perf(serialization): optimized the memory usage when load new format model

GitOrigin-RevId: 2b7313ebe3

imperative/src/impl/ops/opr_attr.cpp

@@ -25,7 +25,11 @@ class OprParamsLoadContext final : public serialization::OprLoadContextRawPOD {
 
     std::shared_ptr<HostTensorND> load_tensor() override { mgb_assert(0); }
 
-    std::shared_ptr<DeviceTensorND> load_tensor_shared() override { mgb_assert(0); }
+    std::shared_ptr<DeviceTensorND> load_tensor_shared(
+            bool copy_immediatly = false) override {
+        (void)copy_immediatly;
+        mgb_assert(0);
+    }
 
     const serialization::GraphLoadConfig& config() const override { mgb_assert(0); }
 

lite/lite-c/src/network.cpp

@@ -245,9 +245,6 @@ int LITE_destroy_network(LiteNetwork network) {
     auto& global_holder = get_gloabl_network_holder();
     if (global_holder.find(network) != global_holder.end()) {
         global_holder.erase(network);
-    } else {
-        //! means the network has been destoryed
-        return -1;
     }
     LITE_CAPI_END();
 }

lite/lite-c/src/tensor.cpp

@@ -75,9 +75,6 @@ int LITE_destroy_tensor(LiteTensor tensor) {
     auto& global_holder = get_global_tensor_holder();
     if (global_holder.find(tensor) != global_holder.end()) {
         global_holder.erase(tensor);
-    } else {
-        //! return -1, means the tensor has been destroyed.
-        return -1;
     }
     LITE_CAPI_END();
 }

lite/load_and_run/src/models/model_lite.cpp

@@ -16,26 +16,8 @@ void ModelLite::create_network() {
 }
 
 void ModelLite::load_model() {
-    if (share_model_mem) {
-        //! WARNNING:maybe not right to share param memmory for this
-        LITE_LOG("enable share model memory");
-
-        FILE* fin = fopen(model_path.c_str(), "rb");
-        LITE_ASSERT(fin, "failed to open %s: %s", model_path.c_str(), strerror(errno));
-        fseek(fin, 0, SEEK_END);
-        size_t size = ftell(fin);
-        fseek(fin, 0, SEEK_SET);
-
-        void* ptr = malloc(size);
-        std::shared_ptr<void> buf{ptr, free};
-        auto nr = fread(buf.get(), 1, size, fin);
-        LITE_ASSERT(nr == size, "read model file failed");
-        fclose(fin);
-
-        m_network->load_model(buf.get(), size);
-    } else {
-        m_network->load_model(model_path);
-    }
+    //! lite shared memory default
+    m_network->load_model(model_path);
 }
 
 void ModelLite::run_model() {

lite/src/parse_model/model_parser.cpp

@@ -128,7 +128,10 @@ std::shared_ptr<void> ModelParser::decrypt_memory(
     const uint8_t* memory_ptr = data;
     if (decryption_name == "NONE") {
         result_length = length;
-        return std::shared_ptr<void>(const_cast<uint8_t*>(memory_ptr), [](void*) {});
+        std::shared_ptr<uint8_t> shptr{
+                new uint8_t[length], [](uint8_t* p) { delete[] p; }};
+        memcpy(shptr.get(), data, length);
+        return shptr;
     }
     LITE_LOCK_GUARD(decryption_static_data().map_mutex);
     auto it = decryption_static_data().decryption_methods.find(decryption_name);

lite/test/test_network_c.cpp

@@ -1032,7 +1032,7 @@ TEST(TestCapiNetWork, GlobalHolder) {
             LITE_make_network(&c_network, *default_config(), *default_network_io()));
     //! make sure destroy_network is destroyed by LITE_make_network
     LITE_destroy_network(destroy_network);
-    ASSERT_EQ(LITE_destroy_network(destroy_network), -1);
+    ASSERT_EQ(LITE_destroy_network(destroy_network), 0);
     LITE_CAPI_CHECK(LITE_destroy_network(c_network));
 }
 

lite/test/test_tensor_c.cpp

@@ -328,7 +328,7 @@ TEST(TestCapiTensor, GlobalHolder) {
     LITE_make_tensor(description, &c_tensor0);
     //! make sure destroy_tensor is destroyed by LITE_make_tensor
     LITE_destroy_tensor(destroy_tensor);
-    ASSERT_EQ(LITE_destroy_tensor(destroy_tensor), -1);
+    ASSERT_EQ(LITE_destroy_tensor(destroy_tensor), 0);
     LITE_destroy_tensor(c_tensor0);
 }
 

src/core/include/megbrain/tensor.h

@@ -332,6 +332,7 @@ class TensorND {
 
 public:
     using ChainReturnType = TensorND<TensorStorage>;
+    using Storage = TensorStorage;
 
     MGE_WIN_DECLSPEC_FUC TensorND();
 

src/core/test/comp_node_helper.cpp

@@ -443,38 +443,41 @@ void run<shape_dep_const_shape>(CompNode cn) {
     HostTensorGenerator<> gen;
     auto host_x = gen({4, 5}, cn);
     auto fname = output_file("test_comp_node_record_shape_dep_const_shape");
+    auto test = [&](serialization::GraphDumpFormat format) {
+        HostTensorND y_expect;
+        {
+            // dump graph
+            auto graph = ComputingGraph::make();
+            auto x = opr::Host2DeviceCopy::make(
+                         *graph, host_x, OperatorNodeConfig{"x"}),
+                 y = x.flatten() +
+                     opr::reduce_sum(opr::GetVarShape::make(x), x.make_scalar(1));
 
-    HostTensorND y_expect;
-    {
-        // dump graph
-        auto graph = ComputingGraph::make();
-        auto x = opr::Host2DeviceCopy::make(*graph, host_x, OperatorNodeConfig{"x"}),
-             y = x.flatten() +
-                 opr::reduce_sum(opr::GetVarShape::make(x), x.make_scalar(1));
-
-        graph->compile({make_callback_copy(y, y_expect)})->execute();
-
-        auto dumper = GraphDumper::make(OutputFile::make_fs(fname.c_str()));
-        dumper->dump({y});
-    }
+            graph->compile({make_callback_copy(y, y_expect)})->execute();
 
-    HostTensorND host_y;
-    {
-        GraphLoadConfig config;
-        config.const_var_shape = true;
-        auto loader = GraphLoader::make(InputFile::make_fs(fname.c_str()));
-        auto load_rst = loader->load(config);
-        load_rst.graph->options().comp_node_seq_record_level = 2;
-        load_rst.graph->options().var_sanity_check_first_run = false;
-        auto x_inp = load_rst.tensor_map.at("x");
-        auto y = load_rst.output_var_list.at(0);
-        auto func = load_rst.graph_compile({make_callback_copy(y, host_y)});
-
-        x_inp->copy_from(*host_x);
-        func->execute();
-    }
+            auto dumper = GraphDumper::make(OutputFile::make_fs(fname.c_str()), format);
+            dumper->dump({y});
+        }
 
-    MGB_ASSERT_TENSOR_EQ(y_expect, host_y);
+        HostTensorND host_y;
+        {
+            GraphLoadConfig config;
+            config.const_var_shape = true;
+            auto loader = GraphLoader::make(InputFile::make_fs(fname.c_str()), format);
+            auto load_rst = loader->load(config);
+            load_rst.graph->options().comp_node_seq_record_level = 2;
+            load_rst.graph->options().var_sanity_check_first_run = false;
+            auto x_inp = load_rst.tensor_map.at("x");
+            auto y = load_rst.output_var_list.at(0);
+            auto func = load_rst.graph_compile({make_callback_copy(y, host_y)});
+
+            x_inp->copy_from(*host_x);
+            func->execute();
+        }
+        MGB_ASSERT_TENSOR_EQ(y_expect, host_y);
+    };
+    test({});
+    test(serialization::GraphDumpFormat::FLATBUFFERS_V2);
 }
 
 //! single thread multi recorder run interleave

src/opr/impl/io.cpp

@@ -367,16 +367,19 @@ MGB_DYN_TYPE_OBJ_FINAL_IMPL(ImmutableTensor);
 
 class ImmutableTensor::Value {
     MGB_MUTEX m_mtx;
-    DeviceTensorND m_dev, m_static_infer;
+    std::shared_ptr<DeviceTensorND> m_dev = std::make_shared<DeviceTensorND>();
+    DeviceTensorND m_static_infer;
     std::string m_summary;
 
 public:
     void setup(CompNode cn, const HostTensorND& val);
 
-    bool initialized() const { return m_dev.shape_valid(); }
+    void setup(std::shared_ptr<DeviceTensorND> val);
+
+    bool initialized() const { return m_dev->shape_valid(); }
 
     //! value on comp node
-    const DeviceTensorND& dev() const { return m_dev; }
+    const DeviceTensorND& dev() const { return *m_dev; }
 
     //! get value on static infer CPU node
     DeviceTensorND& static_infer();
@@ -385,10 +388,17 @@ public:
     const std::string& summary() const { return m_summary; }
 };
 
+void ImmutableTensor::Value::setup(std::shared_ptr<DeviceTensorND> val) {
+    mgb_assert(val);
+    m_dev = val;
+    m_summary = ssprintf("const%s", val->shape().to_string().c_str());
+}
+
 void ImmutableTensor::Value::setup(CompNode cn, const HostTensorND& val) {
-    mgb_assert(m_dev.empty() && !m_dev.shape_valid());
-    m_dev.comp_node(cn).copy_from(val).sync();
-    mgb_assert(val.empty() == m_dev.empty());
+    mgb_assert(m_dev->empty() && !m_dev->shape_valid());
+
+    m_dev->comp_node(cn).copy_from(val).sync();
+    mgb_assert(val.empty() == m_dev->empty());
 
     auto one_elem = [](const TensorShape& shape) {
         for (size_t i = 0; i < shape.ndim; ++i) {
@@ -413,8 +423,8 @@ void ImmutableTensor::Value::setup(CompNode cn, const HostTensorND& val) {
 DeviceTensorND& ImmutableTensor::Value::static_infer() {
     MGB_LOCK_GUARD(m_mtx);
     if (!m_static_infer.shape_valid()) {
-        mgb_assert(m_dev.shape_valid());
-        m_static_infer.comp_node(CompNode::default_cpu()).copy_from(m_dev);
+        mgb_assert(m_dev->shape_valid());
+        m_static_infer.comp_node(CompNode::default_cpu()).copy_from(*m_dev);
     }
     return m_static_infer;
 }
@@ -589,6 +599,19 @@ SymbolVar ImmutableTensor::make(
 }
 
 SymbolVar ImmutableTensor::make(
+        ComputingGraph& graph, std::shared_ptr<DeviceTensorND> val,
+        const OperatorNodeConfig& config) {
+    auto cn = val->comp_node();
+    if (config.has_comp_node_set())
+        cn = config.get_single_comp_node();
+
+    auto value = std::make_shared<Value>();
+    value->setup(val);
+
+    return make_from_value(graph, *value, value, config);
+}
+
+SymbolVar ImmutableTensor::make(
         ComputingGraph& graph, const DTypeScalar& val,
         const OperatorNodeConfig& config) {
     mgb_assert(

src/opr/impl/io.sereg.h

@@ -132,8 +132,10 @@ struct OprLoadDumpImpl<opr::ImmutableTensor, 0> {
             OprLoadContext& ctx, const cg::VarNodeArray& inputs,
             const OperatorNodeConfig& config) {
         mgb_assert(inputs.empty());
-        auto val = ctx.load_tensor();
-        return Opr::make(ctx.graph(), *val, config).node()->owner_opr();
+        //! because ImmutableTensor will used in infer shape or infer value,
+        //! so must copy immediatly
+        auto val = ctx.load_tensor_shared(true);
+        return Opr::make(ctx.graph(), val, config).node()->owner_opr();
     }
 };
 

src/opr/impl/io.sereg.v2.h

@@ -32,8 +32,10 @@ struct OprLoadDumpImplV2<opr::ImmutableTensor, 0> {
         auto fopr = reinterpret_cast<const fbs::v2::Operator*>(
                 fbs_ctx.get_current_opr_data());
         if (fopr->tensors() && fopr->tensors()->size() > 0) {
-            auto val = fbs_ctx.load_tensor();
-            return Opr::make(fbs_ctx.graph(), *val, config).node()->owner_opr();
+            //! because ImmutableTensor will used in infer shape or infer value,
+            //! so must copy immediatly
+            auto val = fbs_ctx.load_tensor_shared(true);
+            return Opr::make(fbs_ctx.graph(), val, config).node()->owner_opr();
         } else {
             mgb_throw(SerializationError, "ImmutableTensor load with no tensor data.");
         }

src/opr/include/megbrain/opr/io.h

@@ -360,6 +360,10 @@ public:
             ComputingGraph& graph, const HostTensorND& val,
             const OperatorNodeConfig& config = {});
 
+    MGE_WIN_DECLSPEC_FUC static SymbolVar make(
+            ComputingGraph& graph, std::shared_ptr<DeviceTensorND> val,
+            const OperatorNodeConfig& config = {});
+
     //! make from DTypeScalar; comp node must be provided in config
     MGE_WIN_DECLSPEC_FUC static SymbolVar make(
             ComputingGraph& graph, const DTypeScalar& val,

src/serialization/impl/file.cpp

@@ -138,6 +138,10 @@ public:
         mgb_assert(m_refhold && size);
     }
 
+    bool is_shared_memory() override { return true; }
+    bool writable() override { return m_writable; }
+    void have_modified() override { m_modified = true; }
+
     void rewind() override {
         if (m_modified) {
             // data has beem modified; can not read again

src/serialization/impl/opr_shallow_copy.cpp

@@ -63,7 +63,11 @@ class OprLoadContextMemory final : public OprLoadContextRawPOD {
 
     std::shared_ptr<HostTensorND> load_tensor() override { mgb_assert(0); }
 
-    std::shared_ptr<DeviceTensorND> load_tensor_shared() override { mgb_assert(0); }
+    std::shared_ptr<DeviceTensorND> load_tensor_shared(
+            bool copy_immediatly = false) override {
+        (void)copy_immediatly;
+        mgb_assert(0);
+    }
 
     const GraphLoadConfig& config() const override {
         mgb_throw(GraphError, "OprLoadContextMemory has no associated config");

src/serialization/impl/serializer_oss.cpp

@@ -483,7 +483,8 @@ class GraphLoaderOSS::OprLoadContextImpl final : public OprLoadContextFlatBuffer
 
     std::shared_ptr<HostTensorND> load_tensor() override;
 
-    std::shared_ptr<DeviceTensorND> load_tensor_shared() override;
+    std::shared_ptr<DeviceTensorND> load_tensor_shared(
+            bool copy_immediatly = false) override;
 
     void load_single_opr(const fbs::Operator* opr);
 
@@ -641,8 +642,8 @@ std::shared_ptr<HostTensorND> GraphLoaderOSS::OprLoadContextImpl::load_tensor()
     return ret;
 }
 
-std::shared_ptr<DeviceTensorND> GraphLoaderOSS::OprLoadContextImpl::
-        load_tensor_shared() {
+std::shared_ptr<DeviceTensorND> GraphLoaderOSS::OprLoadContextImpl::load_tensor_shared(
+        bool copy_immediatly) {
     mgb_assert(
             m_current_opr->tensors() &&
             m_cur_opr_tensor_cnt < m_current_opr->tensors()->size());
@@ -650,6 +651,9 @@ std::shared_ptr<DeviceTensorND> GraphLoaderOSS::OprLoadContextImpl::
     auto comp_node = load_comp_node(tensor->comp_node());
     auto layout = load_tensor_layout(tensor);
     mgb_assert(tensor->data_size());
+    if (m_loader->m_shared_tensor_map.size() <= m_cur_shared_tensor_idx) {
+        m_loader->m_shared_tensor_map.resize(m_cur_shared_tensor_idx + 5);
+    }
     auto&& sh_reg = m_loader->m_shared_tensor_map.at(m_cur_shared_tensor_idx++);
     auto&& sh_ptr_ref = sh_reg.second[comp_node.mem_node()];
     if (sh_ptr_ref) {
@@ -673,6 +677,11 @@ std::shared_ptr<DeviceTensorND> GraphLoaderOSS::OprLoadContextImpl::
         load_tensor_value(&hv, layout, tensor);
         sh_ptr_ref = std::make_shared<DeviceTensorND>();
         *sh_ptr_ref = DeviceTensorND::make_proxy(hv);
+    } else if (copy_immediatly) {
+        HostTensorND hv{CompNode::default_cpu()};
+        load_tensor_value(&hv, layout, tensor);
+        sh_ptr_ref = std::make_shared<DeviceTensorND>();
+        sh_ptr_ref->comp_node(comp_node).copy_from(hv).sync();
     } else {
         // use lazy load for non-CPU devices
         HostTensorND hv{CompNode::default_cpu()};
@@ -803,7 +812,7 @@ GraphLoader::LoadResult GraphLoaderOSS::OprLoadContextImpl::load_oprs() {
         ret.output_var_map_id[out->original_id()] = var;
         ret.output_var_list[i] = var;
     }
-    mgb_assert(m_cur_shared_tensor_idx == m_loader->m_shared_tensor_map.size());
+    mgb_assert(m_cur_shared_tensor_idx <= m_loader->m_shared_tensor_map.size());
     return ret;
 }
 
@@ -880,7 +889,7 @@ GraphLoader::LoadResult GraphLoaderOSS::load(const LoadConfig& config, bool rewi
     if (m_shared_tensor_map.empty()) {
         m_shared_tensor_map.resize(m_graph->nr_shared_tensor());
     } else {
-        mgb_assert(m_shared_tensor_map.size() == m_graph->nr_shared_tensor());
+        mgb_assert(m_shared_tensor_map.size() >= m_graph->nr_shared_tensor());
     }
 
     OprLoadContextImpl ctx{this, m_graph->mgb_version()};

src/serialization/impl/serializer_oss_v2.cpp

@@ -1,5 +1,6 @@
 #if MGB_ENABLE_FBS_SERIALIZATION
 
+#include <map>
 #include "megbrain/comp_node_env.h"
 #include "megbrain/opr/io.h"
 #include "megbrain/serialization/helper.h"
@@ -523,6 +524,77 @@ void GraphDumperOSSV2::dump_buf_with_len(const void* data, uint32_t size) {
 }
 
 // ----------------------------- Loader --------------------------------------
+
+/**
+ * SharedTensorAlignMent will record all shared device tensors, at beginning, the
+ * tensor is not aligned, after all shared device tensor loaded, and the user
+ * provide memory will be wrote, and reorder all the tensor to aligned address
+ * ptr.
+ */
+class GraphLoaderOSSV2::SharedTensorAlignMent {
+public:
+    SharedTensorAlignMent(SharedBuffer buffer, InputFile* file, bool is_enabled)
+            : m_enabled(is_enabled), m_file(file), m_model_buffer(buffer){};
+
+    bool add_device_tensor(std::shared_ptr<DeviceTensorND> tensor) {
+        if (!m_enabled)
+            return false;
+        if (tensor) {
+            m_device_tensors[reinterpret_cast<intptr_t>(tensor->raw_ptr())] = tensor;
+            return true;
+        }
+        return false;
+    }
+
+    /**
+     * record the tensor shared from the m_model_buffer, copy every tensor to
+     * the aligned address, then the model file will be modilfied, so it can't
+     * use again.
+     */
+    bool reorder_and_align_tensor() {
+        if (!m_enabled)
+            return false;
+        bool modilfied = false;
+        intptr_t buffer_start = reinterpret_cast<intptr_t>(m_model_buffer.data());
+        intptr_t write_end = buffer_start;
+        for (auto& iter : m_device_tensors) {
+            auto& tensor = iter.second;
+            size_t tensor_size = tensor->layout().span().dist_byte();
+            size_t alignment = tensor->comp_node().get_mem_addr_alignment();
+            intptr_t tensor_start = reinterpret_cast<intptr_t>(tensor->raw_ptr());
+            intptr_t align_start = static_cast<intptr_t>(
+                    reinterpret_cast<uintptr_t>(tensor->raw_ptr()) & ~(alignment - 1));
+            if (align_start > write_end) {
+                if (tensor_start != align_start) {
+                    memmove(reinterpret_cast<void*>(align_start),
+                            reinterpret_cast<void*>(tensor_start), tensor_size);
+                    modilfied = true;
+                }
+                write_end = align_start + tensor_size;
+                DeviceTensorStorage storage;
+                auto raw_storage = std::shared_ptr<mgb::dt_byte>(
+                        reinterpret_cast<mgb::dt_byte*>(align_start), [](void*) {});
+                storage.reset(tensor->comp_node(), tensor_size, raw_storage);
+                tensor->reset(storage, tensor->layout());
+            } else {
+                DeviceTensorND new_tensor(tensor->comp_node());
+                new_tensor.copy_from(*tensor).sync();
+                *tensor = new_tensor;
+            }
+            if (modilfied) {
+                m_file->have_modified();
+            }
+        }
+        return true;
+    }
+
+private:
+    bool m_enabled = false;
+    InputFile* m_file;
+    SharedBuffer m_model_buffer;
+    std::map<intptr_t, std::shared_ptr<DeviceTensorND>> m_device_tensors;
+};
+
 CompNode GraphLoaderOSSV2::OprLoadContextImpl::load_comp_node(
         const fbs::v2::CompNode* comp_node) {
     mgb_assert(comp_node);
@@ -596,7 +668,9 @@ std::shared_ptr<HostTensorND> GraphLoaderOSSV2::OprLoadContextImpl::load_tensor(
                     "serialization v2 format is pure flatbuffer format, not support "
                     "user tensor value loader callback.");
         }
-        memcpy(ret->raw_ptr(), tensor->data()->data(), tensor->data()->size());
+        fill_tensor_memory(
+                *ret, tensor->data()->data(), tensor->data()->size(),
+                m_loader->m_file->is_shared_memory());
     }
     if (tensor->name()) {
         m_tensor_map[tensor->name()->str()] = ret;
@@ -612,7 +686,7 @@ std::shared_ptr<HostTensorND> GraphLoaderOSSV2::OprLoadContextImpl::load_tensor(
 }
 
 std::shared_ptr<DeviceTensorND> GraphLoaderOSSV2::OprLoadContextImpl::
-        load_tensor_shared() {
+        load_tensor_shared(bool copy_immediatly) {
     mgb_assert(
             m_current_opr->tensors() &&
             m_cur_opr_tensor_cnt < m_current_opr->tensors()->size());
@@ -620,6 +694,9 @@ std::shared_ptr<DeviceTensorND> GraphLoaderOSSV2::OprLoadContextImpl::
     auto comp_node = load_comp_node(tensor->comp_node());
     auto layout = load_tensor_layout(tensor, comp_node);
     mgb_assert(tensor->data());
+    if (m_loader->m_shared_tensor_map.size() <= m_cur_shared_tensor_idx) {
+        m_loader->m_shared_tensor_map.resize(m_cur_shared_tensor_idx + 5);
+    }
     auto&& shared_pair = m_loader->m_shared_tensor_map.at(m_cur_shared_tensor_idx++);
     auto&& shared_tensor_ref = shared_pair.second[comp_node.mem_node()];
     if (shared_tensor_ref) {
@@ -637,19 +714,34 @@ std::shared_ptr<DeviceTensorND> GraphLoaderOSSV2::OprLoadContextImpl::
 
     if (comp_node.mem_node() == CompNode::default_cpu().mem_node()) {
         // directly forward CPU memory
+        shared_tensor_ref = std::make_shared<DeviceTensorND>();
         HostTensorND hv{comp_node};
         if (tensor->data() && tensor->data()->size() > 0) {
             hv.dtype(layout.dtype).resize(layout);
-            memcpy(hv.raw_ptr(), tensor->data()->data(), tensor->data()->size());
+            fill_tensor_memory(
+                    hv, tensor->data()->data(), tensor->data()->size(),
+                    m_loader->m_file->is_shared_memory());
         }
-        shared_tensor_ref = std::make_shared<DeviceTensorND>();
         *shared_tensor_ref = DeviceTensorND::make_proxy(hv);
+        m_tensor_alignment->add_device_tensor(shared_tensor_ref);
+    } else if (copy_immediatly) {
+        HostTensorND hv{CompNode::default_cpu()};
+        shared_tensor_ref = std::make_shared<DeviceTensorND>();
+        if (tensor->data() && tensor->data()->size() > 0) {
+            hv.dtype(layout.dtype).resize(layout);
+            fill_tensor_memory(
+                    hv, tensor->data()->data(), tensor->data()->size(),
+                    m_loader->m_file->is_shared_memory());
+        }
+        shared_tensor_ref->comp_node(comp_node).copy_from(hv).sync();
     } else {
         // use lazy load for non-CPU devices
         HostTensorND hv{CompNode::default_cpu()};
         if (tensor->data() && tensor->data()->size() > 0) {
             hv.dtype(layout.dtype).resize(layout);
-            memcpy(hv.raw_ptr(), tensor->data()->data(), tensor->data()->size());
+            fill_tensor_memory(
+                    hv, tensor->data()->data(), tensor->data()->size(),
+                    m_loader->m_file->is_shared_memory());
         }
         shared_tensor_ref = m_device_value_loader.make(comp_node, std::move(hv));
     }
@@ -784,7 +876,7 @@ GraphLoader::LoadResult GraphLoaderOSSV2::OprLoadContextImpl::load_oprs() {
         ret.output_var_map_id[out->original_id()] = var;
         ret.output_var_list[i] = var;
     }
-    mgb_assert(m_cur_shared_tensor_idx == m_loader->m_shared_tensor_map.size());
+    mgb_assert(m_cur_shared_tensor_idx <= m_loader->m_shared_tensor_map.size());
     return ret;
 }
 
@@ -808,7 +900,6 @@ GraphLoader::LoadResult GraphLoaderOSSV2::load(const LoadConfig& config, bool re
     m_file->read(&size, sizeof(size));
     m_file->skip(-sizeof(size));
     m_model_buf = m_file->read_shared(size + sizeof(size));
-
     {
         flatbuffers::Verifier verifier(
                 static_cast<const uint8_t*>(m_model_buf.data()), m_model_buf.size());
@@ -838,8 +929,10 @@ GraphLoader::LoadResult GraphLoaderOSSV2::load(const LoadConfig& config, bool re
     } else {
         mgb_assert(m_shared_tensor_map.size() == m_model->nr_shared_tensor());
     }
-
-    OprLoadContextImpl ctx{this, m_model->mge_version()};
+    SharedTensorAlignMent tensor_alignment(
+            m_model_buf, m_file.get(),
+            m_file->writable() && m_file->is_shared_memory());
+    OprLoadContextImpl ctx{this, &tensor_alignment, m_model->mge_version()};
     ctx.load_middle_tensor();
     auto metadata = ctx.load_metadata();
     auto result = ctx.load_oprs();
@@ -856,6 +949,7 @@ GraphLoader::LoadResult GraphLoaderOSSV2::load(const LoadConfig& config, bool re
         }
     }
     m_model_loaded = true;
+    tensor_alignment.reorder_and_align_tensor();
     result.graph_compile_ahead();
     return result;
 }

src/serialization/include/megbrain/serialization/file.h

@@ -41,6 +41,15 @@ public:
     //! return current read offset
     virtual size_t tell() = 0;
 
+    //! whether this file format support share memory when load model
+    virtual bool is_shared_memory() { return false; }
+
+    //! whether this can be write
+    virtual bool writable() { return false; }
+
+    //! whether this file have been wrote
+    virtual void have_modified() {}
+
     /*!
      * \brief read into a host tensor
      *

src/serialization/include/megbrain/serialization/opr_load_dump.h

@@ -208,7 +208,8 @@ public:
      *
      * It must be dumped with TensorWriteMethod::VALUE_SHARED
      */
-    virtual std::shared_ptr<DeviceTensorND> load_tensor_shared() = 0;
+    virtual std::shared_ptr<DeviceTensorND> load_tensor_shared(
+            bool copy_immediatly = false) = 0;
 
     //! get associated global configuration
     virtual const GraphLoadConfig& config() const = 0;

src/serialization/include/megbrain/serialization/oss_opr_load_dump.h

@@ -104,6 +104,7 @@ class GraphLoaderOSSV2 final : public GraphLoader {
 
 public:
     class OprLoadContextImpl;
+    class SharedTensorAlignMent;
     friend class OprLoadContextImpl;
 
     GraphLoaderOSSV2(std::unique_ptr<InputFile> input_file)
@@ -136,22 +137,51 @@ class GraphLoaderOSSV2::OprLoadContextImpl final : public OprLoadContextFlatBuff
     size_t m_cur_opr_tensor_cnt;
     size_t m_cur_opr_blob_cnt;
     size_t m_cur_opr_param_cnt;
+    SharedTensorAlignMent* m_tensor_alignment;
 
 public:
+    friend class SharedTensorAlignMent;
+
     ComputingGraph& graph() override { return *m_graph; }
 
     const GraphLoadConfig& config() const override {
         return *m_loader->m_cur_load_config;
     }
 
+    //! shared or copy the loaded flatbuffer memory to the CPU tensor, this can reduce
+    //! the memory used when load model, but should consider the memory
+    //! alignment
+    void fill_tensor_memory(
+            HostTensorND& tensor, const uint8_t* data, size_t size, bool shared) {
+        auto tensor_size = tensor.layout().span().high_byte;
+        mgb_assert(
+                size == tensor_size,
+                "the size is not match when shared the flatbuffer memory\n");
+        auto ptr = reinterpret_cast<void*>(const_cast<uint8_t*>(data));
+        if (shared) {
+            HostTensorStorage storage;
+            auto raw_storage = std::shared_ptr<mgb::dt_byte>(
+                    static_cast<mgb::dt_byte*>(ptr), [](void*) {});
+            storage.reset(tensor.comp_node(), size, raw_storage);
+            tensor.reset(storage, tensor.layout());
+        } else {
+            memcpy(tensor.raw_ptr(), data, size);
+        }
+    }
+
     std::shared_ptr<HostTensorND> load_tensor() override;
 
-    std::shared_ptr<DeviceTensorND> load_tensor_shared() override;
+    std::shared_ptr<DeviceTensorND> load_tensor_shared(
+            bool copy_immediatly = false) override;
 
     void load_single_opr(const fbs::v2::Operator* opr);
 
-    OprLoadContextImpl(GraphLoaderOSSV2* loader, uint32_t version)
-            : OprLoadContextFlatBuffers(version), m_loader{loader} {
+    OprLoadContextImpl(
+            GraphLoaderOSSV2* loader, SharedTensorAlignMent* tensor_alignment,
+            uint32_t version)
+            : OprLoadContextFlatBuffers(version),
+              m_loader{loader},
+              m_tensor_alignment(tensor_alignment) {
         m_graph = loader->m_cur_load_config->comp_graph;
         if (!m_graph) {
             m_graph = ComputingGraph::make();

src/serialization/test/serializer_oss.cpp

@@ -315,7 +315,7 @@ void test_serializer_custom_loader(GraphDumpFormat format) {
     load();
     load();
     ASSERT_EQ(2u, saved_val.size());
-    ASSERT_EQ(1, load_nr_null_ptr);  // immutable tensor is not shared
+    ASSERT_EQ(2, load_nr_null_ptr);  // immutable tensor is also shared
     ASSERT_EQ(4, load_nr_call);
 }
 
@@ -482,10 +482,10 @@ void test_serializer_multiple_param(GraphDumpFormat format) {
         ASSERT_THROW(loader->shared_tensor_id_map(), MegBrainError);
         loader->load();
         auto&& got = loader->shared_tensor_id_map();
-        ASSERT_EQ(values.size(), got.size());
+        ASSERT_EQ(2 * values.size(), got.size());
         for (size_t i = 0; i < values.size(); ++i) {
             ASSERT_EQ(1u, got[i].second.size());
-            auto &&vi = *values[i], &&gi = *got[i].second.begin()->second;
+            auto &&vi = *values[i], &&gi = *got[2 * i].second.begin()->second;
             ASSERT_EQ(vi.shape(), gi.shape());
             ASSERT_EQ(vi.comp_node(), gi.comp_node());
             ASSERT_EQ(vi.dtype(), gi.dtype());
@@ -565,7 +565,7 @@ void test_serializer_const_var_shape(GraphDumpFormat format) {
             }
         };
         run_and_check(config);
-        ASSERT_EQ(2, nr_tensor);
+        ASSERT_EQ(1, nr_tensor);  // immutable tensor is shared tensor
         ASSERT_EQ(1, nr_mod);
     }
 }
@@ -823,6 +823,77 @@ void test_serializer_log_exp(GraphDumpFormat format) {
     load();
 }
 
+void test_serializer_memshare(GraphDumpFormat format) {
+    std::vector<uint8_t> buf;
+    HostTensorGenerator<> gen;
+    constexpr size_t SIZE = 127;
+    auto xval = gen({SIZE}, "cpu0"), bval = gen({1}, "cpu0");
+
+    auto dump = [&]() {
+        auto graph = ComputingGraph::make();
+        auto x0 = opr::SharedDeviceTensor::make(*graph, *xval).rename("x0");
+        auto x1 = opr::SharedDeviceTensor::make(*graph, *xval).rename("x1");
+        auto x2 = opr::SharedDeviceTensor::make(*graph, *xval).rename("x2");
+        auto x3 = opr::SharedDeviceTensor::make(*graph, *xval).rename("x3");
+        auto i4 = opr::ImmutableTensor::make(*graph, *xval).rename("i4");
+        auto i5 = opr::ImmutableTensor::make(*graph, *xval).rename("i5");
+        auto b = opr::SharedDeviceTensor::make(*graph, *bval).rename("b");
+        auto dumper = GraphDumper::make(OutputFile::make_vector_proxy(&buf), format);
+        dumper->dump({((x0 + x1) + b) + (x2 + x3) + i4 + i5, x0, i4});
+    };
+
+    HostTensorND expected;
+    expected.copy_from(*xval);
+    for (size_t i = 0; i < SIZE; ++i) {
+        auto&& v = expected.ptr<float>()[i];
+        v = v * 6 + bval->ptr<float>()[0];
+    }
+
+    std::vector<uint8_t> buf_al;
+    auto load = [&](bool share) {
+        std::unique_ptr<InputFile> fin;
+        if (share) {
+            buf_al.resize(buf.size());
+            memcpy(buf_al.data(), buf.data(), buf.size());
+
+            fin = InputFile::make_mem_proxy(
+                    std::shared_ptr<void>{std::shared_ptr<void>{}, buf_al.data()},
+                    buf.size());
+        } else {
+            fin = InputFile::make_mem_proxy(buf.data(), buf.size());
+        }
+        auto loader = GraphLoader::make(std::move(fin), format);
+        auto rst = loader->load();
+        auto x = rst.output_var_map.at("x0");
+        auto i4 = rst.output_var_map.at("i4");
+        auto&& opr = x.node()->owner_opr()->cast_final_safe<opr::SharedDeviceTensor>();
+        auto&& opr_imm =
+                i4.node()->owner_opr()->cast_final_safe<opr::ImmutableTensor>();
+        HostTensorND val;
+        auto func =
+                rst.graph_compile({make_callback_copy(rst.output_var_list[0], val)});
+        func->execute();
+        return std::make_pair(
+                val, std::vector<DeviceTensorND>{*opr.dev_data(), opr_imm.value()});
+    };
+
+    auto in_range = [](const std::vector<uint8_t>& buf, DeviceTensorND& dv) {
+        auto p0 = reinterpret_cast<uint8_t*>(dv.raw_ptr()),
+             p1 = reinterpret_cast<uint8_t*>(p0 + dv.layout().span().high_byte);
+        return buf.data() <= p0 && p1 <= buf.data() + buf.size();
+    };
+
+    for (bool share : {false, true}) {
+        buf.clear();
+        dump();
+        auto get = load(share);
+        MGB_ASSERT_TENSOR_EQ(*xval, HostTensorND{}.copy_from(get.second[0]).sync());
+        MGB_ASSERT_TENSOR_EQ(expected, get.first);
+        ASSERT_EQ(share, in_range(buf_al, get.second[0]));
+        ASSERT_EQ(share, in_range(buf_al, get.second[1]));
+    }
+}
+
 }  // namespace
 
 TEST(TestSerializer2, GraphDumpLoad) {
@@ -967,6 +1038,10 @@ TEST(TestSerializer2, LOGEXPV2) {
     test_serializer_log_exp(GraphDumpFormat::FLATBUFFERS_V2);
 }
 
+TEST(TestSerializer2, ShareMemv2) {
+    test_serializer_memshare(GraphDumpFormat::FLATBUFFERS_V2);
+}
+
 TEST(TestSerializer2, TestSoftMaxLoadDump) {
     auto fname = GET_OUTPUT_FILE(GraphDumpFormat::FLATBUFFERS_V2);
     TensorShape shape{2, 3};