feat(mge): add jit.trace

GitOrigin-RevId: ec647324c0

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

@@ -17,15 +17,31 @@ from ..ops.builtin import OpDef
 from .core import OpBase, TensorBase, apply
 
 
-class CompiledFunction:
-    def __init__(self, graph, function):
-        self._graph = graph
-        self._function = function
+class Graph(_imperative_rt.ComputingGraph):
+    def __init__(self):
+        super().__init__()
+        self._var_cache = weakref.WeakKeyDictionary()
+        self._op_cache = weakref.WeakKeyDictionary()
+        self._executor = ThreadPoolExecutor(1)
+        self._function = None
         self._future = None
 
+    def _wrap(self, obj):
+        if type(obj) is _imperative_rt.VarNode:
+            wrapper, cache = VarNode, self._var_cache
+        elif type(obj) is _imperative_rt.OperatorNode:
+            wrapper, cache = OpNode, self._op_cache
+        if obj not in cache:
+            cache[obj] = wrapper(obj)
+        return cache[obj]
+
+    def compile(self, *args):
+        self._function = super().compile(_unwrap(args))
+        return self
+
     def execute(self, *args):
         assert self._future is None
-        self._future = self._graph._executor.submit(self._function.execute, *args)
+        self._future = self._executor.submit(self._function.execute, *args)
 
     def wait(self):
         assert self._future is not None
@@ -40,30 +56,23 @@ class CompiledFunction:
         self.execute(*args)
         return self.wait()
 
+    def make_const(self, data, dtype=None, device=None):
+        if isinstance(data, _imperative_rt.DeviceTensorND):
+            assert dtype is None and device is None
+            return self._wrap(_imperative_rt.make_shared(self, data))
+        else:
+            device = as_device(device).to_c()
+            return self._wrap(_imperative_rt.make_const(self, data, device, dtype))
 
-class Graph(_imperative_rt.ComputingGraph):
-    def __init__(self):
-        super().__init__()
-        self._var_cache = weakref.WeakKeyDictionary()
-        self._op_cache = weakref.WeakKeyDictionary()
-        self._executor = ThreadPoolExecutor(1)
-
-    def _wrap(self, obj):
-        if type(obj) is _imperative_rt.VarNode:
-            wrapper, cache = VarNode, self._var_cache
-        elif type(obj) is _imperative_rt.OperatorNode:
-            wrapper, cache = OpNode, self._op_cache
-        if obj not in cache:
-            cache[obj] = wrapper(obj)
-        return cache[obj]
-
-    def compile(self, *args):
-        return CompiledFunction(self, super().compile(_unwrap(args)))
+    def make_input(self, *args: "VarNode", device=None, dtype=None, shape=None):
+        opnode = InputNode(*args, device=device, dtype=dtype, shape=shape, graph=self)
+        return opnode.outputs[0]
 
 
 class VarNode(TensorBase):
     def __init__(self, node: _imperative_rt.VarNode):
         self._node = node
+        self.graph._var_cache[node] = self
 
     @property
     def graph(self) -> Graph:
@@ -81,10 +90,15 @@ class VarNode(TensorBase):
     def device(self):
         return as_device(self._node.comp_node)
 
+    @property
+    def shape(self):
+        return self._node.shape
+
 
 class OpNode:
     def __init__(self, node: _imperative_rt.OperatorNode):
         self._node = node
+        self.graph._op_cache[node] = self
 
     @property
     def graph(self) -> Graph:
@@ -117,21 +131,21 @@ def _(op: OpDef, *args: VarNode):
     return _wrap(outputs)
 
 
-def input_callback(callback, *args, device=None, dtype=None, graph=None):
+def input_callback(callback, *args, device=None, dtype=None, shape=None, graph=None):
     outputs = _imperative_rt.input_callback(
-        callback, as_device(device).to_c(), dtype, _unwrap(args), graph=graph
+        callback, as_device(device).to_c(), dtype, shape, _unwrap(args), graph=graph
     )
     value, dummy = _wrap(outputs)
     return value, dummy
 
 
 class InputNode(OpNode):
-    def __init__(self, *args: VarNode, device=None, dtype=None, graph=None):
+    def __init__(self, *args: VarNode, device=None, dtype=None, shape=None, graph=None):
         r = _imperative_rt.DeviceTensorNDRendezvous()
         if device is not None:
             device = as_device(device).to_c()
         outputs = _imperative_rt.input_callback(
-            r, device, dtype, _unwrap(args), graph=graph
+            r, device, dtype, shape, _unwrap(args), graph=graph
         )
         super().__init__(outputs[0].owner)
         self._rendezvous = r
@@ -169,6 +183,29 @@ class OutputNode(OpNode):
     def get_value(self):
         return self._rendezvous.get()
 
+    def drop_value(self):
+        self._rendezvous.drop()
+
+    def reset(self):
+        self._rendezvous.reset()
+
+
+class ValueOutputNode(OpNode):
+    def __init__(self, var, *args):
+        args = (var,) + args
+        r = _imperative_rt.HostTensorNDRendezvous()
+        dummy = _imperative_rt.value_output_callback(r, _unwrap(args))
+        super().__init__(dummy.owner)
+        self._rendezvous = r
+
+    def get_value(self):
+        hostnd, event = self._rendezvous.get()
+        event.wait()
+        return hostnd.numpy()
+
+    def drop_value(self):
+        self._rendezvous.drop()
+
     def reset(self):
         self._rendezvous.reset()
 
@@ -192,5 +229,8 @@ class AttrOutputNode(OpNode):
         attr = self._rendezvous.get()
         return TensorAttr(attr.shape, attr.dtype, as_device(attr.comp_node))
 
+    def drop_value(self):
+        self._rendezvous.drop()
+
     def reset(self):
         self._rendezvous.reset()

imperative/python/megengine/core/tensor/raw_tensor/__init__.py

@@ -31,11 +31,13 @@ class RawTensor(TensorBase):
 
     _init_cb = None
     _del_cb = None
+    _handle = None
 
-    def __init__(self, handle):
+    def __init__(self, handle=None):
         self._handle = handle
-        if self._init_cb:
-            self._init_cb()
+        if handle is not None:
+            if self._init_cb:
+                self._init_cb()
 
     @property
     def dtype(self):
@@ -61,9 +63,10 @@ class RawTensor(TensorBase):
         )
 
     def __del__(self):
-        if self._del_cb:
-            self._del_cb()
-        delete(self._handle)
+        if self._handle is not None:
+            if self._del_cb:
+                self._del_cb()
+            delete(self._handle)
 
 
 @apply.register()
@@ -89,6 +92,11 @@ def as_raw_tensor(obj, dtype=None, device=None):
     return as_raw_tensor(obj, device=device)
 
 
+@as_raw_tensor.register(DeviceTensorND)
+def _(data: DeviceTensorND):
+    return RawTensor(put(data))
+
+
 @as_raw_tensor.register(np.ndarray)
 def _(array: np.ndarray, dtype=None, device=None):
     device = None if device is None else as_device(device).to_c()

imperative/python/megengine/jit/__init__.py

@@ -0,0 +1 @@
+from .tracing import exclude_from_trace, trace

imperative/python/megengine/jit/tracing.py

@@ -0,0 +1,514 @@
+import contextlib
+import functools
+import typing
+import weakref
+
+from ..core.ops.special import Const
+from ..core.tensor import megbrain_graph as G
+from ..core.tensor.core import OpBase, apply
+from ..core.tensor.raw_tensor import OpDef, RawTensor, as_raw_tensor
+
+
+class TraceMismatchError(RuntimeError):
+    pass
+
+
+active_trace = None
+skip_tracing = False
+
+
+@contextlib.contextmanager
+def exclude_from_trace():
+    global skip_tracing
+    if skip_tracing:
+        yield
+        return
+    try:
+        skip_tracing = True
+        if active_trace is not None:
+            active_trace._begin_excluded_region()
+        yield
+    finally:
+        skip_tracing = False
+
+
+class TensorInfo:
+    __slots__ = (
+        # collected attributes
+        "external",
+        "exported",
+        "data_read",
+        "shape_read",
+        "value_read",
+        "device",
+        "dtype",
+        "bound_data",
+        # resources for execution
+        "varnode",
+        "data_setter",
+        "shape_reader",
+        "value_reader",
+        "data_reader",
+    )
+
+    def __init__(self):
+        self.exported = None
+        self.data_read = None
+        self.shape_read = None
+        self.value_read = None
+        self.bound_data = None
+
+        self.data_setter = None
+        self.shape_reader = None
+        self.value_reader = None
+        self.data_reader = None
+
+
+class trace:
+    def __new__(cls, *args, **kwargs):
+        if not args:
+            return functools.partial(cls, **kwargs)
+        self = super().__new__(cls)
+        self.__init__(*args, **kwargs)
+        return self
+
+    def __init__(self, function, symbolic=False, capture_as_const=False):
+        self.__wrapped__ = function
+        self._symbolic = symbolic
+        self._capture_as_const = capture_as_const
+        self._capture_static_shape = False
+
+        self._untraced = True
+        self._tinfo = []  # handle -> TensorInfo
+        self._seq = []
+        self._pc = 0
+        self._graph = None
+        self._need_reset_nodes = None
+        self._lazy_eval_graph = None
+        self._lazy_eval_tensors = weakref.WeakSet()
+        self._active_tensors = weakref.WeakSet()
+
+    def _new_handle(self):
+        handle = len(self._tinfo)
+        info = TensorInfo()
+        self._tinfo.append(info)
+        return handle, info
+
+    def _apply_op(self, op, args):
+        assert not self._untraced
+        # check against trace
+        if self._pc >= len(self._seq):
+            raise TraceMismatchError("trace should end here, but more op observed")
+        record = self._seq[self._pc]
+        op_, ihandles, ohandles = record
+        if op != op_:
+            raise TraceMismatchError("op different from last time")
+        if len(ihandles) != len(args):
+            raise TraceMismatchError("op input size different from last time")
+
+        for h, x in zip(ihandles, args):
+            info = self._tinfo[h]
+            if info.external:
+                if (
+                    x.__class__ is CompiledTensorProxy
+                    and not self._tinfo[x._CompiledTensorProxy__handle].exported
+                ):
+                    raise TraceMismatchError(
+                        "failed to capture: input was an external tensor "
+                        "last time, got an internal tensor this time"
+                    )
+                if info.bound_data:
+                    if x.__class__ is CompiledTensorProxy:
+                        raise TraceMismatchError(
+                            "const capture violated: was an external tensor "
+                            "last time, got an internal tensor this time"
+                        )
+                    if x._handle != info.bound_data._handle:
+                        raise TraceMismatchError(
+                            "const capture violated: got "
+                            "a different tensor this time"
+                        )
+                else:
+                    if info.dtype != x.dtype:
+                        raise TraceMismatchError(
+                            "failed to capture: different dtype from last time"
+                        )
+                    if info.device != x.device:
+                        raise TraceMismatchError(
+                            "failed to capture: different device from last time"
+                        )
+                    info.data_setter.set_value(x._dev_tensor())
+            else:
+                if x.__class__ is not CompiledTensorProxy:
+                    raise TraceMismatchError(
+                        "unexpected capture: trying to use an external tensor as input, "
+                        "but that input was an internal tensor last time"
+                    )
+                if x._CompiledTensorProxy__handle != h:
+                    raise TraceMismatchError(
+                        "mis-wiring: input edge to an data flow "
+                        "graph node is different from last time"
+                    )
+
+        self._pc += 1
+        outputs = tuple([CompiledTensorProxy(h) for h in ohandles])
+        self._active_tensors.update(outputs)
+        return outputs
+
+    def _record_op(self, op, inputs, outputs):
+        if skip_tracing:
+            for x in inputs:
+                h = getattr(x, "_TraceMixin__handle", None)
+                if h is not None:
+                    self._tinfo[h].data_read = True
+            return
+
+        ihandles = []
+        for x in inputs:
+            h = getattr(x, "_TraceMixin__handle", None)
+            if h is None or (not self._capture_as_const and self._tinfo[h].exported):
+                h, info = self._new_handle()
+                info.external = True
+                info.device = x.device
+                info.dtype = x.dtype
+                if self._capture_as_const:
+                    info.bound_data = x
+
+            ihandles.append(h)
+
+        ohandles = []
+        for x in outputs:
+            h, info = self._new_handle()
+            ohandles.append(h)
+            info.external = False
+            TraceMixin._TraceMixin__inject(x, h)
+
+        self._seq.append((op, tuple(ihandles), tuple(ohandles)))
+        self._active_tensors.update(outputs)
+
+    @contextlib.contextmanager
+    def _setup(self):
+        global active_trace
+        if active_trace:
+            raise NotImplementedError("sorry, not implemented: nested trace")
+        active_trace = self
+
+        if self._untraced:
+            apply.enable(apply_with_tracing)
+            if self._symbolic:
+                apply.enable(apply_symbolic_mode)
+                self._lazy_eval_graph = G.Graph()
+        else:
+            apply.enable(apply_compiled_mode)
+            if self._graph is None:
+                self._compile()
+            self._graph.execute()
+
+        yield
+
+        escaped_tensors = tuple(self._active_tensors)
+        self._active_tensors.clear()
+
+        if self._untraced:
+            for x in escaped_tensors:
+                info = self._tinfo[x._TraceMixin__handle]
+                info.data_read = True
+                x._TraceMixin__restore()
+            if self._symbolic:
+                # eval lazy eval tensors
+                lazy_eval_tensors = tuple(self._lazy_eval_tensors)
+                if lazy_eval_tensors:
+                    readers = [
+                        G.OutputNode(x._LazyEvalTensor__varnode).outputs[0]
+                        for x in lazy_eval_tensors
+                    ]
+                    self._lazy_eval_graph.compile(*readers)
+                    self._lazy_eval_graph()
+                    for r, x in zip(readers, lazy_eval_tensors):
+                        assign_raw_tensor(x, as_raw_tensor(r.op.get_value()))
+                    self._lazy_eval_graph = None
+                    self._lazy_eval_tensors = None
+            self._untraced = False
+        else:
+            if self._pc != len(self._seq):
+                raise TraceMismatchError("premature end")
+            for x in escaped_tensors:
+                assign_raw_tensor(x, as_raw_tensor(x._dev_tensor()))
+            self._graph.wait()
+            self._reset_exec_env()
+            self._pc = 0
+
+        apply.disable(apply_with_tracing)
+        apply.disable(apply_symbolic_mode)
+        apply.disable(apply_compiled_mode)
+        active_trace = None
+
+    def _begin_excluded_region(self):
+        if self._untraced:
+            # conditionally reading a compiled tensor in excluded region
+            # is permitted, so we have to assume every tensor might be read
+            for x in self._active_tensors:
+                info = self._tinfo[x._TraceMixin__handle]
+                info.exported = True
+                info.data_read = True
+
+    def _compile(self):
+        graph = self._graph = G.Graph()
+        # graph.options.graph_opt_level = 0
+        need_reset_nodes = self._need_reset_nodes = []
+        # links enforce ordering of I/O nodes
+        links = ()
+        for op, ihandles, ohandles in self._seq:
+            ivars = []
+            readers = []
+            for h in ihandles:
+                info = self._tinfo[h]
+                if not hasattr(info, "varnode"):
+                    assert info.external
+                    if info.bound_data:
+                        info.varnode = graph.make_const(info.bound_data._dev_tensor())
+                    else:
+                        opnode = info.data_setter = G.InputNode(
+                            *links, device=info.device, dtype=info.dtype, graph=graph
+                        )
+                        need_reset_nodes.append(opnode)
+                        info.varnode, *links = opnode.outputs
+
+                ivars.append(info.varnode)
+            ovars = apply(op, *ivars)
+            assert len(ovars) == len(ohandles)
+            for h, v in zip(ohandles, ovars):
+                info = self._tinfo[h]
+                info.varnode = v
+
+                def add_reader(opnode):
+                    nonlocal links
+                    need_reset_nodes.append(opnode)
+                    readers.append(opnode.outputs[0])
+                    links = opnode.outputs
+
+                if info.data_read:
+                    # Shape can be obtained from data so doesn't need its own
+                    # output node. On the other hand, value is read separately
+                    # to leverage eager h2d copy
+                    info.shape_read = False
+                    opnode = info.data_reader = G.OutputNode(v, *links)
+                    add_reader(opnode)
+                if info.value_read:
+                    opnode = info.value_reader = G.ValueOutputNode(v, *links)
+                    add_reader(opnode)
+                if info.shape_read:
+                    opnode = info.shape_reader = G.AttrOutputNode(v, *links)
+                    add_reader(opnode)
+
+        graph.compile(*readers)
+
+    def _reset_exec_env(self):
+        for opnode in self._need_reset_nodes:
+            opnode.reset()
+
+    def _require_shape(self, handle):
+        info = self._tinfo[handle]
+        info.shape_read = True
+
+    def _require_value(self, handle):
+        info = self._tinfo[handle]
+        info.value_read = True
+
+    def _require_data(self, handle):
+        info = self._tinfo[handle]
+        info.data_read = True
+
+    def __call__(self, *args, **kwargs):
+        with self._setup():
+            return self.__wrapped__(*args, **kwargs)
+
+
+class CompiledTensorProxy(RawTensor):
+    """
+    Duck-typed RawTensor
+    """
+
+    def __init__(self, handle):
+        self.__handle = handle
+        self.__info = active_trace._tinfo[handle]
+        self.__shape = None
+        self.__data = None
+        self.__value = None
+
+    @property
+    def dtype(self):
+        return self.__info.varnode.dtype
+
+    @property
+    def device(self):
+        return self.__info.varnode.device
+
+    @property
+    def shape(self):
+        if self.__shape is None:
+            if self.__info.shape_read:
+                self.__shape = self.__info.shape_reader.get_value().shape
+            elif self.__info.data_read:
+                self.__shape = self._dev_tensor().shape
+            else:
+                raise TraceMismatchError("shape of this tensor is not read in trace")
+        return self.__shape
+
+    def numpy(self):
+        if self.__value is None:
+            if self.__info.value_read:
+                self.__value = self.__info.value_reader.get_value()
+            elif self.__info.data_read:
+                self.__value = self._dev_tensor().numpy()
+            else:
+                raise TraceMismatchError("value of this tensor is not read in trace")
+        return self.__value
+
+    def _dev_tensor(self):
+        if self.__data is None:
+            if not self.__info.data_read:
+                raise TraceMismatchError("raw data of this tensor is not read in trace")
+            self.__data = self.__info.data_reader.get_value()
+        return self.__data
+
+    def __del__(self):
+        if self.__info.shape_read and self.__shape is not None:
+            self.__info.shape_reader.drop_value()
+        if self.__info.value_read and self.__value is not None:
+            self.__info.value_reader.drop_value()
+        if self.__info.data_read and self.__data is not None:
+            self.__info.data_reader.drop_value()
+
+
+class LazyEvalTensor(RawTensor):
+    def __init__(self, varnode):
+        self.__varnode = varnode
+
+    @property
+    def dtype(self):
+        return self.__varnode.dtype
+
+    @property
+    def device(self):
+        return self.__varnode.device
+
+    @property
+    def shape(self):
+        return self.__varnode.shape
+
+    def numpy(self):
+        raise RuntimeError("cannot read value during symbolic tracing")
+
+    def _dev_tensor(self):
+        raise RuntimeError("cannot access data during symbolic tracing")
+
+
+class TraceMixin:
+    __subclass_cache = {}
+
+    def __inject(self, handle):
+        cache = __class__.__subclass_cache
+        cls = self.__class__
+        subcls = cache.get(cls)
+        if subcls is None:
+            subcls = cache[cls] = type("Traced" + cls.__name__, (__class__, cls), {})
+        self.__class__ = subcls
+        self.__handle = handle
+        self.__cls = cls
+        return self
+
+    def __restore(self):
+        cls = self.__cls
+        del self.__handle
+        del self.__cls
+        self.__class__ = cls
+        return self
+
+    @property
+    def shape(self):
+        if not skip_tracing:
+            active_trace._require_shape(self.__handle)
+        return super().shape
+
+    def numpy(self):
+        if not skip_tracing:
+            active_trace._require_value(self.__handle)
+        return super().numpy()
+
+    def _dev_tensor(self):
+        if not skip_tracing:
+            active_trace._require_data(self.__handle)
+        return super()._dev_tensor()
+
+
+class TracedRawTensor(TraceMixin, RawTensor):
+    pass
+
+
+class TracedLazyTensor(TraceMixin, LazyEvalTensor):
+    pass
+
+
+def assign_raw_tensor(lhs, rhs):
+    handle = rhs._handle
+    rhs.__dict__.clear()
+    lhs.__dict__.clear()
+    lhs.__class__ = RawTensor
+    lhs.__init__(handle)
+
+
+# this hook turns RawTensor into LazyEvalTensor
+@apply.register()
+def apply_symbolic_mode(op: OpDef, *args: RawTensor):
+    graph = active_trace._lazy_eval_graph
+    ivars = [
+        getattr(x, "_LazyEvalTensor__varnode", None)
+        or graph.make_const(x._dev_tensor())
+        for x in args
+    ]
+    ovars = apply(op, *ivars)
+    outputs = [LazyEvalTensor(v) for v in ovars]
+    active_trace._lazy_eval_tensors.update(outputs)
+    return outputs
+
+
+apply.disable(apply_symbolic_mode)
+
+
+@apply.register()
+def apply_compiled_mode(op: OpDef, *args: RawTensor):
+    if skip_tracing:
+        args = [
+            as_raw_tensor(x._dev_tensor()) if x.__class__ is CompiledTensorProxy else x
+            for x in args
+        ]
+        return apply.super(op, *args)
+    return active_trace._apply_op(op, args)
+
+
+apply.disable(apply_compiled_mode)
+
+
+# this hook injects TraceMixin
+@apply.register()
+def apply_with_tracing(op: OpDef, *args: RawTensor):
+    outputs = apply.super(op, *args)
+    active_trace._record_op(op, args, outputs)
+    return outputs
+
+
+apply.disable(apply_with_tracing)
+
+
+# @apply.register()
+# def _(op: Const, *args: RawTensor):
+#     return active_trace._apply_const(op, args)
+
+
+class BrokenRawTensor(RawTensor):
+    def __getattribute__(self, _):
+        raise RuntimeError("broken due to misuse of tracing")
+
+    def __setattr__(self, *_):
+        raise RuntimeError("broken due to misuse of tracing")

imperative/python/src/common.cpp

@@ -23,10 +23,29 @@ namespace py = pybind11;
 using namespace mgb;
 using namespace imperative;
 
+namespace {
+
+template<typename XTensorND>
+auto def_TensorND(py::object parent, const char* name) {
+    return py::class_<XTensorND>(parent, name)
+        .def_property_readonly("shape", py::overload_cast<>(&XTensorND::shape, py::const_))
+        .def_property_readonly("dtype", py::overload_cast<>(&XTensorND::dtype, py::const_))
+        .def_property_readonly("comp_node", py::overload_cast<>(&XTensorND::comp_node, py::const_))
+        .def("copy_from", &XTensorND::template copy_from<DeviceTensorStorage>)
+        .def("copy_from", &XTensorND::template copy_from<HostTensorStorage>)
+        .def("copy_from_fixlayout", py::overload_cast<const DeviceTensorND&>(
+            &XTensorND::template copy_from_fixlayout<DeviceTensorStorage>))
+        .def("copy_from_fixlayout", py::overload_cast<const HostTensorND&>(
+            &XTensorND::template copy_from_fixlayout<HostTensorStorage>));
+}
+
+} // namespace
+
 void init_common(py::module m) {
-    py::class_<CompNode>(m, "CompNode")
+    auto&& PyCompNode = py::class_<CompNode>(m, "CompNode")
         .def(py::init())
         .def(py::init(py::overload_cast<const std::string&>(&CompNode::load)))
+        .def("create_event", &CompNode::create_event, py::arg("flags") = 0ul)
         .def("__str__", &CompNode::to_string_logical)
         .def_static("_sync_all", &CompNode::sync_all)
         .def(py::self == py::self)
@@ -40,19 +59,30 @@ void init_common(py::module m) {
                     return CompNode::load(cn);
                 }));
 
+    py::class_<CompNode::Event, std::shared_ptr<CompNode::Event>>(PyCompNode, "Event")
+        .def("record", &CompNode::Event::record)
+        .def("wait", &CompNode::Event::host_wait);
+
     py::implicitly_convertible<std::string, CompNode>();
 
-    py::class_<DeviceTensorND>(m, "DeviceTensorND")
-        .def(py::init())
-        .def_property_readonly("shape", py::overload_cast<>(&DeviceTensorND::shape, py::const_))
-        .def_property_readonly("dtype", py::overload_cast<>(&DeviceTensorND::dtype, py::const_))
-        .def_property_readonly("comp_node", py::overload_cast<>(&DeviceTensorND::comp_node, py::const_))
+    def_TensorND<DeviceTensorND>(m, "DeviceTensorND")
         .def("numpy", [](const DeviceTensorND& self) {
                 HostTensorND hv;
                 hv.copy_from(self).sync();
                 return py::handle(npy::ndarray_from_tensor(hv, npy::ShareType::TRY_SHARE));
             });
 
+    def_TensorND<HostTensorND>(m, "HostTensorND")
+        .def(py::init([](py::array data, CompNode cn, DType dtype) {
+                if (!cn.valid()) {
+                    throw py::type_error("device must not be None");
+                }
+                return npy::np2tensor(data.ptr(), npy::Meth::borrow(cn), dtype);
+            }))
+        .def("numpy", [](const HostTensorND& self) {
+                return py::reinterpret_steal<py::object>(npy::ndarray_from_tensor(self, npy::ShareType::TRY_SHARE));
+            });
+
     py::class_<cg::OperatorNodeConfig>(m, "OperatorNodeConfig")
         .def(py::init())
         .def_property("name",

imperative/python/src/graph_rt.cpp

@@ -12,6 +12,7 @@
 #include "./graph_rt.h"
 
 #include "megbrain/imperative/opr_utility.h"
+#include "megbrain/opr/io.h"
 #include "megbrain/opr/basic_arith.h"
 #include "megbrain/imperative.h"
 #include "./helper.h"
@@ -29,29 +30,44 @@ auto def_rendezvous(py::object m, const char* name) {
         .def(py::init([](){return std::make_shared<Rendezvous<T>>();}))
         .def("set", [](Rendezvous<T>& r, T v) {r.set(std::move(v));})
         .def("get", [](Rendezvous<T>& r) {return r.get();}, py::call_guard<py::gil_scoped_release>())
+        .def("drop", &Rendezvous<T>::drop)
         .def("reset", &Rendezvous<T>::reset);
 }
 
 using TensorAttr = LogicalTensorDesc;
+using HostNDWithEvent = std::pair<HostTensorND, std::shared_ptr<CompNode::Event>>;
 
 void init_graph_rt(py::module m) {
     def_rendezvous<DeviceTensorND>(m, "DeviceTensorNDRendezvous");
 
+    def_rendezvous<HostNDWithEvent>(m, "HostTensorNDRendezvous");
+
     def_rendezvous<TensorAttr>(m, "TensorAttrRendezvous");
 
     py::class_<cg::VarNode, GraphNodePtr<cg::VarNode>>(m, "VarNode")
         .def_property_readonly("owner", [](cg::VarNode* v) {return v->owner_opr();})
         .def_property_readonly("graph", [](cg::VarNode* v) {return v->owner_graph();})
+        .def_property_readonly("name", py::overload_cast<>(&VarNode::name, py::const_))
         .def_property_readonly("dtype", [](cg::VarNode* v) {return v->dtype();})
-        .def_property_readonly("comp_node", [](cg::VarNode* v) {return v->comp_node();});
+        .def_property_readonly("comp_node", [](cg::VarNode* v) {return v->comp_node();})
+        .def_property_readonly("shape", [](cg::VarNode* v) -> const TensorShape* {
+                auto&& mgr = v->owner_graph()->static_infer_manager();
+                auto&& type = mgr.get_infer_type(v);
+                using InferType = cg::static_infer::InferType;
+                if (!(type.shape & (InferType::CONST | InferType::RT_STATIC))) {
+                    return nullptr;
+                }
+                return mgr.infer_shape_fallible(v);
+            });
 
     py::class_<cg::OperatorNodeBase, GraphNodePtr<cg::OperatorNodeBase>>(m, "OperatorNode")
         .def_property_readonly("graph", [](cg::OperatorNodeBase* opr) {return opr->owner_graph();})
+        .def_property_readonly("name", py::overload_cast<>(&cg::OperatorNodeBase::name, py::const_))
         .def_property_readonly("inputs", [](cg::OperatorNodeBase* opr) {
                 return to_tuple(opr->input());
             })
         .def_property_readonly("outputs", [](cg::OperatorNodeBase* opr) {
-                return to_tuple(opr->output());
+                return to_tuple(opr->usable_output());
             });
 
     py::class_<cg::AsyncExecutable>(m, "AsyncExecutable")
@@ -117,7 +133,7 @@ void init_graph_rt(py::module m) {
     common.def("invoke_op", [](const OpDef& def, const std::vector<cg::VarNode*> inputs, cg::ComputingGraph* graph) {
             cg::VarNodeArray vinputs(inputs.begin(), inputs.end());
             auto opr = OpDef::apply_on_var_node(def, vinputs);
-            auto outputs = opr->output();
+            auto outputs = opr->usable_output();
             return to_tuple(outputs);
         },
         py::arg(), py::arg(), py::arg("graph") = py::none());
@@ -125,6 +141,7 @@ void init_graph_rt(py::module m) {
     auto input_callback = [](auto callback,
                              const CompNode& comp_node,
                              const DType& dtype,
+                             const TensorShape& shape,
                              const std::vector<cg::VarNode*>& inputs,
                              cg::ComputingGraph* graph) {
         if (!graph) {
@@ -135,7 +152,7 @@ void init_graph_rt(py::module m) {
             sinputs.emplace_back(i);
         }
         static_assert(!std::is_reference<decltype(callback)>::value);
-        auto soutputs = opr::InputCallback::make(*graph, std::move(callback), comp_node, dtype, sinputs);
+        auto soutputs = opr::InputCallback::make(*graph, std::move(callback), comp_node, dtype, shape, sinputs);
         std::vector<VarNode*> outputs;
         outputs.reserve(soutputs.size());
         for (auto i : soutputs) {
@@ -144,26 +161,40 @@ void init_graph_rt(py::module m) {
         return outputs;
     };
 
+    m.def("make_shared", [](cg::ComputingGraph* graph, const DeviceTensorND& data) {
+            return opr::SharedDeviceTensor::make(*graph, std::make_shared<DeviceTensorND>(data)).node();
+        });
+
+    m.def("make_const", [](cg::ComputingGraph* graph, py::array data, CompNode cn, DType dtype) {
+            if (!cn.valid()) {
+                throw py::type_error("device must not be None");
+            }
+            auto hv = npy::np2tensor(data.ptr(), npy::Meth::borrow(cn), dtype);
+            opr::ImmutableTensor::make(*graph, hv, OperatorNodeConfig(cn)).node();
+        });
+
     m.def("input_callback", [input_callback](std::function<DeviceTensorND(void)> callback,
                                              const CompNode& comp_node,
                                              const DType& dtype,
+                                             const TensorShape& shape,
                                              const std::vector<cg::VarNode*>& inputs,
                                              cg::ComputingGraph* graph) {
-            return input_callback([f=std::move(callback)](){py::gil_scoped_acquire _; return f();}, comp_node, dtype, inputs, graph);
+            return input_callback([f=std::move(callback)](){py::gil_scoped_acquire _; return f();}, comp_node, dtype, shape, inputs, graph);
         },
-        py::arg(), py::arg(), py::arg(), py::arg() = py::tuple(), py::arg("graph") = py::none());
+        py::arg(), py::arg(), py::arg(), py::arg() = py::none(), py::arg() = py::tuple(), py::arg("graph") = py::none());
 
     m.def("input_callback", [input_callback](std::shared_ptr<Rendezvous<DeviceTensorND>> p,
                                              const CompNode& comp_node,
                                              const DType& dtype,
+                                             const TensorShape& shape,
                                              const std::vector<cg::VarNode*>& inputs,
                                              cg::ComputingGraph* graph) {
             auto f = [p]() -> DeviceTensorND {
                 return p->get();
             };
-            return input_callback(std::move(f), comp_node, dtype, inputs, graph);
+            return input_callback(std::move(f), comp_node, dtype, shape, inputs, graph);
         },
-        py::arg(), py::arg(), py::arg(), py::arg() = py::tuple(), py::arg("graph") = py::none());
+        py::arg(), py::arg(), py::arg(), py::arg() = py::none(), py::arg() = py::tuple(), py::arg("graph") = py::none());
 
     auto output_callback = [](auto callback, const std::vector<cg::VarNode*>& inputs, bool borrow = false) {
         SymbolVarArray sinputs;
@@ -193,6 +224,17 @@ void init_graph_rt(py::module m) {
         return output_callback(std::move(f), std::move(inputs));
     });
 
+    m.def("value_output_callback", [output_callback](std::shared_ptr<Rendezvous<HostNDWithEvent>> p, std::vector<cg::VarNode*> inputs) {
+        auto f = [p](DeviceTensorND dv) {
+            HostNDWithEvent hv_with_event;
+            hv_with_event.first.copy_from(dv);
+            hv_with_event.second = dv.comp_node().create_event();
+            hv_with_event.second->record();
+            p->set(std::move(hv_with_event));
+        };
+        return output_callback(std::move(f), std::move(inputs), true);
+    });
+
     m.def("attr_output_callback", [output_callback](std::shared_ptr<Rendezvous<TensorAttr>> p, std::vector<cg::VarNode*> inputs) {
         auto f = [p](DeviceTensorND dv) {
             p->set(TensorAttr{TensorLayout{dv.shape(), dv.dtype()}, dv.comp_node()});

imperative/python/src/graph_rt.h

@@ -39,6 +39,7 @@ template<typename R>
 class Rendezvous {
     std::mutex m_lock;
     int m_read_ahead = 0;
+    bool m_drop_next = false;
     std::promise<R> m_promise;
 public:
     Rendezvous() = default;
@@ -47,6 +48,7 @@ public:
     Rendezvous& operator=(const Rendezvous& rhs) = delete;
     Rendezvous& operator=(Rendezvous&& rhs) {
         MGB_LOCK_GUARD(m_lock);
+        m_drop_next = rhs.m_drop_next;
         m_read_ahead = rhs.m_read_ahead;
         m_promise = std::move(rhs.m_promise);
         return *this;
@@ -67,12 +69,28 @@ public:
         return f.get();
     }
 
+    void drop() {
+        MGB_LOCK_GUARD(m_lock);
+        mgb_assert(m_read_ahead <= 0);
+        mgb_assert(m_read_ahead >= -1);
+        if (m_read_ahead == -1) {
+            m_promise = {};
+        } else {
+            m_drop_next = true;
+        }
+        ++m_read_ahead;
+    }
+
     template<typename T>
     void set(T&& value) {
         MGB_LOCK_GUARD(m_lock);
         mgb_assert(m_read_ahead >= 0);
         mgb_assert(m_read_ahead <= 1);
-        m_promise.set_value(std::forward<T>(value));
+        if (m_drop_next) {
+            m_drop_next = false;
+        } else {
+            m_promise.set_value(std::forward<T>(value));
+        }
         if (m_read_ahead == 1) {
             m_promise = {};
         }
@@ -83,6 +101,7 @@ public:
         MGB_LOCK_GUARD(m_lock);
         m_promise = {};
         m_read_ahead = 0;
+        m_drop_next = false;
     }
 };
 

imperative/python/src/helper.h

@@ -280,9 +280,12 @@ namespace detail {
     public:
         bool load(handle src, bool convert) {
             auto obj = reinterpret_steal<object>(src);
-            if (!isinstance<tuple>(obj)) {
+            if (!convert && !isinstance<tuple>(obj)) {
                 return false;
             }
+            if (obj.is_none()) {
+                return true;
+            }
             value.ndim = len(obj);
             mgb_assert(value.ndim <= mgb::TensorShape::MAX_NDIM);
             size_t i = 0;

imperative/python/src/imperative_rt.cpp

@@ -63,6 +63,7 @@ void init_imperative_rt(py::module m) {
                     return self.put(npy::np2tensor(data.ptr(), npy::Meth::copy_into(&ret), dtype));
                 }
             }, py::arg(), py::arg("dtype") = py::none(), py::arg("device") = py::none())
+        .def("put", py::overload_cast<const DeviceTensorND&>(&Interpreter::Channel::put))
         .def("delete", [](Interpreter::Channel& self, Interpreter::Handle handle) {
                 return self.del(handle);
             })

imperative/python/src/pyext17.h

@@ -24,6 +24,12 @@ constexpr bool has_fastcall = true;
 constexpr bool has_fastcall = false;
 #endif
 
+#ifdef _Py_TPFLAGS_HAVE_VECTORCALL
+constexpr bool has_vectorcall = true;
+#else
+constexpr bool has_vectorcall = false;
+#endif
+
 template<typename... Args>
 struct invocable_with {
     template<typename T>
@@ -55,6 +61,9 @@ private:
 public:
     PyObject_HEAD
     std::aligned_storage_t<sizeof(T), alignof(T)> storage;
+    #ifdef _Py_TPFLAGS_HAVE_VECTORCALL
+    PyObject* vectorcall_slot;
+    #endif
 
     inline T* inst() {
         return reinterpret_cast<T*>(&storage);
@@ -155,6 +164,51 @@ private:
 
     // polyfills
 
+    struct tp_vectorcall {
+        static constexpr bool valid = HAS_MEMBER(T, tp_vectorcall);
+        static constexpr bool haskw = [](){if constexpr (valid)
+                                               if constexpr (std::is_invocable_v<T::tp_vectorcall, T, PyObject*const*, size_t, PyObject*>)
+                                                   return true;
+                                           return false;}();
+
+        template<typename = void>
+        static PyObject* impl(PyObject* self, PyObject*const* args, size_t nargsf, PyObject *kwnames) {
+            auto* inst = reinterpret_cast<wrap_t*>(self)->inst();
+            if constexpr (haskw) {
+                CVT_RET_PYOBJ(inst->tp_vectorcall(args, nargsf, kwnames));
+            } else {
+                if (kwnames && PyTuple_GET_SIZE(kwnames)) {
+                    PyErr_SetString(PyExc_TypeError, "expect no keyword argument");
+                    return nullptr;
+                }
+                CVT_RET_PYOBJ(inst->tp_vectorcall(args, nargsf));
+            }
+        }
+
+        static constexpr Py_ssize_t offset = []() {if constexpr (valid) return offsetof(wrap_t, vectorcall_slot);
+                                                   else return 0;}();
+    };
+
+    struct tp_call {
+        static constexpr bool provided = HAS_MEMBER(T, tp_call);
+        static constexpr bool static_form = invocable_with<T, PyObject*, PyObject*, PyObject*>{}(
+            [](auto&& t, auto... args) -> decltype(std::decay_t<decltype(t)>::tp_call(args...)) {});
+        static constexpr bool valid = provided || tp_vectorcall::valid;
+
+        template<typename = void>
+        static PyObject* impl(PyObject* self, PyObject* args, PyObject* kwargs) {
+            auto* inst = reinterpret_cast<wrap_t*>(self)->inst();
+            CVT_RET_PYOBJ(inst->tp_call(args, kwargs));
+        }
+
+        static constexpr ternaryfunc value = []() {if constexpr (static_form) return T::tp_call;
+                                                   else if constexpr (provided) return impl<>;
+                                                   #ifdef _Py_TPFLAGS_HAVE_VECTORCALL
+                                                   else if constexpr (valid) return PyVectorcall_Call;
+                                                   #endif
+                                                   else return nullptr;}();
+    };
+
     struct tp_new {
         static constexpr bool provided = HAS_MEMBER(T, tp_new);
         static constexpr bool varkw = std::is_constructible_v<T, PyObject*, PyObject*>;
@@ -163,11 +217,14 @@ private:
         template<typename = void>
         static PyObject* impl(PyTypeObject* type, PyObject* args, PyObject* kwargs) {
             auto* self = type->tp_alloc(type, 0);
-            auto* ptr = reinterpret_cast<wrap_t*>(self)->inst();
+            auto* inst = reinterpret_cast<wrap_t*>(self)->inst();
+            if constexpr (has_vectorcall && tp_vectorcall::valid) {
+                reinterpret_cast<wrap_t*>(self)->vectorcall_slot = &tp_vectorcall::template impl<>;
+            }
             if constexpr (varkw) {
-                new(ptr) T(args, kwargs);
+                new(inst) T(args, kwargs);
             } else {
-                new(ptr) T();
+                new(inst) T();
             }
             return self;
         }
@@ -190,22 +247,6 @@ private:
                                                   else return impl<>;}();
     };
 
-    struct tp_call {
-        static constexpr bool valid = HAS_MEMBER(T, tp_call);
-        static constexpr bool static_form = invocable_with<T, PyObject*, PyObject*, PyObject*>{}(
-            [](auto&& t, auto... args) -> decltype(std::decay_t<decltype(t)>::tp_call(args...)) {});
-
-        template<typename = void>
-        static PyObject* impl(PyObject* self, PyObject* args, PyObject* kwargs) {
-            auto* inst = reinterpret_cast<wrap_t*>(self)->inst();
-            CVT_RET_PYOBJ(inst->tp_call(args, kwargs));
-        }
-
-        static constexpr ternaryfunc value = []() {if constexpr (static_form) return T::tp_call;
-                                                   else if constexpr (valid) return impl<>;
-                                                   else return nullptr;}();
-    };
-
 public:
     class TypeBuilder {
         std::vector<PyMethodDef> m_methods;
@@ -228,9 +269,17 @@ public:
                 m_type.tp_name = T::tp_name;
             }
             m_type.tp_dealloc = tp_dealloc::value;
+            #ifdef _Py_TPFLAGS_HAVE_VECTORCALL
+            m_type.tp_vectorcall_offset = tp_vectorcall::offset;
+            #endif
             m_type.tp_call = tp_call::value;
             m_type.tp_basicsize = sizeof(wrap_t);
             m_type.tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE;
+            #ifdef _Py_TPFLAGS_HAVE_VECTORCALL
+            if constexpr (tp_vectorcall::valid) {
+                m_type.tp_flags |= _Py_TPFLAGS_HAVE_VECTORCALL;
+            }
+            #endif
             m_type.tp_new = tp_new::value;
         }
 

imperative/python/test/unit/test_tracing.py

@@ -0,0 +1,65 @@
+import numpy as np
+
+from megengine.core.ops import builtin as ops
+from megengine.core.tensor.core import apply
+from megengine.core.tensor.raw_tensor import as_raw_tensor
+from megengine.jit import exclude_from_trace, trace
+
+
+def test_trace():
+    for symbolic in [False, True]:
+
+        @trace(symbolic=symbolic)
+        def f(x):
+            op = ops.Elemwise(mode="negate")
+            (y,) = apply(op, x)
+            return y
+
+        x = as_raw_tensor([1]).numpy()
+        y = f.__wrapped__(as_raw_tensor(x)).numpy()
+
+        for i in range(3):
+            np.testing.assert_equal(f(as_raw_tensor(x)).numpy(), y)
+
+
+def test_exclude_from_trace():
+    for symbolic in [False, True]:
+
+        @trace(symbolic=symbolic)
+        def f(x):
+            neg = ops.Elemwise(mode="negate")
+            (x,) = apply(neg, x)
+            with exclude_from_trace():
+                if i % 2:
+                    (x,) = apply(neg, x)
+            (x,) = apply(neg, x)
+            return x
+
+        x = as_raw_tensor([1]).numpy()
+
+        for i in range(3):
+            y = f.__wrapped__(as_raw_tensor(x)).numpy()
+            np.testing.assert_equal(f(as_raw_tensor(x)).numpy(), y)
+
+
+def test_print_in_trace():
+    for symbolic in [False]:  # cannot read value in symbolic mode
+
+        @trace(symbolic=symbolic)
+        def f(x):
+            nonlocal buf
+            neg = ops.Elemwise(mode="negate")
+            (x,) = apply(neg, x)
+            buf = x.numpy()
+            (x,) = apply(neg, x)
+            return x
+
+        buf = None
+        x = as_raw_tensor([1]).numpy()
+
+        for i in range(3):
+            y = f.__wrapped__(as_raw_tensor(x)).numpy()
+            z = buf
+            buf = None
+            np.testing.assert_equal(f(as_raw_tensor(x)).numpy(), y)
+            np.testing.assert_equal(z, buf)

imperative/src/impl/interpreter_impl.cpp

@@ -37,6 +37,15 @@ void* ChannelImpl::put(const HostTensorND& value) {
     return info;
 }
 
+void* ChannelImpl::put(const DeviceTensorND& data) {
+    auto info = alloc();
+    info->desc.layout = data.layout();
+    info->desc.comp_node = data.comp_node();
+    info->ptr = Tensor::make(data);
+    m_valid_handle.insert(info);
+    return info;
+}
+
 void ChannelImpl::del(void* handle) {
     mgb_assert(m_valid_handle.erase(handle), "invalid handle: %p", handle);
     m_worker.add_task(Del{reinterpret_cast<TensorInfo*>(handle)});

imperative/src/impl/interpreter_impl.h

@@ -55,6 +55,7 @@ struct ChannelImpl : Interpreter::Channel {
     ~ChannelImpl() override;
 
     Handle put(const HostTensorND& value) override;
+    Handle put(const DeviceTensorND& value) override;
 
     void del(Handle) override;
 

imperative/src/impl/opr_utility.cpp

@@ -31,9 +31,10 @@ MGB_DYN_TYPE_OBJ_FINAL_IMPL(InputCallback);
 
 InputCallback::InputCallback(cg::ComputingGraph& graph, callback_t callback,
                              const VarNodeArray& inputs,
+                             const TensorShape& output_shape,
                              const OperatorNodeConfig& config)
         : Super(&graph, config, "input_callback", inputs),
-          m_callback(callback) {
+          m_output_shape(output_shape), m_callback(callback) {
     for (VarNode* i : inputs) {
         add_input({i});
     }
@@ -48,7 +49,8 @@ InputCallback::InputCallback(cg::ComputingGraph& graph, callback_t callback,
 
 SymbolVarArray InputCallback::make(cg::ComputingGraph& graph,
                                    callback_t callback, CompNode comp_node,
-                                   DType dtype, const SymbolVarArray& inputs) {
+                                   DType dtype, const TensorShape& shape,
+                                   const SymbolVarArray& inputs) {
     mgb_assert(comp_node.valid());
     mgb_assert(dtype.valid());
     OperatorNodeConfig config;
@@ -56,11 +58,22 @@ SymbolVarArray InputCallback::make(cg::ComputingGraph& graph,
     config.output_dtype(dtype);
     auto vinputs = to_var_node_array(inputs);
     auto opr = graph.insert_opr(
-            std::make_unique<InputCallback>(graph, callback, vinputs, config));
+            std::make_unique<InputCallback>(graph, callback, vinputs, shape, config));
     return to_symbol_var_array(opr->output());
 }
 
-void InputCallback::init_output_static_infer_desc() {}
+void InputCallback::init_output_static_infer_desc() {
+    if (m_output_shape.ndim) {
+        using namespace cg::static_infer;
+        auto &&mgr = owner_graph()->static_infer_manager();
+        auto infer_shape = [this](TensorShape &dest, const InpVal &) {
+            dest = m_output_shape;
+            return true;
+        };
+        mgr.register_shape_infer(output(0),
+                {SourceType::CONSTANT, {}, infer_shape});
+    }
+}
 
 cg::OperatorNodeBase::NodeProp* InputCallback::do_make_node_prop() const {
     NodeProp* prop = Super::do_make_node_prop();
@@ -73,9 +86,23 @@ cg::OperatorNodeBase::NodeProp* InputCallback::do_make_node_prop() const {
 
 void InputCallback::scn_do_execute() {
     auto dev_tensor = m_callback();
+    if (m_output_shape.ndim) {
+        mgb_assert(dev_tensor.shape().eq_shape(m_output_shape));
+    }
     output(0)->reset_dev_tensor_from_tensor(dev_tensor);
 }
 
+cg::OperatorNodeBase* InputCallback::shallow_copy(
+        const serialization::OprShallowCopyContext &ctx,
+        const cg::OperatorNodeBase &opr_, const VarNodeArray &inputs,
+        const OperatorNodeConfig &config) {
+    auto &&opr = opr_.cast_final_safe<InputCallback>();
+    auto* graph = ctx.owner_graph(opr, inputs);
+    return graph->insert_opr(std::make_unique<InputCallback>(*graph, opr.m_callback, inputs, opr.m_output_shape, config));
+}
+
+MGB_REG_OPR_SHALLOW_COPY(InputCallback, InputCallback::shallow_copy);
+
 /* ================ OutputCallback ================== */
 
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(OutputCallback);
@@ -122,6 +149,17 @@ void OutputCallback::scn_do_execute() {
     m_param.callback(input(0)->dev_tensor());
 }
 
+cg::OperatorNodeBase* OutputCallback::shallow_copy(
+        const serialization::OprShallowCopyContext &ctx,
+        const cg::OperatorNodeBase &opr_, const VarNodeArray &inputs,
+        const OperatorNodeConfig &config) {
+    auto &&opr = opr_.cast_final_safe<OutputCallback>();
+    auto* graph = ctx.owner_graph(opr, inputs);
+    return graph->insert_opr(std::make_unique<OutputCallback>(opr.m_param, inputs, config));
+}
+
+MGB_REG_OPR_SHALLOW_COPY(OutputCallback, OutputCallback::shallow_copy);
+
 /* ================ NopCallback ================== */
 
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(NopCallback);

imperative/src/include/megbrain/imperative/interpreter.h

@@ -22,6 +22,7 @@ struct Interpreter {
         virtual ~Channel() = default;
 
         virtual Handle put(const HostTensorND& value) = 0;
+        virtual Handle put(const DeviceTensorND& value) = 0;
 
         virtual void del(Handle) = 0;
 

imperative/src/include/megbrain/imperative/opr_utility.h

@@ -17,6 +17,7 @@
 #include "megbrain/opr/internal/param_tag_defs.h"
 #include "megbrain/opr/internal/megdnn_opr_wrapper.h"
 #include "megbrain/opr/param_defs.h"
+#include "megbrain/serialization/sereg.h"
 
 #include "megdnn/oprs/utils.h"
 
@@ -33,17 +34,24 @@ public:
     InputCallback(cg::ComputingGraph& graph,
                   callback_t callback,
                   const VarNodeArray& inputs,
+                  const TensorShape& output_shape,
                   const OperatorNodeConfig &config);
     static SymbolVarArray make(cg::ComputingGraph& graph,
                                callback_t callback,
                                CompNode comp_node,
                                DType dtype,
+                               const TensorShape& shape,
                                const SymbolVarArray& inputs = {});
+    static cg::OperatorNodeBase* shallow_copy(
+            const serialization::OprShallowCopyContext &ctx,
+            const cg::OperatorNodeBase &opr_, const VarNodeArray &inputs,
+            const OperatorNodeConfig &config);
 protected:
     void scn_do_execute() override;
     void init_output_static_infer_desc() override;
     NodeProp* do_make_node_prop() const override;
 private:
+    TensorShape m_output_shape;
     callback_t m_callback;
 };
 
@@ -63,6 +71,10 @@ public:
                           SymbolVar input) {
         return make(std::move(param), SymbolVarArray{input});
     }
+    static cg::OperatorNodeBase* shallow_copy(
+            const serialization::OprShallowCopyContext &ctx,
+            const cg::OperatorNodeBase &opr_, const VarNodeArray &inputs,
+            const OperatorNodeConfig &config);
 protected:
     void scn_do_execute() override;
     void init_output_static_infer_desc() override;