whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h

@@ -0,0 +1,58 @@
+#pragma once
+
+// clang-format off
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#include <cstdint>
+// clang-format on
+
+// ggml_tensor is serialized into apir_rpc_tensor
+struct apir_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];
+
+    char padding[4];
+};
+
+/* frontend */
+
+apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor);
+
+void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output);
+
+/* backend */
+
+void                                      apir_track_backend_buffer(ggml_backend_buffer_t buffer);
+bool                                      apir_untrack_backend_buffer(ggml_backend_buffer_t buffer);
+std::unordered_set<ggml_backend_buffer_t> apir_get_track_backend_buffers();
+
+void apir_add_tensor(ggml_tensor *                       tensor,
+                     std::vector<apir_rpc_tensor> &      tensors,
+                     std::unordered_set<ggml_tensor *> & visited);
+
+ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor);
+
+ggml_tensor * apir_create_node(uint64_t                                                      id,
+                               ggml_context *                                                ctx,
+                               const std::unordered_map<uint64_t, const apir_rpc_tensor *> & tensor_ptrs,
+                               std::unordered_map<uint64_t, ggml_tensor *> &                 tensor_map);
+
+ggml_cgraph * apir_deserialize_graph(uint32_t                n_nodes,
+                                     uint32_t                n_tensors,
+                                     const apir_rpc_tensor * tensors,
+                                     const uint64_t *        nodes);

data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp

@@ -0,0 +1,81 @@
+#include "ggml-remoting.h"
+
+static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                            size_t                     size) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
+    if (!context) {
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
+    }
+
+    context->gpu = gpu;
+
+    bool async__unused, host_buffer__unused, events__unused;
+    bool buffer_from_host_ptr;
+    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused);
+
+    if (buffer_from_host_ptr) {
+        context->apir_context = apir_device_buffer_from_ptr(gpu, size, size);
+        context->base         = context->apir_context.shmem.mmap_ptr;
+        context->is_from_ptr  = true;
+    } else {
+        context->apir_context = apir_buffer_type_alloc_buffer(gpu, gpu->cached_buffer_type.host_handle, size);
+        context->is_from_ptr  = false;
+        context->base         = NULL;
+    }
+
+    ggml_backend_buffer_t buffer =
+        ggml_backend_buffer_init(buft, ggml_backend_remoting_buffer_interface, (void *) context, size);
+
+    return buffer;
+}
+
+static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    // Return the prefixed name that was built once during initialization
+    return gpu->cached_buffer_type.name;
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    return gpu->cached_buffer_type.alignment;
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    return gpu->cached_buffer_type.max_size;
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
+                                                               const ggml_tensor *        tensor) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    if (tensor->buffer == NULL || !tensor->buffer->context ||
+        !buft->device->iface.supports_buft(buft->device, tensor->buffer->buft)) {
+        return ggml_nbytes(tensor);
+    }
+
+    return apir_buffer_type_get_alloc_size(gpu, gpu->cached_buffer_type.host_handle, tensor);
+}
+
+const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_remoting_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_remoting_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_remoting_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_remoting_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_remoting_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_from_ptr_type_interface = {
+    /* .get_name         = */ ggml_backend_remoting_buffer_type_get_name,
+    /* .alloc_buffer     = */ NULL,
+    /* .get_alignment    = */ ggml_backend_remoting_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_remoting_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_remoting_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};

data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp

@@ -0,0 +1,119 @@
+#include "ggml-remoting.h"
+
+#define BUFFER_TO_GPU(name) ((ggml_backend_remoting_buffer_context *) (name)->context)->gpu
+
+static void * ggml_backend_remoting_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) buffer->context;
+    if (context->base) {
+        return context->base;
+    }
+
+    context->base = apir_buffer_get_base(BUFFER_TO_GPU(buffer), BUFFER_TO_APIR_CONTEXT(buffer));
+
+    return context->base;
+}
+
+static void ggml_backend_remoting_buffer_set_tensor(ggml_backend_buffer_t buffer,
+                                                    ggml_tensor *         tensor,
+                                                    const void *          data,
+                                                    size_t                offset,
+                                                    size_t                size) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    if (context->is_from_ptr) {
+        memcpy((char *) tensor->data + offset, data, size);
+    } else {
+        apir_buffer_set_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size);
+    }
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_get_tensor(ggml_backend_buffer_t buffer,
+                                                    const ggml_tensor *   tensor,
+                                                    void *                data,
+                                                    size_t                offset,
+                                                    size_t                size) {
+    virtgpu *                              gpu     = BUFFER_TO_GPU(buffer);
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    if (context->is_from_ptr) {
+        memcpy(data, (const char *) tensor->data + offset, size);
+    } else {
+        apir_buffer_get_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size);
+    }
+}
+
+static void ggml_backend_remoting_buffer_set_tensor_from_ptr(ggml_backend_buffer_t buffer,
+                                                             ggml_tensor *         tensor,
+                                                             const void *          data,
+                                                             size_t                offset,
+                                                             size_t                size) {
+    UNUSED(buffer);
+
+    memcpy((char *) tensor->data + offset, data, size);
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_get_tensor_from_ptr(ggml_backend_buffer_t buffer,
+                                                             const ggml_tensor *   tensor,
+                                                             void *                data,
+                                                             size_t                offset,
+                                                             size_t                size) {
+    UNUSED(buffer);
+
+    memcpy(data, (const char *) tensor->data + offset, size);
+}
+
+static bool ggml_backend_remoting_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
+                                                    const ggml_tensor *   src,
+                                                    ggml_tensor *         dst) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    bool ret = apir_buffer_cpy_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), src, dst);
+
+    return ret;
+}
+
+static void ggml_backend_remoting_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    apir_buffer_clear(gpu, BUFFER_TO_APIR_CONTEXT(buffer), value);
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    apir_buffer_free_buffer(gpu, BUFFER_TO_APIR_CONTEXT(buffer));
+
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    free(context);
+    buffer->context = NULL;
+}
+
+const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_remoting_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_remoting_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_remoting_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+const ggml_backend_buffer_i ggml_backend_remoting_buffer_from_ptr_interface = {
+    /* .free_buffer     = */ ggml_backend_remoting_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_remoting_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor_from_ptr,
+    /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor_from_ptr,
+    /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_remoting_buffer_clear,
+    /* .reset           = */ NULL,
+};

data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp

@@ -0,0 +1,158 @@
+#include "ggml-remoting.h"
+
+static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    // Return the prefixed name that was built once during initialization
+    return gpu->cached_device_info.name;
+}
+
+static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    // Return the pre-cached description from the virtgpu structure
+    return gpu->cached_device_info.description;
+}
+
+static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return (enum ggml_backend_dev_type) gpu->cached_device_info.type;
+}
+
+static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    *free  = gpu->cached_device_info.memory_free;
+    *total = gpu->cached_device_info.memory_total;
+}
+
+static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+#if USE_ALWAYS_TRUE_SUPPORTS_OP == 1
+    /* ggml-rpc cheats it like this */
+    /* with the current implementation of serialize_tensor, the src/view aren't properly passed */
+    UNUSED(dev);
+    UNUSED(op);
+
+    return true;
+#else
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return apir_device_supports_op(gpu, op);
+#endif
+}
+
+static bool ggml_backend_remoting_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    bool supported = buft->device == dev;
+
+    return supported;
+}
+
+static bool ggml_backend_remoting_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    UNUSED(dev);
+    UNUSED(op);
+
+    return false;
+}
+
+static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_remoting_device_get_name(dev);
+    props->description = ggml_backend_remoting_device_get_description(dev);
+    props->type        = ggml_backend_remoting_device_get_type(dev);
+    ggml_backend_remoting_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    virtgpu * gpu = DEV_TO_GPU(dev);
+    apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr,
+                          &props->caps.events);
+
+    props->caps.buffer_from_host_ptr = false;
+    props->caps.async                = false;
+    props->caps.events               = false;
+}
+
+ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    static std::atomic<bool>        initialized = false;
+    static ggml_backend_buffer_type buft;
+
+    if (!initialized) {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (!initialized) {
+            buft = {
+                /* .iface    = */ ggml_backend_remoting_buffer_type_interface,
+                /* .device   = */ dev,
+                /* .context  = */ (void *) gpu->cached_buffer_type.host_handle,
+            };
+            initialized = true;
+        }
+    }
+
+    return &buft;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    static std::atomic<bool>        initialized = false;
+    static ggml_backend_buffer_type buft;
+
+    if (!initialized) {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (!initialized) {
+            buft = {
+                /* .iface    = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
+                /* .device   = */ dev,
+                /* .context  = */ (void *) gpu->cached_buffer_type.host_handle,
+            };
+            initialized = true;
+        }
+    }
+
+    return &buft;
+}
+
+static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_backend_dev_t dev,
+                                                                          void *             ptr,
+                                                                          size_t             size,
+                                                                          size_t             max_tensor_size) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
+    if (!context) {
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
+    }
+
+    context->gpu          = gpu;
+    context->apir_context = apir_device_buffer_from_ptr(gpu, size, max_tensor_size);
+    context->base         = ptr;
+    context->is_from_ptr  = true;
+
+    ggml_backend_buffer_t buffer =
+        ggml_backend_buffer_init(ggml_backend_remoting_device_get_buffer_from_ptr_type(dev),
+                                 ggml_backend_remoting_buffer_from_ptr_interface, (void *) context, size);
+
+    return buffer;
+}
+
+const ggml_backend_device_i ggml_backend_remoting_device_interface = {
+    /* .get_name             = */ ggml_backend_remoting_device_get_name,
+    /* .get_description      = */ ggml_backend_remoting_device_get_description,
+    /* .get_memory           = */ ggml_backend_remoting_device_get_memory,
+    /* .get_type             = */ ggml_backend_remoting_device_get_type,
+    /* .get_props            = */ ggml_backend_remoting_device_get_props,
+    /* .init_backend         = */ ggml_backend_remoting_device_init,
+    /* .get_buffer_type      = */ ggml_backend_remoting_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_remoting_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_remoting_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_remoting_device_supports_buft,
+    /* .offload_op           = */ ggml_backend_remoting_device_offload_op,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};

data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp

@@ -0,0 +1,213 @@
+#include "ggml-remoting.h"
+#include "ggml-virtgpu.h"
+
+#include <iostream>
+#include <mutex>
+
+void ggml_virtgpu_cleanup(virtgpu * gpu);
+
+static virtgpu * apir_initialize() {
+    static virtgpu *         gpu         = NULL;
+    static std::atomic<bool> initialized = false;
+
+    if (initialized) {
+        // fast track
+        return gpu;
+    }
+
+    {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (initialized) {
+            // thread safe
+            return gpu;
+        }
+
+        gpu = create_virtgpu();
+        if (!gpu) {
+            initialized = true;
+            return NULL;
+        }
+
+        // Pre-fetch and cache all device information, it will not change
+        gpu->cached_device_info.description = apir_device_get_description(gpu);
+        if (!gpu->cached_device_info.description) {
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device description", __func__);
+        }
+        gpu->cached_device_info.device_count = apir_device_get_count(gpu);
+        gpu->cached_device_info.type         = apir_device_get_type(gpu);
+
+        {
+            // Get the remote name and create prefixed version
+            char * rmt_device_name = apir_device_get_name(gpu);
+            if (!rmt_device_name) {
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to get the virtgpu device name", __func__);
+            }
+
+            size_t device_name_len       = strlen(rmt_device_name) + 11;  // "[virtgpu] " + null terminator
+            gpu->cached_device_info.name = (char *) malloc(device_name_len);
+            if (!gpu->cached_device_info.name) {
+                free(rmt_device_name);
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to allocate memory for prefixed device name", __func__);
+            }
+            snprintf(gpu->cached_device_info.name, device_name_len, "[virtgpu] %s", rmt_device_name);
+            free(rmt_device_name);
+        }
+
+        apir_device_get_memory(gpu, &gpu->cached_device_info.memory_free, &gpu->cached_device_info.memory_total);
+
+        apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
+        gpu->cached_buffer_type.host_handle             = buft_host_handle;
+        {
+            // Get the remote name and create prefixed version
+            char * rmt_name = apir_buffer_type_get_name(gpu, buft_host_handle);
+            if (!rmt_name) {
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to get the virtgpu buffer type name", __func__);
+            }
+
+            size_t prefixed_len          = strlen(rmt_name) + 11;  // "[virtgpu] " + null terminator
+            gpu->cached_buffer_type.name = (char *) malloc(prefixed_len);
+            if (!gpu->cached_buffer_type.name) {
+                free(rmt_name);
+                GGML_ABORT(GGML_VIRTGPU "%s: failed to allocate memory for prefixed buffer type name", __func__);
+            }
+            snprintf(gpu->cached_buffer_type.name, prefixed_len, "[virtgpu] %s", rmt_name);
+            free(rmt_name);
+        }
+
+        gpu->cached_buffer_type.alignment = apir_buffer_type_get_alignment(gpu, buft_host_handle);
+        gpu->cached_buffer_type.max_size  = apir_buffer_type_get_max_size(gpu, buft_host_handle);
+
+        initialized = true;
+    }
+
+    return gpu;
+}
+
+static int ggml_backend_remoting_get_device_count() {
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        return 0;
+    }
+
+    return gpu->cached_device_info.device_count;
+}
+
+static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+
+    return ggml_backend_remoting_get_device_count();
+}
+
+static std::vector<ggml_backend_dev_t> devices;
+
+ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) {
+    GGML_ASSERT(device < devices.size());
+    return devices[device];
+}
+
+static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
+    if (devices.size() > 0) {
+        GGML_LOG_INFO(GGML_VIRTGPU "%s: already initialized\n", __func__);
+        return;
+    }
+
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: apir_initialize failed\n", __func__);
+        return;
+    }
+
+    static std::atomic<bool> initialized = false;
+
+    if (initialized) {
+        return;  // fast track
+    }
+
+    {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) {
+                ggml_backend_remoting_device_context * ctx       = new ggml_backend_remoting_device_context;
+                char                                   desc[256] = "ggml-virtgpu API Remoting device";
+
+                ctx->device      = i;
+                ctx->name        = GGML_VIRTGPU_NAME + std::to_string(i);
+                ctx->description = desc;
+                ctx->gpu         = gpu;
+
+                ggml_backend_dev_t dev = new ggml_backend_device{
+                    /* .iface   = */ ggml_backend_remoting_device_interface,
+                    /* .reg     = */ reg,
+                    /* .context = */ ctx,
+                };
+                devices.push_back(dev);
+            }
+            initialized = true;
+        }
+    }
+}
+
+static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_t reg, size_t device) {
+    UNUSED(reg);
+
+    return ggml_backend_remoting_get_device(device);
+}
+
+static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+
+    return GGML_VIRTGPU_NAME;
+}
+
+static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
+    /* .get_name         = */ ggml_backend_remoting_reg_get_name,
+    /* .get_device_count = */ ggml_backend_remoting_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_remoting_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
+
+ggml_backend_reg_t ggml_backend_virtgpu_reg() {
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_apir_initialize failed\n", __func__);
+    }
+
+    static ggml_backend_reg reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_remoting_reg_i,
+        /* .context     = */ gpu,
+    };
+
+    static bool initialized = false;
+    if (initialized) {
+        return &reg;
+    }
+    initialized = true;
+
+    ggml_backend_remoting_reg_init_devices(&reg);
+
+    return &reg;
+}
+
+// public function, not exposed in the GGML interface at the moment
+void ggml_virtgpu_cleanup(virtgpu * gpu) {
+    if (gpu->cached_device_info.name) {
+        free(gpu->cached_device_info.name);
+        gpu->cached_device_info.name = NULL;
+    }
+    if (gpu->cached_device_info.description) {
+        free(gpu->cached_device_info.description);
+        gpu->cached_device_info.description = NULL;
+    }
+    if (gpu->cached_buffer_type.name) {
+        free(gpu->cached_buffer_type.name);
+        gpu->cached_buffer_type.name = NULL;
+    }
+
+    mtx_destroy(&gpu->data_shmem_mutex);
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)