FM / autograd

/usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/autograd/

/usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/autograd

Name	Size	Mode	Actions
functions/	-	0755	rm
generated/	-	0755	rm
utils/	-	0755	rm
anomaly_mode.h	1525	0644	edit dl rm
autograd.h	5294	0644	edit dl rm
autograd_not_implemented_fallback.h	201	0644	edit dl rm
cpp_hook.h	518	0644	edit dl rm
custom_function.h	14470	0644	edit dl rm
edge.h	1627	0644	edit dl rm
engine.h	17237	0644	edit dl rm
forward_grad.h	8583	0644	edit dl rm
function.h	25046	0644	edit dl rm
FunctionsManual.h	19904	0644	edit dl rm
function_hook.h	641	0644	edit dl rm
grad_mode.h	207	0644	edit dl rm
InferenceMode.h	182	0644	edit dl rm
input_buffer.h	1597	0644	edit dl rm
input_metadata.h	1384	0644	edit dl rm
profiler.h	112	0644	edit dl rm
profiler_kineto.h	9111	0644	edit dl rm
profiler_legacy.h	16930	0644	edit dl rm
profiler_utils.h	439	0644	edit dl rm
python_anomaly_mode.h	1246	0644	edit dl rm
python_autograd.h	388	0644	edit dl rm
python_cpp_function.h	2456	0644	edit dl rm
python_engine.h	1262	0644	edit dl rm
python_fft_functions.h	128	0644	edit dl rm
python_function.h	4099	0644	edit dl rm
python_hook.h	700	0644	edit dl rm
python_legacy_variable.h	299	0644	edit dl rm
python_linalg_functions.h	131	0644	edit dl rm
python_mode.h	425	0644	edit dl rm
python_nn_functions.h	127	0644	edit dl rm
python_saved_variable_hooks.h	883	0644	edit dl rm
python_special_functions.h	132	0644	edit dl rm
python_torch_functions.h	671	0644	edit dl rm
python_variable.h	1645	0644	edit dl rm
python_variable_indexing.h	313	0644	edit dl rm
record_function_ops.h	587	0644	edit dl rm
saved_variable.h	4322	0644	edit dl rm
saved_variable_hooks.h	272	0644	edit dl rm
symbolic.h	330	0644	edit dl rm
variable.h	33411	0644	edit dl rm
VariableTypeUtils.h	15086	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/torch/csrc/autograd/python_function.h (4099B)

#pragma once

#include <torch/csrc/python_headers.h>

#include <torch/csrc/Exceptions.h>
#include <torch/csrc/autograd/custom_function.h>
#include <torch/csrc/autograd/function.h>
#include <torch/csrc/autograd/variable.h>
#include <torch/csrc/autograd/saved_variable.h>
#include <torch/csrc/utils/object_ptr.h>

#include <c10/util/Optional.h>
#include <c10/core/DeviceGuard.h>

#include <vector>
#include <utility>
#include <memory>

namespace torch { namespace jit { struct Graph; }}
namespace torch { namespace autograd {

// A Function which is implemented by a Python object (i.e., a THPFunction).
// Calls to 'apply' are forwarded to the Python method implementation.
struct PyNode : public Node {
  // NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
  PyNode(THPObjectPtr obj) : obj(obj.release()) {}

variable_list apply(variable_list&& inputs) override;

void release_variables() override;
  std::string name() const override;
  bool is_traceable() override;

// THPFunction this Function is wrapping.  Owning!
  PyObject* obj;

~PyNode() override {
    // Can't use THPObjectPtr as a field in this class; destructor won't take
    // out GIL!  When I forgot to do this by hand
    // TestAutograd.test_inplace_view_python called me out about it.
    // If python is already dead, leak the wrapped python objects
    if (Py_IsInitialized()) {
      pybind11::gil_scoped_acquire gil;
      Py_DECREF(obj);
    }
  }
};

/**
 * Cast an object into a tuple, if it is not a tuple already. Returns true
 * if the original object was not a tuple.
 */
inline bool ensure_tuple(THPObjectPtr& obj) {
  if (PyTuple_Check(obj.get()))
    return false;

PyObject *tuple = PyTuple_New(1);
  if (!tuple) throw python_error();
  PyTuple_SET_ITEM(tuple, 0, obj.release());
  obj = tuple;
  return true;
}

}} // namespace torch::autograd

// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
struct THPFunction {
    PyObject_HEAD

PyObject *needs_input_grad;

// Python tuple of tensors whose variables we should save.  Set
    // by Python with 'save_for_backward'.  If nullptr, no tensors were
    // saved.
    PyObject *to_save;
    // Python tuple of tensors which are not differentiable.  Set by
    // Python with 'mark_non_differentiable'.  If nullptr, no tensors were
    // non-differentiable.
    PyObject *non_differentiable;
    // Python tuple of tensors which had inplace updates in the forward()
    // pass.  Set by Python with 'mark_dirty'.  If nullptr, no tensors were
    // modified inplace.
    PyObject *dirty_tensors;

// boolean indicating whether to materialize undefined output grad tensors
    // into tensors full of zeros. Set by Python with 'set_materialize_grads'.
    // Default is true.
    bool materialize_grads;

std::vector<torch::autograd::VariableInfo> output_info;
    std::vector<torch::autograd::VariableInfo> input_info;
    std::vector<torch::autograd::SavedVariable> saved_variables;
    // For each input, true if the input is a THPVariable
    std::vector<bool> is_variable_input;
    char has_freed_buffers;

// The actual PyNode (in the autograd graph) that this data was
    // saved for.  This field may be NULL (because a user can construct
    // a THPFunction directly from Python), but when this field is non-NULL,
    // it is guaranteed that cdata.lock()->obj == this
    //
    // In most ordinary use, this field should always be non-NULL; e.g.,
    // when we allocate a THPFunction because we are running Node.apply,
    // after constructing a THPFunction, we immediately allocate a PyNode
    // for it.  We can't enforce this directly in the constructor of
    // THPFunction though, because there's no way to keep it live long enough
    // to save an owning reference to PyNode into the grad_fn of a Variable.
    std::weak_ptr<torch::autograd::PyNode> cdata;
};

bool THPFunction_initModule(PyObject *module);
extern PyTypeObject THPFunctionType;
extern PyObject *THPFunctionClass;

inline bool THPFunction_Check(PyObject* obj) {
  return PyObject_IsInstance(obj, (PyObject*)&THPFunctionType);
}

cmd: