feat(src/serialization): add support of serializing metadata

GitOrigin-RevId: b563c94451

imperative/python/megengine/core/tensor/megbrain_graph.py

@@ -11,7 +11,7 @@ import json
 import os
 import weakref
 from concurrent.futures import ThreadPoolExecutor
-from typing import Dict, List, Tuple, Union
+from typing import Dict, List, Optional, Tuple, Union
 
 import numpy as np
 
@@ -256,6 +256,9 @@ def optimize_for_inference(dest_vars, **kwargs):
             * enable_chwn4 --
                 whether to use CHWN4 data layout, currently
                 used in nvidia backend with tensorcore.
+            * enable_nchw64 --
+                whether to use NCHW64 data layout, used for fast int4
+                support on Nvidia GPU.
 
             * enable_fuse_conv_bias_nonlinearity: whether to fuse conv+bias+nonlinearty
                 into one opr.
@@ -273,6 +276,7 @@ def optimize_for_inference(dest_vars, **kwargs):
         "enable_nchw44": GraphOptimizeOptions.LayoutTransform.NCHW44,
         "enable_nchw44_dot": GraphOptimizeOptions.LayoutTransform.NCHW44_DOT,
         "enable_chwn4": GraphOptimizeOptions.LayoutTransform.CHWN4,
+        "enable_nchw64": GraphOptimizeOptions.LayoutTransform.NCHW64,
     }
 
     for k, v in inference_optimize_layout_transform_map.items():
@@ -293,7 +297,46 @@ def optimize_for_inference(dest_vars, **kwargs):
 
     dest_vars = _unwrap(dest_vars)
     res_vars = _imperative_rt.optimize_for_inference(dest_vars, inference_options)
-    return _wrap(res_vars)
+    return _wrap(res_vars), inference_options.serialize()
+
+
+def deserialize_infer_option(x: int) -> Dict[str, bool]:
+    r"""
+    Deserailize optimize options generated by ``imperative_rt.GraphOptimizeOptions``.
+
+    :param x: inference options represented by int.
+    :return: inference options represented by dict.
+    """
+
+    inference_options = GraphOptimizeOptions.deserialize(x)
+
+    inference_optimize_layout_transform_map = {
+        GraphOptimizeOptions.LayoutTransform.NHWCD4: "enable_hwcd4",
+        GraphOptimizeOptions.LayoutTransform.NCHW4: "enable_nchw4",
+        GraphOptimizeOptions.LayoutTransform.NCHW88: "enable_nchw88",
+        GraphOptimizeOptions.LayoutTransform.NCHW32: "enable_nchw32",
+        GraphOptimizeOptions.LayoutTransform.NCHW44: "enable_nchw44",
+        GraphOptimizeOptions.LayoutTransform.NCHW44_DOT: "enable_nchw44_dot",
+        GraphOptimizeOptions.LayoutTransform.CHWN4: "enable_chwn4",
+        GraphOptimizeOptions.LayoutTransform.NCHW64: "enable_nchw64",
+    }
+
+    ret = dict()
+
+    layout = inference_options.layout_transform
+    if layout != GraphOptimizeOptions.LayoutTransform.DEFAULT:
+        ret[inference_optimize_layout_transform_map[layout]] = True
+
+    if inference_options.f16_io_f32_comp:
+        ret["enable_io16xc32"] = True
+    if inference_options.f16_io_comp:
+        ret["enable_ioc16"] = True
+    if inference_options.fuse_conv_bias_nonlinearity:
+        ret["enable_fuse_conv_bias_nonlinearity"] = True
+    if inference_options.fuse_conv_bias_with_z:
+        ret["enable_fuse_conv_bias_with_z"] = True
+
+    return ret
 
 
 def modify_opr_algo_strategy_inplace(dest_vars, strategy: str):
@@ -331,7 +374,8 @@ def dump_graph(
     keep_param_name: bool = False,
     keep_opr_priority: bool = False,
     strip_info_file=None,
-    append_json=False
+    append_json=False,
+    metadata=None
 ) -> Tuple[bytes, CompGraphDumpResult]:
     """
     serialize the computing graph of `output_vars` and get byte result.
@@ -393,6 +437,7 @@ def dump_graph(
         keep_opr_name,
         keep_param_name,
         keep_opr_priority,
+        metadata,
         stat,
         inputs,
         outputs,
@@ -427,7 +472,7 @@ def dump_graph(
 
 
 CompGraphLoadResult = collections.namedtuple(
-    "CompGraphLoadResult", ["graph", "output_vars_dict", "output_vars_list"]
+    "CompGraphLoadResult", ["graph", "output_vars_dict", "output_vars_list", "metadata"]
 )
 
 
@@ -450,8 +495,8 @@ def load_graph(fpath) -> CompGraphLoadResult:
         buf = open(fpath, "rb").read()
     else:
         buf = fpath.read()
-    cg = _imperative_rt.load_graph(buf, output_vars_map, output_vars_list)
-    return CompGraphLoadResult(cg, dict(output_vars_map), output_vars_list)
+    cg, metadata = _imperative_rt.load_graph(buf, output_vars_map, output_vars_list)
+    return CompGraphLoadResult(cg, dict(output_vars_map), output_vars_list, metadata)
 
 
 def _wrap(x):

imperative/python/megengine/jit/tracing.py

@@ -12,10 +12,12 @@ import functools
 import itertools
 import json
 import os
+import pickle
+from typing import Any
 
 import numpy as np
 
-from ..core._imperative_rt import GraphProfiler
+from ..core._imperative_rt import GraphProfiler, SerializationMetadata
 from ..core._imperative_rt.core2 import Tensor as RawTensor
 from ..core._imperative_rt.core2 import (
     TensorWeakRef,
@@ -670,6 +672,8 @@ class trace:
         strip_info_file=None,
         append_json=False,
         optimize_for_inference=True,
+        user_info: Any = None,
+        enable_metadata: bool = True,
         **kwargs
     ):
         r"""
@@ -697,6 +701,8 @@ class trace:
             if set false, will rewrite strip_info_file
         :param optimize_for_inference: enbale optmizations,
             will skip all optimize options if this is False. Default: True
+        :param user_info: any type object, which will be pickled to bytes.
+        :param enable_metadata: whether to save metadata into output file.
 
         :Keyword Arguments:
 
@@ -729,6 +735,9 @@ class trace:
             * enable_chwn4 --
                 whether to use CHWN4 data layout, currently
                 used in nvidia backend with tensorcore.
+            * enable_nchw64 --
+                whether to use NCHW64 data layout, used for fast int4
+                support on Nvidia GPU.
 
             * enable_fuse_conv_bias_nonlinearity: whether to fuse conv+bias+nonlinearty
                 into one opr.
@@ -851,7 +860,15 @@ class trace:
             dest_vars.append(v)
 
         if optimize_for_inference:
-            dest_vars = G.optimize_for_inference(dest_vars, **kwargs)
+            dest_vars, optimize_options = G.optimize_for_inference(dest_vars, **kwargs)
+
+        metadata = SerializationMetadata()
+        if enable_metadata:
+            metadata.user_info = pickle.dumps(user_info)
+            metadata.is_valid = True
+            metadata.graph_modified = False
+            if optimize_for_inference:
+                metadata.optimize_options = optimize_options
 
         if isinstance(file, str):
             permission = "wb" if append == False else "ab"
@@ -864,6 +881,7 @@ class trace:
             keep_opr_priority=keep_opr_priority,
             strip_info_file=strip_info_file,
             append_json=append_json,
+            metadata=metadata,
         )
         file.write(dump_content)
         return dump_info

imperative/python/megengine/tools/load_network_and_run.py

@@ -411,7 +411,8 @@ def main():
         args.embed_input = True
 
     logger.info("loading model ...")
-    graph, _, output_vars = G.load_graph(args.net)
+    ret = G.load_graph(args.net)
+    graph, output_vars = ret.graph, ret.output_vars_list
     input_vars = tools.get_dep_vars(output_vars, "Host2DeviceCopy")
 
     if args.output_name is not None:

imperative/python/megengine/utils/comp_graph_tools.py

@@ -391,7 +391,8 @@ class GraphInference:
         optimize_for_inference: bool = False,
         **kwargs
     ):
-        self._graph, _, output_nodes = G.load_graph(file)
+        ret = G.load_graph(file)
+        self._graph, output_nodes = ret.graph, ret.output_vars_list
         if outputs is not None:
             output_nodes = find_vars_by_name(output_nodes, outputs)
         self._origin_outputs = output_nodes

imperative/python/megengine/utils/network.py

@@ -9,14 +9,12 @@
 import collections
 import fnmatch
 import itertools
+import pickle
 import re
 from collections import OrderedDict
-from typing import Dict, List, Sequence
+from typing import Any, Dict, List, Sequence
 
-import numpy as np
-
-from ..core._imperative_rt import ComputingGraph
-from ..core._imperative_rt.core2 import SymbolVar
+from ..core._imperative_rt import ComputingGraph, SerializationMetadata
 from ..core._trace_option import set_symbolic_shape as _set_symbolic_shape
 from ..core.tensor import megbrain_graph as G
 from ..logger import get_logger
@@ -42,6 +40,30 @@ class Network:
         self.all_oprs_map = OrderedDict()
         self.all_vars_map = OrderedDict()
         self.graph = ComputingGraph()
+        self._metadata = None
+
+    @property
+    def metadata(self):
+        r"""
+        Load metadata as a dict.
+        """
+        if not self._metadata.is_valid:
+            logger.info("metadata is not valid!")
+            return None
+        ret = dict()
+        try:
+            user_info = pickle.loads(self._metadata.user_info)
+        except:  # pylint: disable=bare-except
+            logger.warning(
+                "can't parse user info by pickle, so return the original bytes object!"
+            )
+            user_info = self._metadata.user_info
+        ret["user_info"] = user_info
+        ret["graph_modified"] = self._metadata.graph_modified
+        ret["optimized_for_inference"] = self._metadata.optimized_for_inference
+        if ret["optimized_for_inference"]:
+            ret.update(G.deserialize_infer_option(self._metadata.optimize_options))
+        return ret
 
     @classmethod
     def load(cls, model_path: str, outspec: List[str] = None):
@@ -51,7 +73,8 @@ class Network:
         :param outspec: only load the subgraph with outspec as its endpoints.
         """
         self = cls()
-        _, _, outputs = G.load_graph(model_path)
+        ret = G.load_graph(model_path)
+        outputs, self._metadata = ret.output_vars_list, ret.metadata
         if outspec is not None:
             output_spec = outspec.copy()
             all_vars = get_dep_vars(outputs) + outputs
@@ -125,6 +148,9 @@ class Network:
                 * enable_chwn4 --
                     whether to use CHWN4 data layout, currently
                     used in nvidia backend with tensorcore.
+                * enable_nchw64 --
+                    whether to use NCHW64 data layout, used for fast int4
+                    support on Nvidia GPU.
 
                 * enable_fuse_conv_bias_nonlinearity: whether to fuse conv+bias+nonlinearty
                     into one opr.
@@ -152,6 +178,8 @@ class Network:
         append_json=False,
         optimize_for_inference=True,
         append=False,
+        user_info: Any = None,
+        enable_metadata=True,
         **kwargs
     ):
         """
@@ -176,6 +204,8 @@ class Network:
             if set false, will rewrite strip_info_file
         :param optimize_for_inference: enbale optmizations,
             will skip all optimize options if this is False. Default: True
+        :param user_info: any type object, which will be pickled to bytes.
+        :param enable_metadata: whether to save metadata into output file.
 
         :Keyword Arguments:
 
@@ -201,7 +231,15 @@ class Network:
             )
 
         if optimize_for_inference:
-            out = G.optimize_for_inference(out, **kwargs)
+            out, optimize_options = G.optimize_for_inference(out, **kwargs)
+
+        metadata = SerializationMetadata()
+        if enable_metadata:
+            metadata.is_valid = True
+            metadata.graph_modified = True
+            metadata.user_info = pickle.dumps(user_info)
+            if optimize_for_inference:
+                metadata.optimize_options = optimize_options
 
         dump_content, _ = G.dump_graph(
             out,
@@ -211,6 +249,7 @@ class Network:
             keep_opr_priority=keep_opr_priority,
             strip_info_file=strip_info_file,
             append_json=append_json,
+            metadata=metadata,
         )
         if isinstance(file, str):
             permission = "wb" if append == False else "ab"

imperative/python/src/graph_rt.cpp

@@ -34,6 +34,7 @@ namespace ser = mgb::serialization;
 using _OptimizeForInferenceOptions = mgb::gopt::OptimizeForInferenceOptions;
 using _LayoutTransform = _OptimizeForInferenceOptions::LayoutTransform;
 using _AlgoStrategy = opr::mixin::AlgoChooserHelper::ExecutionPolicy::Strategy;
+using _SerializationMetadata = mgb::serialization::Metadata;
 
 namespace {
 class _CompGraphProfilerImpl {
@@ -240,6 +241,8 @@ void init_graph_rt(py::module m) {
 
     auto GraphOptimizeOptions = py::class_<_OptimizeForInferenceOptions>(m, "GraphOptimizeOptions")
         .def(py::init())
+        .def("serialize", &_OptimizeForInferenceOptions::serialize)
+        .def_static("deserialize", &_OptimizeForInferenceOptions::deserialize)
         .def_readwrite("f16_io_f32_comp", &_OptimizeForInferenceOptions::f16_io_f32_comp)
         .def_readwrite("f16_io_comp", &_OptimizeForInferenceOptions::f16_io_comp)
         .def_readwrite("fuse_conv_bias_nonlinearity", &_OptimizeForInferenceOptions::fuse_conv_bias_nonlinearity)
@@ -256,6 +259,7 @@ void init_graph_rt(py::module m) {
         .value("NCHW44_DOT", _LayoutTransform::NCHW44_DOT)
         .value("NCHW32", _LayoutTransform::NCHW32)
         .value("CHWN4", _LayoutTransform::CHWN4)
+        .value("NCHW64", _LayoutTransform::NCHW64)
         .export_values()
         ;
 
@@ -307,12 +311,24 @@ void init_graph_rt(py::module m) {
         })->to_string();
     });
 
+    py::class_<_SerializationMetadata>(m, "SerializationMetadata")
+        .def(py::init())
+        .def_property("user_info", [](const _SerializationMetadata& meta){return py::bytes(meta.get_user_info()); },
+            &_SerializationMetadata::set_user_info)
+        .def_readonly("optimized_for_inference", &_SerializationMetadata::optimized_for_inference)
+        .def_property("optimize_options", &_SerializationMetadata::get_optimize_options,
+            &_SerializationMetadata::set_optimize_options)
+        .def_readwrite("graph_modified", &_SerializationMetadata::graph_modified)
+        .def_readwrite("is_valid", &_SerializationMetadata::is_valid)
+        ;
+
     m.def("dump_graph", [](
         const std::vector<VarNode*>& dest_vars,
         int keep_var_name,
         bool keep_opr_name,
         bool keep_param_name,
         bool keep_opr_priority,
+        std::optional<_SerializationMetadata> metadata,
         py::list& stat,
         py::list& inputs,
         py::list& outputs,
@@ -325,7 +341,12 @@ void init_graph_rt(py::module m) {
         ser::GraphDumper::DumpConfig config{keep_var_name, keep_param_name,
                                        keep_opr_priority, keep_opr_name};
 
-        auto rst = dumper->dump(symvars, config);
+        ser::GraphDumper::DumpResult rst;
+        if (metadata)
+            rst = dumper->dump(symvars, config, *metadata);
+        else
+            rst = dumper->dump(symvars, config);
+
         for (auto i : rst.inputs) {
             inputs.append(py::cast(i));
         }
@@ -377,8 +398,10 @@ void init_graph_rt(py::module m) {
         for (const auto& var : rst.output_var_list) {
             iter.add(var);
         }
-        return rst.graph;
-
+        auto ret = py::tuple(2);
+        ret[0] = py::cast(rst.graph);
+        ret[1] = py::cast(rst.metadata);
+        return ret;
     });
 
 #define CURRENT_CLASS cg::ComputingGraph::Options

imperative/python/test/unit/jit/test_tracing.py

@@ -239,8 +239,7 @@ def test_dump_volatile():
     file = io.BytesIO()
     f.dump(file, optimize_for_inference=False)
     file.seek(0)
-    cg, _, outputs = G.load_graph(file)
-    (out,) = outputs
+    (out,) = G.load_graph(file).output_vars_list
     assert (
         cgtools.get_owner_opr_type(cgtools.get_owner_opr_inputs(out)[1])
         == "ImmutableTensor"
@@ -337,12 +336,12 @@ def test_goptions_log_exp():
     f(tensor(1.0))
     _, out = mkstemp()
     f.dump(out, optimize_for_inference=False)
-    *_, outputs = G.load_graph(out)
+    outputs = G.load_graph(out).output_vars_list
     oprs_1 = cgtools.get_oprs_seq(outputs)
 
     g(tensor(1.0))
     g.dump(out, optimize_for_inference=False)
-    *_, outputs = G.load_graph(out)
+    outputs = G.load_graph(out).output_vars_list
     oprs_2 = cgtools.get_oprs_seq(outputs)
 
     assert len(oprs_1) - len(oprs_2) == 2

imperative/python/test/unit/utils/test_cgtools.py

@@ -88,7 +88,7 @@ def test_graph_traversal():
     file = io.BytesIO()
     fun.dump(file, optimize_for_inference=False)
     file.seek(0)
-    cg, _, outputs = mgb_graph.load_graph(file)
+    outputs = mgb_graph.load_graph(file).output_vars_list
 
     _, map_vars, var2oprs, *_ = cgtools.graph_traversal(outputs)
     input_var = map_vars[1]
@@ -101,7 +101,9 @@ def test_load_refcnt():
     graph = mgb_graph.Graph()
     varnode = graph.make_const(0)
     buf, _ = mgb_graph.dump_graph([varnode])
-    graph, _, (varnode,) = mgb_graph.load_graph(io.BytesIO(buf))
+    ret = mgb_graph.load_graph(io.BytesIO(buf))
+    graph, (varnode,) = ret.graph, ret.output_vars_list
+    del ret
     del graph
     varnode.owner
 
@@ -132,7 +134,7 @@ def test_get_opr_seq():
     file = io.BytesIO()
     func.dump(file, optimize_for_inference=False)
     file.seek(0)
-    *_, outputs = mgb_graph.load_graph(file)
+    outputs = mgb_graph.load_graph(file).output_vars_list
 
     seq_1 = cgtools.get_oprs_seq(outputs, True)
     assert len(seq_1) == 5

imperative/python/test/unit/utils/test_dump_naming.py

@@ -35,7 +35,7 @@ def _dump_and_load(func, symbolic, keep_opr_name=True):
         keep_var_name=2,
     )
     file.seek(0)
-    *_, outputs = G.load_graph(file)
+    outputs = G.load_graph(file).output_vars_list
     ops = cgtools.get_oprs_seq(outputs)
     return ops
 
@@ -223,7 +223,7 @@ def test_catch_input_name(tensor_name, var_name):
     file = io.BytesIO()
     func.dump(file, optimize_for_inference=False, keep_opr_name=True, keep_var_name=2)
     file.seek(0)
-    *_, outputs = G.load_graph(file)
+    outputs = G.load_graph(file).output_vars_list
     op = cgtools.get_oprs_seq(outputs)[-1]
     assert op.inputs[0].name == var_name
 

imperative/python/test/unit/utils/test_network.py

@@ -14,6 +14,50 @@ from megengine.utils.network import as_oprnode, set_symbolic_shape
 from megengine.utils.network_node import Host2DeviceCopy, VarNode
 
 
+def test_metadata():
+    x = Tensor(0)
+
+    @trace(symbolic=True, capture_as_const=True)
+    def fwd(x):
+        return x * 2
+
+    fwd(x)
+
+    orig_model = io.BytesIO()
+    fwd.dump(orig_model, user_info="test", optimize_for_inference=False)
+    orig_model.seek(0)
+    graph = Net.load(orig_model)
+    assert graph.metadata == {
+        "user_info": "test",
+        "graph_modified": False,  # False: tracing.dump
+        "optimized_for_inference": False,
+    }
+
+    orig_model.seek(0)
+    graph.dump(
+        orig_model,
+        user_info={"str": "x", "tensor": x, "module": M.Module, "none": None},
+        optimize_for_inference=True,
+        enable_nchw4=True,
+        enable_ioc16=True,
+    )
+    orig_model.seek(0)
+    graph = Net.load(orig_model)
+    assert graph.metadata == {
+        "user_info": {"str": "x", "tensor": x, "module": M.Module, "none": None},
+        "graph_modified": True,  # True: Network.dump
+        "optimized_for_inference": True,
+        "enable_nchw4": True,
+        "enable_ioc16": True,
+    }
+
+    orig_model.seek(0)
+    fwd.dump(orig_model, enable_metadata=False)
+    orig_model.seek(0)
+    graph = Net.load(orig_model)
+    assert graph.metadata is None
+
+
 def test_replace_var():
 
     a = Tensor([1, 2])

sdk/load-and-run/dump_with_testcase_mge.py

@@ -170,7 +170,8 @@ def gen_one_testcase(args, inputs, spec):
 
 
 def make_feeds(args):
-    cg_rt, _, outputs = G.load_graph(args.input)
+    ret = G.load_graph(args.input)
+    cg_rt, outputs = ret.graph, ret.output_vars_list
     inputs = cgtools.get_dep_vars(outputs, "Host2DeviceCopy")
 
     inputs = {i.name: i for i in inputs}

src/gopt/include/megbrain/gopt/inference.h

@@ -322,7 +322,31 @@ namespace gopt {
         static std::unique_ptr<EnableNchw44DotPass> make_nchw44_dot_converter();
     };
 
-    struct OptimizeForInferenceOptions : cg::GraphCommonOptimizeOptions {};
+    struct OptimizeForInferenceOptions : cg::GraphCommonOptimizeOptions {
+        uint64_t serialize() {
+            uint64_t ret = 0;
+            ret |= (uint64_t)layout_transform << 32;
+            if (f16_io_f32_comp) ret |= 1u;
+            if (f16_io_comp) ret |= 1u << 1;
+            if (fuse_conv_bias_nonlinearity) ret |= 1u << 2;
+            if (fuse_conv_bias_with_z) ret |= 1u << 3;
+            if (weight_preprocess) ret |= 1u << 4;
+            if (fuse_preprocess) ret |= 1u << 5;
+            return ret;
+        }
+
+        static OptimizeForInferenceOptions deserialize(uint64_t buf) {
+            OptimizeForInferenceOptions ret;
+            ret.f16_io_f32_comp = buf & 1u;
+            ret.f16_io_comp = buf & 1u << 1;
+            ret.fuse_conv_bias_nonlinearity = buf & 1u << 2;
+            ret.fuse_conv_bias_with_z = buf & 1u << 3;
+            ret.weight_preprocess = buf & 1u << 4;
+            ret.fuse_preprocess = buf & 1u << 5;
+            ret.layout_transform = (LayoutTransform)(buf >> 32);
+            return ret;
+        }
+    };
 
     /*!
      * \brief optimize a computing graph for inference

src/serialization/impl/schema.fbs

@@ -128,6 +128,13 @@ table Operator {
     name:string;
 }
 
+table Metadata {
+    is_valid:bool;
+    graph_modified:bool;
+    user_info:string;
+    optimize_options:ulong;
+}
+
 struct OutputVar {
     compact_id:uint;
     original_id:uint;
@@ -141,6 +148,7 @@ table Graph {
     nr_shared_tensor:uint;
     oprs:[Operator];
     output_vars_idx:[OutputVar];
+    metadata:Metadata;
 }
 
 root_type Graph;

src/serialization/impl/serializer_oss.cpp

@@ -30,6 +30,7 @@
 #include "megbrain/serialization/internal/flatbuffers_helper.h"
 #include "megbrain/serialization/internal/schema_generated.h"
 #include "megbrain/serialization/opr_load_dump.h"
+#include "megbrain/serialization/metadata.h"
 #include "megbrain/serialization/serializer.h"
 #include "megbrain/version.h"
 
@@ -115,6 +116,7 @@ class GraphDumperOSS final : public GraphDumper, OprDumpContextFlatBuffers {
     std::vector<flatbuffers::Offset<void>> m_cur_opr_param;
 
     void init_oprs_to_dump(const SymbolVarArray& endpoints);
+    flatbuffers::Offset<fbs::Metadata> build_metadata(const Metadata& metadata);
     flatbuffers::Offset<fbs::Operator> build_single_opr(
             cg::OperatorNodeBase* opr, const OprRegistry* registry);
 
@@ -123,7 +125,8 @@ class GraphDumperOSS final : public GraphDumper, OprDumpContextFlatBuffers {
 public:
     GraphDumperOSS(std::unique_ptr<OutputFile> file) : m_file{std::move(file)} {}
     DumpResult dump(const SymbolVarArray& output_vars,
-                    const DumpConfig& config = {}) override;
+                    const DumpConfig& config = {},
+                    const Metadata& metadata = {}) override;
     const GraphDumpConfig& config() const override { return m_config; }
     void dump_tensor(const std::string& name, const HostTensorND& tensor,
                      TensorWriteMethod method) override;
@@ -185,6 +188,17 @@ void GraphDumperOSS::init_oprs_to_dump(const SymbolVarArray& endpoints) {
     }
 }
 
+flatbuffers::Offset<fbs::Metadata> GraphDumperOSS::build_metadata(
+    const Metadata& metadata) {
+    auto user_info = m_builder.CreateSharedString(metadata.user_info);
+    fbs::MetadataBuilder builder(m_builder);
+    builder.add_is_valid(metadata.is_valid);
+    builder.add_graph_modified(metadata.graph_modified);
+    builder.add_user_info(user_info);
+    builder.add_optimize_options(metadata.optimize_options);
+    return builder.Finish();
+}
+
 flatbuffers::Offset<fbs::Operator> GraphDumperOSS::build_single_opr(
         cg::OperatorNodeBase* opr, const OprRegistry* registry) {
     m_cur_opr = opr;
@@ -282,7 +296,8 @@ flatbuffers::Offset<fbs::Operator> GraphDumperOSS::build_single_opr(
 }
 
 GraphDumper::DumpResult GraphDumperOSS::dump(
-        const SymbolVarArray& output_vars, const DumpConfig& config) {
+        const SymbolVarArray& output_vars,
+        const DumpConfig& config, const Metadata& metadata) {
     mgb_throw_if(output_vars.empty(), SerializationError,
                  "Can't dump empty graph");
 
@@ -323,6 +338,9 @@ GraphDumper::DumpResult GraphDumperOSS::dump(
     uint64_t offset_to_fbs = 0;
     m_file->write(&offset_to_fbs, sizeof(offset_to_fbs));
 
+    // Dump metadata
+    auto fbmeta = build_metadata(metadata);
+
     // Dump operators
     init_oprs_to_dump(output_vars);
     std::vector<flatbuffers::Offset<fbs::Operator>> oprs;
@@ -350,6 +368,7 @@ GraphDumper::DumpResult GraphDumperOSS::dump(
     graph.add_oprs(fb_oprs);
     graph.add_output_vars_idx(fb_output_vars);
     graph.add_nr_shared_tensor(m_nr_shared_tensor);
+    graph.add_metadata(fbmeta);
     m_builder.FinishSizePrefixed(graph.Finish(), fbs::GraphIdentifier());
 
     // Write actual offset_to_fbs
@@ -531,6 +550,7 @@ public:
         mgb_assert(nr == 1);
     }
 
+    Metadata load_metadata();
     LoadResult load_oprs();
     CompNode load_comp_node(const fbs::CompNode* comp_node);
 
@@ -700,6 +720,22 @@ GraphLoaderOSS::OprLoadContextImpl::load_tensor_shared() {
     return sh_ptr_ref;
 }
 
+Metadata GraphLoaderOSS::OprLoadContextImpl::load_metadata() {
+    const auto* fbmeta = m_loader->m_graph->metadata();
+    Metadata ret;
+    ret.is_valid = fbmeta->is_valid();
+    ret.graph_modified = fbmeta->graph_modified();
+    if (fbmeta->user_info()) {
+        ret.user_info = fbmeta->user_info()->str();
+        ret.has_user_info = true;
+    }
+    if (fbmeta->optimize_options()) {
+        ret.optimize_options = fbmeta->optimize_options();
+        ret.optimized_for_inference = true;
+    }
+    return ret;
+}
+
 void GraphLoaderOSS::OprLoadContextImpl::load_single_opr(
         const fbs::Operator* fbopr) {
     m_cur_opr_tensor_cnt = 0;
@@ -872,7 +908,9 @@ GraphLoader::LoadResult GraphLoaderOSS::load(const LoadConfig& config,
     }
 
     OprLoadContextImpl ctx{this, m_graph->mgb_version()};
+    auto metadata = ctx.load_metadata();
     auto result = ctx.load_oprs();
+    result.metadata = metadata;
 
     auto fbs_end = tensor_begin + offset_to_fbs + sizeof(size) + size;
     auto cur = m_file->tell();

src/serialization/include/megbrain/serialization/metadata.h

@@ -0,0 +1,46 @@
+/**
+ * \file src/serialization/include/megbrain/serialization/metadata.h
+ *
+ * This file is part of MegBrain, a deep learning framework developed by Megvii.
+ *
+ * \brief MegEngine model's metadata
+ *
+ * \copyright Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ */
+#pragma once
+
+#include <string>
+
+namespace mgb {
+namespace serialization {
+
+struct Metadata {
+    bool is_valid = false;
+
+    bool graph_modified = false;
+
+    bool has_user_info = false;
+    std::string user_info;
+
+    bool optimized_for_inference = false;
+    uint64_t optimize_options;
+
+#define ADD_PROPERTY(type, name)       \
+    type get_##name() const { return name; } \
+    void set_##name(type x) {          \
+        name = x;                      \
+        has_##name = true;             \
+    }
+ADD_PROPERTY(std::string, user_info)
+#undef ADD_PROPERTY
+
+    uint64_t get_optimize_options() { return optimize_options; }
+    void set_optimize_options(uint64_t value) {
+        optimized_for_inference = true;
+        optimize_options = value;
+    }
+};
+
+}  // namespace serialization
+}  // namespace mgb

src/serialization/include/megbrain/serialization/serializer.h

@@ -15,6 +15,7 @@
 #include "megbrain/serialization/dump_format.h"
 #include "megbrain/serialization/file.h"
 #include "megbrain/serialization/load_dump_config.h"
+#include "megbrain/serialization/metadata.h"
 
 namespace mgb {
 namespace serialization {
@@ -32,6 +33,9 @@ namespace serialization {
                 //! expliit dtor decl to reduce binary size
                 ~LoadResult() noexcept;
 
+                //! metadata
+                Metadata metadata;
+
                 using TensorMap = std::unordered_map<
                     std::string, std::shared_ptr<HostTensorND>>;
 
@@ -178,7 +182,8 @@ namespace serialization {
 
             virtual DumpResult dump(
                     const SymbolVarArray &output_vars,
-                    const DumpConfig &config = {}) = 0;
+                    const DumpConfig &config = {},
+                    const Metadata &metadata = {}) = 0;
             
             virtual GraphDumpFormat format() const = 0;
     };

src/serialization/test/serializer_oss.cpp

@@ -92,6 +92,43 @@ TEST(TestSerializer2, MultiGraphDumpLoad) {
     load();
 }
 
+TEST(TestSerializer2, Metadata) {
+    auto fname = GET_OUTPUT_FILE();
+    TensorShape shape{2, 3};
+
+    auto dump = [&]() {
+        auto cn = CompNode::load("xpu0");
+        auto host_x = std::make_shared<HostTensorND>(cn, shape),
+             host_y = std::make_shared<HostTensorND>(cn, shape);
+        auto graph = ComputingGraph::make();
+        auto x = opr::Host2DeviceCopy::make(*graph, host_x, {"x"}),
+             y = opr::Host2DeviceCopy::make(*graph, host_y, {"y"});
+        using Mode = opr::Elemwise::Mode;
+        auto z = opr::Elemwise::make({x, y}, Mode::ADD, {"add(x, y)"});
+
+        Metadata metadata;
+        metadata.user_info = "TEST_METADATA";
+        metadata.has_user_info = true;
+
+        auto dumper = GraphDumper::make(OutputFile::make_fs(fname.c_str()),
+                                        GraphDumpFormat::FLATBUFFERS);
+        auto rst = dumper->dump({z.rename("z")}, {}, metadata);
+    };
+
+    auto load = [&]() {
+        HostTensorGenerator<> gen;
+        auto loader = GraphLoader::make(InputFile::make_fs(fname.c_str()),
+                                        GraphDumpFormat::FLATBUFFERS);
+        auto rst = loader->load();
+        auto metadata = rst.metadata;
+        int cmp = strcmp(metadata.user_info.c_str(), "TEST_METADATA");
+        EXPECT_EQ(cmp, 0);
+    };
+
+    dump();
+    load();
+}
+
 TEST(TestSerializer2, APlusB) {
     auto fname = GET_OUTPUT_FILE();
     TensorShape shape{2, 3};