ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)

2026-01-28 17:49:40 +08:00
parent 6ad70c5a77
commit b7feacf7f3
47 changed files with 4711 additions and 0 deletions
@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.19)
+cmake_policy(SET CMP0114 NEW)
+
+message(STATUS "Enable the VirtGPU/Virglrenderer backend library")
+
+ggml_add_backend_library(ggml-virtgpu-backend
+                         backend.cpp
+                         backend-dispatched.cpp
+                         backend-dispatched-backend.cpp
+                         backend-dispatched-device.cpp
+                         backend-dispatched-buffer.cpp
+                         backend-dispatched-buffer-type.cpp
+                         shared/api_remoting.h
+                         shared/apir_backend.h
+                         shared/apir_cs.h
+                         apir_cs_ggml-rpc-back.cpp)
+
+target_compile_options(ggml-virtgpu-backend PRIVATE -std=c++20)
+
+# Add include directory for ggml-backend-impl.h and other core headers
+target_include_directories(ggml-virtgpu-backend PRIVATE ../..)
@@ -0,0 +1,115 @@
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "shared/apir_cs_rpc.h"
+
+#include <cinttypes>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+std::unordered_set<ggml_backend_buffer_t> backend_buffers;
+
+void apir_track_backend_buffer(ggml_backend_buffer_t buffer) {
+    backend_buffers.insert(buffer);
+}
+
+bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer) {
+    auto it = backend_buffers.find(buffer);
+    if (it == backend_buffers.end()) {
+        return false;
+    }
+
+    backend_buffers.erase(it);
+    return true;
+}
+
+std::unordered_set<ggml_backend_buffer_t> apir_get_track_backend_buffers() {
+    return backend_buffers;
+}
+
+ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor) {
+    ggml_tensor * result =
+        ggml_new_tensor_4d(ctx, (ggml_type) tensor->type, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = tensor->nb[i];
+    }
+    result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
+    if (result->buffer && backend_buffers.find(result->buffer) == backend_buffers.end()) {
+        printf("WARNING: HOST BUFFER NOT FOUND | %p\n", (void *) result->buffer);
+        result->buffer = nullptr;
+    }
+
+    uint64_t tensor_data = tensor->data;
+    if (result->buffer) {
+        // require that the tensor data does not go beyond the buffer end
+        uint64_t tensor_size  = (uint64_t) ggml_nbytes(result);
+        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+        uint64_t buffer_size  = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+
+        // tensor->data is serialized as an offset to the buffer base address
+        tensor_data += buffer_start;
+
+        GGML_ASSERT(tensor_data + tensor_size >= tensor_data);  // check for overflow
+        GGML_ASSERT(tensor_data >= buffer_start && tensor_data + tensor_size <= buffer_start + buffer_size);
+    }
+
+    result->op = (ggml_op) tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result->op_params[i] = tensor->op_params[i];
+    }
+    result->flags = tensor->flags;
+    result->data  = reinterpret_cast<void *>(tensor_data);
+    ggml_set_name(result, tensor->name);
+    return result;
+}
+
+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) {
+    if (id == 0) {
+        return nullptr;
+    }
+    if (tensor_map.find(id) != tensor_map.end()) {
+        return tensor_map[id];
+    }
+    const apir_rpc_tensor * tensor = tensor_ptrs.at(id);
+    ggml_tensor *           result = apir_deserialize_tensor(ctx, tensor);
+    if (result == nullptr) {
+        return nullptr;
+    }
+    tensor_map[id] = result;
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        result->src[i] = apir_create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+    }
+    result->view_src  = apir_create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+    result->view_offs = tensor->view_offs;
+    return result;
+}
+
+ggml_cgraph * apir_deserialize_graph(uint32_t                n_nodes,
+                                     uint32_t                n_tensors,
+                                     const apir_rpc_tensor * tensors,
+                                     const uint64_t *        nodes) {
+    size_t buf_size = ggml_tensor_overhead() * (n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+    ggml_init_params params = {
+        /*.mem_size   =*/buf_size,
+        /*.mem_buffer =*/NULL,
+        /*.no_alloc   =*/true,
+    };
+    ggml_context * ctx   = ggml_init(params);
+    ggml_cgraph *  graph = ggml_new_graph_custom(ctx, n_nodes, false);
+    graph->n_nodes       = n_nodes;
+    std::unordered_map<uint64_t, const apir_rpc_tensor *> tensor_ptrs;
+    for (uint32_t i = 0; i < n_tensors; i++) {
+        tensor_ptrs[tensors[i].id] = &tensors[i];
+    }
+    std::unordered_map<uint64_t, ggml_tensor *> tensor_map;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        int64_t id;
+        memcpy(&id, &nodes[i], sizeof(id));
+        graph->nodes[i] = apir_create_node(id, ctx, tensor_ptrs, tensor_map);
+    }
+
+    return graph;
+}
@@ -0,0 +1,13 @@
+#include "shared/apir_backend.h"
+
+#define BUFFER_TO_HOST_HANDLE(name) ggml_buffer_to_apir_handle(name)
+
+static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_host_handle_t) buffer;
+}
+
+static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_type_host_handle_t) buft;
+}
@@ -0,0 +1,65 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "shared/apir_backend.h"
+
+#include <cstdint>
+
+uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    static bool async_backend_initialized = false;
+    static bool async_backend;
+
+    if (!async_backend_initialized) {
+        ggml_backend_dev_props props;
+
+        dev->iface.get_props(dev, &props);
+        async_backend             = props.caps.async;
+        async_backend_initialized = true;
+    }
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+    size_t cgraph_size;
+    apir_decode_size_t(dec, &cgraph_size);
+
+    apir_decoder secondary_dec = apir_new_decoder((const char *) shmem_data, cgraph_size);
+
+    ggml_cgraph * cgraph = apir_decode_ggml_cgraph(&secondary_dec, cgraph_size);
+
+    ggml_status status;
+#if APIR_BACKEND_CHECK_SUPPORTS_OP == 1
+    for (int idx = 0; idx < cgraph->n_nodes; idx++) {
+        ggml_tensor * op = ggml_graph_node(cgraph, idx);
+        if (dev->iface.supports_op(dev, op)) {
+            continue;
+        }
+        GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
+
+        status = GGML_STATUS_ABORTED;
+        apir_encode_ggml_status(enc, &status);
+
+        return 0;
+    }
+#endif
+    status = bck->iface.graph_compute(bck, cgraph);
+
+    if (async_backend) {
+        bck->iface.synchronize(bck);
+    }
+
+    apir_encode_ggml_status(enc, &status);
+
+    return 0;
+}
@@ -0,0 +1,89 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    const char * string = buft->iface.get_name(buft);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t value = buft->iface.get_alignment(buft);
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t value = buft->iface.get_max_size(buft);
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    bool is_host = buft->iface.is_host(buft);
+    apir_encode_bool_t(enc, &is_host);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    ggml_backend_buffer_t buffer;
+
+    buffer = buft->iface.alloc_buffer(buft, size);
+
+    apir_encode_ggml_buffer(enc, buffer);
+
+    if (buffer) {
+        apir_track_backend_buffer(buffer);
+    }
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec);
+
+    size_t value = buft->iface.get_alloc_size(buft, op);
+
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
@@ -0,0 +1,131 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    uintptr_t base = (uintptr_t) buffer->iface.get_base(buffer);
+    apir_encode_uintptr_t(enc, &base);
+
+    return 0;
+}
+
+uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    ggml_tensor * tensor;
+    // safe to remove the const qualifier here
+    tensor = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    size_t offset;
+    apir_decode_size_t(dec, &offset);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        return 1;
+    }
+
+    buffer->iface.set_tensor(buffer, tensor, shmem_data, offset, size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    const ggml_tensor * tensor;
+    // safe to remove the const qualifier here
+    tensor = apir_decode_ggml_tensor(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    size_t offset;
+    apir_decode_size_t(dec, &offset);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        return 1;
+    }
+
+    buffer->iface.get_tensor(buffer, tensor, shmem_data, offset, size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    const ggml_tensor * src;
+    // safe to remove the const qualifier here
+    src               = apir_decode_ggml_tensor(dec);
+    ggml_tensor * dst = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec);
+
+    bool ret = buffer->iface.cpy_tensor(buffer, src, (ggml_tensor *) dst);
+
+    apir_encode_bool_t(enc, &ret);
+
+    return 0;
+}
+
+uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    uint8_t value;
+    apir_decode_uint8_t(dec, &value);
+
+    buffer->iface.clear(buffer, value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    if (!apir_untrack_backend_buffer(buffer)) {
+        GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer);
+        return 1;
+    }
+
+    buffer->iface.free_buffer(buffer);
+
+    return 0;
+}
@@ -0,0 +1,148 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    int32_t dev_count = reg->iface.get_device_count(reg);
+    apir_encode_int32_t(enc, &dev_count);
+
+    return 0;
+}
+
+uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    int32_t dev_count = reg->iface.get_device_count(reg);
+    apir_encode_int32_t(enc, &dev_count);
+
+    return 0;
+}
+
+uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    const char * string = dev->iface.get_name(dev);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    const char * string = dev->iface.get_description(dev);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    uint32_t type = dev->iface.get_type(dev);
+    apir_encode_uint32_t(enc, &type);
+
+    return 0;
+}
+
+uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    size_t free, total;
+    dev->iface.get_memory(dev, &free, &total);
+
+    apir_encode_size_t(enc, &free);
+    apir_encode_size_t(enc, &total);
+
+    return 0;
+}
+
+uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+
+    const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec);
+
+    bool supports_op = dev->iface.supports_op(dev, op);
+
+    apir_encode_bool_t(enc, &supports_op);
+
+    return 0;
+}
+
+uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    ggml_backend_buffer_type_t bufft = dev->iface.get_buffer_type(dev);
+
+    apir_encode_ggml_buffer_type(enc, bufft);
+
+    return 0;
+}
+
+uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    ggml_backend_dev_props props;
+    dev->iface.get_props(dev, &props);
+
+    apir_encode_bool_t(enc, &props.caps.async);
+    apir_encode_bool_t(enc, &props.caps.host_buffer);
+    apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr);
+    apir_encode_bool_t(enc, &props.caps.events);
+
+    return 0;
+}
+
+uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_ptr) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+    size_t max_tensor_size;
+    apir_decode_size_t(dec, &max_tensor_size);
+
+    ggml_backend_buffer_t buffer;
+    buffer = dev->iface.buffer_from_host_ptr(dev, shmem_ptr, size, max_tensor_size);
+
+    apir_encode_ggml_buffer(enc, buffer);
+    apir_encode_ggml_buffer_type(enc, buffer->buft);
+
+    if (buffer) {
+        apir_track_backend_buffer(buffer);
+    }
+
+    return 0;
+}
@@ -0,0 +1,46 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+ggml_backend_reg_t reg = NULL;
+ggml_backend_dev_t dev = NULL;
+ggml_backend_t     bck = NULL;
+
+uint64_t timer_start = 0;
+uint64_t timer_total = 0;
+uint64_t timer_count = 0;
+
+uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
+    if (reg != NULL) {
+        GGML_LOG_WARN("%s: already initialized\n", __func__);
+        return APIR_BACKEND_INITIALIZE_ALREADY_INITED;
+    }
+    ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p;
+
+    reg = ggml_backend_reg_fct();
+    if (reg == NULL) {
+        GGML_LOG_ERROR("%s: backend registration failed\n", __func__);
+        return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
+    }
+
+    if (!reg->iface.get_device_count(reg)) {
+        GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__);
+        return APIR_BACKEND_INITIALIZE_NO_DEVICE;
+    }
+
+    dev = reg->iface.get_device(reg, 0);
+
+    if (!dev) {
+        GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__);
+        return APIR_BACKEND_INITIALIZE_NO_DEVICE;
+    }
+
+    bck = dev->iface.init_backend(dev, NULL);
+
+    return APIR_BACKEND_INITIALIZE_SUCCESS;
+}
@@ -0,0 +1,130 @@
+#pragma once
+
+/* device */
+uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* buffer-type */
+uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* buffer */
+uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* backend */
+uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+static inline const char * backend_dispatch_command_name(ApirBackendCommandType type) {
+    switch (type) {
+        /* device */
+        case APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT:
+            return "backend_device_get_device_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_COUNT:
+            return "backend_device_get_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_NAME:
+            return "backend_device_get_name";
+        case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION:
+            return "backend_device_get_description";
+        case APIR_COMMAND_TYPE_DEVICE_GET_TYPE:
+            return "backend_device_get_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY:
+            return "backend_device_get_memory";
+        case APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP:
+            return "backend_device_supports_op";
+        case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE:
+            return "backend_device_get_buffer_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_PROPS:
+            return "backend_device_get_props";
+        case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR:
+            return "backend_device_buffer_from_ptr";
+        /* buffer-type */
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME:
+            return "backend_buffer_type_get_name";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT:
+            return "backend_buffer_type_get_alignment";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
+            return "backend_buffer_type_get_max_size";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
+            return "backend_buffer_type_is_host";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
+            return "backend_buffer_type_alloc_buffer";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
+            return "backend_buffer_type_get_alloc_size";
+        /* buffer */
+        case APIR_COMMAND_TYPE_BUFFER_GET_BASE:
+            return "backend_buffer_get_base";
+        case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR:
+            return "backend_buffer_set_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR:
+            return "backend_buffer_get_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR:
+            return "backend_buffer_cpy_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CLEAR:
+            return "backend_buffer_clear";
+        case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER:
+            return "backend_buffer_free_buffer";
+        /* backend */
+        case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE:
+            return "backend_backend_graph_compute";
+
+        default:
+            return "unknown";
+    }
+}
+
+extern "C" {
+static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {
+
+    /* device */
+
+    /* APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT  = */ backend_device_get_device_count,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_COUNT  = */ backend_device_get_count,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_NAME  = */ backend_device_get_name,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION  = */ backend_device_get_description,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_TYPE  = */ backend_device_get_type,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_MEMORY  = */ backend_device_get_memory,
+    /* APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP  = */ backend_device_supports_op,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE  = */ backend_device_get_buffer_type,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_PROPS  = */ backend_device_get_props,
+    /* APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR  = */ backend_device_buffer_from_ptr,
+
+    /* buffer-type */
+
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME  = */ backend_buffer_type_get_name,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT  = */ backend_buffer_type_get_alignment,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE  = */ backend_buffer_type_get_max_size,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST  = */ backend_buffer_type_is_host,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER  = */ backend_buffer_type_alloc_buffer,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE  = */ backend_buffer_type_get_alloc_size,
+
+    /* buffer */
+
+    /* APIR_COMMAND_TYPE_BUFFER_GET_BASE  = */ backend_buffer_get_base,
+    /* APIR_COMMAND_TYPE_BUFFER_SET_TENSOR  = */ backend_buffer_set_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_GET_TENSOR  = */ backend_buffer_get_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR  = */ backend_buffer_cpy_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_CLEAR  = */ backend_buffer_clear,
+    /* APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER  = */ backend_buffer_free_buffer,
+
+    /* backend */
+
+    /* APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE  = */ backend_backend_graph_compute,
+};
+}
@@ -0,0 +1,23 @@
+#pragma once
+
+#include <cstdint>
+#include <cstddef>
+
+#include <ggml-backend.h>
+
+#include "backend-convert.h"
+#include "backend-virgl-apir.h"
+#include "shared/apir_backend.h"
+#include "shared/apir_cs.h"
+#include "shared/apir_cs_ggml.h"
+
+struct virgl_apir_context {
+    uint32_t               ctx_id;
+    virgl_apir_callbacks * iface;
+};
+
+typedef uint32_t (*backend_dispatch_t)(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+#include "backend-dispatched.gen.h"
+
+uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p);
@@ -0,0 +1,32 @@
+#pragma once
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "shared/api_remoting.h"
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+
+extern ggml_backend_reg_t reg;
+extern ggml_backend_dev_t dev;
+extern ggml_backend_t     bck;
+
+struct virgl_apir_callbacks {
+    const char * (*get_config)(uint32_t virgl_ctx_id, const char * key);
+    void * (*get_shmem_ptr)(uint32_t virgl_ctx_id, uint32_t res_id);
+};
+
+extern "C" {
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs);
+void                      apir_backend_deinit(uint32_t virgl_ctx_id);
+uint32_t                  apir_backend_dispatcher(uint32_t               virgl_ctx_id,
+                                                  virgl_apir_callbacks * virgl_cbs,
+                                                  uint32_t               cmd_type,
+                                                  char *                 dec_cur,
+                                                  const char *           dec_end,
+                                                  char *                 enc_cur,
+                                                  const char *           enc_end,
+                                                  char **                enc_cur_after);
+}
@@ -0,0 +1,148 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+
+#include "shared/api_remoting.h"
+#include "shared/apir_backend.h"
+#include "shared/apir_cs.h"
+
+#include <dlfcn.h>
+#include <ggml-backend.h>
+
+#include <iostream>
+
+#define APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV "APIR_LLAMA_CPP_GGML_LIBRARY_PATH"
+#define APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV  "APIR_LLAMA_CPP_GGML_LIBRARY_REG"
+#define APIR_LLAMA_CPP_LOG_TO_FILE_ENV       "APIR_LLAMA_CPP_LOG_TO_FILE"
+
+#define GGML_DEFAULT_BACKEND_REG "ggml_backend_init"
+
+static void * backend_library_handle = NULL;
+static FILE * apir_logfile = NULL;
+
+static void log_to_file_callback(enum ggml_log_level level, const char * text, void * user_data) {
+    FILE * logfile = (FILE *)user_data;
+    fprintf(logfile, "[%d] %s", level, text);
+    fflush(logfile);
+}
+
+extern "C" {
+void apir_backend_deinit(uint32_t virgl_ctx_id) {
+    GGML_UNUSED(virgl_ctx_id);
+
+    auto buffers = apir_get_track_backend_buffers();
+    for (const auto & buffer : buffers) {
+        apir_untrack_backend_buffer(buffer);
+        buffer->iface.free_buffer(buffer);
+    }
+
+    if (dev) {
+        size_t free, total;
+        dev->iface.get_memory(dev, &free, &total);
+        GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024);
+    }
+
+    if (backend_library_handle) {
+        GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__);
+        dlclose(backend_library_handle);
+        backend_library_handle = NULL;
+    }
+
+    if (apir_logfile) {
+        fclose(apir_logfile);
+        apir_logfile = NULL;
+    }
+}
+
+#define APIR_GGML_LIBRARY_PATH_KEY "ggml.library.path"
+#define APIR_GGML_LIBRARY_REG_KEY "ggml.library.reg"
+
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs) {
+    const char * dlsym_error;
+
+    const char * apir_log_to_file = getenv(APIR_LLAMA_CPP_LOG_TO_FILE_ENV);
+    if (apir_log_to_file) {
+        apir_logfile = fopen(apir_log_to_file, "w");
+        if (apir_logfile) {
+            ggml_log_set(log_to_file_callback, apir_logfile);
+        } else {
+            GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file);
+        }
+    }
+
+    const char * library_name = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_PATH_KEY);
+    const char * virgl_library_reg = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_REG_KEY);
+    const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
+
+    if (!library_name) {
+        GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
+
+        return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
+    }
+
+    backend_library_handle = dlopen(library_name, RTLD_LAZY);
+
+    if (!backend_library_handle) {
+        GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror());
+
+        return APIR_LOAD_LIBRARY_CANNOT_OPEN;
+    }
+
+    if (!library_reg) {
+        GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
+
+        return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
+    }
+
+    void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
+    dlsym_error                 = dlerror();
+    if (dlsym_error) {
+        GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg,
+              APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
+
+        return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
+    }
+
+    uint32_t ret = backend_dispatch_initialize(ggml_backend_reg_fct);
+
+    return (ApirLoadLibraryReturnCode) (APIR_LOAD_LIBRARY_INIT_BASE_INDEX + ret);
+}
+
+uint32_t apir_backend_dispatcher(uint32_t               virgl_ctx_id,
+                                 virgl_apir_callbacks * virgl_cbs,
+                                 uint32_t               cmd_type,
+                                 char *                 dec_cur,
+                                 const char *           dec_end,
+                                 char *                 enc_cur,
+                                 const char *           enc_end,
+                                 char **                enc_cur_after) {
+    apir_encoder enc = {
+        .cur   = enc_cur,
+        .start = enc_cur,
+        .end   = enc_end,
+        .fatal = false,
+    };
+
+    apir_decoder dec = {
+        .cur   = dec_cur,
+        .end   = dec_end,
+        .fatal = false,
+    };
+
+    virgl_apir_context ctx = {
+        .ctx_id = virgl_ctx_id,
+        .iface = virgl_cbs,
+    };
+
+    if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
+        GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
+        return APIR_BACKEND_FORWARD_INDEX_INVALID;
+    }
+
+    backend_dispatch_t forward_fct = apir_backend_dispatch_table[cmd_type];
+    uint32_t           ret         = forward_fct(&enc, &dec, &ctx);
+
+    *enc_cur_after = enc.cur;
+
+    return ret;
+}
+}
@@ -0,0 +1,90 @@
+#pragma once
+
+/* the rest of this file must match virglrenderer/src/apir-protocol.h */
+
+#include <unistd.h>
+
+#include <cstdint>
+
+#define APIR_PROTOCOL_MAJOR 0
+#define APIR_PROTOCOL_MINOR 1
+
+#define APIR_HANDSHAKE_MAGIC 0xab1e
+
+enum ApirCommandType {
+    APIR_COMMAND_TYPE_HANDSHAKE   = 0,
+    APIR_COMMAND_TYPE_LOADLIBRARY = 1,
+    APIR_COMMAND_TYPE_FORWARD     = 2,
+
+    APIR_COMMAND_TYPE_LENGTH      = 3,
+};
+
+typedef uint64_t ApirCommandFlags;
+
+enum ApirLoadLibraryReturnCode {
+    APIR_LOAD_LIBRARY_SUCCESS                        = 0,
+    APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR = 1,
+    APIR_LOAD_LIBRARY_ALREADY_LOADED                 = 2,
+    APIR_LOAD_LIBRARY_ENV_VAR_MISSING                = 3,
+    APIR_LOAD_LIBRARY_CANNOT_OPEN                    = 4,
+    APIR_LOAD_LIBRARY_SYMBOL_MISSING                 = 5,
+    APIR_LOAD_LIBRARY_INIT_BASE_INDEX                = 6,  // anything above this is a APIR backend library initialization return code
+};
+
+enum ApirForwardReturnCode {
+    APIR_FORWARD_SUCCESS         = 0,
+    APIR_FORWARD_NO_DISPATCH_FCT = 1,
+    APIR_FORWARD_TIMEOUT         = 2,
+
+    APIR_FORWARD_BASE_INDEX      = 3,  // anything above this is a APIR backend library forward return code
+} ;
+
+__attribute__((unused)) static inline const char * apir_command_name(ApirCommandType type) {
+    switch (type) {
+        case APIR_COMMAND_TYPE_HANDSHAKE:
+            return "HandShake";
+        case APIR_COMMAND_TYPE_LOADLIBRARY:
+            return "LoadLibrary";
+        case APIR_COMMAND_TYPE_FORWARD:
+            return "Forward";
+        default:
+            return "unknown";
+    }
+}
+
+__attribute__((unused)) static const char * apir_load_library_error(ApirLoadLibraryReturnCode code) {
+#define APIR_LOAD_LIBRARY_ERROR(code_name) \
+    do {                                   \
+        if (code == code_name)             \
+            return #code_name;             \
+    } while (0)
+
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SUCCESS);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ALREADY_LOADED);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ENV_VAR_MISSING);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_CANNOT_OPEN);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SYMBOL_MISSING);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
+
+    return "Unknown APIR_COMMAND_TYPE_LoadLibrary error";
+
+#undef APIR_LOAD_LIBRARY_ERROR
+}
+
+__attribute__((unused)) static const char * apir_forward_error(ApirForwardReturnCode code) {
+#define APIR_FORWARD_ERROR(code_name) \
+    do {                              \
+        if (code == code_name)        \
+            return #code_name;        \
+    } while (0)
+
+    APIR_FORWARD_ERROR(APIR_FORWARD_SUCCESS);
+    APIR_FORWARD_ERROR(APIR_FORWARD_NO_DISPATCH_FCT);
+    APIR_FORWARD_ERROR(APIR_FORWARD_TIMEOUT);
+    APIR_FORWARD_ERROR(APIR_FORWARD_BASE_INDEX);
+
+    return "Unknown APIR_COMMAND_TYPE_FORWARD error";
+
+#undef APIR_FORWARD_ERROR
+}
@@ -0,0 +1,36 @@
+typedef enum ApirBackendCommandType {
+
+    /* device */
+    APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT = 0,
+    APIR_COMMAND_TYPE_DEVICE_GET_COUNT        = 1,
+    APIR_COMMAND_TYPE_DEVICE_GET_NAME         = 2,
+    APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION  = 3,
+    APIR_COMMAND_TYPE_DEVICE_GET_TYPE         = 4,
+    APIR_COMMAND_TYPE_DEVICE_GET_MEMORY       = 5,
+    APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP      = 6,
+    APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE  = 7,
+    APIR_COMMAND_TYPE_DEVICE_GET_PROPS        = 8,
+    APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR  = 9,
+
+    /* buffer-type */
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME       = 10,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT  = 11,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE   = 12,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST        = 13,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER   = 14,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = 15,
+
+    /* buffer */
+    APIR_COMMAND_TYPE_BUFFER_GET_BASE    = 16,
+    APIR_COMMAND_TYPE_BUFFER_SET_TENSOR  = 17,
+    APIR_COMMAND_TYPE_BUFFER_GET_TENSOR  = 18,
+    APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR  = 19,
+    APIR_COMMAND_TYPE_BUFFER_CLEAR       = 20,
+    APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER = 21,
+
+    /* backend */
+    APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = 22,
+
+    // last command_type index + 1
+    APIR_BACKEND_DISPATCH_TABLE_COUNT = 23,
+} ApirBackendCommandType;
@@ -0,0 +1,46 @@
+#pragma once
+
+#include "apir_backend.gen.h"
+
+#include <stdint.h>  // for uintptr_t
+#include <time.h>    // for timespec, clock_gettime
+
+#define APIR_BACKEND_INITIALIZE_SUCCESS                     0
+#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY 1
+#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY    2
+#define APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS     3
+#define APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS        4
+#define APIR_BACKEND_INITIALIZE_BACKEND_FAILED              5
+#define APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED          6
+#define APIR_BACKEND_INITIALIZE_ALREADY_INITED              7
+#define APIR_BACKEND_INITIALIZE_NO_DEVICE                   8
+
+
+// new entries here need to be added to the apir_backend_initialize_error function below
+
+#define APIR_BACKEND_FORWARD_INDEX_INVALID 6
+
+// 0 is fast, 1 avoids the backend to crash if an unsupported tensor is received
+#define APIR_BACKEND_CHECK_SUPPORTS_OP 0
+
+typedef uintptr_t apir_buffer_type_host_handle_t;
+typedef uintptr_t apir_buffer_host_handle_t;
+
+static const char * apir_backend_initialize_error(int code) {
+#define APIR_BACKEND_INITIALIZE_ERROR(code_name) \
+    do {                                         \
+        if (code == code_name)                   \
+            return #code_name;                   \
+    } while (0)
+
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_SUCCESS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_FAILED);
+
+    return "Unknown APIR_BACKEND_INITIALIZE error:/";
+
+#undef APIR_BACKEND_INITIALIZE_ERROR
+}
@@ -0,0 +1,383 @@
+#pragma once
+
+#include "ggml-impl.h"
+
+#include <cassert>
+#include <cstring>
+
+#define likely(x)   __builtin_expect(!!(x), 1)
+#define unlikely(x) __builtin_expect(!!(x), 0)
+
+struct apir_encoder {
+    char *       cur;
+    const char * start;
+    const char * end;
+    bool         fatal;
+
+};
+
+struct apir_decoder {
+    const char * cur;
+    const char * end;
+    bool         fatal;
+};
+
+/*
+ * new encoder and decoder
+ */
+
+static apir_decoder apir_new_decoder(const char * ptr, size_t size) {
+    apir_decoder dec = {
+        .cur = ptr,
+        .end = ptr + size,
+        .fatal = false,
+    };
+
+    return dec;
+}
+
+static apir_encoder apir_new_encoder(char * ptr, size_t size) {
+    apir_encoder enc = {
+        .cur   = ptr,
+        .start = ptr,
+        .end   = ptr + size,
+        .fatal = false,
+    };
+
+    return enc;
+}
+
+/*
+ * fatal flag handling
+ */
+
+static inline void apir_encoder_reset_fatal(apir_encoder * enc) {
+    enc->fatal = false;
+}
+
+static inline void apir_encoder_set_fatal(apir_encoder * enc) {
+    enc->fatal = true;
+}
+
+static inline bool apir_encoder_get_fatal(const apir_encoder * enc) {
+    return enc->fatal;
+}
+
+static inline void apir_decoder_reset_fatal(apir_decoder * dec) {
+    dec->fatal = false;
+}
+
+static inline void apir_decoder_set_fatal(apir_decoder * dec) {
+    dec->fatal = true;
+}
+
+static inline bool apir_decoder_get_fatal(const apir_decoder * dec) {
+    return dec->fatal;
+}
+
+/*
+ * encode peek
+ */
+
+static inline bool apir_decoder_peek_internal(apir_decoder * dec,
+                                              size_t                size,
+                                              void *                val,
+                                              size_t                val_size) {
+    assert(val_size <= size);
+
+    if (unlikely(size > (size_t) (dec->end - dec->cur))) {
+        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        apir_decoder_set_fatal(dec);
+        memset(val, 0, val_size);
+        return false;
+    }
+
+    /* we should not rely on the compiler to optimize away memcpy... */
+    memcpy(val, dec->cur, val_size);
+    return true;
+}
+
+static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val, size_t val_size) {
+    apir_decoder_peek_internal(dec, size, val, val_size);
+}
+
+static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) {
+    if (unlikely(size > (size_t) (dec->end - dec->cur))) {
+        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        apir_decoder_set_fatal(dec);
+        return NULL;
+    }
+    const void * addr = dec->cur;
+    dec->cur += size;
+
+    return addr;
+}
+
+/*
+ * read/write
+ */
+
+static inline void apir_decoder_read(apir_decoder * dec, size_t size, void * val, size_t val_size) {
+    if (apir_decoder_peek_internal(dec, size, val, val_size)) {
+        dec->cur += size;
+    }
+}
+
+static inline char * apir_encoder_write(apir_encoder * enc, size_t size, const void * val, size_t val_size) {
+    assert(val_size <= size);
+    assert(size <= ((size_t) (enc->end - enc->cur)));
+
+    char * write_addr = enc->cur;
+    /* we should not rely on the compiler to optimize away memcpy... */
+    memcpy(write_addr, val, val_size);
+    enc->cur += size;
+
+    return write_addr;
+}
+
+/*
+ * encode/decode
+ */
+
+static inline void apir_decode(apir_decoder * dec, size_t size, void * data, size_t data_size) {
+    assert(size % 4 == 0);
+    apir_decoder_read(dec, size, data, data_size);
+}
+
+static inline void apir_encode(apir_encoder * enc, size_t size, const void * data, size_t data_size) {
+    assert(size % 4 == 0);
+    apir_encoder_write(enc, size, data, data_size);
+}
+
+/*
+ * typed encode/decode
+ */
+
+/* uint8_t */
+
+static inline void apir_encode_uint8_t(apir_encoder * enc, const uint8_t * val) {
+    apir_encode(enc, sizeof(int), val, sizeof(*val));
+}
+
+static inline void apir_decode_uint8_t(apir_decoder * dec, uint8_t * val) {
+    apir_decode(dec, sizeof(int), val, sizeof(*val));
+}
+
+/* uint64_t */
+
+static inline void apir_encode_uint64_t(apir_encoder * enc, const uint64_t * val) {
+    apir_encode(enc, 8, val, sizeof(*val));
+}
+
+static inline void apir_decode_uint64_t(apir_decoder * dec, uint64_t * val) {
+    apir_decode(dec, 8, val, sizeof(*val));
+}
+
+static inline void apir_encode_uint64_t_array(apir_encoder * enc, const uint64_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_uint64_t_array(apir_decoder * dec, uint64_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+static inline const uint64_t * apir_decode_uint64_t_array_inplace(apir_decoder * dec, uint32_t count) {
+    return (uint64_t *) (uintptr_t) apir_decoder_use_inplace(dec, count * sizeof(uint64_t));
+}
+
+/* int32_t */
+
+static inline void apir_encode_int32_t(apir_encoder * enc, const int32_t * val) {
+    apir_encode(enc, 4, val, sizeof(*val));
+}
+
+static inline void apir_decode_int32_t(apir_decoder * dec, int32_t * val) {
+    apir_decode(dec, 4, val, sizeof(*val));
+}
+
+static inline void apir_encode_int32_t_array(apir_encoder * enc, const int32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_int32_t_array(apir_decoder * dec, int32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+/* array size (uint64_t) */
+
+static inline void apir_encode_array_size(apir_encoder * enc, uint64_t size) {
+    apir_encode_uint64_t(enc, &size);
+}
+
+static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expected_size) {
+    uint64_t size;
+    apir_decode_uint64_t(dec, &size);
+    if (size != expected_size) {
+        GGML_LOG_ERROR("Couldn't decode array from the decoder\n");
+        apir_decoder_set_fatal(dec);
+        size = 0;
+    }
+    return size;
+}
+
+static inline uint64_t apir_decode_array_size_unchecked(apir_decoder * dec) {
+    uint64_t size;
+    apir_decode_uint64_t(dec, &size);
+    return size;
+}
+
+/* non-array pointer */
+
+static inline bool apir_encode_simple_pointer(apir_encoder * enc, const void * val) {
+    apir_encode_array_size(enc, val ? 1 : 0);
+    return val;
+}
+
+static inline bool apir_decode_simple_pointer(apir_decoder * dec) {
+    return apir_decode_array_size_unchecked(dec);
+}
+
+/* uint32_t */
+
+static inline void apir_encode_uint32_t(apir_encoder * enc, const uint32_t * val) {
+    apir_encode(enc, 4, val, sizeof(*val));
+}
+
+static inline void apir_decode_uint32_t(apir_decoder * dec, uint32_t * val) {
+    apir_decode(dec, 4, val, sizeof(*val));
+}
+
+static inline void apir_encode_uint32_t_array(apir_encoder * enc, const uint32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_uint32_t_array(apir_decoder * dec, uint32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+/* size_t */
+
+static inline void apir_encode_size_t(apir_encoder * enc, const size_t * val) {
+    const uint64_t tmp = *val;
+    apir_encode_uint64_t(enc, &tmp);
+}
+
+static inline void apir_decode_size_t(apir_decoder * dec, size_t * val) {
+    uint64_t tmp;
+    apir_decode_uint64_t(dec, &tmp);
+    *val = tmp;
+}
+
+static inline void apir_encode_size_t_array(apir_encoder * enc, const size_t * val, uint32_t count) {
+    if (sizeof(size_t) == sizeof(uint64_t)) {
+        apir_encode_uint64_t_array(enc, (const uint64_t *) val, count);
+    } else {
+        for (uint32_t i = 0; i < count; i++) {
+            apir_encode_size_t(enc, &val[i]);
+        }
+    }
+}
+
+static inline void apir_decode_size_t_array(apir_decoder * dec, size_t * val, uint32_t count) {
+    if (sizeof(size_t) == sizeof(uint64_t)) {
+        apir_decode_uint64_t_array(dec, (uint64_t *) val, count);
+    } else {
+        for (uint32_t i = 0; i < count; i++) {
+            apir_decode_size_t(dec, &val[i]);
+        }
+    }
+}
+
+/* opaque blob */
+
+static inline void apir_encode_blob_array(apir_encoder * enc, const void * val, size_t size) {
+    apir_encode(enc, (size + 3) & ~3, val, size);
+}
+
+static inline void apir_decode_blob_array(apir_decoder * dec, void * val, size_t size) {
+    apir_decode(dec, (size + 3) & ~3, val, size);
+}
+
+/* string */
+
+static inline void apir_encode_char_array(apir_encoder * enc, const char * val, size_t size) {
+    assert(size && strlen(val) < size);
+    apir_encode_blob_array(enc, val, size);
+}
+
+static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t size) {
+    apir_decode_blob_array(dec, val, size);
+    if (size) {
+        val[size - 1] = '\0';
+    } else {
+        GGML_LOG_ERROR("Couldn't decode the blog array\n");
+        apir_decoder_set_fatal(dec);
+    }
+}
+
+/* (temp) buffer allocation */
+
+static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
+    size_t alloc_size;
+    if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
+        GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count);
+        return NULL;
+    }
+
+    return malloc(alloc_size);
+}
+
+/* bool */
+
+static inline void apir_encode_bool_t(apir_encoder * enc, const bool * val) {
+    apir_encode(enc, sizeof(int), val, sizeof(bool));
+}
+
+static inline void apir_decode_bool_t(apir_decoder * dec, bool * val) {
+    apir_decode(dec, sizeof(int), val, sizeof(bool));
+}
+
+/* apir_buffer_type_host_handle_t */
+
+static inline void apir_encode_apir_buffer_type_host_handle_t(apir_encoder *                  enc,
+                                                              const apir_buffer_type_host_handle_t * val) {
+    apir_encode(enc, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
+}
+
+static inline void apir_decode_apir_buffer_type_host_handle_t(apir_decoder *            dec,
+                                                              apir_buffer_type_host_handle_t * val) {
+    apir_decode(dec, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
+}
+
+/* apir_buffer_host_handle_t */
+
+static inline void apir_encode_apir_buffer_host_handle_t(apir_encoder *             enc,
+                                                         const apir_buffer_host_handle_t * val) {
+    apir_encode(enc, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t));
+}
+
+static inline void apir_decode_apir_buffer_host_handle_t(apir_decoder * dec, apir_buffer_host_handle_t * val) {
+    apir_decode(dec, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t));
+}
+
+/* uintptr_t */
+
+static inline void apir_encode_uintptr_t(apir_encoder * enc, const uintptr_t * val) {
+    apir_encode(enc, sizeof(*val), val, sizeof(*val));
+}
+
+static inline void apir_decode_uintptr_t(apir_decoder * dec, uintptr_t * val) {
+    apir_decode(dec, sizeof(*val), val, sizeof(*val));
+}
@@ -0,0 +1,211 @@
+#include "ggml-impl.h"
+#include "apir_cs.h"
+#include "apir_cs_rpc.h"
+
+// ggml_buffer_to_apir_host_handle(ggml_backend_buffer_t buffer);
+
+static inline void apir_encode_ggml_buffer_host_handle(apir_encoder *                    enc,
+                                                       const apir_buffer_host_handle_t * handle);
+
+static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec);
+
+/* apir_rpc_tensor */
+
+static inline void apir_encode_rcp_tensor(apir_encoder * enc, const apir_rpc_tensor * apir_rpc_tensor) {
+    size_t apir_rpc_tensor_size = sizeof(*apir_rpc_tensor);
+    apir_encode(enc, apir_rpc_tensor_size, apir_rpc_tensor, apir_rpc_tensor_size);
+}
+
+static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_inplace(apir_decoder * dec) {
+    size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor);
+
+    return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
+}
+
+static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_array_inplace(apir_decoder * dec,
+                                                                          uint32_t       n_tensors) {
+    size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor) * n_tensors;
+
+    return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
+}
+
+/* ggml_tensor */
+
+static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor * tensor) {
+    apir_rpc_tensor serialized = apir_serialize_tensor(tensor);
+
+    apir_encode_rcp_tensor(enc, &serialized);
+}
+
+static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
+    const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
+    ggml_init_params params{
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    ggml_context * ctx = ggml_init(params);
+
+    const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
+
+    return tensor;
+}
+
+/* *** ggml_backend_buffer_type_t *** */
+
+// ggml_backend_buffer_type_t is a POINTER (to a struct).
+// Only the host pointer is shared between the host and guest.
+// The guest stores it in `buft->context`.
+// The host simply writes the pointer address in the buffer variable.
+
+static inline void apir_encode_ggml_buffer_type(apir_encoder * enc, ggml_backend_buffer_type_t buft) {
+    apir_buffer_type_host_handle_t handle = ggml_buffer_type_to_apir_handle(buft);
+    apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
+}
+
+static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decoder * dec) {
+    apir_buffer_type_host_handle_t handle;
+
+    apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle));
+
+    return (ggml_backend_buffer_type_t) handle;
+}
+
+static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
+    apir_buffer_type_host_handle_t handle;
+
+    apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle));
+
+    return handle;
+}
+
+/* *** ggml_backend_type_t *** */
+
+// ggml_backend_buffer_t is a POINTER.
+// same logic as for ggml_backend_buffer_type_t
+
+static inline void apir_encode_ggml_buffer(apir_encoder * enc, const ggml_backend_buffer_t buffer) {
+    apir_buffer_host_handle_t handle = BUFFER_TO_HOST_HANDLE(buffer);
+    apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
+}
+
+static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec) {
+    ggml_backend_buffer_t buffer;
+    size_t                buffer_ptr_size = sizeof(buffer);
+
+    apir_decoder_read(dec, buffer_ptr_size, &buffer, buffer_ptr_size);
+
+    return buffer;
+}
+
+/* enum ggml_status */
+
+static inline void apir_encode_ggml_status(apir_encoder * enc, const ggml_status * status) {
+    apir_encoder_write(enc, sizeof(*status), status, sizeof(*status));
+}
+
+static inline void apir_decode_ggml_status(apir_decoder * dec, ggml_status * status) {
+    apir_decoder_read(dec, sizeof(*status), status, sizeof(*status));
+}
+
+/* virtgpu_shmem */
+
+static inline void apir_encode_virtgpu_shmem_res_id(apir_encoder * enc, uint32_t shmem_res_id) {
+    apir_encode_uint32_t(enc, &shmem_res_id);
+}
+
+static inline void apir_decode_virtgpu_shmem_res_id(apir_decoder * dec, uint32_t * shmem_res_id) {
+    apir_decode_uint32_t(dec, shmem_res_id);
+}
+
+/* ggml_cgraph */
+
+static inline size_t apir_serialize_ggml_cgraph(ggml_cgraph * cgraph, std::vector<uint8_t> & cgraph_data) {
+    apir_serialize_graph(cgraph, cgraph_data);
+
+    return cgraph_data.size();
+}
+
+static inline void apir_encode_cgraph_data(apir_encoder * enc, std::vector<uint8_t> & cgraph_data) {
+    size_t cgraph_size = cgraph_data.size();
+
+    apir_encode(enc, cgraph_size, cgraph_data.data(), cgraph_size);
+}
+
+static inline ggml_cgraph * apir_decode_ggml_cgraph(apir_decoder * dec, size_t cgraph_size) {
+    GGML_UNUSED(cgraph_size);
+
+    uint32_t n_nodes;
+    apir_decode_uint32_t(dec, &n_nodes);
+    const uint64_t * nodes = apir_decode_uint64_t_array_inplace(dec, n_nodes);
+
+    uint32_t n_tensors;
+    apir_decode_uint32_t(dec, &n_tensors);
+    const apir_rpc_tensor * tensors = apir_decode_apir_rpc_tensor_array_inplace(dec, n_tensors);
+
+    return apir_deserialize_graph(n_nodes, n_tensors, tensors, nodes);
+}
+
+static inline void apir_encode_ggml_buffer_handle(apir_encoder * enc, const apir_buffer_host_handle_t * handle) {
+    apir_encoder_write(enc, sizeof(*handle), &handle, sizeof(*handle));
+}
+
+static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml_tensor * tensor) {
+    size_t tensor_size = sizeof(*tensor);
+
+    if (tensor->extra) {
+        GGML_ABORT("Cannot pass tensors with extra");
+    }
+
+    if (tensor->src[0] && tensor->buffer) {
+        static int first = 1;
+        if (first) {
+            GGML_LOG_WARN("Cannot pass tensors with src and buffer\n");
+            first = 0;
+        }
+    }
+
+    apir_encoder_write(enc, tensor_size, tensor, tensor_size);
+
+    // tensor->data is a pointer inside the device buffer. No need to touch it
+    // tensor->buffer is a pointer to a buffer. Encoding the buffer handle in sequence.
+    // (could also make a copy of the tensor, and update locally.)
+
+    if (tensor->buffer) {
+        apir_buffer_host_handle_t buffer_handle = ggml_buffer_to_apir_handle(tensor->buffer);
+        apir_encode_ggml_buffer_handle(enc, &buffer_handle);
+    }
+
+    if (tensor->view_src) {
+        apir_encoder_write(enc, tensor_size, tensor->view_src, tensor_size);
+    }
+
+    for (int i = 0; tensor->src[i]; i++) {
+        const ggml_tensor * tensor_src = tensor->src[i];
+        apir_encoder_write(enc, tensor_size, tensor_src, tensor_size);
+    }
+}
+
+static inline const ggml_tensor * apir_decode_ggml_tensor_inplace(apir_decoder * dec) {
+    // it safe to remove the `const` qualifier here, we *do* want to
+    // modify the shared memory data to fix the `src` pointers.
+    ggml_tensor * tensor = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+
+    // tensor->data is a pointer inside the device buffer. No need to touch it
+    // tensor->buffer is a pointer to a buffer. Decode the buffer handle encoded in sequence.
+    if (tensor->buffer) {
+        tensor->buffer = apir_decode_ggml_buffer(dec);
+    }
+
+    if (tensor->view_src) {
+        ggml_tensor * tensor_view_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+        tensor->view_src              = tensor_view_src;
+    }
+
+    for (int i = 0; tensor->src[i]; i++) {
+        ggml_tensor * tensor_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+        tensor->src[i] = tensor_src;  // overwrite op->src[i] pointer with the actual location of the src tensor
+    }
+
+    return tensor;
+}
@@ -0,0 +1,54 @@
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#include <cstdint>
+
+// 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);