whisper.rn 0.4.0-rc.10 → 0.4.0-rc.11

package/cpp/ggml-backend-reg.cpp

@@ -1,25 +1,43 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
+#include <algorithm>
+#include <codecvt>
 #include <cstring>
+#include <filesystem>
+#include <locale>
+#include <memory>
+#include <string>
+#include <type_traits>
 #include <vector>
 
+#ifdef _WIN32
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#elif defined(__APPLE__)
+#    include <mach-o/dyld.h>
+#    include <dlfcn.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+
 // Backend registry
+#ifdef WSP_GGML_USE_CPU
+#include "ggml-cpu.h"
+#endif
 
 #ifdef WSP_GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
 #ifdef WSP_GGML_USE_METAL
-#include <TargetConditionals.h>
-
-#if !TARGET_OS_SIMULATOR
 #include "ggml-metal.h"
 #endif
 
-#endif
-
 #ifdef WSP_GGML_USE_SYCL
 #include "ggml-sycl.h"
 #endif
@@ -28,6 +46,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef WSP_GGML_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
 #ifdef WSP_GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -36,10 +58,6 @@
 #include "ggml-rpc.h"
 #endif
 
-#ifdef WSP_GGML_USE_AMX
-#  include "ggml-amx.h"
-#endif
-
 #ifdef WSP_GGML_USE_CANN
 #include "ggml-cann.h"
 #endif
@@ -48,8 +66,90 @@
 #include "ggml-kompute.h"
 #endif
 
+// disable C++17 deprecation warning for std::codecvt_utf8
+#if defined(__clang__)
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
+static std::wstring utf8_to_utf16(const std::string & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.from_bytes(str);
+}
+
+static std::string utf16_to_utf8(const std::wstring & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.to_bytes(str);
+}
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static dl_handle * dl_load_library(const std::wstring & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static void * dl_load_library(const std::wstring & path) {
+    dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
+struct wsp_ggml_backend_reg_entry {
+    wsp_ggml_backend_reg_t reg;
+    dl_handle_ptr handle;
+};
+
 struct wsp_ggml_backend_registry {
-    std::vector<wsp_ggml_backend_reg_t> backends;
+    std::vector<wsp_ggml_backend_reg_entry> backends;
     std::vector<wsp_ggml_backend_dev_t> devices;
 
     wsp_ggml_backend_registry() {
@@ -57,18 +157,17 @@ struct wsp_ggml_backend_registry {
         register_backend(wsp_ggml_backend_cuda_reg());
 #endif
 #ifdef WSP_GGML_USE_METAL
-
-#if !TARGET_OS_SIMULATOR
         register_backend(wsp_ggml_backend_metal_reg());
 #endif
-
-#endif
 #ifdef WSP_GGML_USE_SYCL
         register_backend(wsp_ggml_backend_sycl_reg());
 #endif
 #ifdef WSP_GGML_USE_VULKAN
         register_backend(wsp_ggml_backend_vk_reg());
 #endif
+#ifdef WSP_GGML_USE_OPENCL
+        register_backend(wsp_ggml_backend_opencl_reg());
+#endif
 #ifdef WSP_GGML_USE_CANN
         register_backend(wsp_ggml_backend_cann_reg());
 #endif
@@ -78,17 +177,25 @@ struct wsp_ggml_backend_registry {
 #ifdef WSP_GGML_USE_RPC
         register_backend(wsp_ggml_backend_rpc_reg());
 #endif
-#ifdef WSP_GGML_USE_AMX
-        register_backend(wsp_ggml_backend_amx_reg());
-#endif
 #ifdef WSP_GGML_USE_KOMPUTE
         register_backend(wsp_ggml_backend_kompute_reg());
 #endif
-
+#ifdef WSP_GGML_USE_CPU
         register_backend(wsp_ggml_backend_cpu_reg());
+#endif
+    }
+
+    ~wsp_ggml_backend_registry() {
+        // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
+        // since backend threads may still be running and accessing resources from the dynamic library
+        for (auto & entry : backends) {
+            if (entry.handle) {
+                entry.handle.release(); // NOLINT
+            }
+        }
     }
 
-    void register_backend(wsp_ggml_backend_reg_t reg) {
+    void register_backend(wsp_ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -97,7 +204,7 @@ struct wsp_ggml_backend_registry {
         WSP_GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, wsp_ggml_backend_reg_name(reg), wsp_ggml_backend_reg_dev_count(reg));
 #endif
-        backends.push_back(reg);
+        backends.push_back({ reg, std::move(handle) });
         for (size_t i = 0; i < wsp_ggml_backend_reg_dev_count(reg); i++) {
             register_device(wsp_ggml_backend_reg_dev_get(reg, i));
         }
@@ -109,6 +216,76 @@ struct wsp_ggml_backend_registry {
 #endif
         devices.push_back(device);
     }
+
+    wsp_ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
+        dl_handle_ptr handle { dl_load_library(path) };
+        if (!handle) {
+            if (!silent) {
+                WSP_GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
+            }
+            return nullptr;
+        }
+
+        auto score_fn = (wsp_ggml_backend_score_t) dl_get_sym(handle.get(), "wsp_ggml_backend_score");
+        if (score_fn && score_fn() == 0) {
+            if (!silent) {
+                WSP_GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
+            }
+            return nullptr;
+        }
+
+        auto backend_init_fn = (wsp_ggml_backend_init_t) dl_get_sym(handle.get(), "wsp_ggml_backend_init");
+        if (!backend_init_fn) {
+            if (!silent) {
+                WSP_GGML_LOG_ERROR("%s: failed to find wsp_ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
+            }
+            return nullptr;
+        }
+
+        wsp_ggml_backend_reg_t reg = backend_init_fn();
+        if (!reg || reg->api_version != WSP_GGML_BACKEND_API_VERSION) {
+            if (!silent) {
+                if (!reg) {
+                    WSP_GGML_LOG_ERROR("%s: failed to initialize backend from %s: wsp_ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
+                } else {
+                    WSP_GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
+                        __func__, utf16_to_utf8(path).c_str(), reg->api_version, WSP_GGML_BACKEND_API_VERSION);
+                }
+            }
+            return nullptr;
+        }
+
+        WSP_GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, wsp_ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
+
+        register_backend(reg, std::move(handle));
+
+        return reg;
+    }
+
+    void unload_backend(wsp_ggml_backend_reg_t reg, bool silent) {
+        auto it = std::find_if(backends.begin(), backends.end(),
+                               [reg](const wsp_ggml_backend_reg_entry & entry) { return entry.reg == reg; });
+
+        if (it == backends.end()) {
+            if (!silent) {
+                WSP_GGML_LOG_ERROR("%s: backend not found\n", __func__);
+            }
+            return;
+        }
+
+        if (!silent) {
+            WSP_GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, wsp_ggml_backend_reg_name(reg));
+        }
+
+        // remove devices
+        devices.erase(
+            std::remove_if(devices.begin(), devices.end(),
+                            [reg](wsp_ggml_backend_dev_t dev) { return wsp_ggml_backend_dev_backend_reg(dev) == reg; }),
+            devices.end());
+
+        // remove backend
+        backends.erase(it);
+    }
 };
 
 static wsp_ggml_backend_registry & get_reg() {
@@ -126,23 +303,32 @@ void wsp_ggml_backend_device_register(wsp_ggml_backend_dev_t device) {
 }
 
 // Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
 size_t wsp_ggml_backend_reg_count() {
     return get_reg().backends.size();
 }
 
 wsp_ggml_backend_reg_t wsp_ggml_backend_reg_get(size_t index) {
     WSP_GGML_ASSERT(index < wsp_ggml_backend_reg_count());
-    return get_reg().backends[index];
+    return get_reg().backends[index].reg;
 }
 
 wsp_ggml_backend_reg_t wsp_ggml_backend_reg_by_name(const char * name) {
     for (size_t i = 0; i < wsp_ggml_backend_reg_count(); i++) {
         wsp_ggml_backend_reg_t reg = wsp_ggml_backend_reg_get(i);
-        if (std::strcmp(wsp_ggml_backend_reg_name(reg), name) == 0) {
+        if (striequals(wsp_ggml_backend_reg_name(reg), name)) {
             return reg;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Device enumeration
@@ -158,11 +344,11 @@ wsp_ggml_backend_dev_t wsp_ggml_backend_dev_get(size_t index) {
 wsp_ggml_backend_dev_t wsp_ggml_backend_dev_by_name(const char * name) {
     for (size_t i = 0; i < wsp_ggml_backend_dev_count(); i++) {
         wsp_ggml_backend_dev_t dev = wsp_ggml_backend_dev_get(i);
-        if (strcmp(wsp_ggml_backend_dev_name(dev), name) == 0) {
+        if (striequals(wsp_ggml_backend_dev_name(dev), name)) {
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 wsp_ggml_backend_dev_t wsp_ggml_backend_dev_by_type(enum wsp_ggml_backend_dev_type type) {
@@ -172,14 +358,14 @@ wsp_ggml_backend_dev_t wsp_ggml_backend_dev_by_type(enum wsp_ggml_backend_dev_ty
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Convenience functions
 wsp_ggml_backend_t wsp_ggml_backend_init_by_name(const char * name, const char * params) {
     wsp_ggml_backend_dev_t dev = wsp_ggml_backend_dev_by_name(name);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return wsp_ggml_backend_dev_init(dev, params);
 }
@@ -187,7 +373,7 @@ wsp_ggml_backend_t wsp_ggml_backend_init_by_name(const char * name, const char *
 wsp_ggml_backend_t wsp_ggml_backend_init_by_type(enum wsp_ggml_backend_dev_type type, const char * params) {
     wsp_ggml_backend_dev_t dev = wsp_ggml_backend_dev_by_type(type);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return wsp_ggml_backend_dev_init(dev, params);
 }
@@ -198,7 +384,199 @@ wsp_ggml_backend_t wsp_ggml_backend_init_best(void) {
         dev = wsp_ggml_backend_dev_by_type(WSP_GGML_BACKEND_DEVICE_TYPE_CPU);
     }
     if (!dev) {
-        return NULL;
+        return nullptr;
+    }
+    return wsp_ggml_backend_dev_init(dev, nullptr);
+}
+
+// Dynamic loading
+wsp_ggml_backend_reg_t wsp_ggml_backend_load(const char * path) {
+    return get_reg().load_backend(utf8_to_utf16(path), false);
+}
+
+void wsp_ggml_backend_unload(wsp_ggml_backend_reg_t reg) {
+    get_reg().unload_backend(reg, true);
+}
+
+static std::wstring get_executable_path() {
+#if defined(__APPLE__)
+    // get executable path
+    std::vector<char> path;
+    uint32_t size;
+    while (true) {
+        size = path.size();
+        if (_NSGetExecutablePath(path.data(), &size) == 0) {
+            break;
+        }
+        path.resize(size);
+    }
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return utf8_to_utf16(base_path + "/");
+#elif defined(__linux__) || defined(__FreeBSD__)
+    std::string base_path = ".";
+    std::vector<char> path(1024);
+    while (true) {
+        // get executable path
+#    if defined(__linux__)
+        ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+#    elif defined(__FreeBSD__)
+        ssize_t len = readlink("/proc/curproc/file", path.data(), path.size());
+#    endif
+        if (len == -1) {
+            break;
+        }
+        if (len < (ssize_t) path.size()) {
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
+        }
+        path.resize(path.size() * 2);
+    }
+
+    return utf8_to_utf16(base_path + "/");
+#elif defined(_WIN32)
+    std::vector<wchar_t> path(MAX_PATH);
+    DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
+    if (len == 0) {
+        return {};
+    }
+    std::wstring base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + L"\\";
+#else
+    return {};
+#endif
+}
+
+static std::wstring backend_filename_prefix() {
+#ifdef _WIN32
+    return L"ggml-";
+#else
+    return L"libggml-";
+#endif
+}
+
+static std::wstring backend_filename_suffix() {
+#ifdef _WIN32
+    return L".dll";
+#else
+    return L".so";
+#endif
+}
+
+static std::wstring path_separator() {
+#ifdef _WIN32
+    return L"\\";
+#else
+    return L"/";
+#endif
+}
+
+static wsp_ggml_backend_reg_t wsp_ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
+    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
+     // TODO: search system paths
+    std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
+    std::vector<std::wstring> search_paths;
+    if (user_search_path == nullptr) {
+        search_paths.push_back(L"." + path_separator());
+        search_paths.push_back(get_executable_path());
+    } else {
+        search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
+    }
+
+    int best_score = 0;
+    std::wstring best_path;
+
+    namespace fs = std::filesystem;
+    for (const auto & search_path : search_paths) {
+        if (!fs::exists(search_path)) {
+            continue;
+        }
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
+        for (const auto & entry : dir_it) {
+            if (entry.is_regular_file()) {
+                std::wstring filename = entry.path().filename().wstring();
+                std::wstring ext = entry.path().extension().wstring();
+                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
+                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
+                    if (!handle && !silent) {
+                        WSP_GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
+                    }
+                    if (handle) {
+                        auto score_fn = (wsp_ggml_backend_score_t) dl_get_sym(handle.get(), "wsp_ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            WSP_GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().wstring();
+                            }
+                        } else {
+                            if (!silent) {
+                                WSP_GGML_LOG_INFO("%s: failed to find wsp_ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    if (best_score == 0) {
+        // try to load the base backend
+        for (const auto & search_path : search_paths) {
+            std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
+            if (fs::exists(path)) {
+                return get_reg().load_backend(path, silent);
+            }
+        }
+        return nullptr;
+    }
+
+    return get_reg().load_backend(best_path, silent);
+}
+
+void wsp_ggml_backend_load_all() {
+    wsp_ggml_backend_load_all_from_path(nullptr);
+}
+
+void wsp_ggml_backend_load_all_from_path(const char * dir_path) {
+#ifdef NDEBUG
+    bool silent = true;
+#else
+    bool silent = false;
+#endif
+
+    wsp_ggml_backend_load_best("blas", silent, dir_path);
+    wsp_ggml_backend_load_best("cann", silent, dir_path);
+    wsp_ggml_backend_load_best("cuda", silent, dir_path);
+    wsp_ggml_backend_load_best("hip", silent, dir_path);
+    wsp_ggml_backend_load_best("kompute", silent, dir_path);
+    wsp_ggml_backend_load_best("metal", silent, dir_path);
+    wsp_ggml_backend_load_best("rpc", silent, dir_path);
+    wsp_ggml_backend_load_best("sycl", silent, dir_path);
+    wsp_ggml_backend_load_best("vulkan", silent, dir_path);
+    wsp_ggml_backend_load_best("opencl", silent, dir_path);
+    wsp_ggml_backend_load_best("musa", silent, dir_path);
+    wsp_ggml_backend_load_best("cpu", silent, dir_path);
+    // check the environment variable WSP_GGML_BACKEND_PATH to load an out-of-tree backend
+    const char * backend_path = std::getenv("WSP_GGML_BACKEND_PATH");
+    if (backend_path) {
+        wsp_ggml_backend_load(backend_path);
     }
-    return wsp_ggml_backend_dev_init(dev, NULL);
 }

package/cpp/ggml-backend.cpp

@@ -252,6 +252,7 @@ void wsp_ggml_backend_tensor_get_async(wsp_ggml_backend_t backend, const struct
 }
 
 void wsp_ggml_backend_tensor_set(struct wsp_ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    WSP_GGML_ASSERT(tensor);
     wsp_ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -266,6 +267,7 @@ void wsp_ggml_backend_tensor_set(struct wsp_ggml_tensor * tensor, const void * d
 }
 
 void wsp_ggml_backend_tensor_get(const struct wsp_ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    WSP_GGML_ASSERT(tensor);
     wsp_ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -740,7 +742,8 @@ static int wsp_ggml_backend_sched_backend_id_from_cur(wsp_ggml_backend_sched_t s
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        WSP_GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
+        wsp_ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+        WSP_GGML_ABORT("pre-allocated tensor (%s) in a buffer (%s) that cannot run the operation (%s)", tensor->name, wsp_ggml_backend_buffer_name(buffer), wsp_ggml_op_name(tensor->op));
     }
 
     // graph input
@@ -761,7 +764,7 @@ static int wsp_ggml_backend_sched_backend_id_from_cur(wsp_ggml_backend_sched_t s
         if (tensor->op != WSP_GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == WSP_GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = wsp_ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1) {
+            if (src_backend_id == sched->n_backends - 1 && wsp_ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (wsp_ggml_backend_supports_op(sched->backends[b], tensor) && wsp_ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -792,9 +795,12 @@ static void wsp_ggml_backend_sched_print_assignments(wsp_ggml_backend_sched_t sc
     for (int i = 0; i < graph->n_nodes; i++) {
         if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
             wsp_ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
-            WSP_GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs: ", cur_split, wsp_ggml_backend_name(split_backend),
+            WSP_GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs", cur_split, wsp_ggml_backend_name(split_backend),
                 sched->splits[cur_split].n_inputs);
             for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                if (j == 0) {
+                    WSP_GGML_LOG_DEBUG(": ");
+                }
                 WSP_GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
                     fmt_size(wsp_ggml_nbytes(sched->splits[cur_split].inputs[j])));
             }

package/cpp/ggml-backend.h

@@ -190,11 +190,21 @@ extern "C" {
     typedef void                         (*wsp_ggml_backend_set_n_threads_t)(wsp_ggml_backend_t backend, int n_threads);
     // Get additional buffer types provided by the device (returns a NULL-terminated array)
     typedef wsp_ggml_backend_buffer_type_t * (*wsp_ggml_backend_dev_get_extra_bufts_t)(wsp_ggml_backend_dev_t device);
+    // Set the abort callback for the backend
+    typedef void                         (*wsp_ggml_backend_set_abort_callback_t)(wsp_ggml_backend_t backend, wsp_ggml_abort_callback abort_callback, void * abort_callback_data);
+    // Get a list of feature flags supported by the backend (returns a NULL-terminated array)
+    struct wsp_ggml_backend_feature {
+        const char * name;
+        const char * value;
+    };
+    typedef struct wsp_ggml_backend_feature * (*wsp_ggml_backend_get_features_t)(wsp_ggml_backend_reg_t reg);
 
     //
     // Backend registry
     //
 
+    WSP_GGML_API void wsp_ggml_backend_device_register(wsp_ggml_backend_dev_t device);
+
     // Backend (reg) enumeration
     WSP_GGML_API size_t             wsp_ggml_backend_reg_count(void);
     WSP_GGML_API wsp_ggml_backend_reg_t wsp_ggml_backend_reg_get(size_t index);
@@ -214,6 +224,14 @@ extern "C" {
     // = wsp_ggml_backend_dev_init(wsp_ggml_backend_dev_by_type(GPU) OR wsp_ggml_backend_dev_by_type(CPU), NULL)
     WSP_GGML_API wsp_ggml_backend_t wsp_ggml_backend_init_best(void);
 
+    // Load a backend from a dynamic library and register it
+    WSP_GGML_API wsp_ggml_backend_reg_t wsp_ggml_backend_load(const char * path);
+    // Unload a backend if loaded dynamically and unregister it
+    WSP_GGML_API void               wsp_ggml_backend_unload(wsp_ggml_backend_reg_t reg);
+    // Load all known backends from dynamic libraries
+    WSP_GGML_API void               wsp_ggml_backend_load_all(void);
+    WSP_GGML_API void               wsp_ggml_backend_load_all_from_path(const char * dir_path);
+
     //
     // Backend scheduler
     //