fix(mge/subgraph): fix mem leak

GitOrigin-RevId: dc7bfe6265

dnn/src/cuda/eye/eye.cu

@@ -39,6 +39,7 @@ void exec_internal(T* dst, size_t m, size_t n, int k, cudaStream_t stream) {
 #define INST(T)   template void exec_internal<T>(T*, size_t, size_t, int, cudaStream_t);
 #define cb(DType) INST(typename DTypeTrait<DType>::ctype)
 MEGDNN_FOREACH_COMPUTING_DTYPE(cb)
+cb(::megdnn::dtype::Bool)
 
 }  // namespace eye
 }  // namespace cuda

dnn/src/cuda/eye/opr_impl.cpp

@@ -26,6 +26,7 @@ void EyeImpl::exec(_megdnn_tensor_out dst, _megdnn_workspace workspace) {
                 cuda_stream(handle()));                                                \
     }
     MEGDNN_FOREACH_COMPUTING_DTYPE(cb)
+    cb(::megdnn::dtype::Bool)
 #undef cb
 }
 

dnn/src/naive/eye/opr_impl.cpp

@@ -31,6 +31,7 @@ void EyeImpl::exec(_megdnn_tensor_out dst, _megdnn_workspace workspace) {
         MEGDNN_DISPATCH_CPU_KERN_OPR(exec_internal<ctype>(dst.ptr<ctype>(), m, n)); \
     }
     MEGDNN_FOREACH_COMPUTING_DTYPE(cb)
+    cb(::megdnn::dtype::Bool)
 #undef cb
 }
 

dnn/src/rocm/eye/eye.cpp.hip

@@ -44,7 +44,7 @@ void exec_internal(T* dst, size_t m, size_t n, int k, hipStream_t stream) {
     template void exec_internal<T>(T*, size_t, size_t, int, hipStream_t);
 #define cb(DType) INST(typename DTypeTrait<DType>::ctype)
 MEGDNN_FOREACH_COMPUTING_DTYPE(cb)
-
+cb(::megdnn::dtype::Bool)
 }  // namespace eye
 }  // namespace rocm
 }  // namespace megdnn

dnn/src/rocm/eye/opr_impl.cpp

@@ -27,6 +27,7 @@ void EyeImpl::exec(_megdnn_tensor_out dst, _megdnn_workspace workspace) {
                 hip_stream(handle()));                                                 \
     }
     MEGDNN_FOREACH_COMPUTING_DTYPE(cb)
+    cb(::megdnn::dtype::Bool)
 #undef cb
 }
 

imperative/python/megengine/traced_module/__init__.py

@@ -21,10 +21,10 @@ from .traced_module import (
 _register_all_builtin_module()
 set_cpp_apply_module_trace(cpp_apply_module_trace)
 
-__all__ = {
+__all__ = [
     "register_as_builtin",
     "trace_module",
     "wrap",
     "TracedModule",
     "optimize",
-}
+]

imperative/python/megengine/traced_module/node.py

@@ -222,7 +222,7 @@ class TensorNode(Node):
 
     def __init__(
         self,
-        expr: "Expr",
+        expr,
         name: str = None,
         qualname: str = None,
         shape: Tuple[int] = None,
@@ -231,7 +231,7 @@ class TensorNode(Node):
     ):
         super().__init__(expr, name, qualname)
         self._shape = shape
-        self._dtype = shape
+        self._dtype = dtype
         self._qparams = qparams
 
     def __getstate__(self):

imperative/python/src/ops.cpp

@@ -450,6 +450,7 @@ void _init_py_op_def(py::module m) {
     py_type.tp_richcompare = PyOp(OpDef)::tp_richcompare;
     py_type.tp_getset = py_op::py_getsetters;
     py_type.tp_repr = py_op::py_repr;
+    py_type.tp_dealloc = py_dealloc_generic<PyOp(OpDef)>;
     mgb_assert(PyType_Ready(&py_type) >= 0);
     m.add_object("OpDef", reinterpret_cast<PyObject*>(&py_type));
 }

imperative/python/test/unit/functional/test_tensor.py

@@ -24,21 +24,22 @@ from megengine.utils.network_node import VarNode
 
 
 def test_eye():
-    dtype = np.float32
+    dtypes = [np.float32, np.bool]
     cases = [{"input": [10, 20]}, {"input": [30]}]
-    for case in cases:
-        np.testing.assert_allclose(
-            F.eye(case["input"], dtype=dtype).numpy(),
-            np.eye(*case["input"]).astype(dtype),
-        )
-        np.testing.assert_allclose(
-            F.eye(*case["input"], dtype=dtype).numpy(),
-            np.eye(*case["input"]).astype(dtype),
-        )
-        np.testing.assert_allclose(
-            F.eye(tensor(case["input"]), dtype=dtype).numpy(),
-            np.eye(*case["input"]).astype(dtype),
-        )
+    for dtype in dtypes:
+        for case in cases:
+            np.testing.assert_allclose(
+                F.eye(case["input"], dtype=dtype).numpy(),
+                np.eye(*case["input"]).astype(dtype),
+            )
+            np.testing.assert_allclose(
+                F.eye(*case["input"], dtype=dtype).numpy(),
+                np.eye(*case["input"]).astype(dtype),
+            )
+            np.testing.assert_allclose(
+                F.eye(tensor(case["input"]), dtype=dtype).numpy(),
+                np.eye(*case["input"]).astype(dtype),
+            )
 
 
 def test_full():

imperative/src/impl/physical_tensor.cpp

@@ -44,7 +44,7 @@ public:
     }
 
     static CompNodeSyncManager& inst() {
-        static CompNodeSyncManager sl_inst;
+        static CompNodeSyncManager* sl_inst = new CompNodeSyncManager();
 #if MGB_CUDA && defined(WIN32)
         //! FIXME: windows cuda driver shutdown before call atexit function even
         //! register atexit function after init cuda driver! as a workround
@@ -55,7 +55,7 @@ public:
             mgb_assert(!err, "failed to register atexit function");
         }
 #endif
-        return sl_inst;
+        return *sl_inst;
     }
 
     CompNode::Event* get_or_create_event(Blob* blob) {