rpc : add backend registry / device interfaces (#9812)

* rpc : add backend registry / device interfaces

* llama : add llama_supports_rpc API

* ggml_backend_rpc_start_rpc_server -> ggml_backend_rpc_start_server

common/arg.cpp

@@ -1353,15 +1353,15 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.image.emplace_back(value);
         }
     ).set_examples({LLAMA_EXAMPLE_LLAVA}));
-#ifdef GGML_USE_RPC
+    if (llama_supports_rpc()) {
-    add_opt(llama_arg(
+        add_opt(llama_arg(
-        {"--rpc"}, "SERVERS",
+            {"--rpc"}, "SERVERS",
-        "comma separated list of RPC servers",
+            "comma separated list of RPC servers",
-        [](gpt_params & params, const std::string & value) {
+            [](gpt_params & params, const std::string & value) {
-            params.rpc_servers = value;
+                params.rpc_servers = value;
-        }
+            }
-    ).set_env("LLAMA_ARG_RPC"));
+        ).set_env("LLAMA_ARG_RPC"));
-#endif
+    }
     add_opt(llama_arg(
         {"--mlock"},
         "force system to keep model in RAM rather than swapping or compressing",

examples/llama-bench/llama-bench.cpp

@@ -304,9 +304,9 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  --cpu-strict <0|1>                        (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
     printf("  --poll <0...100>                          (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-#ifdef GGML_USE_RPC
+    if (llama_supports_rpc()) {
-    printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+        printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
-#endif
+    }
     printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     printf("  -mg, --main-gpu <i>                       (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
@@ -497,14 +497,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
-#ifdef GGML_USE_RPC
+        } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) {
-        } else if (arg == "-rpc" || arg == "--rpc") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rpc_servers.push_back(argv[i]);
-#endif
         } else if (arg == "-sm" || arg == "--split-mode") {
             if (++i >= argc) {
                 invalid_param = true;

examples/rpc/rpc-server.cpp

@@ -151,7 +151,7 @@ int main(int argc, char * argv[]) {
         get_backend_memory(&free_mem, &total_mem);
     }
     printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
-    start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
     ggml_backend_free(backend);
     return 0;
 }

ggml/include/ggml-rpc.h

@@ -17,7 +17,11 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * en
 
 GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+
+GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend.cpp

@@ -542,6 +542,10 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
 #include "ggml-blas.h"
 #endif
 
+#ifdef GGML_USE_RPC
+#include "ggml-rpc.h"
+#endif
+
 struct ggml_backend_registry {
     std::vector<ggml_backend_reg_t> backends;
     std::vector<ggml_backend_dev_t> devices;
@@ -556,6 +560,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_BLAS
         register_backend(ggml_backend_blas_reg());
 #endif
+#ifdef GGML_USE_RPC
+        register_backend(ggml_backend_rpc_reg());
+#endif
 
         // TODO: sycl, vulkan, kompute, cann
 

ggml/src/ggml-rpc.cpp

@@ -25,7 +25,7 @@
 #  include <netdb.h>
 #  include <unistd.h>
 #endif
-#include <string.h>
+#include <cstring>
 
 #define UNUSED GGML_UNUSED
 
@@ -630,22 +630,6 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
     return (enum ggml_status)output[0];
 }
 
-static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
-    UNUSED(backend);
-    UNUSED(op);
-    //TODO: call the remote backend and cache the results
-    return true;
-}
-
-static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) {
-        return false;
-    }
-    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
-    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
-    return buft_ctx->endpoint == rpc_ctx->endpoint;
-}
-
 static ggml_backend_i ggml_backend_rpc_interface = {
     /* .get_name                = */ ggml_backend_rpc_name,
     /* .free                    = */ ggml_backend_rpc_free,
@@ -659,8 +643,8 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
-    /* .supports_op             = */ ggml_backend_rpc_supports_op,
+    /* .supports_op             = */ NULL,
-    /* .supports_buft           = */ ggml_backend_rpc_supports_buft,
+    /* .supports_buft           = */ NULL,
     /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
@@ -691,7 +675,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * en
 
     ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type {
         /* .iface   = */ ggml_backend_rpc_buffer_type_interface,
-        /* .device  = */ nullptr,
+        /* .device  = */ ggml_backend_rpc_add_device(endpoint),
         /* .context = */ buft_ctx
     };
     buft_map[endpoint] = buft;
@@ -707,7 +691,7 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_rpc_guid(),
         /* .interface = */ ggml_backend_rpc_interface,
-        /* .device    = */ nullptr,
+        /* .device    = */ ggml_backend_rpc_add_device(endpoint),
         /* .context   = */ ctx
     };
     return backend;
@@ -1189,7 +1173,7 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
     }
 }
 
-void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
+void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
     std::string host;
     int port;
     if (!parse_endpoint(endpoint, host, port)) {
@@ -1226,3 +1210,179 @@ void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free
     WSACleanup();
 #endif
 }
+
+// device interface
+
+struct ggml_backend_rpc_device_context {
+    std::string endpoint;
+    std::string name;
+};
+
+static const char * ggml_backend_rpc_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+
+    return ctx->name.c_str();
+}
+
+static const char * ggml_backend_rpc_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_rpc_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+
+    ggml_backend_rpc_get_device_memory(ctx->endpoint.c_str(), free, total);
+
+    UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_rpc_device_get_type(ggml_backend_dev_t dev) {
+    // TODO: obtain value from the server
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
+
+    UNUSED(dev);
+}
+
+static void ggml_backend_rpc_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_rpc_device_get_name(dev);
+    props->description = ggml_backend_rpc_device_get_description(dev);
+    props->type        = ggml_backend_rpc_device_get_type(dev);
+    ggml_backend_rpc_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_rpc_device_init(ggml_backend_dev_t dev, const char * params) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+
+    return ggml_backend_rpc_init(ctx->endpoint.c_str());
+
+    UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_rpc_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_rpc_device_context * ctx = (ggml_backend_rpc_device_context *)dev->context;
+
+    return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
+
+    UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_rpc_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    UNUSED(dev);
+    UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_rpc_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    UNUSED(dev);
+    UNUSED(op);
+    //TODO: call the remote backend and cache the results
+    return true;
+}
+
+static bool ggml_backend_rpc_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) {
+        return false;
+    }
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    ggml_backend_rpc_device_context * dev_ctx = (ggml_backend_rpc_device_context *)dev->context;
+    return buft_ctx->endpoint == dev_ctx->endpoint;
+}
+
+static const struct ggml_backend_device_i ggml_backend_rpc_device_i = {
+    /* .get_name             = */ ggml_backend_rpc_device_get_name,
+    /* .get_description      = */ ggml_backend_rpc_device_get_description,
+    /* .get_memory           = */ ggml_backend_rpc_device_get_memory,
+    /* .get_type             = */ ggml_backend_rpc_device_get_type,
+    /* .get_props            = */ ggml_backend_rpc_device_get_props,
+    /* .init_backend         = */ ggml_backend_rpc_device_init,
+    /* .get_buffer_type      = */ ggml_backend_rpc_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_rpc_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_rpc_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_rpc_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_rpc_reg_get_name(ggml_backend_reg_t reg) {
+    return "RPC";
+
+    UNUSED(reg);
+}
+
+static size_t ggml_backend_rpc_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 0;
+
+    UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_rpc_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ABORT("The RPC backend does not have enumerated devices - use ggml_backend_add_device instead");
+
+    UNUSED(reg);
+    UNUSED(index);
+}
+
+static void * ggml_backend_rpc_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) {
+        return (void *)ggml_backend_rpc_add_device;
+    }
+    return NULL;
+
+    UNUSED(reg);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
+    /* .get_name         = */ ggml_backend_rpc_reg_get_name,
+    /* .get_device_count = */ ggml_backend_rpc_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_rpc_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_rpc_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_rpc_reg(void) {
+    static struct ggml_backend_reg ggml_backend_rpc_reg = {
+        /* .iface   = */ ggml_backend_rpc_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_rpc_reg;
+}
+
+ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
+    static std::unordered_map<std::string, ggml_backend_dev_t> dev_map;
+
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    if (dev_map.find(endpoint) != dev_map.end()) {
+        return dev_map[endpoint];
+    }
+
+    ggml_backend_rpc_device_context * ctx = new ggml_backend_rpc_device_context {
+        /* .endpoint = */ endpoint,
+        /* .name     = */ "RPC[" + std::string(endpoint) + "]",
+    };
+
+    ggml_backend_dev_t dev = new ggml_backend_device {
+        /* .iface   = */ ggml_backend_rpc_device_i,
+        /* .reg     = */ ggml_backend_rpc_reg(),
+        /* .context = */ ctx,
+    };
+
+    dev_map[endpoint] = dev;
+
+    return dev;
+}

include/llama.h

@@ -433,6 +433,7 @@ extern "C" {
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API bool llama_supports_rpc        (void);
 
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);

src/llama.cpp

@@ -8,10 +8,6 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#ifdef GGML_USE_RPC
-#  include "ggml-rpc.h"
-#endif
-
 #if defined(GGML_USE_VULKAN)
 #  include "ggml-vulkan.h"
 #elif defined(GGML_USE_SYCL)
@@ -3404,10 +3400,6 @@ struct llama_lora_adapter {
 static int llama_get_device_count(const llama_model & model) {
     int count = (int) model.devices.size();
 
-#if defined(GGML_USE_RPC)
-    count += (int) model.rpc_servers.size();
-#endif
-
 #if defined(GGML_USE_SYCL)
     count += ggml_backend_sycl_get_device_count();
 #elif defined(GGML_USE_VULKAN)
@@ -3460,15 +3452,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(const llama_mode
 static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int device) {
     ggml_backend_buffer_type_t buft = nullptr;
 
-#if defined(GGML_USE_RPC)
-    int rpc_count = (int)model.rpc_servers.size();
-    if (device < rpc_count) {
-        const char * endpoint = model.rpc_servers[device].c_str();
-        return ggml_backend_rpc_buffer_type(endpoint);
-    }
-    device -= rpc_count;
-#endif
-
     if (device < (int)model.devices.size()) {
         return ggml_backend_dev_buffer_type(model.devices[device]);
     }
@@ -3523,18 +3506,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
 }
 
 static size_t llama_get_device_memory(const llama_model & model, int device) {
-#if defined(GGML_USE_RPC)
-    int rpc_count = (int)model.rpc_servers.size();
-    if (device < rpc_count) {
-        size_t total;
-        size_t free;
-        const char * endpoint = model.rpc_servers[device].c_str();
-        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
-        return free;
-    }
-    device = device - rpc_count;
-#endif
-
     if (device < (int)model.devices.size()) {
         ggml_backend_dev_t dev = model.devices[device];
         size_t total;
@@ -19019,15 +18990,20 @@ bool llama_supports_mlock(void) {
 
 bool llama_supports_gpu_offload(void) {
 #if defined(GGML_USE_VULKAN) || \
-    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
+    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
     // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
     return true;
 #else
     return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr ||
-        ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr;
+           ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL) != nullptr ||
+           llama_supports_rpc();
 #endif
 }
 
+bool llama_supports_rpc(void) {
+    return ggml_backend_reg_by_name("RPC") != nullptr;
+}
+
 void llama_backend_init(void) {
     ggml_time_init();
 
@@ -19102,6 +19078,36 @@ struct llama_model * llama_load_model_from_file(
         model->rpc_servers.push_back(servers);
     }
 
+    // add RPC devices
+    if (!model->rpc_servers.empty()) {
+        ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
+        if (!rpc_reg) {
+            LLAMA_LOG_ERROR("%s: failed to find RPC backend\n", __func__);
+            llama_free_model(model);
+            return nullptr;
+        }
+
+        // ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+        using ggml_backend_rpc_add_device_t = ggml_backend_dev_t (*)(const char *);
+        ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+        if (!ggml_backend_rpc_add_device_fn) {
+            LLAMA_LOG_ERROR("%s: failed to find RPC device add function\n", __func__);
+            llama_free_model(model);
+            return nullptr;
+        }
+
+        for (const std::string & server : model->rpc_servers) {
+            ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+            if (dev) {
+                model->devices.push_back(dev);
+            } else {
+                LLAMA_LOG_ERROR("%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
+                llama_free_model(model);
+                return nullptr;
+            }
+        }
+    }
+
     // create list of devices to use with this model
     // currently, we use all available devices
     // TODO: rework API to give user more control over device selection
@@ -19128,7 +19134,7 @@ struct llama_model * llama_load_model_from_file(
         } else if (status == -2) {
             LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
         }
-        delete model;
+        llama_free_model(model);
         return nullptr;
     }
 
@@ -19311,23 +19317,6 @@ struct llama_context * llama_new_context_with_model(
             main_gpu -= (int)model->devices.size();
         }
 
-#if defined(GGML_USE_RPC)
-        if (model->n_gpu_layers > 0) {
-            for (const auto & endpoint : model->rpc_servers) {
-                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.push_back(backend);
-            }
-        }
-        if (main_gpu >= (int)model->rpc_servers.size()) {
-            main_gpu -= (int)model->rpc_servers.size();
-        }
-#endif
-
 #if defined(GGML_USE_VULKAN)
         if (model->split_mode == LLAMA_SPLIT_MODE_ROW) {
             LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__);