Address review comments

examples/rpc/README.md

@@ -1,6 +1,6 @@
 ## Overview
 
-The `rpc-server` allows  running a `ggml` backend on a remote host.
+The `rpc-server` allows  running `ggml` backend on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:
 
@@ -25,6 +25,7 @@ flowchart TD
 ```
 
 Each host can run a different backend, e.g. one with CUDA and another with Metal.
+You can also run multiple `rpc-server` instances on the same host, each with a different backend.
 
 ## Usage
 
@@ -35,7 +36,7 @@ For example, to build the CUDA backend with RPC support:
 mkdir build-rpc-cuda
 cd build-rpc-cuda
 cmake .. -DLLAMA_CUDA=ON -DLLAMA_RPC=ON
-make -j
+cmake --build . --config Release
 ```
 
 Then, start the `rpc-server` with the backend:
@@ -50,13 +51,20 @@ ggml_cuda_init: found 1 CUDA devices:
 Starting RPC server on 0.0.0.0:50052
 ```
 
+When using the CUDA backend, you can specify the device with the `CUDA_VISIBLE_DEVICES` environment variable, e.g.:
+```bash
+$ CUDA_VISIBLE_DEVICES=0 bin/rpc-server 0.0.0.0 50052
+```
+This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
+
+
 On the main host build `llama.cpp` only with `-DLLAMA_RPC=ON`:
 
 ```bash
 mkdir build-rpc
 cd build-rpc
 cmake .. -DLLAMA_RPC=ON
-make -j
+cmake --build . --config Release
 ```
 
 Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:

examples/rpc/rpc-server.cpp

@@ -7,19 +7,8 @@
 #endif
 
 #include "ggml-rpc.h"
-#include <memory>
 #include <string>
-#ifndef _WIN32
-#  include <sys/socket.h>
-#  include <sys/types.h>
-#  include <arpa/inet.h>
-#  include <netinet/in.h>
-#  include <netinet/tcp.h>
-#  include <netdb.h>
-#  include <unistd.h>
-#endif
 #include <stdio.h>
-#include <stdlib.h>
 
 static ggml_backend_t create_backend() {
     ggml_backend_t backend = NULL;
@@ -55,91 +44,27 @@ static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
 #endif
 }
 
-static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
-#ifdef _WIN32
-    if (fd == INVALID_SOCKET) {
-        return nullptr;
-    }
-#else
-    if (fd < 0) {
-        return nullptr;
-    }
-#endif
-    return std::make_shared<socket_t>(fd);
-}
-
-static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
-    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
-    auto sock = make_socket(sockfd);
-    if (sock == nullptr) {
-        return nullptr;
-    }
-
-    struct sockaddr_in serv_addr;
-    serv_addr.sin_family = AF_INET;
-    serv_addr.sin_addr.s_addr = inet_addr(host);
-    serv_addr.sin_port = htons(port);
-
-    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
-        return nullptr;
-    }
-    if (listen(sockfd, 5) < 0) {
-        return nullptr;
-    }
-    return sock;
-}
-
 int main(int argc, char * argv[])
 {
-#ifdef _WIN32
-    WSADATA wsaData;
-    int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
-    if (res != 0) {
-        fprintf(stderr, "WSAStartup failed: %d\n", res);
-        return 1;
-    }
-#endif
     if (argc < 3) {
         fprintf(stderr, "Usage: %s <host> <port>\n", argv[0]);
         return 1;
     }
     const char * host = argv[1];
     int port = std::stoi(argv[2]);
-
+    if (port <= 0 || port > 65535) {
+        fprintf(stderr, "Invalid port number: %d\n", port);
+        return 1;
+    }
     ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");
         return 1;
     }
-
     printf("Starting RPC server on %s:%d\n", host, port);
-    auto server_socket = create_server_socket(host, port);
-    if (server_socket == nullptr) {
-        fprintf(stderr, "Failed to create server socket\n");
-        return 1;
-    }
-    while (true) {
-        auto client_socket_fd = accept(server_socket->fd, NULL, NULL);
-        auto client_socket = make_socket(client_socket_fd);
-        if (client_socket == nullptr) {
-            fprintf(stderr, "Failed to accept client connection\n");
-            return 1;
-        }
-        // set TCP_NODELAY to disable Nagle's algorithm
-        int flag = 1;
-        int ret = setsockopt(client_socket->fd, IPPROTO_TCP, TCP_NODELAY, (char *) &flag, sizeof(int));
-        if (ret < 0) {
-            fprintf(stderr, "Failed to set TCP_NODELAY\n");
-            continue;
-        }
     size_t free_mem, total_mem;
     get_backend_memory(&free_mem, &total_mem);
-        printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
+    std::string endpoint = std::string(host) + ":" + std::to_string(port);
-        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
+    start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
-        printf("Client connection closed\n");
-    }
-#ifdef _WIN32
-    WSACleanup();
-#endif
     return 0;
 }

ggml-rpc.cpp

@@ -4,9 +4,18 @@
 
 #include <string>
 #include <vector>
+#include <memory>
 #include <unordered_map>
 #include <unordered_set>
-#ifndef _WIN32
+#ifdef _WIN32
+#  define WIN32_LEAN_AND_MEAN
+#  ifndef NOMINMAX
+#     define NOMINMAX
+#  endif
+#  include <windows.h>
+#  include <winsock2.h>
+#else
+#  include <arpa/inet.h>
 #  include <sys/socket.h>
 #  include <sys/types.h>
 #  include <netinet/in.h>
@@ -26,9 +35,57 @@
 #endif
 
 #ifdef _WIN32
+typedef SOCKET sockfd_t;
 using ssize_t = __int64;
+#else
+typedef int sockfd_t;
 #endif
 
+// cross-platform socket
+struct socket_t {
+    sockfd_t fd;
+    socket_t(sockfd_t fd) : fd(fd) {}
+    ~socket_t() {
+#ifdef _WIN32
+        closesocket(this->fd);
+#else
+        close(this->fd);
+#endif
+    }
+};
+
+// ggml_tensor is serialized into rpc_tensor
+struct rpc_tensor {
+    uint64_t id;
+    uint32_t type;
+    uint64_t buffer;
+    uint32_t ne[GGML_MAX_DIMS];
+    uint32_t nb[GGML_MAX_DIMS];
+    uint32_t op;
+    int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+    int32_t  flags;
+    uint64_t src[GGML_MAX_SRC];
+    uint64_t view_src;
+    uint64_t view_offs;
+    uint64_t data;
+    char name[GGML_MAX_NAME];
+};
+
+// RPC commands
+enum rpc_cmd {
+    ALLOC_BUFFER = 0,
+    GET_ALIGNMENT,
+    GET_MAX_SIZE,
+    BUFFER_GET_BASE,
+    FREE_BUFFER,
+    BUFFER_CLEAR,
+    SET_TENSOR,
+    GET_TENSOR,
+    COPY_TENSOR,
+    GRAPH_COMPUTE,
+    GET_DEVICE_MEMORY,
+};
+
 // RPC data structures
 
 static ggml_guid_t ggml_backend_rpc_guid() {
@@ -59,14 +116,6 @@ struct ggml_backend_rpc_buffer_context {
 
 // RPC helper functions
 
-socket_t::~socket_t() {
-#ifdef _WIN32
-    closesocket(this->fd);
-#else
-    close(this->fd);
-#endif
-}
-
 static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
 #ifdef _WIN32
     if (fd == INVALID_SOCKET) {
@@ -80,6 +129,13 @@ static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
     return std::make_shared<socket_t>(fd);
 }
 
+static bool set_no_delay(sockfd_t sockfd) {
+    int flag = 1;
+    // set TCP_NODELAY to disable Nagle's algorithm
+    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
+    return ret >= 0;
+}
+
 static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
     struct sockaddr_in addr;
     auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
@@ -87,10 +143,8 @@ static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
     if (sock_ptr == nullptr) {
         return nullptr;
     }
-    // set TCP_NODELAY to disable Nagle's algorithm
+    if (!set_no_delay(sockfd)) {
-    int flag = 1;
+        fprintf(stderr, "Failed to set TCP_NODELAY\n");
-    int ret = setsockopt(sock_ptr->fd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
-    if (ret < 0) {
         return nullptr;
     }
     addr.sin_family = AF_INET;
@@ -107,6 +161,40 @@ static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
     return sock_ptr;
 }
 
+static std::shared_ptr<socket_t> socket_accept(sockfd_t srv_sockfd) {
+    auto client_socket_fd = accept(srv_sockfd, NULL, NULL);
+    auto client_socket = make_socket(client_socket_fd);
+    if (client_socket == nullptr) {
+        return nullptr;
+    }
+    if (!set_no_delay(client_socket_fd)) {
+        fprintf(stderr, "Failed to set TCP_NODELAY\n");
+        return nullptr;
+    }
+    return client_socket;
+}
+
+static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
+    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    auto sock = make_socket(sockfd);
+    if (sock == nullptr) {
+        return nullptr;
+    }
+
+    struct sockaddr_in serv_addr;
+    serv_addr.sin_family = AF_INET;
+    serv_addr.sin_addr.s_addr = inet_addr(host);
+    serv_addr.sin_port = htons(port);
+
+    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
+        return nullptr;
+    }
+    if (listen(sockfd, 1) < 0) {
+        return nullptr;
+    }
+    return sock;
+}
+
 static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
     size_t bytes_sent = 0;
     while (bytes_sent < size) {
@@ -131,6 +219,17 @@ static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
     return true;
 }
 
+static bool parse_endpoint(const char * endpoint, std::string & host, int & port) {
+    std::string str(endpoint);
+    size_t pos = str.find(':');
+    if (pos == std::string::npos) {
+        return false;
+    }
+    host = str.substr(0, pos);
+    port = std::stoi(str.substr(pos + 1));
+    return true;
+}
+
 // RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
 // RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
 static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
@@ -244,7 +343,7 @@ static ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_ten
     }
     result->flags = tensor->flags;
     result->data = reinterpret_cast<void *>(tensor->data);
-    snprintf(result->name, GGML_MAX_NAME, "%s", tensor->name);
+    ggml_set_name(result, tensor->name);
     return result;
 }
 
@@ -259,7 +358,7 @@ GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t
 GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
-    int input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
+    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
     std::vector<uint8_t> input(input_size, 0);
     rpc_tensor rpc_tensor = serialize_tensor(tensor);
     memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
@@ -530,14 +629,15 @@ static ggml_backend_i ggml_backend_rpc_interface = {
 
 static std::unordered_map<std::string, ggml_backend_t> instances;
 
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const std::string & endpoint) {
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
     ggml_backend_t backend = ggml_backend_rpc_init(endpoint);
     return backend != nullptr ? ggml_backend_rpc_get_default_buffer_type(backend) : nullptr;
 }
 
-GGML_CALL ggml_backend_t ggml_backend_rpc_init(const std::string & endpoint) {
+GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
-    if (instances.find(endpoint) != instances.end()) {
+    std::string endpoint_str(endpoint);
-        return instances[endpoint];
+    if (instances.find(endpoint_str) != instances.end()) {
+        return instances[endpoint_str];
     }
 #ifdef _WIN32
     {
@@ -548,11 +648,12 @@ GGML_CALL ggml_backend_t ggml_backend_rpc_init(const std::string & endpoint) {
         }
     }
 #endif
-    GGML_PRINT_DEBUG("Connecting to %s\n", endpoint.c_str());
+    GGML_PRINT_DEBUG("Connecting to %s\n", endpoint);
-    // split the endpoint into host and port
+    std::string host;
-    size_t pos = endpoint.find(":");
+    int port;
-    std::string host = endpoint.substr(0, pos);
+    if (!parse_endpoint(endpoint, host, port)) {
-    int port = std::stoi(endpoint.substr(pos + 1));
+        return nullptr;
+    }
     auto sock = socket_connect(host.c_str(), port);
     if (sock == nullptr) {
         return nullptr;
@@ -607,7 +708,7 @@ static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * f
     *total = total_mem;
 }
 
-GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const std::string & endpoint, size_t * free, size_t * total) {
+GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
     ggml_backend_t backend = ggml_backend_rpc_init(endpoint);
     if (backend == nullptr) {
         *free = 0;
@@ -781,7 +882,7 @@ static void rpc_graph_compute(ggml_backend_t backend, const std::vector<uint8_t>
     const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
     GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
 
-    static size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead();
+    static size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_size,
         /*.mem_buffer =*/ NULL,
@@ -805,7 +906,7 @@ static void rpc_graph_compute(ggml_backend_t backend, const std::vector<uint8_t>
     ggml_free(ctx);
 }
 
-void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
+static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
     while (true) {
         uint8_t cmd;
         if (!recv_data(sockfd, &cmd, 1)) {
@@ -871,7 +972,7 @@ void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem,
             }
             default: {
                 fprintf(stderr, "Unknown command: %d\n", cmd);
-                break;
+                return;
             }
         }
         uint64_t output_size = output.size();
@@ -883,3 +984,39 @@ void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem,
         }
     }
 }
+
+void start_rpc_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)) {
+        return;
+    }
+#ifdef _WIN32
+    {
+        WSADATA wsaData;
+        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
+        if (res != 0) {
+            fprintf(stderr, "WSAStartup failed: %d\n", res);
+            return 1;
+        }
+    }
+#endif
+    auto server_socket = create_server_socket(host.c_str(), port);
+    if (server_socket == nullptr) {
+        fprintf(stderr, "Failed to create server socket\n");
+        return;
+    }
+    while (true) {
+        auto client_socket = socket_accept(server_socket->fd);
+        if (client_socket == nullptr) {
+            fprintf(stderr, "Failed to accept client connection\n");
+            return;
+        }
+        printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
+        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
+        printf("Client connection closed\n");
+    }
+#ifdef _WIN32
+    WSACleanup();
+#endif
+}

ggml-rpc.h

@@ -2,79 +2,22 @@
 
 #include "ggml.h"
 #include "ggml-backend.h"
-#include <string>
-#include <memory>
-#ifdef _WIN32
-#  define WIN32_LEAN_AND_MEAN
-#  ifndef NOMINMAX
-#     define NOMINMAX
-#  endif
-#  include <windows.h>
-#  include <winsock2.h>
-#endif
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-// cross-platform socket fd
-#ifdef _WIN32
-typedef SOCKET sockfd_t;
-#else
-typedef int sockfd_t;
-#endif
-
-// cross-platform socket
-struct socket_t {
-    sockfd_t fd;
-    socket_t(sockfd_t fd) : fd(fd) {}
-    ~socket_t();
-};
-
-
-// ggml_tensor is serialized into rpc_tensor
-struct rpc_tensor {
-    uint64_t id;
-    uint32_t type;
-    uint64_t buffer;
-    uint32_t ne[GGML_MAX_DIMS];
-    uint32_t nb[GGML_MAX_DIMS];
-    uint32_t op;
-    int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
-    int32_t  flags;
-    uint64_t src[GGML_MAX_SRC];
-    uint64_t view_src;
-    uint64_t view_offs;
-    uint64_t data;
-    char name[GGML_MAX_NAME];
-};
-
-// RPC commands
-enum rpc_cmd {
-    ALLOC_BUFFER = 0,
-    GET_ALIGNMENT,
-    GET_MAX_SIZE,
-    BUFFER_GET_BASE,
-    FREE_BUFFER,
-    BUFFER_CLEAR,
-    SET_TENSOR,
-    GET_TENSOR,
-    COPY_TENSOR,
-    GRAPH_COMPUTE,
-    GET_DEVICE_MEMORY,
-};
-
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const std::string & endpoint);
+GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
 GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const std::string & endpoint);
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const std::string & endpoint, size_t * free, size_t * total);
+GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API GGML_CALL void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem);
+GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
 
 #ifdef  __cplusplus
 }

llama.cpp

@@ -2279,7 +2279,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_
 
 #ifdef GGML_USE_RPC
     std::string endpoint = model.rpc_servers[gpu];
-    buft = ggml_backend_rpc_buffer_type(endpoint);
+    buft = ggml_backend_rpc_buffer_type(endpoint.c_str());
 #elif defined(GGML_USE_METAL)
     buft = ggml_backend_metal_buffer_type();
 #elif defined(GGML_USE_CUDA)
@@ -2348,7 +2348,7 @@ static size_t llama_get_device_memory(const llama_model & model, int device) {
     size_t total;
     size_t free;
     std::string endpoint = model.rpc_servers[device];
-    ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
+    ggml_backend_rpc_get_device_memory(endpoint.c_str(), &free, &total);
     return free;
 #elif defined(GGML_USE_CUDA)
     size_t total;
@@ -15727,7 +15727,7 @@ struct llama_context * llama_new_context_with_model(
         // initialize backends
 #if defined(GGML_USE_RPC)
         for (auto & server : model->rpc_servers) {
-            ggml_backend_t backend = ggml_backend_rpc_init(server);
+            ggml_backend_t backend = ggml_backend_rpc_init(server.c_str());
             if (backend == nullptr) {
                 LLAMA_LOG_ERROR("%s: failed to connect RPC backend to %s\n", __func__, server.c_str());
                 llama_free(ctx);