examples use the backends directly and cannot be built with dynamic loading
add_subdirectory(cvector-generator) add_subdirectory(cvector-generator)
add_subdirectory(export-lora) add_subdirectory(export-lora)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
endif() endif()
endif() endif()

150
tools/rpc/rpc-server.cpp
View File

@ -2,24 +2,6 @@
#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
#endif #endif
#include "ggml-cpu.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#include "ggml-rpc.h" #include "ggml-rpc.h"
#ifdef _WIN32 #ifdef _WIN32
# define NOMINMAX # define NOMINMAX
@ -154,6 +136,7 @@ struct rpc_server_params {
size_t backend_mem = 0; size_t backend_mem = 0;
bool use_cache = false; bool use_cache = false;
int n_threads = std::max(1U, std::thread::hardware_concurrency()/2); int n_threads = std::max(1U, std::thread::hardware_concurrency()/2);
std::string device;
}; };
static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) { static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
@ -161,6 +144,7 @@ static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
fprintf(stderr, "options:\n"); fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n"); fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -t, --threads number of threads for the CPU backend (default: %d)\n", params.n_threads); fprintf(stderr, " -t, --threads number of threads for the CPU backend (default: %d)\n", params.n_threads);
fprintf(stderr, " -d DEV, --device device to use\n");
fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str()); fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str());
fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port); fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port);
fprintf(stderr, " -m MEM, --mem MEM backend memory size (in MB)\n"); 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 &
fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads); fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
return false; return false;
} }
} else if (arg == "-d" || arg == "--device") {
if (++i >= argc) {
return false;
}
params.device = argv[i];
if (ggml_backend_dev_by_name(params.device.c_str()) == nullptr) {
fprintf(stderr, "error: unknown device: %s\n", params.device.c_str());
fprintf(stderr, "available devices:\n");
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
auto * dev = ggml_backend_dev_get(i);
size_t free, total;
ggml_backend_dev_memory(dev, &free, &total);
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);
}
return false;
}
} else if (arg == "-p" || arg == "--port") { } else if (arg == "-p" || arg == "--port") {
if (++i >= argc) { if (++i >= argc) {
return false; return false;
@ -214,66 +214,53 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
} }
static ggml_backend_t create_backend(const rpc_server_params & params) { static ggml_backend_t create_backend(const rpc_server_params & params) {
ggml_backend_t backend = NULL; ggml_backend_t backend = nullptr;
#ifdef GGML_USE_CUDA
fprintf(stderr, "%s: using CUDA backend\n", __func__);
backend = ggml_backend_cuda_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
}
#elif GGML_USE_METAL
fprintf(stderr, "%s: using Metal backend\n", __func__);
backend = ggml_backend_metal_init();
if (!backend) {
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
}
#elif GGML_USE_VULKAN
fprintf(stderr, "%s: using Vulkan backend\n", __func__);
backend = ggml_backend_vk_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
}
#elif GGML_USE_SYCL
fprintf(stderr, "%s: using SYCL backend\n", __func__);
backend = ggml_backend_sycl_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
}
#endif
// if there aren't GPU Backends fallback to CPU backend if (!params.device.empty()) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(params.device.c_str());
if (dev) {
backend = ggml_backend_dev_init(dev, nullptr);
if (!backend) { if (!backend) {
fprintf(stderr, "%s: using CPU backend\n", __func__); fprintf(stderr, "Failed to create backend for device %s\n", params.device.c_str());
backend = ggml_backend_cpu_init(); return nullptr;
ggml_backend_cpu_set_n_threads(backend, params.n_threads);
} }
}
}
// try to initialize a GPU backend first
if (!backend) {
backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
}
// if there aren't GPU backends fallback to CPU backend
if (!backend) {
backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
}
fprintf(stderr, "%s: using %s backend\n", __func__, ggml_backend_name(backend));
// set the number of threads
ggml_backend_dev_t dev = ggml_backend_get_device(backend);
ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
if (reg) {
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");
if (ggml_backend_set_n_threads_fn) {
ggml_backend_set_n_threads_fn(backend, params.n_threads);
}
}
return backend; return backend;
} }
static void get_backend_memory(size_t * free_mem, size_t * total_mem) { static void get_backend_memory(ggml_backend_t backend, size_t * free_mem, size_t * total_mem) {
#ifdef GGML_USE_CUDA ggml_backend_dev_t dev = ggml_backend_get_device(backend);
ggml_backend_cuda_get_device_memory(0, free_mem, total_mem); GGML_ASSERT(dev != nullptr);
#elif GGML_USE_VULKAN ggml_backend_dev_memory(dev, free_mem, total_mem);
ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
#elif GGML_USE_SYCL
ggml_backend_sycl_get_device_memory(0, free_mem, total_mem);
#else
#ifdef _WIN32
MEMORYSTATUSEX status;
status.dwLength = sizeof(status);
GlobalMemoryStatusEx(&status);
*total_mem = status.ullTotalPhys;
*free_mem = status.ullAvailPhys;
#else
long pages = sysconf(_SC_PHYS_PAGES);
long page_size = sysconf(_SC_PAGE_SIZE);
*total_mem = pages * page_size;
*free_mem = *total_mem;
#endif
#endif
} }
int main(int argc, char * argv[]) { int main(int argc, char * argv[]) {
ggml_backend_load_all();
rpc_server_params params; rpc_server_params params;
if (!rpc_server_params_parse(argc, argv, params)) { if (!rpc_server_params_parse(argc, argv, params)) {
fprintf(stderr, "Invalid parameters\n"); fprintf(stderr, "Invalid parameters\n");
@ -301,7 +288,7 @@ int main(int argc, char * argv[]) {
free_mem = params.backend_mem; free_mem = params.backend_mem;
total_mem = params.backend_mem; total_mem = params.backend_mem;
} else { } else {
get_backend_memory(&free_mem, &total_mem); get_backend_memory(backend, &free_mem, &total_mem);
} }
const char * cache_dir = nullptr; const char * cache_dir = nullptr;
std::string cache_dir_str; std::string cache_dir_str;
@ -313,14 +300,21 @@ int main(int argc, char * argv[]) {
} }
cache_dir = cache_dir_str.c_str(); cache_dir = cache_dir_str.c_str();
} }
printf("Starting RPC server v%d.%d.%d\n",
RPC_PROTO_MAJOR_VERSION, ggml_backend_reg_t reg = ggml_backend_reg_by_name("RPC");
RPC_PROTO_MINOR_VERSION, if (!reg) {
RPC_PROTO_PATCH_VERSION); fprintf(stderr, "Failed to find RPC backend\n");
printf(" endpoint : %s\n", endpoint.c_str()); return 1;
printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a"); }
printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024));
ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem); auto start_server_fn = (decltype(ggml_backend_rpc_start_server)*) ggml_backend_reg_get_proc_address(reg, "ggml_backend_rpc_start_server");
if (!start_server_fn) {
fprintf(stderr, "Failed to obtain RPC backend start server function\n");
return 1;
}
start_server_fn(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
ggml_backend_free(backend); ggml_backend_free(backend);
return 0; return 0;
} }