feat(imperative): add external transform

GitOrigin-RevId: e8e3ebe9c86afc9fb97900b5b5af9778cc1354e5

imperative/python/megengine/jit/xla_backend.py

 from collections import OrderedDict, defaultdict
 
+from .. import tensor
 from ..core._imperative_rt import CompNode
 from ..core._imperative_rt.core2 import Tensor as RawTensor
+from ..core._imperative_rt.core2 import (
+    is_external_convert,
+    set_external_convert,
+    set_external_convert_hook,
+    set_py_external_type,
+    unset_external_convert,
+)
 from ..core._trace_option import set_use_xla_backend
 from ..device import get_default_device
 from ..utils.dlpack import from_dlpack, to_dlpack
 from .tracing import trace
 
+# try:
+#     from mge_xlalib.xla_extension import ArrayImpl
+
+#     from ..xla.lib import xla_client as xc
+# except ImportError:
+#     pass
+
+from mge_xlalib.xla_extension import ArrayImpl
+
+from ..xla.lib import xla_client as xc
+
+xla_client_compute_stream = None
+
+
+def apply_external_convert_hook(input, cn):
+    stream = xla_client_compute_stream
+    assert isinstance(input, ArrayImpl)
+    dlpack_capsule = xc._xla.buffer_to_dlpack_managed_tensor(input, take_ownership=True)
+    output = from_dlpack(dlpack_capsule, stream).to(cn, _borrow=True)
+    return output
+
 
 class xla_trace(trace):
     r"""Wraps a callable, and provides accelerated evaluation compiled by xla.
@@ -48,6 +77,12 @@ class xla_trace(trace):
     def __init__(self, function, *, without_host=True, symbolic_shape=False, **kwargs):
         assert without_host, "xla trace only support without host mode"
         assert not symbolic_shape, "xla doesn't support dynamic shape currently"
+
+        set_external_convert_hook(apply_external_convert_hook)
+
+        set_py_external_type(ArrayImpl)
+        set_external_convert()
+
         super().__init__(
             function, without_host=without_host, symbolic_shape=symbolic_shape, **kwargs
         )
@@ -142,8 +177,8 @@ class xla_trace(trace):
         return xc._xla.buffer_to_dlpack_managed_tensor(x, take_ownership=take_ownership)
 
     def execute(self, *args, **kwargs):
-        from ..traced_module.pytree import tree_flatten
         from ..tensor import Tensor
+        from ..traced_module.pytree import tree_flatten
         from ..utils.module_utils import get_expand_structure
 
         inputs, _ = tree_flatten((args, kwargs))
@@ -161,6 +196,8 @@ class xla_trace(trace):
 
         arrays = self.prepare_xla_inputs(arrays)
         outputs = self.xla_exec(*arrays)
+        global xla_client_compute_stream
+        xla_client_compute_stream = xla_stream
         return_vals = []
         for i in self.out_list:
             if i == -1:
@@ -170,28 +207,25 @@ class xla_trace(trace):
                 return_vals.append(outputs[self.outkey2idx[i]])
         keeped_features = []
         for i in self.keeped_activation:
-            capsule = self.to_dlpack(outputs[self.outkey2idx[i]])
-            t = from_dlpack(capsule, xla_stream).to(cn, _borrow=True)
-            keeped_features.append(t)
+            keeped_features.append(outputs[self.outkey2idx[i]])
         out_tensors = []
         for array in return_vals:
             if array is not None:
-                capsule = self.to_dlpack(array)
-                t = from_dlpack(capsule, xla_stream)
-                out_tensors.append(t.to(cn, _borrow=True))
+                t = tensor(array, device=cn)
+                out_tensors.append(t)
             else:
                 out_tensors.append(array)
         if self.overall:
             for attr, key in self.update_param_dict.items():
                 param = get_expand_structure(attr[0], attr[1])
                 xla_array = outputs[self.outkey2idx[key]]
-                capsule = self.to_dlpack(xla_array)
-                param._reset(from_dlpack(capsule).to(cn, _borrow=True))
+                t = tensor(xla_array, device=cn)
+                param._reset(t)
 
             for state, key in self.update_opt_param_dict.items():
                 xla_array = outputs[self.outkey2idx[key]]
-                capsule = self.to_dlpack(xla_array)
-                state._reset(from_dlpack(capsule).to(cn, _borrow=True))
+                t = tensor(xla_array, device=cn)
+                state._reset(t)
         rst = (
             self.outdef.unflatten(out_tensors)
             if hasattr(self, "outdef")

imperative/python/megengine/xla/compile.py

@@ -4,6 +4,7 @@ from typing import Any, Callable, Dict, List, Optional, Protocol, Sequence, Set,
 
 import numpy as np
 
+from .. import tensor
 from ..distributed import is_distributed
 from ..utils.dlpack import from_dlpack, to_dlpack
 from . import ir_utils
@@ -68,10 +69,20 @@ class InputsHandler:
 
     def __call__(self, input_buffers):
         rst = []
-        for ibuf in input_buffers:
-            capsule = to_dlpack(ibuf)
-            xla_array = self.from_dlpack(capsule)
-            rst.append([xla_array])
+        for idx, i in enumerate(input_buffers):
+            if i._is_external_value():
+                rst.append([i._external_obj()])
+            else:
+                if "gpu" in i.device.physical_name:
+                    capsule = to_dlpack(i)
+                    xla_array = self.from_dlpack(capsule)
+                    rst.append([xla_array])
+                else:
+                    r = self.handler(
+                        self.local_devices, [self.input_indices[idx],], [i,]
+                    )[0]
+                    rst.append(r)
+                    i._reset(tensor(r[0]))
         return rst
 
     def __str__(self):

imperative/python/src/external_convert.h

+#pragma once
+
+#include <list>
+#include "megbrain/imperative/dispatch.h"
+#include "megbrain/imperative/ops/autogen.h"
+#include "megbrain/imperative/utils/map.h"
+
+#include "./tensor.h"
+
+namespace mgb::imperative::python {
+
+namespace py = pybind11;
+
+class CreateExternalWrapper final : public OperatorImpl<CreateExternalWrapper> {
+private:
+    py::object m_object;
+    CompNode m_device;
+
+public:
+    CreateExternalWrapper(py::object obj, CompNode device)
+            : m_object(obj), m_device(device) {}
+
+    py::object object() const { return m_object; }
+
+    CompNode device() const { return m_device; }
+
+    std::string raw_type() const { return "CreateExternalWrapper"; }
+
+    std::string to_string() const { return "CreateExternalWrapper"; };
+};
+
+class GetExternalVal final
+        : public OperatorImpl<GetExternalVal, Operator::GetAttrLike> {
+public:
+    std::string to_string() const { return "GetExternalVal"; };
+    std::string raw_type() const { return "GetExternalVal"; }
+};
+
+class PyobjectStorage {
+private:
+    py::object m_object;
+
+public:
+    PyobjectStorage() = default;
+    PyobjectStorage(py::object object) : m_object(object) {}
+    py::object object() const { return m_object; }
+    std::string to_string() const { return "PyobjectStorage"; }
+};
+
+class PyobjectValue final : public PrimitiveValue<PyobjectValue, PyobjectStorage> {
+public:
+    using PrimitiveValue::PrimitiveValue;
+
+    std::string to_string() const override { return PyobjectStorage::to_string(); }
+};
+
+class ExternalValue final : public ObjectValue<ExternalValue> {
+private:
+    py::object m_obj;
+    mutable CompNodeValue::ref_t m_device;
+
+public:
+    ExternalValue(py::object obj, CompNode device)
+            : m_obj(obj), m_device(CompNodeValue::make(device)) {}
+
+    py::object object() const { return m_obj; }
+
+    CompNodeValue::ref_t device() const { return m_device; }
+
+    std::string to_string() const override { return "ExternalValue"; }
+
+    void clear() override {}
+};
+
+class ExternalConvertTransformation final : public Transformation {
+private:
+    py::function m_hook_fn;
+    int m_enabled = 0;
+    ObjectType<ExternalValue> m_value_type{"ExternalValue"};
+
+public:
+    ValueRefList apply_external_imperative_hook(
+            const Operator& op, Span<ValueRef> input_values) {
+        for (int i = 0; i < input_values.size(); i++) {
+            if (auto* val = input_values[i].as(m_value_type)) {
+                CompNode cn = *(val->device());
+                py::object fn_res = m_hook_fn(val->object(), cn);
+                auto* tw = TensorWrapper::try_cast(fn_res.ptr());
+                mgb_assert(tw, "expect Tensor");
+                auto external_input = input_values[i].as_ref(m_value_type);
+                external_input.reset(tw->m_tensor->data());
+            }
+        }
+        auto outputs = imperative::apply(op, input_values);
+        return outputs;
+    }
+
+    ExternalConvertTransformation(py::function hook_fn) : m_hook_fn(hook_fn) {}
+    ValueRefList apply_transformation(
+            const Operator& op, Span<ValueRef> inputs) override {
+        if (!m_enabled) {
+            return imperative::apply(op, inputs);
+        }
+        bool has_external_inp = false;
+        if (auto* obj_value = op.as<CreateExternalWrapper>()) {
+            return m_value_type.make(obj_value->object(), obj_value->device());
+        }
+        for (auto&& input : inputs) {
+            if (input.is(m_value_type)) {
+                has_external_inp = true;
+                break;
+            }
+        }
+        if (!has_external_inp) {
+            return imperative::apply(op, inputs);
+        } else if (op.is<GetExternalVal>()) {
+            py::object m_object = inputs.item().cast(m_value_type).object();
+            PyobjectStorage inp_obj = PyobjectStorage(m_object);
+            return {PyobjectValue::make(inp_obj)};
+        } else if (op.is<RenameValue>()) {
+            return {inputs[0]};
+        } else if (auto* get_attr = op.as<GetAttr>()) {
+            auto& input = inputs.item().cast(m_value_type);
+            ValueRefList outputs;
+            switch (get_attr->attr()) {
+                case GetAttr::Device:
+                    outputs = {input.device()};
+                    break;
+                default:
+                    outputs = apply_external_imperative_hook(op, inputs);
+                    break;
+            }
+            return outputs;
+        } else {
+            auto outputs = apply_external_imperative_hook(op, inputs);
+            return outputs;
+        }
+    }
+
+    void enable() { m_enabled = 1; }
+
+    void disable() { m_enabled = 0; }
+
+    bool enabled() const { return m_enabled; }
+
+    ValueRef unwrap(ValueRef value) override { return value; }
+
+    const Type<ExternalValue>& value_type() const { return m_value_type; }
+
+    std::string name() const override { return "ExternalConvertTransformation"; }
+};
+
+}  // namespace mgb::imperative::python
\ No newline at end of file

imperative/python/src/tensor.cpp

@@ -28,6 +28,7 @@
 #include "./common.h"
 #include "./dlpack.h"
 #include "./dlpack_convertor.h"
+#include "./external_convert.h"
 #include "./grad.h"
 #include "./graph_rt.h"
 #include "./helper.h"
@@ -61,6 +62,7 @@ namespace mgb::imperative::python {
 interpreter::Interpreter::Channel* interpreter_for_py = nullptr;
 PyTypeObject* py_tensor_type = nullptr;
 PyTypeObject* py_varnode_type = nullptr;
+PyTypeObject* py_external_type = nullptr;
 pybind11::handle py_device_type = nullptr;
 PyObject* cpp_use_symbolic_shape;
 
@@ -589,7 +591,13 @@ TensorWrapper::TensorWrapper(PyObject* args, PyObject* kwargs) {
                                     : no_cache ? CreateTensor::Unique
                                                : CreateTensor::Common;
             ValueRef val;
-            if (py::isinstance(data, Py_Varnode)) {
+            bool use_external_inp = py_external_type != nullptr;
+            if (use_external_inp &&
+                PyObject_TypeCheck(py::handle(data).ptr(), py_external_type)) {
+                val = imperative::apply(
+                        CreateExternalWrapper(data, cn),
+                        Span<ValueRef>(nullptr, nullptr))[0];
+            } else if (py::isinstance(data, Py_Varnode)) {
                 cg::VarNode* m_node = py::handle(data).cast<cg::VarNode*>();
                 val = imperative::apply(
                         CreateNode(m_node), Span<ValueRef>(nullptr, nullptr))[0];
@@ -750,6 +758,27 @@ PyObject* TensorWrapper::_graph() {
     return py::cast(graph).release().ptr();
 }
 
+PyObject* TensorWrapper::_external_obj() {
+    TypedValueRef<PyobjectValue> value =
+            imperative::apply(GetExternalVal(), m_tensor->data())[0]
+                    .as_ref<PyobjectValue>();
+    return value->object().release().ptr();
+}
+
+PyObject* TensorWrapper::_is_external_value() {
+    auto&& external_tsf =
+            TransformationManager::get_instance()
+                    .segments[TransformationManager::Segment::ExternalConvert];
+    auto* tsf = reinterpret_cast<ExternalConvertTransformation*>(external_tsf[0].get());
+    mgb_assert(tsf->enabled());
+    auto valueref = m_tensor->data();
+    if (valueref.is(tsf->value_type())) {
+        Py_RETURN_TRUE;
+    } else {
+        Py_RETURN_FALSE;
+    }
+}
+
 void dlpack_capsule_destructor(PyObject* data) {
     if (!PyCapsule_IsValid(data, "dltensor")) {
         // early out, see DLPack spec: if a consuming library sets the capsule
@@ -931,6 +960,8 @@ void init_tensor(py::module m) {
                     .def<&TensorWrapper::_var>("var")
                     .def<&TensorWrapper::_graph>("graph")
                     .def<&TensorWrapper::value_id>("value_id")
+                    .def<&TensorWrapper::_is_external_value>("_is_external_value")
+                    .def<&TensorWrapper::_external_obj>("_external_obj")
                     .def_getset<
                             &TensorWrapper::module_trace_info,
                             &TensorWrapper::set_module_trace_info>("_NodeMixin__node")
@@ -1150,6 +1181,10 @@ void init_tensor(py::module m) {
         py_varnode_type = reinterpret_cast<PyTypeObject*>(type_obj.inc_ref().ptr());
     });
 
+    m.def("set_py_external_type", [](py::object type_obj) {
+        py_external_type = reinterpret_cast<PyTypeObject*>(type_obj.inc_ref().ptr());
+    });
+
     m.def("set_py_device_type",
           [](py::object type_obj) { py_device_type = type_obj.inc_ref(); });
 
@@ -1705,6 +1740,24 @@ void init_tensor(py::module m) {
         return module_trace_transformation;
     };
 
+    static py::function external_convert_hook;
+
+    static auto get_external_convert = [] {
+        static std::shared_ptr<ExternalConvertTransformation>
+                external_convert_transformation;
+        if (!external_convert_transformation) {
+            mgb_assert(external_convert_hook);
+            external_convert_transformation =
+                    std::make_shared<ExternalConvertTransformation>(
+                            external_convert_hook);
+            MGB_MARK_USED_VAR(transformations
+                                      .register_at<Segment::ExternalConvert>(
+                                              external_convert_transformation)
+                                      .release());
+        }
+        return external_convert_transformation;
+    };
+
     m.def("set_cpp_use_symbolic_shape", &set_cpp_use_symbolic_shape);
 
     m.def("set_module_tracing", [=] { get_module_trace()->enable(); });
@@ -1712,6 +1765,12 @@ void init_tensor(py::module m) {
     m.def("unset_module_tracing", [=] { get_module_trace()->disable(); });
 
     m.def("is_tracing_module", [=] { return get_module_trace()->enabled(); });
+
+    m.def("set_external_convert", [=] { get_external_convert()->enable(); });
+
+    m.def("unset_external_convert", [=] { get_external_convert()->disable(); });
+
+    m.def("is_external_convert", [=] { return get_external_convert()->enabled(); });
     m.def("set_python_backtrace_enabled", &set_python_backtrace_enabled);
     m.def("set_transformation_backtrace_enabled",
           &set_transformation_backtrace_enabled);
@@ -1723,8 +1782,16 @@ void init_tensor(py::module m) {
         module_trace_hook.inc_ref();
     });
 
+    m.def("set_external_convert_hook", [](py::function function) {
+        external_convert_hook = function;
+        external_convert_hook.inc_ref();
+    });
+
     auto atexit = py::module::import("atexit");
-    atexit.attr("register")(py::cpp_function([]() { module_trace_hook = {}; }));
+    atexit.attr("register")(py::cpp_function([]() {
+        module_trace_hook = {};
+        external_convert_hook = {};
+    }));
     m.def("begin_record_values", [] { Value::begin_record_values(); });
 
     m.def("end_record_values", [] {

imperative/python/src/tensor.h

@@ -31,6 +31,7 @@ namespace mgb::imperative::python {
 extern interpreter::Interpreter::Channel* interpreter_for_py;
 extern PyTypeObject* py_tensor_type;
 extern PyTypeObject* py_varnode_type;
+extern PyTypeObject* py_external_type;
 extern pybind11::handle py_device_type;
 extern PyObject* cpp_use_symbolic_shape;
 extern PyObject* cpp_astensor1d;
@@ -142,6 +143,8 @@ public:
     PyObject* _detail();
     PyObject* _var();
     PyObject* _graph();
+    PyObject* _is_external_value();
+    PyObject* _external_obj();
     void _watch();
 };
 

imperative/python/src/transformation.h

@@ -22,6 +22,7 @@ public:
         Complex,
         Format,
         Grad,
+        ExternalConvert,
         Scalar,
         Symbol,
         Trace,

imperative/python/test/unit/xla/functional/test_xla_convert.py

+import platform
+
+import numpy as np
+import pytest
+
+import megengine.functional as F
+import megengine.jit as jit
+import megengine.tensor as tensor
+from megengine import autodiff, is_cuda_available
+from megengine.autodiff.grad_manager import GradManager
+from meg_xlalib.xla_extension import ArrayImpl
+
+
+def test_external_flag_set():
+
+    @xla_trace(capture_as_const=True)
+    def test_fun():
+        pass
+
+
+
+
+
+
+def test_external_value():
+    m = Conv2d(9,9, 3,groups=9)
+    gm = GradManager()
+    gm.attach(m.parameters())
+
+    @xla_trace(capture_as_const=True)
+    def conv_grad(inp, model):
+        with gm:
+            gm.attach(inp)
+            rst = model(inp)
+            gm.backward(rst.mean())
+        ig = inp.grad
+        wg = model.weight.grad
+        inp.grad = None
+        model.weight.grad = None
+        return ig, wg
+
+    inp = tensor(np.random.random((9,9, 32, 32)))*100
+
+    a, b = conv_grad(inp, m)
+    a1, b1 = conv_grad(inp, m)
+    np.testing.assert_allclose(a.numpy(), a1.numpy())
\ No newline at end of file