include llama.cpp since there were some edits

llama.cpp/tools/rpc/CMakeLists.txt

@@ -0,0 +1,20 @@
+set(TARGET ggml-rpc-server)
+add_executable(${TARGET} rpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE ggml)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+if (LLAMA_BUILD_TESTS AND UNIX AND NOT GGML_BACKEND_DL)
+    add_executable(test-rpc-multi-server ${PROJECT_SOURCE_DIR}/tests/test-rpc-multi-server.cpp)
+    target_link_libraries(test-rpc-multi-server PRIVATE ggml ggml-rpc)
+    target_include_directories(test-rpc-multi-server PRIVATE ${PROJECT_SOURCE_DIR}/ggml/src)
+    add_test(
+        NAME test-rpc-multi-server
+        COMMAND bash ${PROJECT_SOURCE_DIR}/tests/test-rpc-multi-server.sh
+            $<TARGET_FILE:ggml-rpc-server>
+            $<TARGET_FILE:test-rpc-multi-server>)
+    set_property(TEST test-rpc-multi-server PROPERTY LABELS main)
+endif()
+
+if(LLAMA_TOOLS_INSTALL)
+    install(TARGETS ${TARGET} RUNTIME)
+endif()

llama.cpp/tools/rpc/README.md

@@ -0,0 +1,137 @@
+## Overview
+
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
+The `ggml-rpc-server` allows exposing `ggml` devices on a remote host.
+The RPC backend communicates with one or several instances of `ggml-rpc-server` and offloads computations to them.
+This can be used for distributed LLM inference with `llama.cpp` in the following way:
+
+```mermaid
+flowchart TD
+    rpcb<-->|TCP|srva
+    rpcb<-->|TCP|srvb
+    rpcb<-.->|TCP|srvn
+    subgraph hostn[Host N]
+    srvn[ggml-rpc-server]<-.->dev4["CUDA0"]
+    srvn[ggml-rpc-server]<-.->dev5["CPU"]
+    end
+    subgraph hostb[Host B]
+    srvb[ggml-rpc-server]<-->dev3["Metal"]
+    end
+    subgraph hosta[Host A]
+    srva[ggml-rpc-server]<-->dev["CUDA0"]
+    srva[ggml-rpc-server]<-->dev2["CUDA1"]
+    end
+    subgraph host[Main Host]
+    local["Local devices"]<-->ggml[llama-cli]
+    ggml[llama-cli]<-->rpcb[RPC backend]
+    end
+    style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
+    classDef devcls fill:#5B9BD5
+    class local,dev,dev2,dev3,dev4,dev5 devcls
+```
+
+By default, `ggml-rpc-server` exposes all available accelerator devices on the host.
+If there are no accelerators, it exposes a single `CPU` device.
+
+## Usage
+
+### Remote hosts
+
+On each remote host, build the backends for each accelerator by adding `-DGGML_RPC=ON` to the build options.
+For example, to build the `ggml-rpc-server` with support for CUDA accelerators:
+
+```bash
+mkdir build-rpc-cuda
+cd build-rpc-cuda
+cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON
+cmake --build . --config Release
+```
+
+When started, the `ggml-rpc-server` will detect and expose all available `CUDA` devices:
+
+```bash
+$ bin/ggml-rpc-server
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
+Starting RPC server v3.0.0
+  endpoint       : 127.0.0.1:50052
+  local cache    : n/a
+Devices:
+  CUDA0: NVIDIA GeForce RTX 5090 (32109 MiB, 31588 MiB free)
+```
+
+You can control the set of exposed CUDA devices with the `CUDA_VISIBLE_DEVICES` environment variable or the `--device` command line option. The following two commands have the same effect:
+```bash
+$ CUDA_VISIBLE_DEVICES=0 bin/ggml-rpc-server -p 50052
+$ bin/ggml-rpc-server --device CUDA0 -p 50052
+```
+
+### Main host
+
+On the main host build `llama.cpp` with the backends for the local devices and add `-DGGML_RPC=ON` to the build options.
+Finally, when running `llama-cli` or `llama-server`, use the `--rpc` option to specify the host and port of each `ggml-rpc-server`:
+
+```bash
+$ llama-cli -hf ggml-org/gemma-3-1b-it-GGUF -ngl 99 --rpc 192.168.88.10:50052,192.168.88.11:50052
+```
+
+By default, llama.cpp distributes model weights and the KV cache across all available devices -- both local and remote -- in proportion to each device's available memory.
+You can override this behavior with the `--tensor-split` option and set custom proportions when splitting tensor data across devices.
+
+### Local cache
+
+The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
+This can speed up model loading significantly, especially when using large models.
+To enable the cache, use the `-c` option:
+
+```bash
+$ bin/ggml-rpc-server -c
+```
+
+By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.
+
+### RDMA transport
+
+The RPC backend can use RDMA instead of TCP for lower latency and higher throughput. The transport is negotiated during the initial handshake -- no changes to command-line usage are required, and the connection falls back to TCP unless both peers can use RDMA.
+
+Two providers are supported, each enabled by default when its library is found at build time:
+
+- **Linux**: RoCEv2-capable NICs (e.g. Mellanox ConnectX), via `libibverbs`.
+- **macOS**: RDMA over Thunderbolt on Apple silicon Macs with Thunderbolt 5, via `librdma`. Requires macOS 26.2 or later, with RDMA enabled once from macOS Recovery via `rdma_ctl enable`. See [TN3205](https://developer.apple.com/documentation/technotes/tn3205-low-latency-communication-with-rdma-over-thunderbolt).
+
+RDMA is point-to-point, so each side uses the local device whose GID matches the address the connection was made on. Connect over the RDMA-capable link -- with Thunderbolt, use the peer's Thunderbolt address in `--rpc`; a connection made over another interface stays on TCP.
+
+To force plain TCP without rebuilding, set `GGML_RPC_NO_RDMA` on either peer:
+```bash
+$ GGML_RPC_NO_RDMA=1 bin/ggml-rpc-server
+```
+
+### Troubleshooting
+
+The `GGML_RPC_DEBUG` environment variable controls the verbosity of the logs emitted by the RPC backend.
+It can be set on the server, on the client (e.g. `llama-cli`), or both. Larger values produce more detailed output:
+
+- unset / `0` - disabled (only warnings and errors are printed)
+- `1` - high-level events: connections, handshake, buffer allocations, graph computes, cache evictions
+- `2` - per-command trace: every RPC message sent/received, tensor transfers, cache hits/misses
+- `3` - fine detail: per-command timings, graph node dumps, transport byte counts
+
+Non-numeric values are treated as `1`.
+
+```bash
+$ GGML_RPC_DEBUG=1 bin/ggml-rpc-server
+```
+
+> [!NOTE]
+> The debug levels are emitted as `ggml` debug records, which `llama.cpp` applications drop unless the
+> global verbosity threshold is raised. On the client, pass `-lv 5` together with `GGML_RPC_DEBUG`:
+
+```bash
+$ GGML_RPC_DEBUG=2 bin/llama-cli -hf ggml-org/gemma-3-1b-it-GGUF -ngl 99 --rpc 192.168.88.10:50052 -lv 5
+```
+

llama.cpp/tools/rpc/rpc-server.cpp

@@ -0,0 +1,268 @@
+#include "ggml-backend.h"
+#include "ggml-rpc.h"
+#ifdef _WIN32
+#  define NOMINMAX
+#  include <windows.h>
+#  include <fcntl.h>
+#  include <io.h>
+#else
+#  include <unistd.h>
+#endif
+#include <algorithm>
+#include <clocale>
+#include <codecvt>
+#include <filesystem>
+#include <regex>
+#include <stdio.h>
+#include <string>
+#include <thread>
+#include <vector>
+
+#if !defined(_WIN32)
+#include <sys/types.h>
+#include <pwd.h>
+#endif
+
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::string fs_path_to_utf8(const std::filesystem::path & path) {
+    const auto value = path.u8string();
+    return std::string(value.begin(), value.end());
+}
+
+// common_get_path_from_env() is adapted to avoid utf8_to_wstring
+static std::filesystem::path common_get_path_from_env(const std::string & name) {
+#ifdef _WIN32
+    std::wstring wname;
+    for (const char * p = name.c_str(); *p; ++p) {
+        wname.push_back((wchar_t)*p);
+    }
+    const wchar_t * wvalue = _wgetenv(wname.c_str());
+    return wvalue ? std::filesystem::path(wvalue) : std::filesystem::path();
+#else
+    const char * value = std::getenv(name.c_str());
+    return value ? std::filesystem::path(value) : std::filesystem::path();
+#endif
+}
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+#if !defined(_WIN32)
+static std::filesystem::path get_home_directory() {
+    std::filesystem::path home = common_get_path_from_env("HOME");
+    if (!home.empty()) {
+        return home;
+    }
+    const struct passwd * pw = getpwuid(getuid());
+    if (!pw || !pw->pw_dir || !*pw->pw_dir) {
+        throw std::runtime_error("Failed to find $HOME directory");
+    }
+    return pw->pw_dir;
+}
+#endif
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::filesystem::path fs_get_cache_directory() {
+    std::filesystem::path cache_directory = common_get_path_from_env("LLAMA_CACHE");
+    if (!cache_directory.empty()) {
+        return cache_directory;
+    }
+#if defined(_WIN32)
+    cache_directory = common_get_path_from_env("LOCALAPPDATA");
+    if (cache_directory.empty()) {
+        throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
+    }
+#elif defined(__APPLE__)
+    cache_directory = get_home_directory() / "Library/Caches";
+#else
+    cache_directory = common_get_path_from_env("XDG_CACHE_HOME");
+    if (cache_directory.empty()) {
+        cache_directory = get_home_directory() / ".cache";
+    }
+#endif
+    return cache_directory / "llama.cpp";
+}
+
+// NOTE: this is copied from common.h to avoid linking with libcommon
+static bool common_create_directories(const std::filesystem::path & path, std::error_code & ec) {
+#if defined(__linux__)
+    return std::filesystem::create_directories(path / "", ec);
+#else
+    return std::filesystem::create_directories(path, ec);
+#endif
+}
+
+struct rpc_server_params {
+    std::string              host        = "127.0.0.1";
+    int                      port        = 50052;
+    bool                     use_cache   = false;
+    int                      n_threads   = std::max(1U, std::thread::hardware_concurrency()/2);
+    std::vector<std::string> devices;
+};
+
+static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
+    fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                       show this help message and exit\n");
+    fprintf(stderr, "  -t, --threads N                  number of threads for the CPU device (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -d, --device <dev1,dev2,...>     comma-separated list of devices\n");
+    fprintf(stderr, "  -H, --host HOST                  host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p, --port PORT                  port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -c, --cache                      enable local file cache\n");
+    fprintf(stderr, "\n");
+}
+
+static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & params) {
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg == "-H" || arg == "--host") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.host = argv[i];
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
+                return false;
+            }
+        } else if (arg == "-d" || arg == "--device") {
+            if (++i >= argc) {
+                return false;
+            }
+            const std::regex regex{ R"([,/]+)" };
+            std::string dev_str = argv[i];
+            std::sregex_token_iterator iter(dev_str.begin(), dev_str.end(), regex, -1);
+            std::sregex_token_iterator end;
+            for ( ; iter != end; ++iter) {
+                try {
+                    params.devices.push_back(*iter);
+                } catch (const std::exception & ) {
+                    fprintf(stderr, "error: invalid device: %s\n", iter->str().c_str());
+                    return false;
+                }
+            }
+        } else if (arg == "-p" || arg == "--port") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.port = std::stoi(argv[i]);
+            if (params.port <= 0 || params.port > 65535) {
+                return false;
+            }
+        } else if (arg == "-c" || arg == "--cache") {
+            params.use_cache = true;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+    return true;
+}
+
+static std::vector<ggml_backend_dev_t> get_devices(const rpc_server_params & params) {
+    std::vector<ggml_backend_dev_t> devices;
+    if (!params.devices.empty()) {
+        for (auto device : params.devices) {
+            ggml_backend_dev_t dev = ggml_backend_dev_by_name(device.c_str());
+            if (dev) {
+                devices.push_back(dev);
+            } else {
+                fprintf(stderr, "error: unknown device: %s\n", 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 {};
+            }
+        }
+    }
+
+    // Try non-CPU devices first
+    if (devices.empty()) {
+        for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            enum ggml_backend_dev_type dev_type = ggml_backend_dev_type(dev);
+            if (dev_type != GGML_BACKEND_DEVICE_TYPE_CPU && dev_type != GGML_BACKEND_DEVICE_TYPE_ACCEL) {
+                devices.push_back(dev);
+            }
+        }
+    }
+
+    // If there are no accelerators, fallback to CPU device
+    if (devices.empty()) {
+        ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (dev) {
+            devices.push_back(dev);
+        }
+    }
+
+    return devices;
+}
+
+int main(int argc, char * argv[]) {
+    std::setlocale(LC_NUMERIC, "C");
+
+    ggml_backend_load_all();
+
+    rpc_server_params params;
+    if (!rpc_server_params_parse(argc, argv, params)) {
+        fprintf(stderr, "Invalid parameters\n");
+        return 1;
+    }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
+    auto devices = get_devices(params);
+    if (devices.empty()) {
+        fprintf(stderr, "No devices found\n");
+        return 1;
+    }
+    std::string endpoint = params.host + ":" + std::to_string(params.port);
+    const char * cache_dir = nullptr;
+    std::string cache_dir_str;
+    if (params.use_cache) {
+        const std::filesystem::path cache_dir_path = fs_get_cache_directory() / "rpc";
+        std::error_code ec;
+        common_create_directories(cache_dir_path, ec);
+        if (ec) {
+            fprintf(stderr, "Failed to create cache directory: %s\n", fs_path_to_utf8(cache_dir_path).c_str());
+            return 1;
+        }
+        cache_dir_str = fs_path_to_utf8(cache_dir_path);
+        cache_dir = cache_dir_str.c_str();
+    }
+
+    ggml_backend_reg_t reg = ggml_backend_reg_by_name("RPC");
+    if (!reg) {
+        fprintf(stderr, "Failed to find RPC backend\n");
+        return 1;
+    }
+
+    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(endpoint.c_str(), cache_dir, params.n_threads, devices.size(), devices.data());
+    return 0;
+}