feat(autodiff): support higher order grad

GitOrigin-RevId: 86390d2179

imperative/python/megengine/autodiff/grad_manager.py

@@ -20,6 +20,9 @@ class AttachSpec:
     __slots__ = "tensor", "callbacks"
 
 
+_global_priority = 0
+
+
 class GradManager:
     r"""
     GradManager computes gradients or more generally, vector-Jacobian product, by reverse mode
@@ -118,6 +121,7 @@ class GradManager:
         self._grad = None
         self._after_backward_callback = []
         self._gradients = {}
+        self._priority = None
 
     def attach(self, tensors: Iterable[Tensor], callbacks=None):
         r"""
@@ -293,6 +297,7 @@ class GradManager:
 
         After this call, you will be able to call :meth:`backward`.
         """
+        global _global_priority
         if self._recording:
             raise RuntimeError("already recording")
         grad = Grad()
@@ -300,6 +305,9 @@ class GradManager:
         self._grad = grad
         for spec in self._attach_specs.values():
             self._do_record(spec)
+        if self._priority is None:
+            grad._priority = _global_priority
+            _global_priority -= 1
         grad.__enter__()
 
     def _do_record(self, spec):
@@ -321,11 +329,14 @@ class GradManager:
 
         After this call, you will not be able to call :meth:`backward`.
         """
+        global _global_priority
         if self._grad is not None:
             self._grad.__exit__(None, None, None)
             self._grad = None
         self._recording = False
         self._gradients = dict()
+        if self._priority is None:
+            _global_priority += 1
 
     def __enter__(self):
         self.record()
@@ -333,3 +344,41 @@ class GradManager:
 
     def __exit__(self, exc_type, exc_val, exc_tb):
         self.release()
+
+    def __and__(self, other):
+        if isinstance(other, GradManager):
+            return GradManagerGroup([self, other])
+        return NotImplemented
+
+    __rand__ = __and__
+
+
+class GradManagerGroup:
+    def __init__(self, gms) -> None:
+        self._gms = list(gms)
+
+    def merge_with(self, other):
+        if isinstance(other, GradManager):
+            other = GradManagerGroup([other])
+        elif not isinstance(other, GradManagerGroup):
+            return NotImplemented
+        return GradManagerGroup([*self._gms, *other._gms])
+
+    __and__ = merge_with
+    __rand__ = merge_with
+    __or__ = merge_with
+    __ror__ = merge_with
+
+    def __enter__(self):
+        global _global_priority
+        _global_priority += 1
+        for gm in self._gms:
+            gm._priority = _global_priority
+            gm.record()
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        global _global_priority
+        _global_priority -= 1
+        for gm in self._gms:
+            gm.release()
+            gm._priority = None

imperative/python/megengine/core/autodiff/grad.py

@@ -47,6 +47,14 @@ class Grad:
         self._impl = GradKey(name)
         _grad_manager_dict[self._name] = self
 
+    @property
+    def _priority(self):
+        return self._impl.priority
+
+    @_priority.setter
+    def _priority(self, priority):
+        self._impl.priority = priority
+
     @property
     def _name(self):
         return self._impl.name

imperative/python/src/grad.cpp

@@ -54,7 +54,7 @@ std::shared_ptr<OptimizedBackwardGraphResult> make_backward_graph(
     for (size_t i = 0; i < ctx.nargs; ++i) {
         *(size_t_ptr++) = mgb::hash(ctx.args[i]->dtype().handle());
         *(size_t_ptr++) = mgb::hash(ctx.args[i]->comp_node());
-        *(bool_ptr++) = bool(ctx.args[i]->m_grad_info.grad_fn);
+        *(bool_ptr++) = !ctx.args[i]->m_grad_info_dict.empty();
     }
     mgb_assert(bool_ptr0 == reinterpret_cast<bool*>(size_t_ptr) &&
                bool_ptr == reinterpret_cast<bool*>(buf + buf_size));
@@ -321,7 +321,7 @@ apply_result_t backward_graph_grad_rule(ApplyContext& ctx, GradFnHelper& ret_gra
     for (size_t i = 0; i < ctx.nargs; ++i) {
         inputs_copy.push_back(python::apply(FastpathCopy::make(), ctx.args[i]->shared_from_this())[0]);
         inputs_copy_weak.push_back(inputs_copy.back().get());
-        inputs_copy.back()->m_grad_info = ctx.args[i]->m_grad_info;
+        inputs_copy.back()->m_grad_info_dict = ctx.args[i]->m_grad_info_dict;
     }
     ApplyContext ctx_dup = ctx;
     ctx_dup.args = inputs_copy_weak.data();
@@ -365,25 +365,19 @@ apply_result_t python_grad_rule(ApplyContext& ctx, GradFnHelper& ret_grad_fn) {
 } // namespace
 
 apply_result_t apply_grad(ApplyContext& ctx) {
-    std::shared_ptr<GradKey> grad_key;
+    std::unordered_set<std::shared_ptr<GradKey>> grad_keys;
     for (size_t i = 0; i < ctx.nargs; ++i) {
         auto* tensor = ctx.args[i];
-        if (tensor->m_grad_info.grad_fn) {
-            auto&& input_grad_key = tensor->m_grad_info.grad_fn->key.lock();
-            // tensor is attached to a live GradKey
-            if (input_grad_key && input_grad_key->active) {
-                if (grad_key) {
-                    if (grad_key != input_grad_key) {
-                        PyErr_SetString(PyExc_NotImplementedError, "second order grad");
-                        throw pyext17::py_err_set();
-                    }
-                } else {
-                    grad_key = std::move(input_grad_key);
+        if (!tensor->m_grad_info_dict.empty()) {
+            size_t grad_cnt = 0;
+            for (auto&& grad_info: tensor->m_grad_info_dict) {
+                auto input_grad_key = grad_info.grad_fn->key.lock();
+                if (input_grad_key && input_grad_key->active && !input_grad_key->is_blocked()) {
+                    grad_keys.insert(input_grad_key);
+                    grad_cnt++;
                 }
-            } else {
-                // cleanup stale grad info
-                // under what condition?
-                tensor->m_grad_info = {};
+            }
+            if (!grad_cnt) {
                 tensor->m_flags &= ~Flags::GRAD;
             }
         } else {
@@ -393,7 +387,7 @@ apply_result_t apply_grad(ApplyContext& ctx) {
 
     ctx.flags &= ~Flags::GRAD;
 
-    if (!grad_key) {
+    if (grad_keys.empty()) {
         return apply(ctx);
     }
 
@@ -418,54 +412,65 @@ apply_result_t apply_grad(ApplyContext& ctx) {
         return backward_graph_grad_rule(ctx, grad_fn_holder);
     }();
 
-    auto& grad_fn = grad_fn_holder.grad_fn;
-    if (!grad_fn) {
+    if (!grad_fn_holder.grad_fn) {
         return outputs;
     }
 
-    grad_fn->key = grad_key;
-    grad_fn->slots.resize(outputs.size());
-    grad_fn->dsts.reserve(ctx.nargs);
+    for (auto&& grad_key: grad_keys) {
+        auto grad_fn = std::make_shared<GradFn>();
+        grad_fn->backward = grad_fn_holder.grad_fn->backward;
+        grad_fn->key = grad_key;
+        grad_fn->slots.resize(outputs.size());
+        grad_fn->dsts.reserve(ctx.nargs);
 
-    std::visit([&](auto& backward) {
-        using T = std::decay_t<decltype(backward)>;
-        if constexpr (std::is_same_v<T, std::monostate>) {
-            mgb_assert(0);
-        } else {
-            for (size_t i = 0; i < ctx.nargs; ++i) {
-                if (backward.input_has_grad(i) && input_requires_grad(ctx, i)) {
-                    auto& input_grad_info = ctx.args[i]->m_grad_info;
-                    grad_fn->dsts.emplace_back(input_grad_info);
-                    // register as grad producer
-                    grad_fn->dsts.back().producer_record.insert_after(input_grad_info->producer_head);
-                } else {
-                    grad_fn->dsts.emplace_back();
+        std::visit([&](auto& backward) {
+            using T = std::decay_t<decltype(backward)>;
+            if constexpr (std::is_same_v<T, std::monostate>) {
+                mgb_assert(0);
+            } else {
+                for (size_t i = 0; i < ctx.nargs; ++i) {
+                    if (backward.input_has_grad(i) && input_requires_grad(ctx, i) && ctx.args[i]->m_grad_info_dict.count(grad_key.get())) {
+                        auto& input_grad_info = ctx.args[i]->m_grad_info_dict.at(grad_key.get());
+                        grad_fn->dsts.emplace_back(input_grad_info);
+                        // register as grad producer
+                        grad_fn->dsts.back().producer_record.insert_after(input_grad_info->producer_head);
+                    } else {
+                        grad_fn->dsts.emplace_back();
+                    }
                 }
-            }
-            for (size_t i = 0; i < outputs.size(); ++i) {
-                if (backward.output_requires_grad(i)) {
-                    if (backward.output_captured(i)) {
-                        // avoid reference cycle [Tensor <-> GradFn]
-                        static std::shared_ptr<OpDef> op = std::shared_ptr<OpDef>(new FastpathCopy());
-                        outputs[i] = python::apply(op, outputs[i])[0];
+                for (size_t i = 0; i < outputs.size(); ++i) {
+                    if (backward.output_requires_grad(i)) {
+                        if (backward.output_captured(i)) {
+                            // avoid reference cycle [Tensor <-> GradFn]
+                            static std::shared_ptr<OpDef> op = std::make_shared<FastpathCopy>();
+                            outputs[i] = python::apply(op, outputs[i])[0];
+                        }
+                        // populate grad info of output tensor
+                        auto& grad_info = outputs[i]->m_grad_info_dict[grad_key.get()];
+                        grad_info.grad_fn = grad_fn;
+                        grad_info.idx = i;
+                        grad_info.insert_after(grad_key->free_vars_head);
+                        outputs[i]->m_flags |= Flags::GRAD;
                     }
-                    // populate grad info of output tensor
-                    auto& grad_info = outputs[i]->m_grad_info;
-                    grad_info.grad_fn = grad_fn;
-                    grad_info.idx = i;
-                    grad_info.insert_after(grad_key->free_vars_head);
-                    outputs[i]->m_flags |= Flags::GRAD;
                 }
             }
-        }
-    }, grad_fn->backward);
+        }, grad_fn->backward);
 
-    // record forward history
-    grad_key->tape.emplace_back(grad_fn);
+        // record forward history
+        grad_key->tape.emplace_back(grad_fn);
+    }
 
     return outputs;
 }
 
+PyObject* GradKeyWrapper::get_priority() {
+    return py::cast(m_key->priority).release().ptr();
+}
+
+void GradKeyWrapper::set_priority(pybind11::handle priority) {
+    m_key->name = py::cast<int>(priority);
+}
+
 void GradKeyWrapper::attach(PyObject*const* args, size_t nargs) {
     if (nargs != 2) {
         throw py::type_error("expect 2 arguments");
@@ -488,24 +493,21 @@ void GradKey::attach(Tensor* tensor, pybind11::object callback) {
         throw py::value_error("grad key finalized");
     }
 
-    if (tensor->m_grad_info.grad_fn) {
-        if (tensor->m_grad_info.grad_fn->key.lock().get() != this) {
-            PyErr_SetString(PyExc_NotImplementedError, "second order grad");
-            throw pyext17::py_err_set();
-        }
-        if (tensor->m_grad_info->callback) {
+    if (tensor->m_grad_info_dict.count(this)) {
+        if (tensor->m_grad_info_dict.at(this)->callback) {
             throw py::value_error("callback already set on this tensor");
         }
     } else {
-        tensor->m_grad_info.idx = 0;
-        auto& grad_fn = tensor->m_grad_info.grad_fn;
+        auto& grad_info = tensor->m_grad_info_dict[this];
+        grad_info.idx = 0;
+        auto& grad_fn = grad_info.grad_fn;
         grad_fn = std::make_shared<GradFn>();
         grad_fn->key = shared_from_this();
         grad_fn->slots.resize(1);
-        tensor->m_grad_info.insert_after(free_vars_head);
+        grad_info.insert_after(free_vars_head);
         tensor->m_flags |= Flags::GRAD;
     }
-    tensor->m_grad_info.grad_fn->slots[0].callback = std::move(callback);
+    tensor->m_grad_info_dict.at(this).grad_fn->slots[0].callback = std::move(callback);
 }
 
 template<typename T>
@@ -530,8 +532,15 @@ void GradKey::backward(std::vector<TensorWrapper*> tensors, std::vector<TensorWr
     active = false;
     struct CleanupGuard {
         GradKey* owner;
-        CleanupGuard(GradKey* this_) : owner(this_) {}
-        ~CleanupGuard() {owner->cleanup();}
+        size_t priority_backup;
+        CleanupGuard(GradKey* this_) : owner(this_) {
+            priority_backup = sm_min_priority;
+            sm_min_priority = owner->priority;
+        }
+        ~CleanupGuard() {
+            owner->cleanup();
+            sm_min_priority = priority_backup;
+        }
     } _cleanup_guard(this);
 
     if (tape.empty()) return;
@@ -542,14 +551,16 @@ void GradKey::backward(std::vector<TensorWrapper*> tensors, std::vector<TensorWr
     }
 
     for (size_t i = 0; i < tensors.size(); ++i) {
-        auto& grad_info = tensors[i]->m_tensor->m_grad_info;
-        if (grad_info.grad_fn && grad_info.grad_fn->key.lock().get() == this) {
-            grad_info->grad = grads[i]->m_tensor;
+        if (tensors[i]->m_tensor->m_grad_info_dict.count(this) == 0) {
+            continue;
         }
+        auto& grad_info = tensors[i]->m_tensor->m_grad_info_dict.at(this);
+        grad_info->grad = grads[i]->m_tensor;
     }
 
     std::vector<std::shared_ptr<GradFn>> ref_keeper;
     ref_keeper.reserve(tape.size());
+
     // back-propagation in reverse order
     for (std::ptrdiff_t k = tape.size() - 1; k >= 0; --k) {
         auto&& grad_fn = tape[k].lock();
@@ -619,13 +630,14 @@ PyObject* GradKeyWrapper::is_attached_to(PyObject*const* args, size_t nargs) {
         PyErr_SetString(PyExc_TypeError, "expect Tensor");
         return nullptr;
     }
-    auto&& grad_fn = tw->m_tensor->m_grad_info.grad_fn;
-    if (grad_fn && grad_fn->key.lock() == m_key) {
+    if (tw->m_tensor->m_grad_info_dict.count(m_key.get())) {
         Py_RETURN_TRUE;
     }
     Py_RETURN_FALSE;
 }
 
+int GradKey::sm_min_priority = 0;
+
 GradKey::~GradKey() {
     cleanup();
 }
@@ -635,4 +647,41 @@ std::unordered_map<Typeinfo*, GradRuleFn>& grad_rule_registry() {
     return registry;
 }
 
+void GradInfoCollection::_shrink() {
+    auto pred = [](GradInfo& info){ return !(info.grad_fn) || info.grad_fn->key.expired(); };
+    auto iter = std::remove_if(m_storage.begin(), m_storage.end(), pred);
+    m_storage.erase(iter, m_storage.end());
+}
+
+bool GradInfoCollection::contains(GradKey* key) {
+    _shrink();
+    for (auto&& grad_info: m_storage) {
+        if (grad_info.grad_fn->key.lock().get() == key) {
+            return true;
+        }
+    }
+    return false;
+}
+
+GradInfo& GradInfoCollection::operator[](GradKey* key) {
+    _shrink();
+    for (auto&& grad_info: m_storage) {
+        if (grad_info.grad_fn->key.lock().get() == key) {
+            return grad_info;
+        }
+    }
+    m_storage.emplace_back();
+    return m_storage.back();
+}
+
+GradInfo& GradInfoCollection::at(GradKey* key) {
+    _shrink();
+    for (auto&& grad_info: m_storage) {
+        if (grad_info.grad_fn->key.lock().get() == key) {
+            return grad_info;
+        }
+    }
+    mgb_assert(false);
+}
+
 } // namespace mgb::imperative::python

imperative/python/src/grad.h

@@ -26,12 +26,18 @@ struct GradKey : std::enable_shared_from_this<GradKey>, NonCopyableObj {
     bool active = true;
     GradInfo::head_t free_vars_head;
     std::vector<std::weak_ptr<GradFn>> tape;
+    int priority = 0;
 
     ~GradKey();
 
     void attach(Tensor* tensor, pybind11::object callback);
     void backward(std::vector<TensorWrapper*>, std::vector<TensorWrapper*>);
     void cleanup();
+    bool is_blocked() const {
+        return priority < sm_min_priority;
+    }
+private:
+    static int sm_min_priority;
 };
 
 struct GradKeyWrapper {
@@ -44,6 +50,8 @@ struct GradKeyWrapper {
 
     PyObject* get_name();
     void set_name(pybind11::handle name);
+    PyObject* get_priority();
+    void set_priority(pybind11::handle priority);
     void attach(PyObject*const* args, size_t nargs);
     void backward(std::vector<TensorWrapper*>, std::vector<TensorWrapper*>);
     PyObject* is_attached_to(PyObject*const* args, size_t nargs);
@@ -150,7 +158,7 @@ using GradRuleFn = std::function<apply_result_t(ApplyContext&, CustomBackward::M
 std::unordered_map<Typeinfo*, GradRuleFn>& grad_rule_registry();
 
 inline bool input_requires_grad(const ApplyContext& ctx, size_t i) {
-    return bool(ctx.args[i]->m_grad_info.grad_fn);
+    return !ctx.args[i]->m_grad_info_dict.empty();
 }
 
 struct GradRuleFallback : std::exception {};

imperative/python/src/grad_info.h

@@ -15,6 +15,7 @@
 
 namespace mgb::imperative::python {
 
+struct GradKey;
 struct GradFn;
 struct GradSlot;
 
@@ -32,6 +33,10 @@ struct GradInfo : GradSlotPtr, intrusive_list::Node<GradInfo, intrusive_list::be
     GradInfo(GradInfo&&) = default;
     GradInfo& operator=(GradInfo&) = default;
     GradInfo& operator=(GradInfo&&) = default;
+    GradInfo(const GradInfo& rhs): GradInfo(const_cast<GradInfo&>(rhs)){}
+    GradInfo& operator=(const GradInfo& rhs) {
+        return *this = const_cast<GradInfo&>(rhs);
+    }
 };
 
 } // namespace mgb::imperative::python

imperative/python/src/tensor.cpp

@@ -182,7 +182,7 @@ PyObject* py_apply(PyObject* self, PyObject*const* args, size_t nargs/* , PyObje
         if (py::isinstance<PySymbolVar>(py::handle(args[0]))){
             SmallVector<cg::VarNode*> vinputs(nargs);
             for (size_t i = 0; i < nargs; ++i) {
-                    vinputs[i] = py::handle(args[i]).cast<PySymbolVar*>()->m_node;   
+                vinputs[i] = py::handle(args[i]).cast<PySymbolVar*>()->m_node;
             }
             auto op = ctx.op.get();
             auto rst = OpDef::apply_on_var_node(*op, vinputs);

imperative/python/src/tensor.h

@@ -17,6 +17,7 @@
 #include "megbrain/imperative/interpreter.h"
 #include "pybind11/pybind11.h"
 #include <string>
+#include <unordered_map>
 
 #include "./pyext17.h"
 
@@ -36,6 +37,8 @@ struct ObjectPtr : B {
 
 namespace mgb::imperative::python {
 
+struct GradKey;
+
 extern interpreter::Interpreter::Channel* interpreter_for_py;
 
 class SharedHandle {
@@ -58,6 +61,34 @@ public:
 };
 
 
+// impl in grad.cpp
+class GradInfoCollection {
+private:
+    SmallVector<GradInfo> m_storage;
+protected:
+    void _shrink();
+public:
+    bool contains(GradKey* key);
+    GradInfo& operator[](GradKey* key);
+    GradInfo& at(GradKey* key);
+    bool empty() {
+        _shrink();
+        return m_storage.empty();
+    }
+    auto begin() {
+        _shrink();
+        return m_storage.begin();
+    }
+    auto end() {
+        _shrink();
+        return m_storage.end();
+    }
+    size_t count(GradKey* key) {
+        return contains(key) ? 1 : 0;
+    }
+};
+
+
 struct Tensor : std::enable_shared_from_this<Tensor>, NonCopyableObj {
     using flags_t = uint64_t;
 
@@ -69,7 +100,7 @@ struct Tensor : std::enable_shared_from_this<Tensor>, NonCopyableObj {
 
     flags_t m_flags = 0;
 
-    GradInfo m_grad_info;
+    GradInfoCollection m_grad_info_dict;
     TraceInfo m_trace_info;
     SharedHandle m_handle;
     std::string user_custom_name;
@@ -88,7 +119,7 @@ struct Tensor : std::enable_shared_from_this<Tensor>, NonCopyableObj {
     inline std::shared_ptr<Tensor> copy() {
         auto ret = std::make_shared<Tensor>(m_handle);
         ret->m_flags = m_flags;
-        ret->m_grad_info = m_grad_info;
+        ret->m_grad_info_dict = m_grad_info_dict;
         ret->m_trace_info = m_trace_info;
         ret->m_var = m_var;
         return ret;

imperative/python/test/unit/core/test_autodiff.py

@@ -108,21 +108,24 @@ def test_grad_2():
     np.testing.assert_almost_equal(x.grad.numpy(), 4 * x_np ** 3, decimal=6)
 
 
-@pytest.mark.skip(reason="high order gradient was not implemented yet")
 def test_2nd_grad():
     x_np = np.random.rand(10).astype("float32")
     x = as_tensor(x_np)
     ones = as_tensor(np.ones_like(x_np))
 
     grad = Grad().wrt(x, callback=save_to(x))
+    grad._priority = -1
     grad2 = Grad().wrt(x, callback=save_to(x))
+    grad2._priority = 0
 
     y = cos(x)
 
     grad(y, ones)
+    z = x.grad
     np.testing.assert_almost_equal(x.grad.numpy(), -np.sin(x_np), decimal=5)
 
-    grad2(x.grad, ones)
+    x.grad = None
+    grad2(z, ones)
     np.testing.assert_almost_equal(x.grad.numpy(), -np.cos(x_np))
 
 

imperative/src/impl/ops/specializations.cpp

@@ -398,20 +398,6 @@ OP_TRAIT_REG(Copy, Copy)
     .fallback();
 }} // copy
 
-namespace { namespace identity {
-auto apply_on_var_node(
-        const OpDef& def,
-        const VarNodeArray& inputs) {
-    auto&& op = def.cast_final_safe<Identity>();
-    mgb_assert(inputs.size() == 1);
-    OperatorNodeConfig config{op.make_name()};
-    return opr::Identity::make(inputs[0], config);
-}
-OP_TRAIT_REG(Identity, Identity)
-    .apply_on_var_node(apply_on_var_node)
-    .fallback();
-}} // identity
-
 namespace { namespace assert_equal {
 auto apply_on_var_node(
         const OpDef& def,

imperative/src/impl/ops/utility.cpp

@@ -9,6 +9,7 @@
  * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  */
 
+#include "megbrain/imperative/ops/autogen.h"
 #include "megbrain/imperative/ops/utility.h"
 #include "megbrain/imperative/ops/opr_attr.h"
 #include "megbrain/opr/utility.h"
@@ -32,4 +33,25 @@ OP_TRAIT_REG(FastpathCopy,FastpathCopy)
 
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(FastpathCopy);
 
+namespace { namespace identity {
+auto apply_on_var_node(
+        const OpDef& def,
+        const VarNodeArray& inputs) {
+    auto&& op = def.cast_final_safe<Identity>();
+    mgb_assert(inputs.size() == 1);
+    OperatorNodeConfig config{op.make_name()};
+    return opr::Identity::make(inputs[0], config);
+}
+
+auto apply_on_physical_tensor(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs) {
+    return SmallVector<TensorPtr>{inputs[0]};
+}
+OP_TRAIT_REG(Identity, Identity)
+    .apply_on_var_node(apply_on_var_node)
+    .apply_on_physical_tensor(apply_on_physical_tensor)
+    .fallback();
+}} // identity
+
 } // namespace mgb::imperative